# Contents

TWO USEFUL SUBSTITUTIONS 2
ALWAYS CAUCHY-SCHWARZ 11
EQUATIONS AND BEYOND 25
LOOK AT THE EXPONENT! 38
PRIMES AND SQUARES 53
T
2
’S LEMMA 65
ONLY GRAPHS, NO SUBGRAPHS! 81
COMPLEX COMBINATORICS 90
FORMAL SERIES REVISITED 101
NUMBERS AND LINEAR ALGEBRA 117
ARITHMETIC PROPERTIES OF POLYNOMIALS 130
LAGRANGE INTERPOLATION 144
HIGHER ALGEBRA IN COMBINATORICS 166
GEOMETRY AND NUMBERS 184
THE SMALLER, THE BETTER 195
DENSITY AND REGULAR DISTRIBUTION 204
THE SUM OF DIGITS OF A POSITIVE INTEGER 218
ANALYSIS AGAINST NUMBER THEORY? 233
SOLVING ELEMENTARY INEQUALITIES WITH
INTEGRALS 264
1
TWO USEFUL SUBSTITUTIONS
We know that in most inequalities with a constraint such as abc = 1
the substitution a =
x
y
, b =
y
z
, c =
z
x
simpliﬁes the solution (don’t kid
yourself, not all problems of this type become easier!). But have you ever
the conditions x, y, z > 0 and xyz = x + y + z + 2? Or x, y, z > 0 and
xy + yz + zx + 2xyz = 1? There are numerous problems that reduce
to these conditions and to their corresponding substitutions. You will
be probably surprised when ﬁnding out that the ﬁrst set of conditions
implies the existence of positive real numbers a, b, c such that
x =
b +c
a
, y =
c +a
b
, z =
a +b
c
.
Let us explain why. The condition xyz = x+y+z +2 can be written
in the following equivalent way:
1
1 +x
+
1
1 +y
+
1
1 +z
= 1.
Proving this is just a matter of simple computations. Take now
a =
1
1 +x
, b =
1
1 +y
, c =
1
1 +z
.
Then a + b + c = 1 and x =
1 −a
a
=
b +c
a
. Of course, in the same
way we ﬁnd y =
c +a
b
, z =
a +b
c
. The converse (that is,
b +c
a
,
c +a
b
,
a +b
c
satisfy xyz = x+y +z = 2) is much easier and is settled again by
basic computations. Now, what about the second set of conditions? If
you look carefully, you will see that it is closely related to the ﬁrst one.
Indeed, x, y, z > 0 satisfy xy + yz + zx + 2xyz = 1 if and only if
1
x
,
1
y
,
1
z
verify
1
xyz
=
1
x
+
1
y
+
1
z
+ 2, so the substitution here is
x =
a
b +c
, y =
b
c +a
, z =
c
a +b
.
2
So, let us summarize: we have seen two nice substitutions, with even
nicer proofs, but we still have not seen any applications. We will see
them in a moment ... and there are quite a few inequalities that can be
solved by using these ”tricks”.
First, an easy and classical problem, due to Nesbitt. It has so many
extensions and generalizations, that we must discuss it ﬁrst.
Example 1. Prove that
a
b +c
+
b
c +a
+
c
a +b

3
2
for all a, b, c > 0.
Solution. With the ”magical” substitution, it suﬃces to prove that
if x, y, z > 0 satisfy xy + yz + zx + 2xyz = 1, then x + y + z =
3
2
.
Let us suppose that this is not the case, i.e. x + y + z <
3
2
. Because
xy + yz + zx ≤
(x +y +z)
2
3
, we must have xy + yz + zx <
3
4
and
since xyz ≤
_
x +y +z
3
_
3
, we also have 2xyz <
1
4
. It follows that
1 = xy +yz +zx + 2xyz <
3
4
+
1
4
= 1, a contradiction, so we are done.
Let us now increase the level of diﬃculty and make an experiment:
imagine that you did not know about these substitutions and try to
solve the following problem. Then look at the solution provided and you
will see that sometimes a good substitution can solve a problem almost
alone.
Example 2. Let x, y, z > 0 such that xy + yz + zx + 2xyz = 1.
Prove that
1
x
+
1
y
+
1
z
≥ 4(x +y +z).
Mircea Lascu, Marian Tetiva
Solution. With our substitution the inequality becomes
b +c
a
+
c +a
b
+
a +b
c
≥ 4
_
a
b +c
+
b
c +a
+
c
a +b
_
.
3
But this follows from
4s
b +c

a
b
+
a
c
,
4b
c +a

b
c
+
b
a
,
4c
a +b

c
a
+
c
b
.
Simple and eﬃcient, these are the words that characterize this sub-
stitution.
Here is a geometric application of the previous problem.
Example 3. Prove that in any acute-angled triangle ABC the fol-
lowing inequality holds
cos
2
Acos
2
B+cos
2
Bcos
2
C+cos
2
C cos
2
A ≤
1
4
(cos
2
A+cos
2
B+cos
2
C).
Titu Andreescu
Solution. We observe that the desired inequality is equivalent to
cos Acos B
cos C
+
cos Bcos C
cos A
+
cos Acos C
cos B

1
4
_
cos A
cos Bcos C
+
cos B
cos C cos A
+
cos C
cos Acos B
_
Setting
x =
cos Bcos C
cos A
, y =
cos Acos C
cos B
, z =
cos Acos B
cos C
,
the inequality reduces to
4(x +y +z) ≤
1
x
+
1
y
+
1
z
.
But this is precisely the inequality in the previous example. All that
remains is to show that xy +yz +zx + 2xyz = 1. This is equivalent to
cos
2
A+ cos
2
B + cos
2
C + 2 cos Acos Bcos C = 1,
a well-known identity, proved in the chapter ”Equations and beyond”.
The level of diﬃculty continues to increase. When we say this, we
refer again to the proposed experiment. The reader who will try ﬁrst to
solve the problems discussed without using the above substitutions will
certainly understand why we consider these problems hard.
4
Example 4. Prove that if x, y, z > 0 and xyz = x +y +z +2, then
2(

xy +

yz +

zx) ≤ x +y +z + 6.
Solution. This is tricky, even with the substitution. There are two
main ideas: using some identities that transform the inequality into
an easier one and then using the substitution. Let us see. What does
2(

xy +

yz +

zx) suggest? Clearly, it is related to
(

x +

y +

z)
2
−(x +y +z).
Consequently, our inequality can be written as

x +

y +

z ≤
_
2(x +y +z + 3).
The ﬁrst idea that comes to mind (that is using the Cauchy-
Schwarz inequality in the form

x +

y +

z ≤
_
3(x +y +z) ≤
_
2(x +y +z + 3)) does not lead to a solution. Indeed, the last inequal-
ity is not true: setting x+y+z = s, we have 3s ≤ 2(s+3). This is because
from the AM-GM inequality it follows that xyz ≤
s
3
27
, so
s
3
27
≥ s + 2,
which is equivalent to (s −6)(s + 3)
2
≥ 0, implying s ≥ 6.
Let us see how the substitution helps. The inequality becomes
_
b +c
a
+
_
c +a
b
+
_
a +b
c

¸
2
_
b +c
a
+
c +a
b
+
a +b
c
+ 3
_
The last step is probably the most important. We have to change
the expression
b +c
a
+
c +a
b
+
a +b
c
+ 3 a little bit.
We see that if we add 1 to each fraction, then a +b +c will appear
as common factor, so in fact
b +c
a
+
c +a
b
+
a +b
c
+ 3 = (a +b +c)
_
1
a
+
1
b
+
1
c
_
.
5
And now we have ﬁnally solved the problem, amusingly, by employ-
ing again the Cauchy-Schwarz inequality:
_
b +c
a
+
_
c +a
b
+
_
a +b
c

¸
(b +c +c +a +a +b)
_
1
a
+
1
b
+
1
c
_
.
We continue with a 2003 USAMO problem. There are many proofs
for this inequality, none of them easy. The following solution is again not
easy, but it is natural for someone familiar with this kind of substitution.
Example 5. Prove that for any positive real numbers a, b, c the
following inequality holds
(2a +b +c)
2
2a
2
+ (b +c)
2
+
(2b +c +a)
2
2b
2
+ (c +a)
2
+
(2c +a +b)
2
2c
2
+ (a +b)
2
≤ 8.
Titu Andreescu, Zuming Feng, USAMO 2003
Solution. The desired inequality is equivalent to
_
1 +
b +c
a
_
2
2 +
_
b +c
a
_
2
+
_
2 +
c +a
b
_
2
2 +
_
c +a
b
_
2
+
_
1 +
a +b
c
_
2
2 +
_
a +b
c
_
2
≤ 8.
Taking our substitution into account, it suﬃces to prove that if xyz =
x +y +z + 2, then
(2 +x)
2
2 +x
2
+
(2 +y)
2
2 +y
2
+
(2 +z)
2
2 +z
2
≤ 8.
This is in fact the same as
2x + 1
x
2
+ 2
+
2y + 1
y
2
+ 2
+
2z + 1
z
2
+ 2

5
2
.
Now, we transform this inequality into
(x −1)
2
x
2
+ 2
+
(y −1)
2
y
2
+ 2
+
(z −1)
2
z
2
+ 2

1
2
.
This last form suggests using the Cauchy-Schwarz inequality to prove
that
(x −1)
2
x
2
+ 2
+
(y −1)
2
y
2
+ 2
+
(z −1)
2
z
2
+ 2

(x +y +z −3)
2
x
2
+y
2
+z
2
+ 6
.
6
So, we are left with proving that 2(x+y +z −3)
2
≥ x
2
+y
2
+z
2
+6.
But this is not diﬃcult. Indeed, this inequality is equivalent to
2(x +y +z −3)
2
≥ (x +y +z)
2
−2(xy +yz +zx) + 6.
Now, from xyz ≥ 8 (recall who x, y, z are and use the AM-GM
inequality three times), we ﬁnd that xy +yz +zx ≥ 12 and x+y +z ≥ 6
(by the same AM-GM inequality). This shows that it suﬃces to prove
that 2(s−3)
2
≥ s
2
−18 for all s ≥ 6, which is equivalent to (s−3)(s−6) ≥
0, clearly true. And this diﬃcult problem is solved!
The following problem is also hard. We have seen a diﬃcult solution
in the chapter ”Equations and beyond”. Yet, there is an easy solution
using the substitutions described in this unit.
Example 6. Prove that if x, y, z ≥ 0 satisfy xy +yz +zx +xyz = 4
then x +y +z ≥ xy +yz +zx.
India, 1998
Solution. Let us write the given condition as
x
2

y
2
+
y
2

z
2
+
z
2

x
2
+ 2
x
2

y
2

z
2
= 1.
Hence there are positive real numbers a, b, c such that
x =
2a
b +c
, y =
2b
c +a
, z =
2c
a +b
.
But now the solution is almost over, since the inequality
x +y +z ≥ xy +yz +zx
is equivalent to
a
b +c
+
b
c +a
+
c
a +b

2ab
(c +a)(c +b)
+
2bc
(a +b)(a +c)
+
2ca
(b +a)(b +c)
.
After clearing denominators, the inequality becomes
a(a +b)(a +c) +b(b +a)(b +c) +c(c +a)(c +b) ≥
7
≥ 2ab(a +b) + 2bc(b +c) + 2ca(c +a).
After basic computations, it reduces to
a(a −b)(a −c) +b(b −a)(b −c) +c(c −a)(c −b) ≥ 0.
But this is Schur’s inequality!
We end the discussion with a diﬃcult problem, in which the substi-
tution described plays a key role. But this time using the substitution
only will not suﬃce.
Example 7. Prove that if x, y, z > 0 satisfy xyz = x + y + z + 2,
then xyz(x −1)(y −1)(z −1) ≤ 8.
Gabriel Dospinescu
Solution. Using the substitution
x =
b +c
a
, y =
c +a
b
, z =
a +b
c
,
the inequality becomes
(a +b)(b +c)(c +a)(a +b −c)(b +c −a)(c +a −b) ≤ 8a
2
b
2
c
2
(1)
for any positive real numbers a, b, c. It is readily seen that this form is
stronger than Schur’s inequality (a +b −c)(b +c −a)(c +a −b) ≤ abc.
First, we may assume that a, b, c are the sides of a triangle ABC, since
otherwise the left-hand side in (1) is negative. This is true because no
more than one of the numbers a+b−c, b+c−a, c+a−b can be negative.
Let R be the circumradius of the triangle ABC. It is not diﬃcult to ﬁnd
the formula
(a +b −c)(b +c −a)(c +a −b) =
a
2
b
2
c
2
(a +b +c)R
2
.
Consequently, the desired inequality can be written as
(a +b +c)R
2

(a +b)(b +c)(c +a)
8
.
8
But we know that in any triangle ABC, 9R
2
≥ a
2
+ b
2
+ c
2
. Hence
it suﬃces to prove that
8(a +b +c)(a
2
+b
2
+c
2
) ≥ 9(a +b)(b +c)(c +a).
This inequality follows from the following ones:
8(a +b +c)(a
2
+b
2
+c
2
) ≥
8
3
(a +b +c)
3
and
9(a +b)(b +c)(c +a) ≤
1
3
(a +b +c)
3
.
The ﬁrst inequality reduces to
a
2
+b
2
+c
2

1
3
(a +b +c)
2
,
while the second is a consequence of the AM-GM inequality. By com-
bining these two results, the desired inequality follows.
Problems for training
1. Prove that if x, y, z > 0 satisfy xy +yz +zx + 2xyz = 1, then
xyz ≤
1
8
and xy +yz +zx ≥
3
4
.
2. Prove that for any positive real numbers a, b, c the following in-
equality holds
b +c
a
+
c +a
b
+
a +b
c

a
b +c
+
b
c +a
+
c
a +b
+
9
2
.
J. Nesbitt
3. Prove that if x, y, z > 0 and xyz = x +y +z + 2, then
xy +yz +zx ≥ 2(x +y +z) and

x +

y +

z ≤
3
2

xyz.
4. Let x, y, z > 0 such that xy +yz +zx = 2(x +y +z). Prove that
xyz ≤ x +y +z + 2.
Gabriel Dospinescu, Mircea Lascu
9
5. Prove that in any triangle ABC the following inequality holds
cos A+ cos B + cos C ≥
1
4
(3 + cos(A−B) + cos(B −C) + cos(C −A)).
Titu Andreescu
6. Prove that in every acute-angled triangle ABC,
(cos A+ cos B)
2
+ (cos B + cos C)
2
+ (cos C + cos A)
2
≤ 3.
7. Prove that if a, b, c > 0 and x = a +
1
b
, y = b +
1
c
, z = c +
1
a
, then
xy +yz +zx ≥ 2(x +y +z).
Vasile Cartoaje
8. Prove that for any a, b, c > 0,
(b +c −a)
2
(b +c)
2
+a
2
+
(c +a −b)
2
(c +a)
2
+b
2
+
(a +b −c)
2
(a +b)
2
+c
2

3
5
.
Japan, 1997
10
ALWAYS CAUCHY-SCHWARZ
In recent years the Cauchy-Schwarz inequality has become one of
the most used results in elementary mathematics, an indispensable tool
of any serious problem solver. There are countless problems that reduce
readily to this inequality and even more problems in which the Cauchy-
Schwarz inequality is the key idea of the solution. In this unit we will
not focus on the theoretical results, since they are too well-known. Yet,
seeing the Cauchy-Schwarz inequality at work is not so well spread out.
This is the reason why we will see this inequality in action in several
simple examples ﬁrst, employing then gradually the Cauchy-Schwarz
inequality in some of the most diﬃcult problems.
Let us begin with a very simple problem, a direct application of the
inequality. Yet, it underlines something less emphasized: the analysis of
the equality case.
Example 1. Prove that the ﬁnite sequence a
0
, a
1
, . . . , a
n
of positive
real numbers is a geometrical progression if and only if
(a
2
0
+a
2
1
+ +a
2
n−1
)(a
2
1
+a
2
2
+ +a
2
n
) = (a
0
a
1
+a
1
a
2
+ +a
n−1
a
n
)
2
.
Solution. We see that the relation given in the problem is in fact
the equality case in the Cauchy-Schwarz inequality. This is equivalent to
the proportionality of the n-tuples (a
0
, a
1
, . . . , a
n−1
) and (a
1
, a
2
, . . . , a
n
),
that is
a
0
a
1
+
a
1
a
2
= =
a
n−1
a
n
.
But this is just actually the deﬁnition of a geometrical progression.
Hence the problem is solved. Note that Lagrange’s identity allowed us
to work with equivalences.
Another easy application of the Cauchy-Schwarz inequality is the
following problem. This time the inequality is hidden in a closed form,
11
which suggests using calculus. There exists a solution by using deriva-
tives, but it is not as elegant as the featured one:
Example 2. Let p be a polynomial with positive real coeﬃcients.
Prove that p(x
2
)p(y
2
) ≥ p
2
(xy) for any positive real numbers x, y.
Solution. If we work only with the closed expression p(x
2
)p(y
2
) ≥
p
2
(xy), the chances of seeing a way to proceed are small. So, let us write
p(x) = a
0
+a
1
x + +a
n
x
n
. The desired inequality becomes
(a
0
+a
1
x
2
+ +a
n
x
2n
)(a
0
+a
1
y
2
+ +a
n
y
2n
)
≥ (a
0
+a
1
xy + +a
n
x
n
y
n
)
2
.
And now the Cauchy-Schwarz inequality comes into the picture:
(a
0
+a
1
xy + +a
n
x
n
y
n
)
2
= (

a
0

a
0
+
_
a
1
x
2

_
a
2
y
2
+ +

a
n
x
n

a
n
y
n
)
2
≤ (a
0
+a
1
x
2
+ +a
n
x
2n
)(a
0
+a
1
y
2
+ +a
n
y
2n
).
And the problem is solved. Moreover, we see that the conditions
x, y > 0 are useless, since we have of course p
2
(xy) ≤ p
2
tionally, note an interesting consequence of the problem: the function
f : (0, ∞) → (0, ∞), f(x) = ln p(e
x
) is convex, that is why we said in
the introduction to this problem that it has a solution based on calculus.
The idea of that solution is to prove that the second derivative of is non-
negative. We will not prove this here, but we note a simple consequence:
the more general inequality
p(x
k
1
)p(x
k
2
) . . . p(x
k
k
) ≥ p
k
(x
1
x
2
. . . x
k
),
which follows the Jensen’s inequality for the convex function f(x) =
ln p(e
x
).
12
Here is another application of the Cauchy-Schwarz inequality, though
this time you might be surprised why the ”trick” fails at a ﬁrst approach:
Example 3. Prove that if x, y, z > 0 satisfy
1
x
+
1
y
+
1
z
= 2, then

x −1 +
_
y −1 +

z −1 ≤

x +y +z.
Iran, 1998
Solution. The obvious and most natural approach is to apply the
Cauchy-Schwarz inequality in the form

x −1 +
_
y −1 +

z −1 ≤
_
3(x +y +z −3)
and then to try to prove the inequality
_
3(x +y +z −3) ≤

x +y +z,
which is equivalent to x + y + z ≤
9
2
. Unfortunately, this inequality is
not true. In fact, the reversed inequality holds, that is x + y + z ≥
9
2
,
since 2 =
1
x
+
1
y
+
1
z

9
x +y +z
. Hence this approach fails. Then, we
try another approach, using again the Cauchy-Schwarz inequality, but
this time in the form

x −1 +
_
y −1 +

z −1 =

a
_
x −1
a
+

b
_
y −1
b
+

c
_
z −1
c

¸
(a +b +c)
_
x −1
a
+
y −1
b
+
z −1
c
_
.
We would like to have the last expression equal to

x +y +z. This
encourages us to take a = x, b = y, c = z, since in this case
x −1
a
+
y −1
b
+
z −1
c
= 1 and a +b +c = x +y +z.
So, this idea works and the problem is solved.
We continue with a classical result, the not so well-known inequality
of Aczel. We will also see during our trip through the exciting world of
the Cauchy-Schwarz inequality a nice application of Aczel’s inequality.
13
Example 4. Let a
1
, a
2
, . . . , a
n
, b
1
, b
2
, . . . , b
n
be real numbers and let
A, B > 0 such that
A
2
≥ a
2
1
+a
2
2
+ +a
2
n
or B
2
≥ b
2
1
+b
2
2
+ +b
2
n
.
Then
(A
2
−a
2
1
−a
2
2
− −a
2
n
)(B
2
−b
2
1
−b
2
2
− −b
2
n
)
≤ (AB −a
1
b
1
−a
2
b
2
− −a
n
b
n
)
2
.
Solution. We observe ﬁrst that we may assume that
A
2
> a
2
1
+a
2
2
+ +a
2
n
and B
2
> b
2
1
+b
2
2
+ +b
2
n
.
Otherwise the left-hand side of the desired inequality is smaller than
or equal to 0 and the inequality becomes trivial. From our assumption
and the Cauchy-Schwarz inequality, we infer that
a
1
b
1
+a
2
b
2
+ +a
n
b
n

_
a
2
1
+a
2
2
+ +a
2
n

_
b
2
1
+b
2
2
+ +b
2
n
< AB
Hence we can rewrite the inequality in the more appropriate form
a
1
b
1
+a
2
b
2
+ +a
n
b
n
+
_
(A
2
−a)(B
2
−b) ≤ AB,
where a = a
2
1
+ a
2
2
+ + a
2
n
and b = b
2
1
+ b
2
2
+ + b
2
n
. Now, we can
apply the Cauchy-Schwarz inequality, ﬁrst in the form
a
1
b
1
+a
2
b
2
+ +a
n
b
n
+
_
(A
2
−a)(B
2
−b) ≤

ab+
_
(A
2
−a)(B
2
−b)
and then in the form

ab +
_
(A
2
−a)(B
2
−b) ≤
_
(a +A
2
−a)(b +B
2
−b) = AB.
And by combining the last two inequalities the desired inequality
follows.
As a consequence of this inequality we discuss the following problem,
in which the condition seems to be useless. In fact, it is the key that
suggests using Aczel’s inequality.
14
Example 5. Let a
1
, a
2
, . . . , a
n
, b
1
, b
2
, . . . , b
n
be real numbers such
that
(a
2
1
+a
2
2
+ +a
2
n
−1)(b
2
1
+b
2
2
+ +b
2
n
−1) > (a
1
b
1
+a
2
b
2
+ +a
n
b
n
−1)
2
.
Prove that a
2
1
+a
2
2
+ +a
2
n
> 1 and b
2
1
+b
2
2
+ +b
2
n
> 1.
Titu Andreescu, Dorin Andrica, TST 2004, USA
Solution. At ﬁrst glance, the problem does not seem to be related
to Aczel’s inequality. Let us take a more careful look. First of all, it
is not diﬃcult to observe that an indirect approach is more eﬃcient.
Moreover, we may even assume that both numbers a
2
1
+a
2
2
+ +a
2
n
−1
and b
2
1
+ b
2
2
+ + b
2
n
− 1 are negative, since they have the same sign
(this follows immediately from the hypothesis of the problem). Now, we
want to prove that
(a
2
1
+a
2
2
+ +a
2
n
−1)(b
2
1
+b
2
2
+ +b
2
n
−1)
≤ (a
1
b
1
+a
2
b
2
+ +a
n
b
n
−1)
2
(1)
in order to obtain the desired contradiction. And all of a sudden we
arrived at the result in the previous problem. Indeed, we have now the
conditions 1 > a
2
1
+ a
2
2
+ + a
2
n
and 1 > b
2
1
+ b
2
2
+ + b
2
n
, while the
conclusion is (1). But this is exactly Aczel’s inequality, with A = 1 and
B = 1. The conclusion follows.
Of a diﬀerent kind, the following example shows that an apparently
very diﬃcult inequality can become quite easy if we do not complicate
things more than necessary. It is also a reﬁnement of the Cauchy-Schwarz
inequality, as we can see from the solution.
Example 6. For given n > k > 1 ﬁnd in closed form the best con-
stant T(n, k) such that for any real numbers x
1
, x
2
, . . . , x
n
the following
15
inequality holds:

1≤i<j≤n
(x
i
−x
j
)
2
≥ T(n, k)

1≤i<j≤k
(x
i
−x
j
)
2
.
Gabriel Dospinescu
Solution. In this form, we cannot make any reasonable conjecture
about T(n, k), so we need an eﬃcient transformation. We observe that

1≤i<j≤n
(x
i
−x
j
)
2
is nothing else than n
n

i=1
x
2
i

_
n

i=1
x
i
_
2
and also

1≤i<j≤k
(x
i
−x
j
)
2
= k
k

i=1
x
2
i

_
k

i=1
x
i
_
2
,
according to Lagrange’s identity. Consequently, the inequality can be
written in the equivalent form
n
n

i=1
x
2
i

_
n

i=1
x
i
_
2
≥ T(n, k)
_
_
k
k

i=1
x
2
i

_
k

i=1
x
i
_
2
_
_
.
And now we see that it is indeed a reﬁnement of the Cauchy-Schwarz
inequality, only if in the end it turns out that T(n, k) > 0. We also
observe that in the left-hand side there are n − k variables that do not
appear in the right-hand side and that the left-hand side is minimal
when these variables are equal. So, let us take them all to be zero. The
result is
n
k

i=1
x
2
i

_
k

i=1
x
i
_
2
≥ T(n, k)
_
_
k
k

i=1
x
2
i

_
k

i=1
x
i
_
2
_
_
,
which is equivalent to
(T(n, k) −1)
_
k

i=1
x
i
_
2
≥ (kT(n, k) −n)
k

i=1
x
2
i
(1)
16
Now, if kT(n, k) −n > 0, we can take a k-tuple (x
1
, x
2
, . . . , x
k
) such
that
k

i=1
x
i
= 0 and
k

i=1
x
2
i
,= 0 and we contradict the inequality (1).
Hence we must have kT(n, k) − n ≤ 0 that is T(n, k) ≤
n
k
. Now, let us
proceed with the converse, that is showing that
n
n

i=1
x
2
i

_
n

i=1
x
i
_
2

n
k
_
_
k
k

i=1
x
2
i

_
k

i=1
x
i
_
2
_
_
(2)
for any real numbers x
1
, x
2
, . . . , x
n
. If we manage to prove this inequality,
then it will follow that T(n, k) =
n
k
. But (2) is of course equivalent to
n
n

i=k+1
x
2
i

_
n

i=1
x
i
_
2

n
k
_
k

i=1
x
i
_
2
.
Now, we have to apply the Cauchy-Schwarz inequality, because we
need
n

i=k+1
x
i
. We ﬁnd that
n
n

i=k+1
x
2
i

n
n −k
_
n

i=k+1
x
i
_
2
and so it suﬃces to prove that
n
n −k
A
2
≥ (A+B)
2

n
k
B
2
, (3)
where we have taken A =
n

i=k+1
x
i
and B =
k

i=1
x
i
. But (3) is straight-
forward, since it is equivalent to
(kA−(n −k)B)
2
+k(n −k)B
2
≥ 0,
which is clear. Finally, the conclusion is settled: T(n, k) =
n
k
is the best
constant.
We continue the series of diﬃcult inequalities with a very nice prob-
lem of Murray Klamkin. This time, one part of the problem is obvious
17
from the Cauchy-Schwarz inequality, but the second one is not immedi-
ate. Let us see.
Example 7. Let a, b, c be positive real numbers. Find the extreme
values of the expression
_
a
2
x
2
+b
2
y
2
+c
2
z
2
+
_
b
2
x
2
+c
2
y
2
+a
2
z
2
+
_
c
2
x
2
+a
2
y
2
+b
2
z
2
where x, y, z are real numbers such that x
2
+y
2
+z
2
= 1.
Murray Klamkin, Crux Mathematicorum
Solution. Finding the upper bound does not seem to be too diﬃcult,
since from the Cauchy-Schwarz inequality it follows that
_
a
2
x
2
+b
2
y
2
+c
2
z
2
+
_
b
2
x
2
+c
2
y
2
+a
2
z
2
+
_
c
2
x
2
+a
2
y
2
+b
2
z
2

_
3(a
2
x
2
+b
2
y
2
+c
2
z
2
+c
2
y
2
+a
2
z
2
+c
2
x
2
+a
2
y
2
+b
2
z
2
)
=
_
3(a
2
+b
2
+c
2
).
We have used here the hypothesis x
2
+ y
2
+ z
2
= 1. Thus,
_
3(a
2
+b
2
+c
2
) is the upper bound and this value if attained for
x = y = z =

3
3
.
But for the lower bound things are not so easy. Investigating what
happens when xyz = 0, we conclude that the minimal value should be
a +b +c, attained when two variables are zero and the third one is 1 or
−1. Hence, we should try to prove the inequality
_
a
2
x
2
+b
2
y
2
+c
2
z
2
+
_
b
2
x
2
+c
2
y
2
+a
2
z
2
+
_
c
2
x
2
+a
2
y
2
+b
2
z
2
≥ a +b +c.
Why not squaring it? After all, we observe that
a
2
x
2
+b
2
y
2
+c
2
z
2
+b
2
x
2
+c
2
y
2
+a
2
z
2
+c
2
x
2
+a
2
y
2
+b
2
z
2
= a
2
+b
2
+c
2
,
so the new inequality cannot have a very complicated form. It becomes
_
a
2
x
2
+b
2
y
2
+c
2
z
2

_
b
2
x
2
+c
2
y
2
+a
2
z
2
18
+
_
b
2
x
2
+c
2
y
2
+a
2
z
2

_
c
2
x
2
+a
2
y
2
+b
2
z
2
+
_
c
2
x
2
+a
2
y
2
+b
2
z
2

_
a
2
x
2
+b
2
y
2
+c
2
z
2
≥ ab +bc +ca
which has great chances to be true. And indeed, it is true and it follows
from what else?, the Cauchy-Schwarz inequality:
_
a
2
x
2
+b
2
y
2
+c
2
z
2

_
b
2
x
2
+c
2
y
2
+a
2
z
2
≥ abx
2
+bxy
2
+caz
2
and the other two similar inequalities. This shows that the minimal value
is indeed a +b +c, attained for example when (x, y, z) = (1, 0, 0).
It is now time for the champion inequalities. We will discuss two
hard inequalities and after that we will leave for the reader the pleasure
of solving many other problems based on these techniques.
Example 8. Prove that for any nonnegative numbers a
1
, a
2
, . . . , a
n
such that
n

i=1
a
i
=
1
2
, the following inequality holds

1≤i<j≤n
a
i
a
j
(1 −a
i
)(1 −a
j
)

n(n −1)
2(2n −1)
2
.
Vasile Cartoaje
Solution. This is a very hard problem, in which intuition is better
than technique. We will concoct a solution using a combination between
the Cauchy-Schwarz inequality and Jensen’s inequality, but we warn the
reader that such a solution cannot be invented easily. Fasten your seat
belts! Let us write the inequality in the form
_
n

i=1
a
i
1 −a
i
_
2

n

i=1
a
2
i
(1 −a
i
)
2
+
n(n −1)
(2n −1)
2
.
We apply now the Cauchy-Schwarz inequality to ﬁnd that
_
n

i=1
a
i
1 −a
i
_
2

_
n

i=1
a
i
__
n

i=1
a
i
(1 −a
i
)
2
_
=
n

i=1
a
i
2
(1 −a
i
)
2
.
19
Thus, it remains to prove the inequality
n

i=1
a
i
2
(1 −a
i
)
2

n

i=1
a
2
i
(1 −a
i
)
2
+
n(n −1)
(2n −1)
2
.
The latter can be written of course in the following form:
n

i=1
a
i
(1 −2a
i
)
(1 −a
i
)
2

2n(n −1)
(2n −1)
2
.
This encourages us to study the function
f :
_
0,
1
2
_
→R, f(x) =
x(1 −2x)
(1 −x)
2
and to see if it is concave. This is not diﬃcult, for a short computa-
tion shows that f
tt
(x) =
−6x
(1 −x)
4
≤ 0. Hence we can apply Jensen’s
inequality to complete the solution.
We end this discussion with a remarkable solution, found by the
member of the Romanian Mathematical Olympiad Committee, Claudiu
Raicu, to the diﬃcult problem given in 2004 in one of the Romanian
Team Selection Tests.
Example 9. Let a
1
, a
2
, . . . , a
n
be real numbers and let S be a non-
empty subset of ¦1, 2, . . . , n¦. Prove that
_

i∈S
a
i
_
2

1≤i≤j≤n
(a
i
+ +a
j
)
2
.
Gabriel Dospinescu, TST 2004, Romania
Solution. Let us deﬁne s
i
= a
1
+a
2
+ +a
i
for i ≥ 1 and s
0
= 0.
Now, partition S into groups of consecutive numbers. Then

i∈S
a
i
is of
the forms
j
1
−s
i
1
+s
j
2
−s
i
2
+ +s
j
k
−s
i
k
, with 0 ≤ i
1
< i
2
< < i
k
≤ n,
j
1
< j
2
< < j
k
and also i
1
< j
1
, . . . , i
k
< j
k
. Now, let us observe that
20
the left-hand side is nothing else than
n

i=1
s
2
i
+

1≤i<j≤n
(s
j
−s
i
)
2
=

1≤i<j≤n+1
(s
j
−s
i
)
2
.
Hence we need to show that
(s
j
1
−s
i
1
+s
j
2
−s
i
2
+ +s
j
k
−s
i
k
)
2

0≤i<j≤n+1
(s
j
−s
i
)
2
.
Let us take a
1
= s
i
1
, a
2
= s
j
1
, . . . , a
2k−1
= s
i
k
, a
2k
= s
j
k
and observe
the obvious (but important) inequality

0≤i<j≤n+1
(s
j
−s
i
)
2

1≤i<j≤2k
(a
i
−a
j
)
2
.
And this is how we arrived at the inequality
(a
1
−a
2
+a
3
− +a
2k−1
−a
2k
)
2

1≤i<j≤2k
(a
i
−a
j
)
2
(1)
The latter inequality can be proved by using the Cauchy-Schwarz
inequality k-times:
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
(a
1
−a
2
+a
3
− +a
2k−1
−a
2k
)
2
≤ k((a
1
−a
2
)
2
+ (a
3
−a
4
)
2
+ + (a
2k−1
−a
2k
)
2
)
(a
1
−a
2
+a
3
− +a
2k−1
−a
2k
)
2
≤ k((a
1
−a
4
)
2
+ (a
3
−a
6
)
2
+ + (a
2k−1
−a
2
)
2
)
. . .
(a
1
−a
2
+a
3
− +a
2k−1
−a
2k
)
2
≤ k((a
1
−a
2k
)
2
+ (a
3
−a
2
)
2
+ + (a
2k−1
−a
2k−2
)
2
)
and by summing up all these inequalities. In the right-hand side we
obtain an even smaller quantity than

1≤i<j≤2k
(a
i
− a
j
)
2
, which proves
that (1) is correct. The solution ends here.
21
Problems for training
1. Let a, b, c be nonnegative real numbers. Prove that
(ax
2
+bx +c)(cx
2
+bx +a) ≥ (a +b +c)
2
x
2
for all nonnegative real numbers x.
Titu Andreescu, Gazeta Matematica
2. Let p be a polynomial with positive real coeﬃcients. Prove that
if p
_
1
x
_

1
p(x)
is true for x = 1, then it is true for all x > 0.
Titu Andreescu, Revista Matematica Timisoara
3. Prove that for any real numbers a, b, c ≥ 1 the following inequality
holds:

a −1 +

b −1 +

c −1 ≤
_
a(bc + 1).
4. For any positive integer n ﬁnd the number of ordered n-tuples of
integers (a
1
, a
2
, . . . , a
n
) such that
a
1
+a
2
+ +a
n
≥ n
2
and a
2
1
+a
2
2
+ +a
2
n
≤ n
3
+ 1.
China, 2002
5. Prove that for any positive real numbers a, b, c,
1
a +b
+
1
b +c
+
1
c +a
+
1
2
3

abc

(a +b +c +
3

abc)
2
(a +b)(b +c)(c +a)
.
Titu Andreescu, MOSP 1999
6. Let a
1
, a
2
, . . . , a
n
, b
1
, b
2
, . . . , b
n
be real numbers such that

1≤i<j≤n
a
i
a
j
> 0.
Prove the inequality
_
_

1≤i,=j≤n
a
i
b
j
_
_
2

_
_

1≤i,=j≤n
a
i
a
j
_
_
_
_

1≤i,=j≤n
b
i
b
j
_
_
Alexandru Lupas, AMM
22
7. Let n ≥ 2 be an even integer. We consider all polynomials of the
form x
n
+a
n−1
x
n−1
+ +a
1
x +1, with real coeﬃcients and having at
least one real zero. Determine the least possible value of a
2
1
+a
2
2
+ +
a
2
n−1
.
Czech-Polish-Slovak Competition, 2002
8. The triangle ABC satisﬁes the relation
_
cot
A
2
_
2
+
_
2 cot
B
2
_
2
+
_
3 cot
C
2
_
2
=
_
6s
7r
_
2
.
Show that ABC is similar to a triangle whose sides are integers and
ﬁnd the smallest set of such integers.
Titu Andreescu, USAMO 2002
9. Let x
1
, x
2
, . . . , x
n
be positive real numbers such that
1
1 +x
1
+
1
1 +x
2
+ +
1
1 +x
n
= 1.
Prove the inequality

x
1
+

x
2
+ +

x
n
≥ (n −1)
_
1

x
1
+
1

x
2
+ +
1

x
n
_
.
Vojtech Jarnik Competition, 2002
10. Given are real numbers x
1
, x
2
, . . . , x
10

_
0,
π
2
_
such that
sin
2
x
1
+ sin
2
x
2
+ + sin
2
x
10
= 1.
Prove that
3(sin x
1
+ sin x
2
+ + sin x
10
) ≤ cos x
1
+ cos x
2
+ + cos x
10
.
Saint Petersburg, 2001
11. Prove that for any real numbers a, b, c, x, y, z the following in-
equality holds
ax +by +cz +
_
(a
2
+b
2
+c
2
)(x
2
+y
2
+z
2
) ≥
2
3
(a +b +c)(x +y +z).
Vasile Cartoaje, Kvant
23
12. Prove that for any real numbers x
1
, x
2
, . . . , x
n
the following in-
equality holds
_
n

i=1
n

i=1
[x
i
−x
j
[
_
2

2(n
2
−1)
3
_
_
n

i=1
n

j=1
[x
i
−x
j
[
2
_
_
.
IMO 2003
13. Let n > 2 and x
1
, x
2
, . . . , x
n
be positive real numbers such that
(x
1
+x
2
+ +x
n
)
_
1
x
1
+
1
x
2
+ +
1
x
n
_
= n
2
+ 1.
Prove that
(x
2
1
+x
2
2
+ +x
2
n
)
_
1
x
2
1
+
1
x
2
2
+ +
1
x
2
n
_
> n
2
+ 4 +
2
n(n −1)
.
Gabriel Dospinescu
14. Prove that for any positive real numbers a, b, c, x, y, z such that
xy +yz +zx = 3,
a
b +c
(y +z) +
b
c +a
(x +z) +
c
a +b
(x +y) ≥ 3.
Titu Andreescu, Gabriel Dospinescu
15. Prove that for any positive real numbers a
1
, a
2
, . . . , a
n
, x
1
,
x
2
, . . . , x
n
such that

i≤i<j≤n
x
i
x
j
=
_
n
2
_
,
the following inequality holds
a
1
a
2
+ +a
n
(x
2
+ +x
n
) + +
a
n
a
1
+ +a
n−1
(x
1
+ +x
n−1
) ≥ n.
Vasile Cartoaje, Gabriel Dospinescu
24
EQUATIONS AND BEYOND
Real equations with multiple unknowns have in general inﬁnitely
many solutions if they are solvable. In this case, an important task char-
acterizing the set of solutions by using parameters. We are going to
discuss two real equations and two parameterizations, but we will go
beyond, showing how a simple idea can generate lots of nice problems,
some of them really diﬃcult.
We begin this discussion with a problem. It may seem unusual, but
this problem is in fact the introduction that leads to the other themes
in this discussion.
Example 1. Consider three real numbers a, b, c such that abc = 1
and write
x = a +
1
a
, y = b +
1
b
, z = c +
1
c
(1)
Find an algebraic relation between x, y, z, independent of a, b, c.
Of course, without any ideas, one would solve the equations from
(1) with respect to a, b, c and then substitute the results in the relation
abc = 1. But this is a mathematical crime! Here is a nice idea. To
generate a relation involving x, y, z, we compute the product
xyz =
_
a +
1
a
__
b +
1
b
__
c +
1
c
_
=
_
a
2
+
1
a
2
_
+
_
b
2
+
1
b
2
_
+
_
c
2
+
1
c
2
_
+ 2
= (x
2
−2) + (y
2
−2) + (z
2
−2) + 2.
Thus,
x
2
+y
2
+z
2
−xyz = 4 (2)
and this is the answer to the problem.
Now, another question appears: is the converse true? Obviously not
(take for example the numbers (x, y, z) = (1, 1, −1)). But looking again
25
at (1), we see that we must have min¦[x[, [y[, [z[¦ ≥ 2. We will prove the
following result.
Example 2. Let x, y, z be real numbers with max¦[x[, [y[, [z[¦ > 2.
Prove that there exist real numbers a, b, c with abc = 1 satisfying (1).
Whenever we have a condition of the form max¦[x[, [y[, [z[¦ > 2, it is
better to make a choice. Here, let us take [x[ > 2. This shows that there
exists a nonzero real number u such that x = u+
1
u
, (we have used here
the condition [x[ > 2). Now, let us regard (2) as a second degree equation
with respect to z. Since this equation has real roots, the discriminant
must be nonnegative, which means that (x
2
−4)(y
2
−4) ≥ 0. But since
[x[ > 2, we ﬁnd that y
2
≥ 4 and so there exist a non-zero real number
v for which y = v +
1
v
. How do we ﬁnd the corresponding z? Simply by
solving the second degree equation. We ﬁnd two solutions:
z
1
= uv +
1
uv
, z
2
=
u
v
+
v
u
and now we are almost done. If z = uv+
1
uv
we take (a, b, c) =
_
u, v,
1
uv
_
and if z =
u
v
+
v
u
, then we take (a, b, c) =
_
1
u
, v,
u
v
_
. All the conditions
are satisﬁed and the problem is solved.
A direct consequence of the previous problem is the following:
If x, y, z > 0 are real numbers that verify (2), then there exist
α, β, χ ∈ R such that
x = 2ch(α), y = 2ch(β), z = 2ch(χ),
where ch : R → (0, ∞), ch(x) =
e
x
+e
−x
2
. Indeed, we write (1), in
which this time it is clear that a, b, c > 0 and we take α = ln a, β = ln b,
χ = ln c.
Inspired by the previous equation, let us consider another one
x
2
+y
2
+z
2
+xyz = 4, (3)
26
where x, y, z > 0. We will prove that the set of solutions of this equation
is the set of triples (2 cos A, 2 cos B, 2 cos C) where A, B, C are the angles
of an acute triangle. First, let us prove that all these triples are solutions.
This reduces to the identity
cos
2
A+ cos
2
B + cos
2
C + 2 cos Acos Bcos C = 1.
This identity can be proved readily by using the sum-to-product
formulas, but here is a nice proof employing geometry and linear algebra.
We know that in any triangle we have the relations
_
¸
¸
_
¸
¸
_
a = c cos B +b cos C
b = a cos C +c cos A
c = b cos A+a cos B
which are simple consequences of the Law of Cosines. Now, let us con-
sider the system
_
¸
¸
_
¸
¸
_
x −y cos C −z cos B = 0
−xcos C +y −z cos A = 0
−xcos B +y cos A−z = 0
From the above observation, it follows that this system has a non-
trivial solution, that is (a, b, c) and so we must have
¸
¸
¸
¸
¸
¸
¸
¸
1 −cos C −cos B
−cos C 1 −cos A
−cos B −cos A 1
¸
¸
¸
¸
¸
¸
¸
¸
= 0,
which expanded gives
cos
2
A+ cos
2
B + cos
2
C + 2 cos Acos Bcos C = 1.
For the converse, we see ﬁrst that 0 < x, y, z < 2, hence there are
numbers A, B ∈
_
0,
π
2
_
such that x = 2 cos A, y = 2 cos B. Solving the
equation with respect to z and taking into account that z ∈ (0, 2) we
27
obtain z = −2 cos(A + B). Thus we can take C = π − A − B and we
will have (x, y, z) = (2 cos A, 2 cos B, 2 cos C). All in all we have solved
the following problem.
Example 3. The positive real numbers x, y, z satisfy (3) if and only
if there exists an acute-angled triangle ABC such that
x = 2 cos A, y = 2 cos B, z = 2 cos C.
With the introduction and the easy problems over it is now time to
see some nice applications of the above results.
Example 4. Let x, y, z > 2 satisfying (2). We deﬁne the sequences
(a
n
)
n≥1
, (b
n
)
n≥1
, (c
n
)
n≥1
by
a
n+1
=
a
2
n
+x
2
−4
a
n−1
, b
n+1
=
b
2
n
+y
2
−4
b
n−1
, c
n+1
=
c
2
n
+z
2
−4
c
n−1
,
with a
1
= x, b
1
= y, c
1
= z and a
2
= x
2
− 2, b
2
= y
2
− 2, c
2
= z
2
− 2.
Prove that for all n ≥ 1 the triple (a
n
, b
n
, c
n
) also satisﬁes (2).
Solution. Let us write x = a+
1
a
, y = b+
1
b
, z = c+
1
c
, with abc = 1.
Then
a
2
= a
2
+
1
a
2
, b
2
= b
2
+
1
b
2
, c
2
= c
2
+
1
c
2
.
So, a reasonable conjecture is that
(a
n
, b
n
, c
n
) =
_
a
n
+
1
a
n
, b
n
+
1
b
n
, c
n
+
1
c
n
_
.
Indeed, this follows by induction from
_
a
n
+
1
a
n
_
2
+a
2
+
1
a
2
−2
a
n−1
+
1
a
n−1
= a
n+1
+
1
a
n+1
and two similar identities. We have established that
(a
n
, b
n
, c
n
) =
_
a
n
+
1
a
n
, b
n
+
1
b
n
, c
n
+
1
c
n
_
28
But if abc = 1, then certainly a
n
b
n
c
n
= 1, which shows that indeed
the triple (a
n
, b
n
, c
n
) satisﬁes (2).
The following problem is a nice characterization of the equation (2)
by polynomials and also teaches us some things about polynomials in
two or three variables.
Example 5. Find all polynomials f(x, y, z) with real coeﬃcients
such that
f
_
a +
1
a
, b +
1
b
, c +
1
c
_
= 0
whenever abc = 1.
Gabriel Dospinescu
Solution. From the introduction, it is now clear that the polyno-
mials divisible by x
2
+ y
2
+ z
2
− xyz − 4 are solutions to the problem.
But it is not obvious why any desired polynomial should be of this form.
To show this, we use the classical polynomial long division. There are
polynomials g(x, y, z), h(y, z), k(y, z) with real coeﬃcients such that
f(x, y, z) = (x
2
+y
2
+z
2
−xyz −4)g(x, y, z) +xh(y, z) +k(y, z)
Using the hypothesis, we deduce that
0 =
_
a +
1
a
_
h
_
b +
1
b
, c +
1
c
_
+k
_
b +
1
b
, c +
1
c
_
whenever abc = 1. Well, it seems that this is a dead end. Not exactly.
Now we take two numbers x, y such that min¦[x[, [y[¦ > 2 and we write
x = b +
1
b
, y = c +
1
c
with b =
x +

x
2
−4
2
, c =
y +
_
y
2
−4
2
.
Then it is easy to compute a+
1
a
. It is exactly xy+
_
(x
2
−4)(y
2
−4).
So, we have found that
(xy +
_
(x
2
−4)(y
2
−4))h(x, y) +k(x, y) = 0
whenever min¦[x[, [y[¦ > 2. And now? The last relation suggests that we
should prove that for each y with [y[ > 2, the function x →

x
2
−4 is
29
not rational, that is, there aren’t polynomials p, q such that

x
2
−4 =
p(x)
q(x)
. But this is easy because if such polynomials existed, than each
zero of x
2
−4 should have even multiplicity, which is not the case. Con-
sequently, for each y with [y[ > 2 we have h(x, y) = k(x, y) = 0 for all
x. But this means that h(x, y) = k(x, y) = 0 for all x, y, that is our
polynomial is divisible with x
2
+y
2
+z
2
−xyz −4.
O a diﬀerent kind, the following problem and the featured solution
prove that sometimes an eﬃcient substitution can help more than ten
complicated ideas.
Example 6. Let a, b, c > 0. Find all triples (x, y, z) of positive real
numbers such that
_
x +y +z = a +b +c
a
2
x +b
2
y +c
2
z +abc = 4xyz
Titu Andreescu, IMO Shortlist, 1995
Solution. We try to use the information given by the second equa-
tion. This equation can be written as
a
2
yz
+
b
2
zx
+
c
2
xy
+
abc
xyz
= 4
and we already recognize the relation
u
2
+v
2
+w
2
+uvw = 4
where u =
a

yz
, v =
b

zx
, w =
c

xy
. According to example 3, we can
ﬁnd an acute-angled triangle ABC such that
u = 2 cos A, v = 2 cos B, w = 2 cos C.
We have made use of the second condition, so we use the ﬁrst one
to deduce that
x +y +z = 2

xy cos C + 2

yz cos A+ 2

zxcos B.
30
Trying to solve this as a second degree equation in

x, we ﬁnd the
discriminant
−4(

y sin C −

z sin B)
2
.
Because this discriminant is nonnegative, we infer that

y sin C =

z sin B and

x =

y cos C +

z cos B.
Combining the last two relations, we ﬁnd that

x
sin A
=

y
sin B
=

z
sin C
Now we square these relations and we use the fact that
cos A =
a
2

yz
, cos B =
b
2

zx
, cos C =
c
2

xy
.
The conclusion is:
x =
b +c
2
, y =
c +a
2
, z =
a +b
2
and it is immediate to see that this triple satisﬁes both conditions. Hence
there is a unique triple that is solution to the given system. Notice that
the condition
x +y +z = 2

xy cos C + 2

yz cos A+ 2

zxcos B
is the equality case in the lemma stated in the solution of the following
problem. This could be another possible solution of the problem.
We have discussed the following very diﬃcult problem in the chapter
”An useful substitution”. We will see that example 3 helps us ﬁnd a nice
geometric solution to this inequality.
Example 7. Prove that if the positive real numbers x, y, z satisfy
xy +yz +zx +xyz = 4, then
x +y +z ≥ xy +yz +zx.
India, 1998
31
Solution. It is not diﬃcult to observe that at ﬁrst glance, the con-
dition xy +yz +zx +xyz = 4 it’s not the same as the equation (3). Let
us write the condition xy +yz +zx +xyz = 4 in the form

xy
2
+

yz
2
+

zx
2
+

xy

yz

zx = 4.
Now, we can use the result from example 3 and we deduce the exis-
tence of an acute-angled triangle ABC such that
_
¸
¸
_
¸
¸
_

yz = 2 cos A

zx = 2 cos B

xy = 2 cos C
We solve the system and we ﬁnd the triplet
(x, y, z) =
_
2 cos Bcos C
cos A
,
2 cos Acos C
cos B
,
2 cos Acos B
cos C
_
Hence we need to prove that
2 cos Bcos C
cos A
+
2 cos Acos C
cos B
+
2 cos Acos B
cos C
≥ 2(cos
2
A+cos
2
B+cos
2
C).
This one is a hard inequality and it follows from a more general
result.
Lemma. If ABC is a triangle and x, y, z are arbitrary real numbers,
then
x
2
+y
2
+z
2
≥ 2yz cos A+ 2zxcos B + 2xy cos C.
Proof of the lemma. Let us consider points P, Q, R on the lines
AB, BC, CA, respectively, such that AP = BQ = CR = 1 and P, Q, R
and do not lie on the sides of the triangle. Then we see that the inequality
is equivalent to
(x
−→
AP +y
−−→
BQ+z
−→
CR)
2
≥ 0,
which is obviously true.
32
The lemma being proved, we just have to take
x =
_
2 cos Bcos C
cos A
y =
_
2 cos Acos C
cos B
, z =
_
2 cos Acos B
cos C
in the above lemma and the problem will be solved.
But of course, this type of identities does not appear only in inequal-
ities. We are going to discuss two problems in which the identity is very
Example 8. Find all continuous functions f : (0, ∞) → (0, ∞)
satisfying
f(x)f(y) = f(xy) +f
_
x
y
_
.
Sankt Petersburg
Solution. First of all, observe that by symmetry in x, y we must
have f
_
x
y
_
= f
_
y
x
_
and so f(x) = f
_
1
x
_
. Next, by taking x = y = 1
we obtain f(1) = 2 and then f(x
2
) = f
2
(x) −2. These relations should
now ring a bell! It seems that we are searching for something like f(x) =
x
k
+
1
x
k
. We are right, but still far from the solution. Let’s make another
small step: proving that f(x) ≥ 2 for all x. Indeed, this is going to be
easy, since f(x
2
) = f
2
(x) − 2 implies that f(x) >

2 for all x. Thus,
f
2
(x) = f(x
2
) + 2 > 2 +

2. Repeating this argument, we ﬁnd that for
all x we have
f(x) >
_
2 +
_
2 +

2 +. . . = 2
(the last equality being immediate for a beginner in analysis).
Yet, till now nothing related to our theme. Wrong! Let’s observe that
f(x
2
) +f(y
2
) = f(xy)f
_
x
y
_
for all x, y. Indeed, it suﬃces to write
x
2
= xy
x
y
, y
2
=
xy
x
y
.
33
With this information, let us make one more step and write
f
2
(x) +f
2
(y) −4 = f(x
2
) +f(y
2
) = f(xy)(f(x)f(y) −f(xy)).
We are now on the right track, since we ﬁnd that
f
2
(x) +f
2
(y) +f
2
(xy) = f(x)f(y)f(xy) + 4.
Using also the fact that f(x) ≥ 2, we deduce the existence of a con-
tinuous function g : (0, ∞) → [1, ∞) such that f(x) = g(x) +
1
g(x)
. The
above relation implies of course that g(xy) = g(x)g(y). By considering
h(x) = ln g(e
x
), we obtain that h is a continuous solution of Cauchy’s
functional equation f(x+y) = f(x) +f(y), thus h(x) = kx for a certain
k. This shows that g(x) = x
k
and that our thoughts were right; these are
all solutions of the equation (the veriﬁcation of the identity is immediate
for this class of functions).
And ﬁnally, an apparently inextricable recursive relation.
Example 9. Let (a
n
)
n≥0
be a non-decreasing sequence of positive
integers such that
a
0
= a
1
= 47 and a
2
n−1
+a
2
n
+a
2
n+1
−a
n−1
a
n
a
n+1
= 4 for all n ≥ 1.
Prove that 2 +a
n
and 2 +

2 +a
n
are perfect squares for all n ≥ 0.
Titu Andreescu
Solution. Using the idea from the chapter with real equations, we
write a
n
= x
n
+
1
x
n
, with x
n
> 1. The the given condition becomes
x
n+1
= x
n
x
n−1
(we have used here explicitly that x
n
> 1), which shows
that (ln x
n
)
n≥0
is a Fibonacci-type sequence. Since x
0
= x
1
, we deduce
that x
n
= x
F
n
0
, where F
0
= F
1
= 1, F
n+1
= F
n
+F
n−1
. Now, we have to
do more: who is x
0
0
=
47 +

47
2
−1
2
won’t suﬃces.
Let us remark that
_

x
0
+
1

x
0
_
2
= 49
34
from where we ﬁnd that

x
0
+
1

x
0
= 7.
Similarly, we obtain that
4

x
0
+
1
4

x
0
= 3.
Solving the equation, we obtain
4

x
0
=
_
1 +

5
2
_
2
= λ
2
that is x
0
= λ
8
. And so we have found the general formula a
n
= λ
8F
n
+
λ
−8F
n
. And now the problem becomes easy, since
a
n
+ 2 = (λ
4F
n

−4F
n
)
2
and 2 +

2 +a
n
= (λ
2F
n

−2F
n
)
2
.
All we are left to prove is that λ
2k
+
1
λ
2k
∈ R for all k ∈ R. But this
isn’t diﬃcult, since
λ
2
+
1
λ
2
∈ R, λ
4
+
1
λ
4
∈ R
and
λ
2(k+1)
+
1
λ
2(k+1)
=
_
λ
2
+
1
λ
2
__
λ
2k
+
1
λ
2k
_

_
λ
2(k−1)
+
1
λ
2(k−1)
_
.
Problems for training
1. Find all triples x, y, z of positive real numbers, solutions to the
system:
_
x
2
+y
2
+z
2
= xyz + 4
xy +yz +zx = 2(x +y +z)
2. Let x, y, z > 0 such that x
2
+y
2
+z
2
+xyz = 4. Prove that
¸
(2 −a)(2 −b)
(2 +a)(2 +b)
+
¸
(2 −b)(2 −c)
(2 +b)(2 +c)
+
¸
(2 −c)(2 −a)
(2 +c)(2 +a)
= 1.
Cristinel Mortici, Romanian Inter-county Contest
35
3. Prove that if a, b, c ≥ 0 satisfy the condition [a
2
+b
2
+c
2
−4[ = abc,
then
(a −2)(b −2) + (b −2)(c −2) + (c −2)(a −2) ≥ 0.
Titu Andreescu, Gazeta Matematica
4. Find all triples (a, b, c) of positive real numbers, solutions to the
system
_
a
2
+b
2
+c
2
+abc = 4
a +b +c = 3
Cristinel Mortici, Romanian Inter-county Contest
5. Prove that in any triangle the following inequality holds
_
sin
A
2
+ sin
B
2
+ sin
C
2
_
2
≤ cos
2
A
2
+ cos
2
B
2
+ cos
2
C
2
.
6. Let x, y, z > 0 such that xy +yz +zx +xyz = 4. Prove that
3
_
1

x
+
1

y
+
1

z
_
2
≥ (x + 2)(y + 2)(z + 2).
Gabriel Dospinescu
7. Prove that in any acute-angled triangle the following inequality
holds
_
cos A
cos B
_
2
+
_
cos B
cos C
_
2
+
_
cos C
cos A
_
2
+ 8 cos Acos Bcos C ≥ 4.
Titu Andreescu, MOSP 2000
8. Solve in positive integers the equation
(x + 2)(y + 2)(z + 2) = (x +y +z + 2)
2
.
Titu Andreescu
9. Let n > 4 be a given positive integer. Find all pairs of positive
integers (x, y) such that
xy −
(x +y)
2
n
= n −4.
36
Titu Andreescu
10. Let the sequence (a
n
)
n≥0
, where a
0
= a
1
= 97 and a
n+1
=
a
n−1
a
n
+
_
(a
2
n
−1)(a
2
n−1
−1) for all n ≥ 1. Prove that 2 +

2 + 2a
n
is
a perfect square for all n ≥ 0.
Titu Andreescu
11. Find all triplets of positive integers (k, l, m) with sum 2002 and
for which the system
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
_
x
y
+
y
x
= k
y
z
+
z
y
= l
z
x
+
x
y
= m
has real solutions.
Titu Andreescu, proposed for IMO 2002
12. Find all functions f : (0, ∞) → (0, ∞) with the following prop-
erties:
a) f(x)+f(y)+f(z)+f(xyz) = f(

xy)f(

yz)f(

zx) for all x, y, z;
b) if 1 ≤ x < y then f(x) < f(y).
Hojoo Lee, IMO Shortlist 2004
13. Prove that if a, b, c ≥ 2 satisfy the condition a
2
+b
2
+c
2
= abc+4,
then
a +b +c +ac +bc ≥ 2
_
(a +b +c + 3)(a
2
+b
2
+c
2
−3).
Marian Tetiva
14. Prove that if a, b, c ≥ 0 satisfy a
2
+b
2
+c
2
+abc = 4 then
0 ≤ ab +bc +ca −abc ≤ 2.
Titu Andreescu, USAMO 2001
37
LOOK AT THE EXPONENT!
Most of the times, proving divisibility reduces to congruences and
the famous theorems from this ﬁeld, such as Fermat, Euler, or Wil-
son. But what do we do when we have to prove for example that
lcm(a, b, c)
2
[lcm(a, b) lcm(b, c) lcm(c, a) for any positive integers a, b, c?
Then one thing is sure: the above methods fail. Yet, another smart idea
appears: if we have to prove that a[b, then it is enough to prove that
the exponent of any prime number in the decomposition of a is at least
the exponent of that prime number in the decomposition of b. For sim-
plicity, let us denote by v
p
(a) the exponent of the prime number p in
the decomposition of a. Of course, if p doesn’t divide a, then v
p
(a) = 0.
Also, it is easy to prove the following properties of v
p
(a):
1) min¦v
p
(a), v
p
(b)¦ ≤ v
p
(a +b) ≤ max¦v
p
(a), v
p
(b)¦
2) v
p
(ab) = v
p
(a) +v
p
(b)
for any positive integer numbers a, b. Now, let us repeat the above idea
in terms of v
p
(a): we have a[b if and only if for any prime number p
we have v
p
(a) ≤ v
p
(b) and we have a = b if and only if for any prime
number p, v
p
(a) = v
p
(b).
Some other useful properties of v
p
(a) are:
3) v
p
(gcd(a
1
, a
2
, . . . , a
n
)) = min¦v
p
(a
1
), v
p
(a
2
), . . . , v
p
(a
n
)¦,
4) v
p
(lcm(a
1
, a
2
, . . . , a
n
)) = max¦v
p
(a
1
), v
p
(a
2
), . . . , v
p
(a
n
)¦ and
5) v
p
(n!) =
_
n
p
_
+
_
n
p
2
_
+
_
n
p
3
_
+ =
n −s
p
(n)
p −1
whenever p[n. Here,
s
p
(n) is the sum of digits of n when written in base b. Observe that 3)
and 4) are simple consequences of the deﬁnitions. Less straightforward is
5). This follows from the fact that there are
_
n
p
_
multiples of p,
_
n
p
2
_
are
multiples of p
2
and so on. The other equality is not diﬃcult. Indeed, let
us write n = a
0
+a
1
p+ +a
k
p
k
, where a
0
, a
1
, . . . , a
k
∈ ¦0, 1, . . . , p−1¦
38
and a
k
,= 0. Then
_
n
p
_
+
_
n
p
2
_
+ = a
1
+a
2
p+ +a
k
p
k−1
+a
2
+a
3
p+ +a
k
p
k−2
+ +a
k
and now using the formula
1 +p + +p
i
=
p
i+1
−1
p −1
,
we ﬁnd exactly 5). Enough with the introduction, let’s see some concrete
results. We have chosen with intention the ﬁrst problem (the classical
one) a very nasty one, so that the reader doesn’t think that all the above
formulas were for nothing and because it oﬀers us the opportunity to
prove a very nice inequality. There are hundreds of variants of it in all
contests around the world and in all elementary magazines. Let us see.
Example 1. Prove that
(3a + 3b)!(2a)!(3b)!(2b)!
(2a + 3b)!(a + 2b)!(a +b)!a!(b!)
2
∈ Z for
any positive integers a, b.
Solution. First, let us clearify something. When we write
_
n
p
_
+
_
n
p
2
_
+
_
n
p
3
_
+. . . ,
we write in fact

k≥1
_
n
p
k
_
and this sum has clearly a ﬁnite number of
non-zero terms. Now, let us take a prime p and let us apply formula 5),
as well as the ﬁrst observations. We ﬁnd that
v
p
((3a+3b)!(2a)!(3b)!(2b)!) =

k≥1
__
3a + 3b
p
k
_
+
_
2a
p
k
_
+
_
3b
p
k
_
+
_
2b
p
k
__
and also
v
p
= ((2a + 3b)!(a + 2b)!(a +b)!a!(b!)
2
)
=

k≥1
__
2a + 3b
p
k
_
+
_
a + 2b
p
k
_
+
_
a +b
p
k
_
+
_
a
p
k
_
+ 2
_
b
p
k
__
39
Of course, it is enough to prove that for each k ≥ 1 the term cor-
responding to k in the ﬁrst sum is greater than or equal to the term
corresponding to k in the second sum. With the substitution x =
a
p
k
,
y =
b
p
k
, we have to prove that for any nonnegative real numbers x, y we
have
[3x +3y] +[2x] +[3y] +[2y] ≥ [2x +3y] +[x +2y] +[x +y] +[x] +2[y].
This isn’t easy, but with another useful idea the inequality will be-
come easy. The idea is that
[3x + 3y] = 3[x] + 3[y] + [3¦x¦ + 3¦y¦]
and similar relations for the other terms of the inequality. After this
operation, we see that it suﬃces to prove the inequality only for 0 ≤
x, y < 1. Why is the new inequality easy? Because we can easily compute
all terms, after splitting in some cases, so that to see when [2¦x¦], [3¦y¦],
[2¦y¦] are 0, 1 or 2.
We won’t continue studying these cases, since another beautiful
problem is waiting.
Example 2. Let a, b be positive integers such that a[b
2
, b
3
[a
4
, a
5
[b
6
,
b
7
[a
8
, . . . . Prove that a = b.
Solution. Let us take a prime p and try to prove that v
p
(a) = v
p
(b).
We see that the hypothesis a[b
2
, b
3
[a
4
, a
5
[b
6
, b
7
[a
8
, . . . is the same as
a
4n+1
[b
4n+2
and b
4n+3
[a
4n+4
for all natural number n. But the relation
a
4n+1
[b
4n+2
can be interpreted as (4n + 1)v
p
(a) ≤ (4n + 2)v
p
(b) for all
n, that is
v
p
(a) ≤ lim
n→∞
4n + 2
4n + 1
v
p
(b) = v
p
(b).
Similarly, the condition b
4n+3
[a
4n+4
implies v
p
(a) ≥ v
p
(b) and so
v
p
(a) = v
p
(b). The conclusion follows: a = b.
40
We have mentioned in the beginning of the discussion a nice and
easy problem, so probably it’s time to solve it, although for sure the
Example 3. Prove that lcm(a, b, c)
2
[lcm(a, b) lcm(b, c) lcm(c, a)
for any positive integers a, b, c.
Solution. Let p an arbitrary prime number. We have
v
p
(lcm(a, b, c)
2
) = 2 max¦x, y, z¦
and
v
p
(lcm(a, b) lcm(b, c) lcm(c, a)) = max¦x, y¦ +max¦y, z¦ +max¦z, x¦,
where x = v
p
(a), y = v
p
(b), z = v
p
(c). So, we need to prove that
max¦x, y¦ + max¦y, z¦ + max¦z, x¦ ≥ 2 max¦x, y, z¦
for any nonnegative integers x, y, z. But this is easy, since we may assume
that x ≥ y ≥ z (the symmetry allows us this supposition) and the
inequality becomes 2x +y ≥ 2x, obviously true.
It is time for some diﬃcult problems, which are all based on the
observations from the beginning of the discussion.
Example 4. Prove that there exists a constant c such that for any
positive integers a, b, n that verify a! b![n! we have a +b < n +c ln n.
Paul Erdos
Solution. This time the other formula for v
p
(n!) is useful. Of course,
there is no reasonable estimation of this constant, so we should better
see what happens if a! b![n!. Then v
2
(a!) + v
2
(b!) ≤ v
2
(n!), which can
be translated as a − s
2
(a) + b − s
2
(b) ≤ n − s
2
(n) < n. So, we have
found almost exactly what we needed: a + b < n + s
2
(a) + s
2
(b). Now,
we need another observation: the sum of digits of a number A when
written in binary is at most the number of digits of A in base 2, which
is 1 + [log
2
A] (this follows from the fact that 2
k−1
≤ A < 2
k
, where
41
k is the number of digits of A in base 2). So, we have the estimations
a + b < n + s
2
(a) + s
2
(b) ≤ n + 2 + log
2
ab ≤ n + 2 + 2 log
2
n (since we
have of course a, b ≤ n). And now the conclusion is immediate.
The following problem appeared in Kvant as a hard problem. It took
quite a long time before an olympic found an extraordinary solution. We
shall not present his solution; but another one, even easier.
Example 5. Is there an inﬁnite set of positive integers such that
no matter how we choose some elements of this set, their sum is not an
integer power of exponent at least 2?
Kvant
Solution. Let us take A = ¦2
n
3
n+1
[n ≥ 1¦ If we consider some
diﬀerent numbers from this set, their sum will be of the form 2
x
3
x+1
y,
where (y, 6) = 1. This is surely not a power of exponent at least 2, since
otherwise the exponent should divide both x and x +1. Thus this set is
actually a good choice.
The following problem shows the beauty of elementary number-
theory. It combines diverse ideas and techniques and the result is at
least beautiful. This one is also a classic problem, that appeared in lots
of mathematics competitions.
Example 6. Prove that for any natural number n, n! is a divisor of
n−1

k=0
(2
n
−2
k
).
Solution. So, let us take a prime number p. Of course, for the ar-
gument to be non-trivial, we take p ≤ n (otherwise doesn’t divide n!).
First, let us see what happens with p = 2. We have
v
2
(n!) = n −s
2
(n) ≤ n −1
42
and also
v
2
_
n−1

k=0
(2
n
−2
k
)
_
=
n−1

k=0
v
2
(2
n
−2
k
) ≥ n −1
(since 2
n
−2
k
is even for k ≥ 1), so we are done with this case. Now, let
us assume that p > 2. We have p[2
p−1
−1 from Fermat’s theorem, so we
also have p[2
k(p−1)
−1 for all k ≥ 1. Now,
n−1

k=0
(2
n
−2
k
) = 2
n(n−1)
2
n

k=1
(2
k
−1)
and so, from the above remarks we infer that
v
2
_
n−1

k=0
(2
n
−2
k
)
_
=
n

k=1
v
2
(2
k
−1)

1≤k(p−1)≤n
v
2
(2
k(p−1)
−1) ≥ card¦k[1 ≤ k(p −1) ≤ n¦
Since
card¦k[1 ≤ k(p −1) ≤ n¦ =
_
n
p −1
_
,
we have found that
v
2
_
n−1

k=0
(2
n
−2
k
)
_

_
n
p −1
_
.
But we know that
v
2
(n!) =
n −s
p
(n)
p −1

n −1
p −1
<
n
p −1
and since v
2
(n!) ∈ R, we must have
v
2
(n!) ≤
_
n
p −1
_
.
From these two inequalities, we conclude that
v
2
_
n−1

k=0
(2
n
−2
k
)
_
≥ v
2
(n!)
and now the problem is solved.
43
Diophantine equations can also be solved using the methods em-
ployed in this topic. Here is a diﬃcult one, given in a russian olympiad.
Example 7. Prove that the equation
1
10
n
=
1
n
1
!
+
1
n
2
!
+ +
1
n
k
!
does not have integer solutions such that 1 ≤ n
1
< n
2
< < n
k
.
Solution. Suppose we have found a solution of the equation and let
us consider
P = n
1
!n
2
! . . . n
k
!.
We have
10
n
((n
1
+ 1) . . . (n
k
−1)n
k
+ + (n
k−1
+ 1) . . . (n
k
−1)n
k
+ 1) = n
k
!
which shows that n
k
divides 10
n
. Let us write n
k
= 2
x
5
y
. First of
all, suppose that x, y are positive. Thus, (n
1
+ 1) . . . (n
k
− 1)n
k
+ +
(n
k−1
+ 1) . . . (n
k
− 1)n
k
+ 1 is relatively prime with 10 and it follows
that v
2
(n
k
!) = v
5
(n
k
!). This implies of course that
_
n
k
2
j
_
=
_
n
k
5
j
_
for all
j (because we clearly have
_
n
k
2
j
_
>
_
n
k
5
j
_
) and so n
k
≤ 3. A veriﬁcation
by hand shows that there is no solution in this case.
Next, suppose that y = 0. Then (n
1
+1) . . . (n
k
−1)n
k
+ +(n
k−1
+
1) . . . (n
k
− 1)n
k
+ 1 is odd and thus v
2
(n
k
!) = n ≤ v
5
(n
k
!). Again this
implies v
2
(n
k
!) = v
5
(n
k
!) and we have seen that this gives no solution.
So, actually x = 0. A crucial observation is that if n
k
> n
k−1
+ 1,
then (n
1
+1) . . . (n
k
−1)n
k
+ +(n
k−1
+1) . . . (n
k
−1)n
k
+1 is again
odd and thus we ﬁnd again that v
2
(n
k
!) = n ≤ v
5
(n
k
!), impossible. So,
n
k
= n
k−1
+1. But then, taking into account that n
k
is a power of 5, we
deduce that (n
1
+1) . . . (n
k
−1)n
k
+ +(n
k−1
+1) . . . (n
k
−1)n
k
+1 is
congruent to 2 modulo 4 and thus v
2
(n
k
!) = n+1 ≤ v
5
(n
k
!)+1. It follows
that
_
n
k
2
_
≤ 1 +
_
n
k
5
_
and thus n
k
≤ 6. Since n
k
is a power of 5, we
44
ﬁnd that n
k
= 5, n
k−1
= and a quick research of all possibilities shows
that there are no solutions. Thus, the given equation has no solution in
natural numbers.
A tricky APMO problem asked once upon a time to prove there is a
number 2 < n < 2000 such that n[2
n
+ 2. We will let to the reader the
job to verify that 2 11 43 is a solution (and especially the job to ﬁnd
how we arrived at this number) and also the exercise to prove that there
are actually inﬁnitely many such numbers. Yet... small veriﬁcations show
that all such numbers are even. Proving this turns out to be a diﬃcult
problem and this was proved for the ﬁrst time by Sierpinski.
Note. After the quadratic reciprocity law topic, it will be proved that
2 11 43 is a solution of the problem.
Example 8. Prove that for any n > 1 we cannot have n[2
n−1
+ 1.
Solution. Although very short, the proof is tricky. Let n =
s

i=1
p
k
i
i
where p
1
< < p
s
are prime numbers. The idea is to look at v
2
(p
i
−1).
Choose that p
i
which minimizes this quantity and write p
i
= 1 + 2
r
i
m
i
with m
i
odd. Then of course we have n ≡ 1 (mod 2
m
i
). Hence we can
write n − 1 = 2
m
t. We have 2
2
m
t
≡ −1 (mod p
i
) thus we surely have
−1 ≡ 2
2
m
tm
i
≡ 2
(p
i
−1)t
≡ 1 (mod p
i
) (the last congruence being derived
from Fermat’s theorem). Thus p
i
= 2, which is clearly impossible.
We continue with a very nice and hard problem, in which the idea of
looking at the exponents really saves us. This problem seemed to appear
for the ﬁrst time in AMM , proposed by Armond E. Spencer. In the last
years, it appeared in various contests.
Example 9. Prove that for any integers a
1
, a
2
, . . . , a
n
the number

1≤i<j≤n
a
i
−a
j
i −j
is an integer.
Armond Spencer, AMM E 2637
45
Solution. This time, we consider a prime number p and we prove
that for each k ≥ 1, there are more numbers divisible by p
k
in the se-
quence of diﬀerences (a
i
−a
j
)
1≤i<j≤n
than in the sequence (i−j)
1≤i<j≤n
.
Since
v
p
_
_

1≤i<j≤n
(a
i
−a
j
)
_
_
=

k≥1
N
p
k
_
_

1≤i<j≤n
(a
i
−a
j
)
_
_
(here N
x
_
_

y∈A
y
_
_
is the number of terms from the sequence A that are
multiples of x) and
v
p
_
_

1≤i<j≤n
(i −j)
_
_
=

k≥1
N
p
k
_
_

1≤i<j≤n
(i −j)
_
_
,
the problem will be solved if we prove our claim. Now, let us ﬁx k ≥ 1
and let us suppose that there are exactly b
i
indices j ∈ ¦1, 2, . . . , n¦ such
that a
j
≡ i (mod p
k
), for each i ∈ ¦0, 1, . . . , p
k
−1¦. Then we have
N
p
k
_
_

1≤i<j≤n
(a
i
−a
j
)
_
_
=
p
k
−1

i=0
_
b
i
2
_
.
We see that if a
i
= i, then b
i
=
_
n +i
p
k
_
(there are
_
n +i
p
k
_
numbers
congruent with i (mod p) between 1 and n; any of them is of the form
i +jp, with 0 ≤ j ≤
n −i
p
, of course, if i = 0 we have 1 ≤ j ≤
n
p
). So,
N
p
k
_
_

1≤i<j≤n
(i −j)
_
_
=
p
k
−1

i=0
_
_
_
n
i
p
k
_
2
_
_
and it suﬃces to prove that
p
k
−1

i=0
_
b
i
2
_

p
k
−1

i=0
_
_
_
n
i
p
k
_
2
_
_
.
46
Now, observe that we are practically asked to ﬁnd the minimal
value of
p
k
−1

i=0
_
x
i
2
_
, when
p
k
−1

i=0
x
i
= n (it is clear that
p
k
−1

i=0
b
i
= n =
p
k
−1

i=0
_
n
i
p
k
_
from the deﬁnition of b
i
). For this, let us suppose that
x
1
≤ x
2
≤ ≤ x
p
k
−1
is the n-tuple which attains the minimal
value (such a n-tuple exists since the equation
p
k
−1

i=0
x
i
= n has a ﬁnite
number of solutions). If x
p
k
−1
> x
0
+ 1, then we consider the n-tuple
(x
0
+ 1, x
1
, . . . , x
p
k
−2
, x
p
k
−1
− 1) which has the sum of the components
n, but for which
_
x
0
+ 1
2
_
+
_
x
1
2
_
+ +
_
x
p
k
−2
2
_
+
_
x
p
k
−1
−1
2
_
<
_
x
0
2
_
+
_
x
1
2
_
+ +
_
x
p
k
−2
2
_
+
_
x
p
k
−1
2
_
.
The last inequality is true, since it is equivalent with x
p
k
−1
> x
0
+1, so it
is true. But this contradicts the minimality of (x
0
, x
1
, . . . , x
2
, . . . , x
p
k
−1
).
So, we must have x
p
k
−1
≤ x
0
+ 1 and from here it follows that
x
i
∈ ¦x
0
, x
0
+ 1¦ for all i ∈ ¦0, 1, 2, . . . , p
k
− 1¦. Thus, there is
j ∈ ¦0, 1, 2, . . . , p
k
− 1¦ such that x
0
= x
1
= = x
j
and x
j+1
=
x
j+2
= = x
p
k
−1
= x
0
+ 1. This easily implies that the minimal
n-tuple is in fact
__
n +i
p
k
__
i=0,p
k
−1
and the problem is solved.
Finally, it is time for a challenge.
Example 10. Let a, b two diﬀerent positive rational numbers such
that for inﬁnitely many numbers n, a
n
−b
n
is integer. Then prove that
a, b are also integers.
Solution. Let us start by writing a =
x
z
, b =
y
z
, where x, y, z are
diﬀerent natural numbers relatively prime. We know thus that z
n
[x
n
−y
n
47
for inﬁnitely many numbers n. Let M be the set of those numbers n.
Now, assume that z > 1 and take p a prime divisor of z. Assuming that
p does not divide x, it obviously follows that it can’t divide y. We have
thus two cases:
i) If p = 2, then let n such that 2
n
[x
n
− y
n
and write n = 2
u
n
v
n
,
where v
n
is odd. From the identity
x
2
u
n
v
n
−y
2
u
n
v
n
= (x
v
n
−y
v
n
)(x
v
n
+y
v
n
) . . . (x
2
u
n
−1
v
n
−y
2
u
n
−1
v
n
)
it follows that
v
2
(x
n
−y
n
) = v
2
(x
v
n
−y
v
n
) +
u
n
−1

k=0
v
2
(x
2
k
v
n
+y
2
k
v
n
).
But x
v
n
−1
+ x
v
n
−2
+ + xy
v
n
−2
+ y
v
n
−1
is obviously odd (since
v
n
, x, y are odd), hence
v
2
(x
v
n
−y
v
n
) = v
2
(x −y).
Similarly, we can prove that
v
2
(x
v
n
+y
v
n
) = v
2
(x +y).
Since for k > 0 we have
x
2
k
v
n
+y
2
k
v
n
≡ 2 (mod 4),
we ﬁnally deduce that
2
u
n
v
n
≤ v
2
(x
n
−y
n
) ≤ v
2
(x +y) +v
2
(x −y) +u
n
−1 (∗)
Consequently, (2
u
n
)
n∈M
is bounded, a simple reason being the in-
equality 2
u
n
≤ v
2
(x+y) +v
2
(x−y) +u
n
−1. Hence (u
n
)
n∈M
takes only
a ﬁnite number of values and from (∗) it follows that (v
n
)
n∈M
also takes
a ﬁnite number of values, that is M is ﬁnite.
ii) If p is odd, then let d the smallest positive integer k such that
p[x
k
−y
k
. Then for any n in M we will have p[x
n
−y
n
. Let x = tu, y = tv,
48
where (u, v) = 1. Obviously, tuv is not a multiple of p. It follows then
that p[(u
d
−v
d
, u
n
−v
n
) = u
(n,d)
−v
(n,d)
[x
(n,d)
−y
(n,d)
and by the choice of
d, we must have d[n. Take now n in M and write it in the form n = md,
with m natural. Let A = x
d
, B = y
d
. Then p
m
[p
n
[x
n
− y
n
= A
m
− B
m
and this happens for inﬁnitely many numbers m. Moreover, p[A − B.
Let R the inﬁnite set of those numbers m. For any m in R we have
m ≤ v
p
(A
m
− B
m
). Now, let us write m = p
i
j, where j is relatively
prime with p. We clearly have
A
m
−B
m
= (A
j
−B
j
)
A
pj
−B
pj
A
j
−B
j
. . .
A
jp
i
−B
jp
i
A
jp
i−1
−B
jp
i−1
(we have assumed that i > 1, since the ﬁnal conclusion will be obvi-
ous in any other case). An essential observation is that we cannot have
p
2
[
A
jp
k
−B
jp
k
A
jp
k−1
−B
jp
k−1
for a certain k > 1. Indeed, otherwise we would have
p
2
[A
jp
k
− B
jp
k
⇒ p
2
[A
pj
− B
pj
(Euler’s theorem). Yet, we also have
p
2
[A
jp
k−1
(p−1)
+ A
jp
k−1
(p−2)
B
jp
k−1
+ + b
jp
k−1
(p−1)
. From p
2
[A
j
−B
j
we have
A
jp
k−1
(p−1)
+A
jp
k−1
(p−2)
B
jp
k−1
+ +B
jp
k−1
(p−1)
≡ pA
jp
k−1
(p−1)
(mod p
2
),
so we should have p[A, that is p[x, false.
Let us prove now that we cannot have p
2
[
A
pj
−B
pj
A
j
−B
j
. Indeed, oth-
erwise (since p[A − B), we can write A
j
= B
j
+ wp and then a simple
computation using Newton’s binomial formula shows that
A
pj
−B
pj
A
j
−B
j
= A
j(p−1)
+A
j(p−2)
+ +B
j(p−1)
≡ pB
j(p−1)
+
p −1
2
B
j(p−2)
p
2
≡ pB
j(p−1)
(mod p
2
)
and thus it would follow that p[B, that is p[y, false.
49
After all, we have shown that in this case we must have
m ≤ v
p
(A
m
−B
m
) ≤ v
p
(A
j
−B
j
) +i.
Using again the fact that A ≡ B (mod p), we infer that
A
j−1
+A
j−2
B + +B
j−1
≡ jA
p−1
≡ j (mod p),
which shows that
v
p
(A
j
−B
j
) = v
p
(A−B).
Thus, for inﬁnitely many numbers m we have
m ≤ v
p
(A−B) + [log
2
m],
which is clearly impossible.
Thus, we must have p[x and p[y, contradiction with the fact that
x, y, z are relatively prime. This shows that z = 1 and a, b are integers.
Problems for training
1. Prove the identity
lcm(a, b, c)
2
lcm(a, b) lcm(b, c) lcm(c, a)
=
gcd(a, b, c)
2
gcd(a, b) gcd(b, c) gcd(c, a)
for any positive integers a, b, c.
USAMO, 1972
2. Let a, b, c, d be positive integers such that ab = cd. Prove that
gcd(a, c) gcd(a, d) = a gcd(a, b, c, d).
3. Let a
1
, a
2
, . . . , a
k
, b
1
, b
2
, . . . , b
k
be positive integers such that
gcd(a
i
, b
i
) = 1 for all i ∈ ¦1, 2, . . . , k¦. Let m−lcm[b
1
, b
2
, . . . , b
k
]. Prove
that
gcd
_
a
1
m
1
b
1
,
a
2
m
2
b
2
, . . . ,
a
k
m
k
b
k
_
= gcd(a
1
, a
2
, . . . , a
k
).
IMO Shortlist 1974
50
4. Let n such that 2
n−2005
[n!. Prove that this number has at most
2005 non-zero digits when written in base 2.
5. Prove that for any natural number n we have
(n
2
)!
_
n
n
__
n + 1
n
_
. . .
_
2n −1
n
_
n!
n
∈ R.
R.M Grassl, T. Porter, AMM E 3123
6. Prove the identity
(n + 1)lcm
k=0,n
_
n
k
_
= lcm(1, 2, . . . , n + 1)
for any positive integer n.
Peter L Montgomery, AMM E 2686
7. Let 0 < a
1
< < a
n
be integers. Find the maximal value of the
number m for which we can ﬁnd the integers 0 < b
1
< < b
m
such
that
n

k=1
2
a
k
=
m

k−1
b
k
and
n

k=1
(2
a
k
)! =
m

k=1
b
k
!.
Gabriel Dospinescu
8. Prove that the least common multiple of the numbers 1, 2, . . . , n
equals the least common multiple of the numbers
_
n
1
_
,
_
n
2
_
, . . . ,
_
n
n
_
if and only if n + 1 is a prime.
Laurentiu Panaitopol, TST 1990 Romania
9. Prove that for any n ∈ N we have n!(n + 1)!(n + 2)![(3n)!.
Komal
10. Prove that the product of the numbers between 2
1917
+ 1 and
2
1991
−1 is not a perfect square.
Tournament of the Towns, 1991
11. Show that if n is a positive integer and a and b are integers, then
n! divides a(a +b)(a + 2b) . . . (a + (n −1)b)bn −1.
51
IMO Shortlist, 1985
12. Prove that k!
k
2
+k+1
divides (k
3
)!.
13. Let x, y be relatively prime diﬀerent natural numbers. Prove that
for inﬁnitely many primes p the exponent of p in x
p−1
−y
p−1
is odd.
AMM
14. Let a
1
, . . . , a
n
> 0 such that whenever k is a prime number of a
power of a prime number, we have
_
a
1
k
_
+ +
_
a
n
k
_
< 1.
Prove that there is a unique index i ∈ ¦1, 2, . . . , n¦ such that a
1
+
+a
n
< 1 + [a
i
].
16. Find the exponent of 2 in the decomposition of the number
_
2
n+1
2
n
_

_
2
n
2
n−1
_
.
AMM
17. Prove that (x
n
)
n≥1
the exponent of 2 in the decomposition of
the numerator of
2
1
+
2
2
2
+ +
2
n
n
, goes to inﬁnity as n → ∞. Even
more, prove that x
2
n ≥ 2
n
− n + 1 (hint: try to prove ﬁrst the identity
2
1
+
2
2
2
+ +
2
n
n
=
2
n
n
n−1

k=0
1
_
n −1
k
_
).
18. Prove that the product of at most 25 consecutive integers is not
a square.
Narumi’s theorem
52
PRIMES AND SQUARES
The study of the properties of the prime numbers is so well developed
(yet, many old conjectures and open questions wait their solution), that
some properties have become classical and need to be known. In this unit,
we will try to present a unitary view over the properties of some classes of
primes and also some classical results related to representations as sum
of two squares. These things are not new, but they must be included
in the mathematical culture of a serious problem-solver. Yet, in the end
of the unit, we will discuss as usual some non-classical and surprising
problems. So, don’t skip this unit!
Since we will use some facts several times in this paper, we prefer to
make some notations before discussing the problems. So, we will consider
A, B the sets of all prime numbers of the form 4k + 1 and 4k + 3,
respectively. Also, let C be the set of all numbers which can be written as
the sum of two perfect squares. Our purpose is to present some classical
things related to A, B, C. The most spectacular property of the set A
is surely the fact that any element is the sum of two squares of positive
integers. This is not a trivial property and we will see a beautiful proof
for this theorem of Fermat, which is far from easy.
Example 1. Prove that A is a subset of C.
Solution. Thus, we need to prove that any prime number of the form
4k+1 is the sum of two squares. We will use a very nice theorem of Thue,
which says that if n is a positive integer and a is relatively prime with
n, then there exist integers 0 < x, y ≤

n such that xa ≡ ±y (mod n)
for a suitable choice of the signs + and −. The proof is simple, but the
theorem itself is a diamond. Indeed, let us consider all the pairs xa −y,
with 0 ≤ x, y ≤ [

n]. So, we have a list of ([

n] +1)
2
> n numbers and
it follows that two numbers among them give the same remainder when
divided by n, let them be ax
1
−y
1
and ax
2
−y
2
. It is not diﬃcult to see
53
that we may assume that x
1
> x
2
(we certainly cannot have x
1
= x
2
or
y
1
= y
2
). If we take x = x
1
− x
2
, y = [y
1
− y
2
[, all the conditions are
satisﬁed and the theorem is proved.
We will use now Wilson’s theorem to ﬁnd an integer n such that
p[n
2
+ 1. Indeed, let us write p = 4k + 1 and observe that we can take
n = (2k)!. Why? Because from Wilson’s theorem we have
−1 ≡ (p −1)! (mod p) ≡ 1 2 . . .
_
p −1
2
__
p −
p −1
2
_
. . . (p −1)
≡ (−1)
p−1
2
_
p −1
2
_
!
2
≡ (2k)!
2
(mod p)
and the claim is proved. Now, since p[n
2
+ 1, it is clear that p and n
are relatively prime. Hence we can apply Thue’s theorem and we ﬁnd
the existence of positive integers 0 < x, y <

p (since

p ,∈ R) such
that p[n
2
x
2
−y
2
. Because p[n
2
+ 1, we ﬁnd that p[x
2
+ y
2
and because
0 < x, y <

p, we conclude that we have in fact p = x
2
+ y
2
. The
theorem is proved.
Now, it is time now to study some properties of the set B. Since they
are easier, we will discuss them all in a single example.
Example 2. Let p ∈ B and suppose that x, y are integers such that
p[x
2
+ y
2
. Then p[(x, y). Consequently, any number of the form n
2
+ 1
has only prime factors that belong to A or are equal to 2. Conclude that
A is inﬁnite and then that B is inﬁnite.
Solution. Let us focus on the ﬁrst question. Suppose that p[(x, y)
is not true. Then, it is obvious that xy is not a multiple of p. Because
p[x
2
+ y
2
, we can write x
2
≡ −y
2
(mod p). Combining this with the
observation that (x, p) = (y, p) = 1 and with Fermat’s theorem, we
ﬁnd that 1 ≡ x
p−1
≡ (−1)
p−1
2
≡ (−1)
p−1
2
≡ −1 (mod p), which is
clearly impossible. This settles the ﬁrst question. The second one follows
clearly from the ﬁrst one. Now, it remains to prove the third assertion.
54
Proving that B is inﬁnite is almost identical with the proof that there
exist inﬁnitely many primes. Indeed, suppose that p
1
, p
2
, . . . , p
n
are all
the elements of B greater than 3 and consider the odd number N =
4p
1
p
2
. . . p
n
+ 3. Because N ≡ 3 (mod 4), N must have a prime factor
that belongs to B. But since p
i
is not a divisor of N for any i = 1, n
the contradiction is reached and thus B is inﬁnite. In the same manner
we can prove that A is inﬁnite, but this time we must use the second
question. Indeed, we consider this time the number M = (q
1
q
2
. . . q
m
)
2
+
1, where q
1
, q
2
, . . . , q
m
are all the elements of A and then simply apply
the result from the second question. The conclusion is plain.
It is not diﬃcult to characterize the elements of the set C. A number
is a sum of two squares if and only if any prime factor of it that also
belongs to B appears at an even exponent in the decomposition of that
number. The proof is just a consequence of the ﬁrst examples and we
will not insist. Having presented some basic results that we will use in
this unit, it is time to see how many applications these two examples
have. An easy consequence of the previous observations is the following.
As a simple application of the ﬁrst example, we consider the following
problem, which is surely easy for someone who knows Fermat’s theorem
regarding the elements of A and very diﬃcult otherwise.
Example 3. Find the number of integers x ∈ ¦−1997, . . . , 1997¦ for
which 1997[x
2
+ (x + 1)
2
.
India, 1998
Solution. We know that any congruence of the second degree re-
duces to the congruence x
2
≡ a (mod p). So, let us proceed and reduce
the given congruence to this special form. This is not diﬃcult, since
x
2
+(x+1)
2
≡ 0 (mod 1997) is of course equivalent to 2x
2
+2x+1 ≡ 0
(mod 1997), which in turn becomes (2x + 1)
2
+ 1 ≡ 0 (mod 1997).
Since 1997 ∈ A, the congruence n
2
≡ −1 (mod 1997) surely has at
55
least a solution. More precisely, there are exactly two solutions that
belong to ¦1, 2, . . . , 1996¦ because if n
0
is a solution, so is 1997 − n
0
and it is clear that it has at most two non-congruent solutions mod
1997. Because (2, 1997) = 1, the function x → 2x + 1 is a permutation
of R
1997
and so the initial congruence has exactly two solutions with
x ∈ ¦1, 2, . . . , 1996¦. In a similar way, we ﬁnd that there are exactly two
solutions with x ∈ ¦−1997, −1996, . . . , −1¦. Therefore there are exactly
four numbers x ∈ ¦−1997, . . . , 1997¦ such that 1997[x
2
+ (x + 1)
2
.
From a previous observation, we know that the condition that a
number is a sum of two squares is quite restrictive. This suggests that
the set X is quite RARA. This conclusion can be translated in the
following nice problem.
Example 4. Prove that C doesn’t have bounded gaps, that is there
are arbitrarily long sequences of integers, no term of which can be written
as the sum of two perfect squares.
AMM
Solution. The statement of the problem suggests using the Chinese
Remainder Theorem, but here the main idea is to use the complete
characterization of the set C, that we have just discussed: C = ¦n ∈
R[ if p[n and p ∈ B, then v
p
(n) ∈ 2R¦. Hence we know what we have
to do. We will take long sequences of consecutive integers, each of them
having a prime factor that belongs to B and has exponent 1. More
precisely, we take diﬀerent elements of B, let them p
1
, p
2
, . . . , p
n
(we can
take as many as we need, since B is inﬁnite) and then we look for a
56
solution of the system of congruences
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
x ≡ p
1
−1 (mod p
2
1
)
x ≡ p
2
−2 (mod p
2
2
)
. . .
x ≡ p
n
−n (mod p
2
n
)
The existence of such a solution follows from the Chinese Remainder
Theorem. Thus, the numbers x+1, x+2, . . . , x+n cannot be written as
the sum of two perfect squares, since p
i
[x
i
, but p
2
i
does not divide x +i.
Since n is as large as we want, the conclusion follows.
The Diophantine equation x(x +1)(x +2) . . . (x +n) = y
k
has been
extensively studied by many mathematicians and great results have been
obtained. But these results are very diﬃcult to prove and we prefer to
present a related problem, with a nice ﬂavor of elementary mathematics.
Example 5. Prove that a set of p −1 consecutive positive integers,
where p ∈ B, cannot be partitioned into two subsets, each having the
same product of the elements.
Solution. Let us suppose that the positive integers x + 1, x +
2, . . . , x + p − 1 have been partitioned into two classes X, Y , each of
them having the same product of the elements. If at least one of them is
a multiple of p, then there must be another one divisible by p (since in
this case both the products of elements from X and Y must be multi-
ples of p), which is clearly impossible. Thus, none of these numbers is a
multiple of p, which means that the set of remainders of these numbers
when divided by p is exactly 1, 2, . . . , p −1. Also, from the hypothesis it
follows that there exists a positive integer n such that
(x + 1)(x + 2) . . . (x +p −1) = n
2
.
Hence n
2
≡ 1 2(p − 1) ≡ −1 (mod p), the last congruence being
true by Wilson’s theorem. But from the second example we know that
57
the congruence n
2
≡ −1 (mod p) is impossible for p ∈ B and this is the
The results stated in the second example are an useful tool in solving
non-standard Diophantine equations. The technique is better explained
in the following two examples.
Example 6. Prove that the equation x
4
= y
2
+z
2
+4 does not have
integer solutions.
Reid Barton, Rookie Contest, 1999
Solution. Practically, we have to show that x
4
−4 does not belong
to C. Hence we need to ﬁnd an element of B that has an odd exponent
in the decomposition of x
4
−4. The ﬁrst case is when x is odd. Using the
factorization x
4
−4 = (x
2
−2)(x
2
+2) and the observation that x
2
+2 ≡ 3
(mod 4), we deduce that there exists p ∈ B such that v
p
(x
2
+2) is odd.
But since p cannot divide x
2
−2 (otherwise p[x
2
+2 −(x
2
−2), which is
not the case), we conclude that v
p
(x
4
−4) is odd and so x
4
−4 does not
belong to C. We have thus shown that in any solution of the equation x
is even, let us say x = 2k. Then, we must also have 4k
4
−1 ∈ C, which
is clearly impossible since 4k
4
− 1 ≡ 3 (mod 4) and thus 4k
4
− 1 has
a prime factor that belongs to B and has odd exponent. Moreover, it
worth noting that the equation x
2
+y
2
= 4k+3 can be solved directly, by
working modulo 4. We leave to the reader the details, which are trivial.
The following problem is much more diﬃcult, but the basic idea is
the same. Yet, the details are not so obvious and, most important, it
is not clear how to begin. It has become a classical problem due to its
beauty and diﬃculty.
Example 7. Let p ∈ B and suppose that x, y, z, t are integers such
that x
2p
+y
2p
+z
2p
= t
2p
. Prove that at least one of the numbers x, y, z, t
is a multiple of p.
58
Barry Powel, AMM
Solution. First of all, we observe that it is enough to assume that
x, y, z, t are relatively prime. Next, we prove that t is odd. Supposing
the contrary, we obtain that x
2p
+ y
2p
+ z
2p
≡ 0 (mod 4). Since a
2
(mod 4) ∈ ¦0, 1¦, the latter implies that x, y, z are even, contradicting
the assumption that (x, y, z, t) = 1. Hence t is odd. This implies that
at least one of the numbers x, y, z is odd. Suppose that it is z. Now,
another step is required. We write the equation in the form
x
2p
+y
2p
=
t
2p
−z
2p
t
2
−z
2
(t
2
−z
2
)
and we look for a prime number q ∈ B with an odd exponent in the
decomposition of a factor that appears in the right-hand side. The best
candidate for this factor seems to be
t
2p
−z
2p
t
2
−z
2
= (t
2
)
p−1
+ (t
2
)
p−2
z
2
+ + (z
2
)
p−1
,
which is congruent to 3 (mod 4). This follows from the hypothesis p ∈ B
and the fact that a
2
≡ 1 (mod 4) for any odd number a. Thus, there
exists q ∈ B such that v
q
_
t
2p
−z
2p
t
2
−z
2
_
is odd. Since x
2p
+ y
2p
∈ C, it
follows that v
q
(x
2p
+y
2p
) is even and so v
q
(t
2
−z
2
) is odd. In particular
q[t
2
−z
2
and, because q[(t
2
)
p−1
+(t
2
)
p−2
z
2
+ +(z
2
)
p−1
, we deduce that
q[pt
2(p−1)
. If q ,= p, then q[t, hence q[z and also q[x
2p
+y
2p
. Because q ∈
B, we infer that q[(x, y, z, t) = 1, which is clearly impossible. Therefore
q = p and so p[x
2p
+y
2p
. Because p ∈ B, we ﬁnd that p[x and p[y. The
conclusion follows.
It’s time for a hard problem.
Example 8. Find the smallest nonnegative integer n for which there
exists a non-constant function f : Z → [0, ∞) with the following prop-
erties:
a) f(xy) = f(x)f(y);
59
b) 2f(x
2
+y
2
) −f(x) −f(y) ∈ ¦0, 1, . . . , n¦ for all integers x and y.
For this n, ﬁnd all the functions with the above properties.
Solution. We will use all results proved in the beginning of the note.
First, we will prove that for n = 1 there are functions which verify a)
and b). We remind that A and B are the sets of all primes of the form
4k + 1 and 4k + 3, respectively. For any p ∈ B we deﬁne:
f
p
: Z →Z, f
p
(x) =
_
0, if p[x
1, otherwise
Using properties of sets A and B, one can easily verify that f
p
veriﬁes
the restrictions of the problem. Hence f
p
is a solution of the problem for
any p ∈ B.
We will prove now that if f is non-constant and veriﬁes the conditions
of the problem, then n > 0. Suppose not. Then 2f(x
2
+y
2
) = f(x)+f(y)
and hence 2f
2
(x) = 2f(x
2
+ 0
2
) = f(x) + f(0). It is clear that we
have f
2
(0) = f(0). Since f is non-constant, we must have f(0) = 0.
Consequently, we must have 2f
2
(x) = f(x) for every integer x. But if
there exists x such that f(x) =
1
2
, then f
2
(x
2
) ,= 2f(x
2
Thus, f(x) = 0 for any integer x and f is constant, contradiction. So,
n = 1 is the smallest number for which there are non-constant functions
which verify a) and b).
We will prove now that any non-constant function f which veriﬁes
a) and b) must be of the form f
p
. We have already seen that f(0) = 0.
Since f
2
(1) = f(1) and f is non-constant, we must have f(1) = 1. Also,
2f
2
(x) −f(x) = 2f(x
2
+ 0
2
) −f(x) −f(0) ∈ ¦0, 1¦ for every integer x.
Thus, f(x) ∈ ¦0, 1¦.
Since
f
2
(−1) = f(1) = 1 and f(−1) ∈ [0, ∞), we must have f(−1) = 1
and f(−x) = f(−1)f(x) = f(x) for any integer x. Then, since f(xy) =
f(x)f(y), it is enough to ﬁnd f(p) for any prime p. We prove that there is
60
exactly one prime number p for which f(p) = 0. Since f is non-constant,
there exists a prime number p for which f(p) = 0. Suppose there is
another prime q for which f(q) = 0. Then 2f(p
2
+ q
2
) ∈ ¦0, 1¦, which
means f(p
2
+q
2
) = 0. Then for any integers a and b we must have: 0 =
2f(a
2
+b
2
)f(p
2
+q
2
) = 2f((ap+bq)
2
+(aq−bp)
2
). Since 0 ≤ f(x)+f(y) ≤
2f(x
2
+y
2
) for any x and y, we must have f(ap +bq) = f(aq −bp) = 0.
Since p and q are relatively prime, there are integers a and b such that
aq −bp = 1. Then we have 1 = f(1) = f(aq −bp) = 0, contradiction. So
,there is exactly one prime number p for which f(p) = 0. Let us suppose
that p = 2. Then f(x) = 0 for any even x and 2f(x
2
+y
2
) = 0 for any odd
numbers x and y. This implies that f(x) = f(y) = 0 for any odd numbers
x and y and thus f is constant, contradiction. Therefore p ∈ A ∪ B.
Suppose p ∈ A. According to proposition 2, there are positive integers
a and b such that p = a
2
+b
2
. Then we must have f(a) = f(b) = 0. But
max¦a, b¦ > 1 and there is a prime number q such that q[ max¦a, b¦ and
f(q) = 0 (otherwise, we would have f(max¦a, b¦ = 1). But it is clear
that q < p and thus we have found two distinct primes p and q such
that f(p) = f(q) = 0, which, as we have already seen, is impossible.
Consequently, p ∈ B and we have f(x) = 0 for any x divisible by p and
f(x) = 1 for any x which is not divisible by p. Hence, f must be f
p
and
the conclusion follows.
Problems for training
1. Prove that if p ∈ A, then it can be represented in exactly one way
as the sum of the squares of two integers, except for the order of the
terms.
2. Prove that a positive integer can be written as the sum of two
perfect squares if and only if it can be written as the sum of the squares
of two rational numbers.
61
Euler
3. Find all positive integers n with the property that the equation
n = x
2
+y
2
, where 0 ≤ x ≤ y and (x, y) = 1 has exactly one solution.
4. Here is another proof of the theorem from example 1. Suppose
that p = 4k + 1 ∈ A and let x, y ∈ Z such that max¦[x[, [y[¦ <
p
2
and
2xε
_
2k
k
_
(mod p), y ≡ (2k)!x (mod p). Prove that p = x
2
+y
2
.
Gauss
5. Find all pairs of positive integers (m, n) such that
m
2
−1[3
m
+ (n! −1)
m
.
Gabriel Dospinescu
6. The positive integers a, b have the property that the numbers
15a + 16b and 16a − 15b are both perfect squares. What is the least
possible value that can be taken on by the smallest of the two squares?
IMO CE AN?
7. Prove that the number 4mn −m−n cannot be a perfect square
if m, n are positive integers.
IMO 1984 Shortlist
8. Find all n-tuples of positive integers (a
1
, a
2
, . . . , a
n
) such that
(a
1
! −1)(a
2
! −1) . . . (a
n
! −1) −16
is a perfect square.
Gabriel Dospinescu
9. Find all pairs of positive integers (x, y) such that the number
x
2
+y
2
x −y
is a divisor of 1995.
Bulgaria, 1995
62
10. Prove that the equation y
2
= x
5
− 4 does not have integer
solutions.
Balkan, 1998
11. Solve in integer numbers the equation x
2
= y
7
+ 7.
ROMOP, 2001
12. Find all positive integers n such that the number 2
n
− 1 has a
multiple of the form m
2
+ 9.
IMO Shortlist, 1999
13. Prove that there exists inﬁnitely many pairs of consecutive num-
bers, no two of them having any prime factor that belongs to B.
14. Prove that if n
2
+a ∈ C for any positive integer n, then a ∈ C.
Gabriel Dospinescu
15. Let T the set of the positive integers n for which the equation
n
2
= a
2
+b
2
has solutions in positive integers. Prove that T has density
1.
Moshe Laub, 6583
16. a) Prove that for any real number x and any natural number N
one can ﬁnd integer numbers p, q such that [qx −p[ ≤
1
N + 1
.
b) Suppose that a ∈ Z is a divisor of a number of the form n
2
+ 1.
Then prove that a ∈ C.
17. Find all functions f : N →Z with the properties:
1. if a[b then f(a) ≥ f(b)
2. for any natural numbers a, b we have
f(ab) +f(a
2
+b
2
) = f(a) +f(b).
63
18. (for the die hards) Let L
0
= 2, L
1
= 1 and L
n+2
= L
n+1
+ L
n
be the famous Lucas’s sequence. Then the only n > 1 such that L
n
is a
perfect square is n = 3.
Cohn’s theorem
64
T
2
’S LEMMA
T
2
’s lemma is clearly a direct application of the Cauchy-Schwarz in-
equality. Some will say that it is actually the Cauchy-Schwarz inequality
and they are not wrong. Anyway, this particular lemma has become very
popular among the American students who attended the training of the
USA IMO team. This happened after a lecture delivered by the ﬁrst
author at the Mathematical Olympiad Summer Program (MOSP) held
at Georgetown University in June, 2001.
But what exactly does this lemma say? It says that for any real
numbers a
1
, a
2
, . . . , a
n
and any positive real numbers x
1
, x
2
, . . . , x
n
the
inequality
a
2
1
x
1
+
a
2
2
x
2
+ +
a
2
n
x
n

(a
1
+a
2
+ +a
n
)
2
x
1
+x
2
+ +x
n
(1)
holds. And now we see why calling it also the Cauchy-Schwarz inequality
is natural, since it is practically an equivalent form of this inequality:
_
a
2
1
x
1
+
a
2
2
x
2
+ +
a
2
n
x
n
_
(x
1
+x
2
+ +x
n
)

_
_
¸
a
2
1
x
1

x
1
+
¸
a
2
2
x
2

x
2
+ +
¸
a
2
n
x
n

x
n
_
_
2
.
But there is another nice proof of (1), by induction. The inductive
step is reduced practically to the case n = 2, which is immediate. Indeed,
it boils down to (a
1
x
2
−a
2
x
1
)
2
≥ 0 and the equality occurs if and only
if
a
1
x
1
=
a
2
x
2
. Applying this result twice it follows that
a
2
1
x
1
+
a
2
2
x
2
+
a
2
3
x
3

(a
1
+a
2
)
2
x
1
+x
2
+
a
2
3
x
3

(a
1
+a
2
+a
3
)
2
x
1
+x
2
+x
3
and we see that a simple inductive argument ﬁnishes the proof. With this
brief introduction, let us discuss some problems. And there are plenty
65
of them given in mathematical contests or proposed in mathematical
magazines!
First, an old problem, that became classical. We will see that with
T
2
’s lemma it becomes straightforward and even more, we will obtain a
reﬁnement of the inequality.
Example 1. Prove that for any positive real numbers a, b, c
a
3
a
2
+ab +b
2
+
b
3
b
2
+bc +c
2
+
c
3
c
2
+ca +a
2

a +b +c
3
.
Tournament of the Towns, 1998
Solution. We will change the left-hand side of the inequality so that
we could apply T
2
’s lemma. This is not diﬃcult: we just have to write it
in the form
a
4
a(a
2
+ab +b
2
)
+
b
4
b(b
2
+bc +c
2
)
+
c
4
c(c
2
+ca +a
2
)
.
It follows that the left-hand side is greater than or equal to
(a
2
+b
2
+c
2
)
2
a
3
+b
3
+c
3
+ab(a +b) +bc(b +c) +ca(c +a)
But we can easily observe that
a
3
+b
3
+c
3
+ab(a +b) +bc(b +c) +ca(c +a) = (a +b +c)(a
2
+b
2
+c
2
),
so we have proved an even stronger inequality, that is
a
3
a
2
+ab +b
2
+
b
3
b
2
+bc +c
2
+
c
3
c
2
+ca +a
2

a
2
+b
2
+c
2
a +b +c
.
The second example also became representative for a whole class of
problems. There are countless examples of this type in numerous contests
and mathematical magazines, so we ﬁnd it necessary to discuss it at this
point.
66
Example 2. For arbitrary positive real numbers a, b, c, d prove the
inequality
a
b + 2c + 3d
+
b
c + 2d + 3a
+
c
d + 2a + 3b
+
d
a + 2b + 3c

2
3
.
Titu Andreescu, IMO 1993 Shortlist
Solution. If we write the left-hand side in the form
a
2
a(b + 2c + 3d)
+
b
2
b(c + 2d + 3a)
+
c
2
c(d + 2a + 3b)
+
d
2
d(a + 2b + 3c)
,
then the way to continue is clear, since from the lemma we obtain
a
b + 2c + 3d
+
b
c + 2d + 3a
+
c
d + 2a + 3b
+
d
a + 2b + 3c

(a +b +c +d)
2
4(ab +bc +cd +da +ac +bd)
.
Hence it suﬃces to prove the inequality
3(a +b +c +d)
2
≥ 8(ab +bc +cd +da +ac +bd).
But it is not diﬃcult to see that
(a +b +c +d)
2
= a
2
+b
2
+c
2
+d
2
+ 2(ab +bc +cd +da +ac +bd),
implies
8(ab +bc +cd +da +ac +bd) = 4(a +b +c +d)
2
−4(a
2
+b
2
+c
2
+d
2
).
Consequently, we are left with the inequality
4(a
2
+b
2
+c
2
+d
2
) ≥ (a +b +c +d)
2
,
which is just the Cauchy-Schwarz inequality for four variables.
The problem below, given at the IMO 1995, was discussed exten-
sively in many publications. It could be also solved by using the above
lemma.
67
Example 3. Let a, b, c be positive real numbers such that abc = 1.
Prove that
1
a
3
(b +c)
+
1
b
3
(c +a)
+
1
c
3
(a +b)

3
2
.
Solution. We have:
1
a
3
(b +c)
+
1
b
3
(c +a)
+
1
c
3
(a +b)
=
1
a
2
a(b +c)
+
1
b
2
b(c +a)
+
1
c
2
c(c +a)

_
1
a
+
1
b
+
1
c
_
2
2(ab +bc +ca)
=
(ab +bc +ca)
2
2(ab +bc +ca)
=
ab +bc +ca
2

3
2
,
the last inequality following from the AM-GM inequality.
The following problem is also not diﬃcult, but it uses a nice combi-
nation between this lemma and the Power-Mean inequality. It is another
example in which proving the intermediate inequality (that is, the in-
equality that remains to be proved after using the lemma) is not diﬃcult.
Example 4. Let n ≥ 2. Find the minimal value of the expression
x
5
1
x
2
+x
3
+ +x
n
+
x
5
2
x
1
+x
3
+ +x
n
+ +
x
5
n
x
1
+x
2
+ +x
n−1
,
where x
1
, x
2
, . . . , x
n
are positive real numbers satisfying x
2
1
+x
2
2
+ +
x
2
n
= 1.
Turkey, 1997
Solution. Usually, in such problems the minimal value is attained
when the variables are equal. So, we conjecture that the minimal value
is
1
n(n −1)
attained when x
1
= x
2
= = x
n
=
1

n
. Indeed, by using
the lemma, it follows that the left-hand side is greater than or equal to
_
n

i=1
x
3
i
_
2
n

i=1
x
i
(x
1
+ +x
i−1
+x
i+1
+ +x
n
)
.
68
But it is not diﬃcult to observe that
n

i=1
x
i
(x
1
+ +x
i−1
+x
i+1
+ +x
n
) =
_
n

i=1
x
i
_
2
−1.
So, proving that
x
5
1
x
2
+x
3
+ +x
n
+
x
5
2
x
1
+x
3
+ +x
n
+ +
x
5
n
x
1
+x
2
+ +x
n−1

1
n(n −1)
reduces to proving the inequality
_
n

i=1
x
3
i
_
2

_
n

i=1
x
i
_
2
−1
n(n −1)
.
But this is a simple consequence of the Power-Mean inequality. In-
deed, we have
_
_
_
_
_
_
n

i=1
x
3
i
n
_
_
_
_
_
_
1
3

_
_
_
_
_
_
n

i=1
x
2
i
n
_
_
_
_
_
_
1
2

n

i=1
x
i
n
,
implying
n

i=1
x
3
i

1

n
and
n

i=1

x
i

n.
The conclusion follows.
In 1954, H.S.Shapiro asked whether the following inequality is true
for any positive real numbers a
1
, a
2
, . . . , a
n
:
a
1
a
2
+a
3
+
a
2
a
3
+a
4
+ +
a
n
a
1
+a
2

n
2
.
The question turned out to be extremely diﬃcult. The answer is
really unexpected: one can prove that the inequality is true for all n =
3, 4, 5, 6, 7 (and for all small values of n the shortest proof is based on
69
this lemma), but it is false for all even numbers n ≥ 14 as well as for
suﬃciently large odd numbers n. Let us examine the case n = 5, a
problem proposed for MOSP 2001.
Example 5. Prove that for any positive real numbers a
1
, a
2
, a
3
, a
4
,
a
5
,
a
1
a
2
+a
3
+
a
2
a
3
+a
4
+
a
3
a
4
+a
5
+
a
4
a
5
+a
1
+
a
5
a
1
+a
2

5
2
.
Solution. Again, we apply the lemma and we conclude that it suf-
ﬁces to prove the inequality
(a
1
+a
2
+a
3
+a
4
+a
5
)
2

5
2
[a
1
(a
2
+a
3
) +a
2
(a
3
+a
4
) +a
3
(a
4
+a
5
) +a
4
(a
5
+a
1
) +a
5
(a
1
+a
2
)]
Let us denote a
1
+a
2
+a
3
+a
4
+a
5
= S. Then we observe that
a
1
(a
2
+a
3
) +a
2
(a
3
+a
4
) +a
3
(a
4
+a
5
) +a
4
(a + 5 +a
1
) +a
5
(a
1
+a
2
)
=
a
1
(S −a
1
) +a
2
(S −a
2
) +a
3
(S −a
3
) +a
4
(S −a
4
) +a
5
(S −a
5
)
2
=
S
2
−a
2
1
−a
2
2
−a
2
3
−a
2
4
−a
2
5
2
.
With this identity, we infer that the intermediate inequality is in fact
(a
1
+a
2
+a
3
+a
4
+a
5
)
2

5
4
(S
2
−a
2
1
−a
2
2
−a
2
3
−a
2
4
−a
2
5
),
equivalent to 5(a
2
1
+a
2
2
+a
2
3
+a
2
4
+a
2
5
) ≥ S
2
, which is nothing else then
the Cauchy-Schwarz inequality.
Another question arises: is there a positive real number such that
for any positive real numbers a
1
, a
2
, . . . , a
n
and any n ≥ 3 the following
inequality holds:
a
1
a
2
+a
3
+
a
2
a
3
+a
4
+ +
a
n
a
1
+a
2
≥ cn.
This time, the answer is positive, but ﬁnding the best such constant
is an extremely diﬃcult task. It was ﬁrst solved by Drinﬁeld (who, by
the way, is a Fields’ medalist). The answer is quite complicated and we
70
will not discuss it here (for a detailed presentation of Drinﬁeld’s method
the interested reader can consult the written examination given at ENS
in 1997). The following problem, given at the Moldavian TST in 2005,
shows that c =

2 −1 is such a constant (not optimal).
For any a
1
, a
2
, . . . , a
n
and any n ≥ 3 the following inequality holds:
a
1
a
2
+a
3
+
a
2
a
3
+a
4
+ +
a
n
a
1
+a
2
≥ (

2 −1)n.
The proof is completely elementary, yet very diﬃcult to ﬁnd. An in-
genious argument using the arithmetic-geometric means inequality does
the job: let us write the inequality in the form
a
1
+a
2
+a
3
a
2
+a
3
+
a
2
+a
3
+a
4
a
3
+a
4
+ +
a
n
+a
1
+a
2
a
1
+a
2

2 n.
Now, using the AM-GM inequality, we see that it suﬃces to prove
the stronger inequality:
a
1
+a
2
+a
3
a
2
+a
3

a
2
+a
3
+a
4
a
3
+a
4
. . .
a
n
+a
1
+a
2
a
1
+a
2
≥ (

2)
n
.
Observe that
(a
i
+a
i+1
+a
i+2
)
2
=
_
a
i
+
a
i+1
2
+
a
i+1
2
+a
i+2
_
2
≥ 4
_
a
i
+
a
i+1
2
__
a
i+1
2
+a
i+2
_
(the last inequality being again a consequence of the AM-GM inequal-
ity). Thus,
n

i=1
(a
i
+a
i+1
+a
i+2
)
2

n

i=1
(2a
i
+a
i+1
)
n

i=1
(2a
i+2
+a
i+1
).
Now, the real trick is to rewrite appropriately the last products. Let
us observe that
n

i=1
(2a
i+2
+a
i+1
) =
n

i=1
(2a
i+1
+a
i
),
71
so
n

i=1
(2a
i
+a
i+1
)
n

i=1
(2a
i+2
+a
i+1
) =
n

i=1
[(2a
i
+a
i+1
)(a
i
+ 2a
i+1
)]

n

i=1
(2(a
i
+a
i+1
)
2
) = 2
n
_
n

i=1
(a
i
+a
i+1
)
_
2
.
The conclusion now follows.
This lemma came handy even at the IMO 2005 (problem 3). In order
to prove that for any positive real numbers x, y, z such that xyz ≥ 1 the
following inequality holds

x
2
+y
2
+z
2
x
5
+y
2
+z
2
≤ 3,
a few students successfully used the above mentioned lemma. For exam-
ple, a student from Ireland applied this result and called it ”SQ Lemma”.
During the coordination, the Irish deputy leader explained what ”SQ”
stood for: ”...escu”. A typical solution using this lemma is as follows:
x
5
+y
2
+z
2
=
x
4
1
x
+
y
4
y
2
+
z
4
z
2

(x
2
+y
2
+z
2
)
2
1
x
+y
2
+z
2
,
hence

x
2
+y
2
+z
2
x
5
+y
2
+z
2

1
x
+y
2
+z
2
x
2
+y
2
+z
2
= 2 +
xy +yz +zx
xyz(x
2
+y
2
+z
2
)
≤ 3.
It is now time for the champions. We begin with a diﬃcult geometric
inequality for which we have found a direct solution using T
2
’s lemma.
Here it is.
Example 6. Prove that in any triangle ABC the following inequality
holds
r
a
r
b
m
a
m
b
+
r
b
r
c
m
b
m
c
+
r
c
r
a
m
c
m
a
≥ 3.
Ji Chen, Crux Mathematicorum
72
Solution. Of course, we start by translating the inequality into an
algebraic one. Fortunately, this is not diﬃcult, since using Heron’s rela-
tion and the formulas
r
a
=
K
s −a
, m
a
=

2b
2
+ 2c
2
−a
2
2
and the likes the desired inequality takes the equivalent form
(a +b +c)(b +c −a)

2a
2
+ 2b
2
−c
2

2a
2
+ 2c
2
−b
2
+
(a +b +c)(c +a −b)

2b
2
+ 2a
2
−c
2

2b
2
+ 2c
2
−a
2
+
(a +b +c)(a +b −c)

2c
2
+ 2b
2
−a
2

2c
2
+ 2a
2
−b
2
≥ 3.
In this form, the inequality is more that monstrous, so we try to
see if a weaker form holds, by applying the AM-GM inequality to each
denominator. So, let us try to prove the stronger inequality
2(a +b +c)(c +b −a)
4a
2
+b
2
+c
2
+
2(a +b +c)(c +a −b)
4b
2
+c
2
+a
2
+
2(a +b +c)(a +b −c)
4c
2
+a
2
+b
2
≥ 3.
Written in the more appropriate form
c +b −a
4a
2
+b
2
+c
2
+
c +a −b
4b
2
+c
2
+a
2
+
a +b −c
4c
2
+a
2
+b
2

3
2(a +b +c)
we see that by T
2
’s lemma the left-hand side is at least
(a + b + c)
2
(b + c −a)(4a
2
+ b
2
+ c
2
) + (c + a −b)(4b
2
+ a
2
+ c
2
) + (a + b −c)(4c
2
+ a
2
+ b
2
)
.
Basic computations show that the denominator of the last expression
is equal to
4a
2
(b +c) + 4b
2
(c +a) + 4c
2
(a +b) −2(a
3
+b
3
+c
3
)
and consequently the intermediate inequality reduces to the simpler form
3(a
3
+b
3
+c
3
) + (a +b +c)
3
≥ 6[a
2
(b +c) +b
2
(c +a) +c
2
(a +b)].
73
Again, we expand (a +b +c)
3
and obtain the equivalent inequality
4(a
3
+b
3
+c
3
) + 6abc ≥ 3[a
2
(b +c) +b
2
(c +a) +c
2
(a +b)],
which is not diﬃcult at all. Indeed, it follows from the inequalities
4(a
3
+b
3
+c
3
) ≥ 4[a
2
(b +c) +b
2
(c +a) +c
2
(a +b)] −12abc
and
a
2
(b +c) +b
2
(c +a) +c
2
(a +b) ≥ 6abc.
The ﬁrst one is just an equivalent form of Schur’s inequality, while
the second follows immediately from the identity
a
2
(b +c) +b
2
(c +a) +c
2
(a +b) −6abc = a(b −c)
2
+b(c −a)
2
+c(a −b)
2
.
After all, we have managed to prove the intermediate inequality,
hence the problem is solved.
The journey continues with a very diﬃcult problem, given at the
Japanese Mathematical Olympiad in 1997 and which became famous due
to its diﬃculty. We will present two solutions for this inequality. The ﬁrst
one uses a nice combination between this lemma and the substitution
discussed in the unit ”Two useful substitutions”.
Example 7. Prove that for any positive real numbers a, b, c the
following inequality holds
(b +c −a)
2
a
2
+ (b +c)
2
+
(c +a −b)
2
b
2
+ (c +a)
2
+
(a +b −c)
2
c
2
+ (a +b)
2

3
5
.
Japan, 1997
Solution. Of course, from the introduction to this problem, the
reader has already noticed that it is useless to try a direct application of
the lemma, since any such approach is doomed. But with the substitution
x =
b +c
a
, y =
c +a
b
, z =
a +b
c
,
74
we have to prove that for any positive real numbers x, y, z satisfying
xyz = x +y +z + 2, the inequality
(x −1)
2
x
2
+ 1
+
(y −1)
2
y
2
+ 1
+
(z −1)
2
z
2
+ 1

3
5
holds. It is now time to use T
2
’s lemma in the form
(x −1)
2
x
2
+ 1
+
(y −1)
2
y
2
+ 1
+
(z −1)
2
z
2
+ 1

(x +y +z −3)
2
x
2
+y
2
+z
2
+ 3
.
Hence it is enough to prove the inequality
(x +y +z −3)
2
x
2
+y
2
+z
2
+ 3

3
5
.
But this is equivalent to
(x +y +z)
2
−15(x +y +z) + 3(xy +yz +zx) + 18 ≥ 0.
This is not an easy inequality. We will use the proposed problem 3
from the unit ”Two useful substitutions” to reduce the above inequality
to the form
(x +y +z)
2
−9(x +y +z) + 18 ≥ 0,
which follows from the inequality x + y + z ≥ 6. And the problem is
solved.
But here is another original solution.
Alternative solution. Let us apply T
2
’s lemma in the following
form:
(b +c −a)
2
a
2
+ (b +c)
2
+
(c +a −b)
2
b
2
+ (c +a)
2
+
(a +b −c)
2
c
2
+ (a +b)
2
=
((b +c)
2
−a(b +c))
2
a
2
(b +c)
2
+ (b +c)
4
+
((c +a)
2
−b(c +a))
2
b
2
(c +a)
2
+ (c +a)
4
+
((a +b)
2
−c(a +b))
2
c
2
(a +b)
2
+ (a +b)
4

4(a
2
+b
2
+c
2
)
2
a
2
(b +c)
2
+b
2
(c +a)
2
+c
2
(a +b)
2
+ (a +b)
4
+ (b +c)
4
+ (c +a)
4
.
75
Consequently, it suﬃces to prove that the last quantity is greater
than or equal to
3
5
. This can be done by expanding everything, but here
is an elegant proof using the observation that
a
2
(b +c)
2
+b
2
(c +a)
2
+c
2
(a +b)
2
+ (a +b)
4
+ (b +c)
4
+ (c +a)
4
= [(a +b)
2
+ (b +c)
2
+ (c +a)
2
](a
2
+b
2
+c
2
)
+2ab(a +b)
2
+ 2bc(b +c)
2
+ 2ca(c +a)
2
.
Because
(a +b)
2
+ (b +c)
2
+ (c +a)
2
≤ 4(a
2
+b
2
+c
2
),
we observe that the desired inequality reduces to
2ab(a +b)
2
+ 2bc(b +c)
2
+ 2ca(c +a)
2

8
3
(a
2
+b
2
+c
2
)
2
.
But this inequality is not so diﬃcult. Indeed, ﬁrst we observe that
2ab(a +b)
2
+ 2bc(b +c)
2
+ 2ca(c +a)
2
≤ 4ab(a
2
+b
2
) + 4bc(b
2
+c
2
) + 4ca(c
2
+a
2
).
Then, we also ﬁnd that
4ab(a
2
+b
2
) ≤ a
4
+b
4
+ 6a
2
b
2
,
since (a −b)
4
≥ 0. Hence
4ab(a
2
+b
2
) + 4bc(b
2
+c
2
) + 4ca(c
2
+a
2
) ≤ 2(a
2
+b
2
+c
2
)
2
+2(a
2
b
2
+b
2
c
2
+c
2
a
2
) ≤
8
3
(a
2
+b
2
+c
2
)
2
and so the problem is solved. With minor changes, we can readily see
that this solution works without the assumption that a, b, c are positive.
We end this discussion (which remains probably permanently open)
with a diﬃcult problem, based on two hidden applications of T
2
’s lemma.
76
Example 8. Let a
1
, a
2
, . . . , a
n
> 0 such that a
1
+a
2
+ +a
n
= 1.
Prove that:
(a
1
a
2
+a
2
a
3
+ +a
n
a
1
)
_
a
1
a
2
2
+a
2
+
a
2
a
2
3
+a
3
+ +
a
n
a
2
1
+a
1
_

n
n + 1
.
Gabriel Dospinescu
Solution. How can we get to a
1
a
2
+ a
2
a
3
+ + a
n
a
1
? Probably
from
a
2
1
a
1
a
2
+
a
2
2
a
2
a
3
+ +
a
2
n
a
n
a
1
after we use the lemma. So, let us try
this the following estimation:
a
1
a
2
+
a
2
a
3
+ +
a
n
a
1
=
a
2
1
a
1
a
2
+
a
2
2
a
2
a
3
+ +
a
2
n
a
n
a
1

1
a
1
a
2
+a
2
a
3
+ +a
n
a
1
.
The new problem, proving that
a
1
a
2
2
+a
2
+
a
2
a
2
3
+a
3
+ +
a
n
a
2
1
+a
1

n
n + 1
_
a
1
a
2
+
a
2
a
3
+ +
a
n
a
1
_
seems even more diﬃcult, but we will see that we have to make one more
step in order to solve it. Again , we look at the right-hand side and we
write
a
1
a
2
+
a
2
a
3
+ +
a
n
a
1
as
_
a
1
a
2
+
a
2
a
3
+ +
a
n
a
1
_
2
a
1
a
2
+
a
2
a
3
+ +
a
n
a
1
.
After applying T
2
’s lemma, we ﬁnd that
a
1
a
2
2
+a
2
+
a
2
a
2
3
+a
3
+ +
a
n
a
2
1
+a
1
=
_
a
1
a
2
_
2
a
1
+
a
1
a
2
+
_
a
2
a
3
_
2
a
2
+
a
2
a
3
+ +
_
a
n
a
1
_
2
a
n
+
a
n
a
1

_
a
1
a
2
+
a
2
a
3
+ +
a
n
a
1
_
2
1 +
a
1
a
2
+
a
2
a
3
+ +
a
n
a
1
.
77
And we are left with an easy problem: if t =
a
1
a
2
+ +
a
n
a
1
, then
t
2
1 +t

nt
n + 1
, or t ≥ n. But this follows immediately from the AM-GM
inequality.
Problems for training
1. Let a, b, c, d be positive real numbers such that a +b +c +d = 1.
Prove that
a
2
a +b
+
b
2
b +c
+
c
2
c +d
+
d
2
d +a

1
2
.
Ireland, 1999
2. Let a, b, c, be positive real numbers satisfying a
2
+b
2
+c
2
= 3abc.
Prove that
a
b
2
c
2
+
b
c
2
a
2
+
c
a
2
b
2

9
a +b +c
.
India
3. Let x
1
, x
2
, . . . , x
n
, y
1
, y
2
, . . . , y
n
be positive real numbers such that
x
1
+x
2
+ +x
n
≥ x
1
y
1
+x
2
y
2
+ +x
n
y
n
.
Prove that
x
1
+x
2
+ +x
n

x
1
y
1
+
x
2
y
2
+ +
x
n
y
n
.
4. For arbitrary positive real numbers a, b, c prove the inequality
a
b + 2c
+
b
c + 2a
+
c
a + 2b
≥ 1.
Czech-Slovak Competition, 1999
5. Prove that for any positive real numbers a, b, c satisfying a+b+c =
1,
a
1 +bc
+
b
1 +ca
+
c
1 +ab

9
10
.
India
78
6. Prove that for any positive real numbers a, b, c, d satisfying ab +
bc +cd +da = 1 the following inequality is true
a
3
b +c +d
+
b
3
c +d +a
+
c
3
d +a +b
+
d
3
a +b +c

1
3
.
IMO 1990 Shortlist
7. Prove that if the positive real numbers a, b, c satisfy abc = 1, then
a
b +c + 1
+
b
c +a + 1
+
c
a +b + 1
≥ 1.
Vasile Cartoaje, Gazeta Matematica
8. Prove that for any positive real numbers a, b, c the following in-
equality holds
a
2
+bc
b +c
+
b
2
+ca
c +a
+
c
2
+ab
a +b
≥ a +b +c.
Cristinel Mortici, Gazeta Matematica
9. Prove that for any nonnegative real numbers x
1
, x
2
, . . . , x
n
,
x
1
x
n
+x
2
+
x
2
x
1
+x
3
+ +
x
n
x
n−1
+x
1
≥ 2.
Tournament of the Towns, 1982
10. Prove that for any positive real numbers a, b, c, d, e satisfying
abcde = 1,
a +abc
1 +ab +abcd
+
b +bcd
1 +bc +bcde
+
c +cde
1 +cd +cdea
+
d +dea
1 +de +deab
+
e +eab
1 +ea +eabc

10
3
.
Waldemar Pompe, Crux Mathematicorum
11. Prove that for any positive real numbers a, b, c the following
inequality holds
_
a
b +c
_
2
+
_
b
c +a
_
2
+
_
c
a +b
_
2

3
4

a
2
+b
2
+c
2
ab +bc +ca
.
Gabriel Dospinescu
79
12. Let n ≥ 4 an integer and let a
1
, a
2
, . . . , a
n
be positive real num-
bers such that a
2
1
+a
2
2
+ +a
2
n
= 1. Prove that
a
1
a
2
2
+ 1
+
a
2
a
2
3
+ 1
+ +
a
n
a
2
1
+ 1

4
5
(a
1

a
1
+a
2

a
2
+ +a
n

a
n
)
2
.
Mircea Becheanu, Bogdan Enescu, TST 2002, Romania
13. Find the best constant k(n) such that for any positive real
numbers a
1
, a
2
, . . . , a
n
satisfying a
1
a
2
. . . a
n
= 1 the following inequality
holds
a
1
a
2
(a
2
1
+a
2
)(a
2
2
+a
1
)
+
a
2
a
3
(a
2
2
+a
3
)(a
2
3
+a
2
)
+ +
a
n
a
1
(a
2
n
+a
1
)(a
2
1
+a
2
)
≤ k
n
.
Gabriel Dospinescu, Mircea Lascu
14. Prove that for any positive real numbers a, b, c,
(2a +b +c)
2
2a
2
+ (b +c)
2
+
(2b +c +a)
2
2b
2
+ (c +a)
2
+
(2c +a +b)
2
2c
2
+ (a +b)
2
≤ 8.
Titu Andreescu, Zuming Feng, USAMO 2003
80
ONLY GRAPHS, NO SUBGRAPHS!
There were so many strategies and useful ideas till now, that the
reader might say: enough with this game of tricks! When shall we go
to serious facts? Not only that we will ”dissapoint” him again, but we
will try also to convince him that these are more than simple tools and
tricks. They help to create a good base, which is absolutely indispensable
for someone who enjoys mathematics and moreover, they are the ﬁrst
step to some really beautiful and diﬃcult theorems or problems. And the
reader must admit that the last problems discussed in the previous units
are quite serious facts. It is worth mentioning that they are not panacea.
This assertion is proved by the fact that each year problems that are
based on well-known ”tricks” prove to be very diﬃcult in contests.
We will focus in this unit on a very familiar theme: graphs without
complete subgraphs. Why do we say familiar? Because there are hun-
dreds of problems proposed to diﬀerent contests around the world and in
mathematical magazines that deal with this subject and each one seems
to add something. Before passing to the ﬁrst problem, we will assume
that the basic knowledge about graphs is known and we will denote by
d(A) and C(A) the number, respectively the set of vertices adjacent to
A. Also, we will say that a graph does not have a complete k subgraph
if there aren’t k vertices any two of them connected. For simplicity, we
will say that G is k-free. First, we will discuss probably the ﬁrst classical
result about triangles-free graphs, the famous Turan’ theorem. But be-
fore that, an useful lemma, which is also known as Zarankiewicz lemma
and which is the main idea in Turan’ theorem’ proof.
Example 1. If G is a k-free graph, then there exists a vertex having
degree at most
_
k −2
k −1
n
_
.
Zarankiewicz
81
Solution. Suppose not and take an arbitrary vertex A
1
. Then
[C(A
1
)[ >
_
k −2
k −1
n
_
,
so there exists A
2
∈ C(A
1
). Moreover,
[C(A
1
) ∩ C(A
2
)[ = d(A
1
) +d(A
2
) −[C(A
1
∪ A
2
)[
≥ 2
_
1 +
_
k −2
k −1
n
__
−n > 0.
Pick a vertex A
3
∈ C(A
1
) ∩ C(A
2
). A similar argument shows that
[C(A
1
) ∩ C(A
2
) ∩ C(A
3
)[ ≥ 3
_
1 +
_
k −2
k −1
n
__
−2n.
Repeating this argument, we ﬁnd
A
4
∈ C(A
1
) ∩ C(A
2
) ∩ C(A
3
), . . . , A
k−1

k−2

i=1
C(A
i
).
Also, we have
¸
¸
¸
¸
¸
j

i=1
C(A
i
)
¸
¸
¸
¸
¸
≥ j
_
1 +
_
k −2
k −1
n
__
−(j −1)n.
This can be proved easily by induction. Thus,
¸
¸
¸
¸
¸
k−1

i=1
C(A
i
)
¸
¸
¸
¸
¸
≥ (k −1)
_
1 +
_
k −2
k −1
n
__
−(k −2)n > 0
and consequently we can choose
A
k

k−1

i=1
C(A
i
).
But it is clear that A
1
, A
2
, . . . , A
k
form a complete k graph, which
contradicts the assumption that G is k-free.
We are now ready to prove Turan’ theorem.
Example 2. The maximal number of edges of a k-free graph with
vertices is
k −2
2

n
2
−r
2
k −1
+
_
r
2
_
,
82
where r = n (mod k −1).
Turan’ theorem
Solution. The theorem will be proved by induction on n. Since the
ﬁrst case is trivial, let us suppose the theorem true for all k-free graphs
having n − 1 vertices and let G a k-free graph with n vertices. Using
Zarankiewicz’ lemma, we can ﬁnd a vertex A such that
d(A) ≤
_
k −2
k −1
n
_
.
Since the subgraph determined by the other n−1 vertices is obviously
k-free, using the inductive hypothesis we ﬁnd that G has at most
_
k −2
k −1
n
_
+
k −2
k −1

(n −1)
2
−r
2
1
2
+
_
r
1
2
_
edges, where r
1
= n −1 (mod k −1).
Let n = q(k −1) +r = q
1
(k −1) +r
1
+1. Then r
1
∈ ¦r −1, r +k −2¦
(this is because r −r
1
≡ 1 (mod k −1)) and it is easy to check that
_
k −2
k −1
n
_
+
k −2
k −1

(n −1)
2
−r
2
1
2
+
_
r
1
2
_
=
k −2
2

n
2
−r
2
k −1
+
_
r
2
_
and the inductive step is proved. Now, it remains to construct a k-
free graph with n vertices and
k −2
2

n
2
−r
2
k −1
+
_
r
2
_
edges. This is
not diﬃcult. Just consider k − 1 classes of vertices, r of them having
q + 1 elements and the rest q elements and join the vertices situated in
diﬀerent groups. It is immediate to prove that this graph is k-free, has
k −2
2

n
2
−r
2
k −1
+
_
r
2
_
edges and also the minimal degree of the vertices
is
_
k −2
k −1
n
_
. This graph is called Turan’ graph and it is denoted by
T(n, k).
These two examples generate lots of beautiful and diﬃcult problems.
For example, knowing them means a straightforward solution for the
following bulgarian problem.
83
Example 3. There are 2001 towns in a country, every one of which is
connected with at least 1600 towns by a direct bus line. Find the largest
n for which it is always possible to ﬁnd n towns, any two of which are
connected by a direct bus line.
Spring Mathematics Tournament, 2001
Solution. Practically, the problem asks to ﬁnd the maximal n such
that any graph G with 2001 vertices and minimum degree at least 1600
is not n-free. But Zarankiewicz’ lemma implies that if G is n-free, then
at least one vertex has degree at most
_
n −2
n −1
2001
_
. So, we need the
maximal n for which
_
n −2
n −1
2001
_
< 1600. It is immediate to see that
it is n = 5. Thus, if n = 5 then any such graph G is not n-free. It
suﬃces to construct a graph with all degrees of the vertices at least
1600, which is 6-free. We will take of course T(2001, 6), whose minimal
degree is
_
4
4
2001
_
= 1600 and which is of course 6-free. Thus, the answer
is n = 5.
Here is a beautiful application of Turan’ theorem in combinatorial
geometry.
Example 4. Given are 21 points on a circle. Show that at least 100
pairs of points subtend an angle smaller than or equal to 120 at the
center.
Tournament of the Towns, 1986
Solution. In such problems, it is more important to choose the
right graph than to apply the theorem, because as soon as the graph is
appropriately chosen, the solution is more or less straightforward. Here,
we will consider the graph with vertices in the points and we will connect
two points if they subtend an angle smaller than or equal to 120 at the
center. Therefore, we need to prove that this graph has at least 100
edges. It seems that this is a reversed form of Turan’ theorem, which
84
maximizes the number of edges in a k-free graph. Yet, the reversed form
of a reversed form is the natural one. In the aim of this principle, let
us look at the ”reversed” graph, the complementary one. We must show
that it has at most
_
21
2
_
−100 = 110 edges. But this is immediate, since
it is clear that this new graph does not have triangles and so, by Turan’
theorem it has at most
21
2
−1
4
= 110 edges. And the problem is solved.
At ﬁrst glance, the following problem seem to have no relation with
the previously examples, but, as we will see, it is a simple consequence
of Zarankiewicz’ lemma. This problem is an adaptation of a USAMO
1978 problem. Anyway, this is trickier than the contest problem.
Example 5. There are n delegates at a conference, each of them
knowing at most k languages. Anyway, among any three delegates, at
least two speak a common language. Find the smallest number n (in
terms of k) such that it is always possible to ﬁnd a language spoken by
at least three delegates.
Solution. We will prove that n = 2k + 3. First, we prove that if
there are 2k + 3 delegates,then the conclusion of the problem holds.
The condition ”among any three of them there are at least two who
can communicate” suggests us to take the 3-free graph with vertices
in the persons and whose edges join persons that cannot communicate.
From Zarankiewicz’ lemma, there exists a vertex whose degree is at most
_
n
2
_
= k +1. Thus, it is not connected with at least k +1 other vertices.
Therefore, there exists a person A and k + 1 persons A
1
, A
2
, . . . , A
k+1
that can communicate with A. Since A knows at most k languages, there
are two persons among A
1
, A
2
, . . . , A
p
that know a language also known
by A. But that language is known by at least three delegates and we are
done. It remains to prove now that we can create a situation in which
there are 2k + 2 delegates, but no language is known by more than two
delegates. We use again Turan’ graph, by creating two groups of k + 1
85
delegates. In each group a person will have a common language with
each other person from the group and will not have common languages
with the members of the other group. Of course, any language is spoken
by at most two delegates and there are no triangles.
The following problem turned out to be a surprise at one of the Team
Selection Tests for 2004 IMO, being solved by 4 contestants. The idea is
even easier than in the previous problems, but this time we need a little
observation, that is not so obvious.
Example 6. Let A
1
, A
2
, . . . , A
101
be diﬀerent subsets of the set
¦1, 2, . . . , n¦. Suppose that the union of any 50 subsets has more than
50
51
n elements. Prove that there are three subsets among them, any two
of them having common elements.
Gabriel Dospinescu, TST 2004 Romania
Solution. Of course, as the conclusion suggests, we should take a
graph with vertices in the subsets, connecting two subsets if they have
common elements. Let us assume that this graph is 3-free. The main
idea is not to use Zarankiewicz’ lemma, but to ﬁnd much more vertices
with small degrees. In fact, we will prove that there are at least 51
vertices whose degree are smaller than or equal to 50. Suppose this is
not the case, thus there are at least 51 vertices whose degrees are greater
than 51. Let us pick such a vertex A. It is connected with at least 51
vertices, thus it must be adjacent to a vertex B, whose degree is at
least 51. Since A and B are each connected with at least 51 vertices,
there is a vertex adjacent to both, so we have a triangle, contradicting
our assumption. Therefore, we can ﬁnd A
i
1
, . . . , A
i
51
, all of them having
degrees at most 50. Consequently, A
i
1
is disjoint from at least 50 subsets.
Since the union of these subsets has more than
50
51
n elements, we infer
that [A
i
1
[ < n −
50
51
n =
n
51
. In a similar way, we obtain that [A
i
j
[ ≤
n
51
86
for all j ∈ ¦1, 2, . . . , 51¦ and so
[A
i
1
∪ A
i
2
∪ ∪ A
i
50
[ ≤ [A
i
1
[ + +[A
i
50
[ <
50
51
n,
which contradicts the hypothesis. And the solution ends here.
We end the discussion with an adaptation of a very nice and quite
challenging problem from the American Mathematical Monthly.
Example 7. Prove that the complementary of any 3-free graph with
n vertices and m edges has at least
n(n −1)(n −5)
24
+
2
n
_
m−
n
2
−n
4
_
2
triangles.
A.W Goodman, AMM
Solution. Believe it or not, the number of triangles from the com-
plementary graph can be expressed only in terms of the degrees of the
vertices of the graph. More precisely, if G is the graph, then the number
of triangles from the complementary graph is
_
n
3
_

1
2

x∈X
d(x)(n −1 −d(x)),
where X is the set of vertices of G. Indeed, consider all triples (x, y, z) of
vertices of G. We will count the triples that do not form a triangle in the
complementary graph G. Indeed, consider the sum

x∈X
d(x)(n−1−d(x)).
It counts twice every triple (x, y, z) in which are connected, while z is not
adjacent to any of x, y: once for x and once for y. But it also counts twice
every triple (x, y, z) in which y is connected with both x, z: once for x and
once for z. Therefore,
1
2

x∈X
d(x)(n −1 −d(x)) is exactly the number of
triples (x, y, z) that do not form a triangle in the complementary graph
(here we have used the fact that G is 3-free). Now, it is enough to prove
87
that
_
n
3
_

1
2

x∈X
d(x)(n−1−d(x)) ≥
n(n −1)(n −5)
24
+
2
n
_
m−
n
2
−n
4
_
2
.
Using the observation that

x∈X
d(x) = 2m, after a few computations
we ﬁnd the equivalent form of the inequality

x∈X
d
2
(x) ≥
4m
2
n
.
But this is exactly the Cauchy-Schwarz inequality combined with
the observation that

x∈X
d(x) = 2m.
Problems for training
1. In a country there are 1998 cities. In each group of three cities,
at least two are not directly connected. What is the maximal number of
direct ﬂights?
Japan, 1998
2. Let x
1
, x
2
, . . . , x
n
be real numbers. Prove that there are at most
n
2
4
pairs (i, j) ∈ ¦1, 2, . . . , n¦ ¦1, 2, . . . , n¦ such that 1 < [x
i
−x
j
[ < 2.
MOSP, CE AN?
3. If n points lie on a circle, then at most
n62
3
segments connecting
them have length greater than

2.
Poland, 1997
4. Let G be a graph with no triangles and such that no point is
adjacent to all the other vertices. Also, if A and B are not joined by
an edge, then there exists a vertex C such that AC and BC are edges.
Prove that all vertices have the same degree.
APMO 1990
88
5. Show that a graph with n vertices and k edges has at least
k
3n
(4k−
n
2
) triangles.
APMO 1989
6. Let A be a subset of the set S = ¦1, 2, . . . , 1000000¦ having exactly
101 elements. Prove that there exist t
1
, t
2
, . . . , t
100
∈ S such that the sets
A
j
= ¦x +t
j
[x ∈ A¦ are pairwise disjoint.
IMO 2003
6. There are 1999 people participating in an exhibition. Out of any
50 people, at least 2 do not know each other. Prove that we can ﬁnd at
least 41 people who each know at most 1958 other people.
Taiwan, 1999
7. A graph with n vertices and k edges is 3-free. Prove that we
can choose a vertex such that the subgraph induced by the remaining
vertices has at most k
_
1 −
4k
n
2
_
vertices.
USAMO 1995
8. Prove that for every n one can construct a graph with no triangles
and whose chromatic number is at least n.
9. A graph with n
2
+1 edges and 2n vertices is given. Prove that it
contains two triangles sharing a common edge.
China TST, 1987
10. We are given 5n points in a plane and we connect some of them,
so that 10n
2
+ 1 segments are drawn. We color these segments in 2
colours. Prove that we can ﬁnd a monochromatic triangle.
89
COMPLEX COMBINATORICS
When reading the title, one will surely expect a hard unit, which will
show what a complex ﬁeld is combinatorics. Unfortunately, this was not
our intention. We ”just” want to discuss some combinatorial problems
that can be solved elegantly using complex numbers. In this moment,
the reader will probably say we are crazy, but we will continue our
idea and say that complex numbers can play a very important role in
counting problems and also in problems related to tilings. There are also
numerous applications in combinatorial number theory, so our purpose
is to present a little bit from each of these situations. After that, the
reader will surely have the pleasure of solving the proposed problems
using this technique. For fear of useless repetition, we will present in the
beginning of the discussion a useful result
Lemma. If p is a prime number and a
0
, a
1
, . . . , a
p−1
∈ Q satisfy the
relation
a
0
+a
1
ε +a
2
ε
2
+ +a
p−1
ε
p−1
= 0,
where
ε = cos

p
+i sin

p
,
then a
0
= a
1
= = a
p−1
.
We will say just a few words about the proof, which is not diﬃcult. It
is enough to observe that the polynomials a
0
+a
1
x+a
2
x
2
+ +a
p−1
x
p−1
and 1+x+x
2
+ +x
p−1
cannot be relatively prime-because they share
a common root-and since 1 + x + x
2
+ + x
p−1
is irreducible over Q,
1 + x + x
2
+ + x
p−1
must divide a
0
+ a
1
x + a
2
x
2
+ + a
p−1
x
p−1
,
which can only happen if a
0
= a
1
= = a
p−1
. Therefore, the lemma
is proved and it is time to solve some nice problems. Not before saying
that in the following examples m(A) will denote the sum of the elements
of the set A. By convention m(∅) = 0.
90
The ﬁrst example is an adaptation from a problem given in the Inter-
County Contest ”Traian Lalescu”. Of course, there is a solution using
recurrent sequences, but it is by far less elegant than the following one.
Example 1. How many numbers with n digits, all equal to 1, 3, 4,
6, 7, 9 are divisible by 7?
Solution. Let a
(k)
n
be the number of n-digits numbers, formed using
only the digits 1, 3, 4, 6, 7, 9 and which are congruent to k modulo 7.
It is clear that
6

k=0
a
(k)
n
ε
k
=

x
1
,x
2
,...,x
n
∈¡1,3,4,6,7,9¦
ε
x
1
+x
2
++x
n
= (ε +ε
3

4

6

7

9
)
n
,
where ε = cos

7
+ i sin

7
. The remark that 1 + ε + ε
2
+ + ε
6
= 0
helps us to bring (ε+ε
3

4

6

7

9
)
n
to the simpler form (−ε
5
)
n
.
Let us assume that n is divisible by 7, for example (the other cases can
be discussed similarly). Then
6

k=0
a
(k)
n
ε
k
= (−1)
n
and from the lemma we infer that a
(0)
n
−(−1)
n
= a
(1)
n
= = a
(6)
n
. Let
k be the common value. Then 7k =
6

k=0
a
(k)
n
−(−1)
n
= 6
n
−(−1)
n
- this
is because exactly 6
n
numbers have n digits, all equal to 1, 3, 4, 6, 7, 9.
Thus, in this case we have a
(0)
n
= (−1)
n
+
6
n
−(−1)
n
7
. We leave to the
reader the study of the other cases: n ≡ 12, 3, 4, 5, 6 (mod 7).
The same simple, but tricky idea can oﬀer probably the most beau-
tiful solution for the diﬃcult IMO 1995 problem 6. It worth saying that
Nikolai Nikolov won a special prize for the following magniﬁcent solu-
tion.
91
Example 2. Let p > 2 be a prime number and A = ¦1, 2, . . . , 2p¦.
Find the number of subsets of A, each having p elements and the sum
of the elements divisible by p.
Marcin Kuczma, IMO 1995
Solution. Consider ε = cos

p
+ i sin

p
and let x
j
the number of
subsets x ⊂ A such that [X[ = p and m(X) ≡ j (mod p). Then it is
clear that
p−1

j=0
x
j
ε
j
=

B⊂A,[B[=p
ε
m(B)
=

1≤c
1
<c
2
<<c
p
≤2p
ε
c
1
+c
2
++c
p
.
But

1≤c
1
<c
2
<<c
p
≤2p
ε
c
1
+c
2
++c
p
is exactly the coeﬃcient of x
p
in
the expansion (X + ε)(X + ε
2
) . . . (X + ε
2p
). Since X
p
− 1 = (X −
1)(X − ε) . . . (X − ε
p−1
), we easily ﬁnd that (X + ε)(X + ε
2
) . . . (X +
ε
2p
) = (X
p
+ 1)
2
. Thus,
p−1

j=0
x
j
ε
j
= 2 and lemma implies the equality
x
0
−2 = x
1
= = x
p−1
. Since there are
_
2p
p
_
subsets with p elements,
we have
x
0
+x
1
+ +x
p−1
=
_
2p
p
_
.
Therefore,
x
0
= 2 +
1
p
__
2p
p
_
−2
_
.
With a somewhat diﬀerent, but closely related idea we can solve the
following nice problem.
Example 3. Let a
1
, a
2
, . . . , a
m
be natural numbers and let f(k) the
number of m-tuples (c
1
, c
2
, . . . , c
m
) such that 1 ≤ c
i
≤ a
i
, i = 1, m and
c
1
+c
2
+ +c
m
≡ k (mod n), where n > 1 is a natural number.
Prove that f(0) = f(1) = = f(n − 1) if and only if there exists
an index i ∈ ¦1, 2, . . . , m¦ such that n[a
i
.
Reid Burton, Rookie Contest ,1999
92
Solution. It is not diﬃcult to observe that
n−1

k=0
f(k)ε
k
=

1≤c
i
≤a
i
ε
c
1
+c
2
++c
m
=
m

i=1
(ε +ε
2
+ +ε
a
i
)
for any complex number ε such that ε
n−1

n−2
+ +ε +1 = 0. Thus,
one part of the problem is already proved, since if f(0) = f(1) = =
f(n −1) then of course we can ﬁnd i ∈ ¦1, 2, . . . , m¦ such that ε +ε
2
+

a
i
= 0, where we have chosen here a primitive root of the unity ε.
We infer that ε
a
i
= 1 and so n[a
i
. Now, suppose there exists an index
i ∈ ¦1, 2, . . . , m¦ such that n[a
i
. Then for any root ε of the polynomial
n−1

k=0
X
k
we have
n−1

k=0
f(k)ε
k
and so the polynomial
n−1

k=0
X
k
divides the
polynomial
n−1

k=0
f(k)X
k
. This is because the polynomial
n−1

k=0
X
k
has only
simple roots. By a simple degree consideration, this is possible only if
f(0) = f(1) = = f(n −1). The solution ends here.
The enthusiasm determined by the above solutions will surely be
ATENUAT by the following problem, in which e need some tricky ma-
nipulations.
Example 4. Let p > 2 be a prime number and let m, n be multiples
of p such that n is odd. For any function f : ¦1, 2, . . . , m¦ → ¦1, 2, . . . , n¦
that satisﬁes p[f(1) +f(2) + +f(m), consider the product f(1)f(2)
f(m). Prove that the sum of these products is divisible by
_
n
p
_
m
.
Gabriel Dospinescu
Solution. Let ε = cos

p
+ i sin

p
and x
k
be the sum of all
the numbers f(1)f(2) . . . f(m), after all functions f : ¦1, 2, . . . , m¦ →
¦1, 2, . . . , n¦ that satisfy f(1) + f(2) + + f(m) ≡ k (mod p). It is
93
clear that:
p−1

k=0
x
k
ε
k
=

c
1
,c
2
,...,c
m
∈¡1,2,...,n¦
c
1
c
2
. . . c
m
ε
c
1
+c
2
++c
m
= (ε + 2ε
2
+ +nε
n
)
m
.
Recall the identity
1 + 2x + 3x
2
+ +nx
n−1
=
nx
n+1
−(n + 1)x
n
+ 1
(x −1)
2
.
Plugging ε in the previous identity, we ﬁnd that
ε + 2ε
2
+ +nε
n
=

n+2
−(n + 1)ε
n+1

(ε −1)
2
=

ε −1
.
Consequently,
p−1

k=0
x
k
ε
k
=
n
m
(ε −1)
m
.
On the other hand, it is not diﬃcult to deduce the relations
ε
p−1

p−2
+ +ε + 1 = 0 ⇔
1
ε −1
= −
1
p

p−2
+ 2ε
p−3
+ + (p −2)ε +p −1).
Thus, if we consider
(X
p−2
+2X
p−3
+ +(p−2)X+p−1)
m
= b
0
+b
1
X+ +b
m(p−2)
X
m(p−2)
,
then we have
n
m
(ε −1)
m
=
_

n
p
_
m
(c
0
+c
1
ε + +c
p−1
ε
p−1
),
where
c
k
=

k≡k (mod p)
b
j
.
If r =
_

n
p
_
m
, then we have the relation
x
0
−rc
0
+ (x
1
rc
1
)ε + + (x
p−1
−rc
p−1

p−1
= 0.
94
From the lemma, it follows that x
0
−rc
0
= x
1
−rc
1
= = x
p−1

rc
p−1
= k. Because clearly c
0
, c
1
, . . . , c
p−1
∈ R, it remains to prove that
r[k. Since
pk = x
0
+x
1
+ +x
p−1
−r(c
0
+c
1
+ +c
p−1
)
= (1 + 2 + +n)
m
−r(b
0
+b
1
+ +b
m(p−2)
)
=
_
n(n + 1)
2
_
m
−r
_
p(p −1)
2
_
m
,
it is clear that r[k. Here we have used the hypothesis. The problem is
solved.
It is time now to leave this kind of problems and to speak a little bit
about some nice applications of complex numbers in tilings. The idea is
to put a complex number in each square of a table and then to translate
the hypothesis and the conclusion in terms of complex numbers. But
we will better see how this technique works by solving a few problems.
First, some easy problems.
Example 5. Consider a rectangle which can be tiled with a ﬁnite
combination of 1m or n1 rectangles, where m, n are natural numbers.
Prove that it is possible to tile this rectangle using only rectangles 1m
or only with rectangles n 1.
Gabriel Carrol ,BMC Contest,2000
Solution. It is obvious that the rectangle has natural dimensions,
let them be a, b. Now, let us partition the rectangle into 1 1 squares
and denote this squares
(1, 1), (1, 2), . . . , (a, 1), . . . , (a, 1), (a, 2), . . . , (a, b).
Next, put the number ε
i
1
ε
j
2
in the square whose label is (i, j), where
ε
1
= cos

n
+i sin

n
, ε
2
= cos

m
+i sin

m
.
95
The main observation is that the sum of the numbers in any 1 m
or n 1 rectangle is 0. This is immediate, but the consequence of this
simple observation is really surprising. Indeed, it follows that the sum
of the numbers from all the squares is 0 and so
0 =

1≤i≤a
1≤j≤b
ε
i
1
ε
j
2
=
a

i=1
ε
i
1
b

j=1
ε
j
2
.
Thus, at least one of the numbers
a

i=1
ε
i
1
and
b

j−1
ε
j
2
is 0. But this
means that n[a or m[b. In any of these cases, it is clear that we can tile
the rectangle using only horizontal or vertical rectangles.
The idea in the previous problem is quite useful, many tilings prob-
lems having straightforward solutions by using it. An example is the
following problem, given in Baltic Contest in 1998.
Example 6. Can we tile a 13 13 table using only 1 4, 4 1
rectangles, such that only the center of the table does not belong to any
rectangle?
Baltic Contest,1998
Solution. Suppose such a tiling is possible and label the squares of
the table as in the previous problem. Next, associate to square (k, j) the
number i
k+2j
. Obviously, the sum of the numbers from each 1 4, 4 1
rectangle is 0. Therefore, the sum of all numbers from the squares of the
table is equal to the number in the square situated at the center of the
table. Thus,
i
21
= (i +i
2
+ +i
13
)(i
2
+i
4
+ +i
26
) = i
i
13
−1
i −1
i
2

i
26
−1
i
2
−1
= i
3
,
which clearly cannot hold. Thus, the assumption was wrong and such a
tiling does not exist.
96
The following example we are going to discuss is based on the same
idea, but here complex numbers are more involved.
Example 7. On a 8 9 table we put rectangles 3 1 and ﬁgures
formed by rectangles 1 3 by cutting the median 1 1 square. The
rectangles and the ﬁgures do not intersect and cannot be rotated. Prove
that there exists a set S of 18 squares of the table such that if there are
exactly two uncovered squares, then they belong to S.
Gabriel Dospinescu
Solution. Again, we label the squares of the table (1, 1), (1, 2), . . . ,
(8, 9) by starting from the up-left corner. In the square labeled (k, j) we
will put the number i
j
ε
k
, where i
2
= −1 and ε
2
+ε+1 = 0. The sum of
the numbers from any ﬁgure or rectangle is 0. The sum of the numbers
from the table is
_
8

k=1
ε
k
_
_
_
9

j=1
i
j
_
_
= −i.
Let us suppose that the squares (a
1
, b
1
), (a
2
, b
2
) are the only un-
covered squares. Then we have of course i
b
1
ε
a
1
+ i
b
2
ε
a
2
= −i. Let
z
1
= i
b
1
−1
ε
a
1
, z
2
= i
b
2
−1
ε
a
2
. We have [z
1
[ = [z
2
[ = 1 and z
1
+ z
2
= −1.
It follows that
1
z
1
+
1
z
2
= −1 and so z
3
1
= z
3
2
= 1. This in turn im-
plies the equalities i
3(b
1
−1)
= i
3(b
2
−1)
= 1, from where we conclude that
b
1
≡ b
2
≡ 1 (mod 4). Therefore, the relation z
1
+ z
2
= −1 becomes
ε
a
1
+ ε
a
2
= −1, which is possible if and only if the remainders of the
numbers a
1
, a
2
when divided by 3 are 1 and 2. Thus, we can take S the
set of squares that lie at the intersection of the lines 1, 2, 4, 5, 7, 8 with
the columns 1, 5, 9. From the above argument, if two squares remain
uncovered, then surely they belong to S. The conclusion is immediate.
97
Problems for training
1. Three persons A, B, C play the following game: a subset with k
elements of the set ¦1, 2, . . . , 1986¦ is selected randomly, all selections
having the same probability. The winner is A, B or C, according to the
case when the sum of the elements of the selected subset is congruent
to 0, 1, or 2 modulo 3. Find all values of k for which A, B, C have equal
chances of winning.
Imo Shortlist, 1987
2. The faces of a die are labeled with the numbers 1, 2, 3, 4, 5, 6.
We throw the die n times. What is the probability that the sum of the
numbers shown by the die is a multiple of 5?
IMC, 1999
3. Let a
k
, b
k
, c
k
∈ R, k = 1, n. Let f(p) be the number of ordered
triples (A, B, C) of subsets (not necessarily non-empty) of the set M =
¦1, 2, . . . , n¦ whose union is M and for which

i∈M\A
a
i
+

i∈M\B
b
i
+

i∈M\C
c
i
≡ 3 (mod p).
We assume that

i∈∅
x
i
= 0 and f(0) = f(1) = f(2).
Prove that there exists i ∈ M such that 3[a
i
+b
i
+c
i
.
Gabriel Dospinescu, Recreat ¸ii Matematice
4. How many subsets with 100 elements of the set ¦1, 2, . . . , 2000¦
have the sum of their elements divisible by 5?
Qihong Xie, High School-Mathematics
5. There are 2000 white balls in a box. There are also unlimited
supplies of white, green and red balls, initially outside the box. At each
step, we can replace two balls in the box with one or two balls according
98
to the following rules: two whites or two reds with a green, two greens
with a white and red, a white and green with a red or a green and red
with a white.
a) After some ﬁnite number of steps, in the box there are exactly
three balls. Prove that at least one of them is green.
b) Is it possible that after a ﬁnite number of steps there is just one
ball in the box?
Bulgaria, 2000
6. A 7 7 table is tiled with 16 rectangles 1 3 such that only one
square remains uncovered. What is the position of this square?
Tournament of the Towns, 1984
7. Let k > 2 be an integer. For which odd natural numbers n can
we tile a n n table with 1 k or k 1 rectangles such that only the
square in the center of the table does not belong to any rectangle?
Arhimede Magazine, Gabriel Dospinescu
8. Let n ≥ 2 be an integer. In each point (i, j) having integer coordi-
nates we write the number i +j (mod n). Find all pairs (a, b) of natural
numbers such that any residue modulo n appears the same number of
times on the frontier of the rectangle of vertices (0, 0), (a, 0), (a, b), (0, b)
and also any residue modulo n appears the same number of times in the
interior of the same rectangle.
Bulgaria, 2001
9. Let F be the family of the subsets of the set A = ¦1, 2, . . . , 3n¦
which have the sum of their elements divisible by 3. For each element of
F, compute the square of sum of its elements. What is the value of the
sum of all the obtained numbers?
Gabriel Dospinescu
99
10. Let p > 3 be a prime number and let h be the number of se-
quences (a
1
, a
2
, . . . , a
p−1
) ⊂ ¦0, 1, 2¦
p−1
such that p[
p−1

j=0
ja
j
. Also, let
k be the number of sequences (a
1
, a
2
, . . . , a
p−1
) ⊂ ¦0, 1, 3¦
p−1
such that
p[
p−1

j=0
ja
j
. Prove that h ≤ k and that the equality appears only for p = 5.
IMO 1999 Shortlist
100
FORMAL SERIES REVISITED
connection between the following problems:
1. The set of natural numbers (including 0) is partitioned into a
ﬁnite number n ≥ 2 of inﬁnite arithmetic progressions having ratios
r
1
, r
2
, . . . , r
n
and ﬁrst term a
1
, a
2
, . . . , a
n
. Then the following relation is
satisﬁed:
a
1
r
1
+
a
2
r
2
+ +
a
n
r
n
=
n −1
2
.
2. The vertices of a regular polygon are colored in some fashion so
that each set of vertices having the same colour is the set of vertices of
a regular polygon. Then there are two congruent polygons among them.
The ﬁrst problem was discussed during the preparation for IMO of
the USA team, but it seems it is a classical result. As for the second one,
well, it is a famous problem given in a Russian olympiad and proposed
by N. Vasiliev.
If the reader has no clue, then let’s give him one small hint: the
methods used to solve both problems are very similar and can be in-
cluded into a larger ﬁeld, that of formal series. What is that? Well,
given a commutative ring A, we can deﬁne another ring, called the ring
of formal series with coeﬃcients in A and denoted A[X]. An element of
A[X] is of the form

n≥0
a
n
X
n
, where a
n
∈ A. As we are going to see
in what follows, these formal series have some very nice applications in
diﬀerent ﬁelds: algebra, combinatorics, number theory. But let’s start
working now, reminding that the reader is supposed to be familiar with
some basic analysis tools:
Example 1. Let a
1
, a
2
, . . . , a
n
be some complex numbers such that
for any 1 ≤ k ≤ n we have a
k
1
+a
k
2
+ +a
k
n
= 0. Then all numbers are
equal to 0.
101
this problem is a trivial consequence of Newton’s relations. But what
can we do if don’t know them? Here is a nice way to solve the problem
(and a way to prove Newton’s relations too).
First of all, observe that the given condition implies that
a
k
1
+a
k
2
+ +a
k
n
= 0
for all positive integer k. Indeed, let
f(X) = X
n
+b
n−1
X
n−1
+ +b
1
X +b
0
the polynomial
n

i−1
(x −a
i
). Then for all k ≥ n + 1 we have
a
k
i
+b
n−1
a
k−1
i
+ +b
0
a
k−n
i
= 0.
Then it suﬃces to add these relations and to prove the statement by
strong induction.
Now, let us consider the function
f(z) =
n

i=1
1
1 −za
i
.
Developing it by using
1
1 −x
= 1 +x +x
2
+. . . (for [x[ < 1),
we obtain that f(z) = n for all suﬃciently small z (which means that
[z[ max([a
i
[) < 1). Assume that not all numbers are zero and take
a
1
, . . . , a
s
(s ≥ 1) to be the collection of numbers of maximal modu-
lus among the n numbers. Let the common value of the modulus be r.
By taking a sequence z
p

1
r
such that
¸
¸
¸
z
p
r
¸
¸
¸ < 1, we obtain a contra-
diction with the relation
n

i=1
1
1 −z
p
a
i
= n (indeed, it suﬃces to observe
that the left-hand side is unbounded, while the second one is bounded).
This shows that all numbers are equal to 0.
102
We are going to discuss a nice number theory problem, whose solu-
tion is practically based on the same idea. Yet, there are some details
that make the problem more diﬃcult.
Example 2. Let a
1
, a
2
, . . . , a
q
, x
1
, x
2
, . . . , x
q
and m some integers
such that m[a
1
x
k
1
+a
2
x
k
2
+ +a
q
x
k
q
for all k ≥ 0. Then
m[a
1
q

i=2
(x
1
−x
i
).
Gabriel Dospinescu
Solution. Consider this time the formal series
f(z) =
q

i=1
a
i
1 −zx
i
.
By using the same formula as in the ﬁrst problem, we deduce immedi-
ately that
f(z) =
q

i=1
a
i
+
_
q

i=1
a
i
x
i
_
z +. . . ,
which shows that all coeﬃcients of this formal series are integers multi-
ples of m. Obviously, it follows that the formal series

a
1
(1 −x
2
z) . . . (1 −x
q
z)
also has all coeﬃcients multiples of m. Now, consider S
(i)
t
the i-th
fundamental symmetric sum in x
j
(j ,= i). Since all coeﬃcients of

a
1
(1 − x
2
z) . . . (1 − x
q
z) are multiples of m, a simple computation
shows that we have the divisibility relation:
m[x
q−1
1
q

i=1
a
i
−x
q−2
i
q

i=1
a
i
S
(i)
1
+ + (−1)
q−1
q

i=1
a
i
S
(i)
q−1
.
This can also be rewritten in the nicer form
m[
q

i=1
a
i
(x
q−1
1
−x
q−2
1
S
(i)
1
+ + (−1)
q−1
S
(i)
q−1
).
103
Now, the trivial identity
(x
1
−x
1
) . . . (x
1
−x
i−1
(x
1
−x
i+1
) . . . (x
1
−x
n
) = 0
gives us the not-so obvious relation
x
q−1
1
−x
q−2
1
S
(i)
1
+ + (−1)
q−1
S
(i)
q−1
= 0
for i ≥ 2. Therefore we can conclude, since
x
q−1
1
−x
q−2
q
S
(1)
1
+ + (−1)
q−1
S
(1)
q−1
= (x
1
−x
2
) . . . (x
1
−x
n
).
In order to solve the problem announced in the very beginning of
the presentation, we need a little lemma, which is interesting itself and
which we prefer to present as a separate problem:
Example 3. Suppose that the set of natural numbers (including
0) is partitioned into a ﬁnite number of inﬁnite arithmetic progressions
of ratios r
1
, r
2
, . . . , r
n
and ﬁrst term a
1
, a
2
, . . . , a
n
. Then the following
relation is satisﬁed:
1
r
1
+
1
r
2
+ +
1
r
n
= 1.
Solution. Let us observe that for any [x[ < 1 we have the identity:

k≥0
x
a
1
+kr
1
+

k≥0
x
a
2
+kr
2
+ +

k≥0
x
a
n
+kr
n
=

k≥0
x
k
.
Indeed, all we did was to write the fact that each natural number is
exactly in one of the arithmetic progressions. The above relation becomes
of course the very useful relation:
x
a
1
1 −x
r
1
+
x
a
2
1 −x
r
2
+ +
x
a
n
1 −x
r
n
=
1
1 −x
(1)
Let us multiply the relation (1) with 1 − x and use the fact that
lim
x→1
1 −x
a
1 −x
= a. We ﬁnd of course the desired relation
1
r
1
+
1
r
2
+ +
1
r
n
= 1.
104
It’s time to solve the ﬁrst problem. We will just a small, but not
obvious step and we’ll be done. The fundamental relation is again (1).
So:
Example 4. The set of natural numbers (including 0) is partitioned
into a ﬁnite number n ≥ 2 of inﬁnite arithmetic progressions having
ratios r
1
, r
2
, . . . , r
n
and ﬁrst term a
1
, a
2
, . . . , a
n
. Then the following re-
lation is satisﬁed:
a
1
r
1
+
a
2
r
2
+ +
a
n
r
n
=
n −1
2
.
MOSP
Solution. Let us write the relation (1) in the more appropriate form:
x
a
1
1 +x + +x
r
1
−1
+ +
x
a
n
1 +x + +x
r
n
−1
= 1 (2)
Now, let us derive the relation (2) and then make x → 1 in the
resulting expression. A small computation let to the reader will show
that
n

i=1
a
i
r
i

r
i
(r
i
−1)
2
r
2
i
= 0.
But it suﬃces to use the result proved in example 3 in order to
conclude that we must have
a
1
r
1
+
a
2
r
2
+ +
a
n
r
n
=
n −1
2
.
Some commentaries about these two relations are necessary. First of
all, using a beautiful and hard result due to Erdos, we can say that the
relation
1
r
1
+
1
r
2
+ +
1
r
n
= 1
implies that max(r
1
, r
2
, . . . , r
n
) < 2
2
n−1
. Indeed, this remarkable theo-
rem due to Erdos asserts that if x
1
, x
2
, . . . , x
k
are natural numbers whose
105
sum of inverses is strictly smaller than 1, then
1
x
1
+
1
x
2
+ +
1
x
k

1
u
1
+
1
u
2
+ +
1
u
k
,
where u
1
= 2, u
n+1
= u
2
n
−u
n
+1. But the reader can verify immediately
that
1
u
1
+
1
u
2
+ +
1
u
k
= 1 −
1
u
1
u
2
. . . u
k
(it is trivial by induction). Thus we can write
1 −
1
r
n
≤ 1 −
1
u
1
u
2
. . . u
n−1
,
or even better r
n
≤ u
1
u
2
. . . u
n−1
= u
n
− 1 (the last relation being
again a simple induction). Once again, the reader will do a short in-
duction to prove that u
n
≤ 2
2
n−1
. And here is how we can prove
that max(r
1
, r
2
, . . . , r
n
) < 2
2
n−1
(since of course any number among
r
1
, r
2
, . . . , r
n
can be taken as r
n
). Using the relation proved in example
4, we also deduce that max(a
1
, a
2
, . . . , a
n
) < (n−1) 2
2
n−1
−1
. This shows
that for ﬁxed n not only there is a ﬁnite number of ways to partition
the set of natural numbers into n arithmetic progressions, but we also
have some explicit (even though huge) bounds on ratios and ﬁrst terms.
It is now time to solve the remarkable problem discussed in the
beginning of this note. We will see that in the framework of the previous
results proved here, the solution becomes natural. However, it is not at
all true, the problem is really diﬃcult.
Example 5. The vertices of a regular polygon are colored such that
vertices having the same colour form regular polygons. Prove that there
are at least two congruent polygons among them.
Solution. Let us assume that the initial polygon (which we will call
big from now on) has n edges and that it is inscribed in the unit circle,
the vertices having as aﬃxes the numbers 1, ε, ε
2
, . . . , ε
n−1
, where ε =
106
e
2iπ
n
(of course, we will not loose generality with all these restrictions).
Consider now n
1
, n
2
, . . . , n
k
the number of edges of each monochromatic
polygon and let us assume that all these numbers are diﬀerent. Let
ε
j
= e
2iπ
n
j
; observe that the aﬃxes of the vertices of each monochromatic
polygon are z
j
, z
j
ε
j
, . . . , z
j
ε
n
j
−1
j
, for some z
j
complex numbers on the
unit circle. First, a technical result.
Lemma 1. For any complex number z, if ζ = e
2iπ
p
then we have the
identity
1
1 −z
+
1
1 −zζ
+ +
1
1 −zζ
p−1
=
p
1 −z
p
.
Proving this lemma is a very simple task. Indeed, it suﬃces to observe
that z, zζ, . . . , zζ
p−1
are exactly the roots of P(X) = X
p
− z
p
. Or, we
know that
P
t
(x)
P(X)
=
1
X −z
+ +
1
X −zζ
p−1
.
By taking X = 1, we obtain exactly the desired identity.
Now, the hypothesis of the problem and lemma 1 allow to write
n
1
1 −(zz
1
)
n
1
+ +
n
k
1 −(zz
k
)
n
k
=
n
1 −z
n
.
Also, the simple observation that n
1
+n
2
+ +n
k
= n allows the
new identity
n
1
z
n
1
1
1 −(zz
1
)
n
1
z
n
1
+
n
2
z
n
2
2
1 −(zz
2
)
n
2
z
n
2
+ +
n
k
z
n
k
k
1 −(zz
k
)
n
k
z
n
k
=
nz
n
1 −z
n
. (1)
Let us assume now that n
1
< min(n
2
, . . . , n
k
) and let us divide (1)
by z
n
1
. It follows that for non-zero z we have
n
1
z
n
1
1
1 −(zz
1
)
n
1
+
n
2
z
n
2
2
1 −(zz
2
)
n
2
z
n
2
−n
1
+ +
n
k
z
n
k
k
1 −(zz
k
)
n
k
z
n
k
−n
1
=
nz
n−n
1
1 −z
n
.
(2)
107
Well, we are done: it suﬃces to observe that if we make z → 0
(by non-zero values) in (2) , we obtain that z
n
1
1
= 0, which is surely
impossible since [z
1
[ = 1. The proof ends here.
The following problem that we are going to discuss appeared in var-
ious contests under diﬀerent forms. It is a very nice identity that can be
proved elementary in a quite messy way. Here is a magical proof using
formal series.
Example 6. For any complex numbers a
1
, a
2
, . . . , a
n
∈ C, the fol-
lowing identity holds:
_
n

i=1
a
i
_
n

n

i=1
_
_

j,=i
a
j
_
_
n
+

1≤i<j≤n
_
_

k,=i,j
a
k
_
_
n
− + (−1)
n−1
n

i=1
a
n
i
= n!
n

i=1
a
i
.
Solution. Consider the formal series
f(z) =
n

i=1
(e
za
i
−1).
We are going to compute it in two diﬀerent ways. First of all, it is
clear that
f(z) =
n

i=1
_
za
i
+
z
2
a
2
i
2!
+. . .
_
thus we can say that the coeﬃcient of z
n
in this formal series is
n

i=1
a
i
.
On the other hand, we can write
f(z) = e
z
n

i=1
a
i

n

i=1
e
z

j=i
a
j
+ + (−1)
n−1
n

i=1
e
za
i
+ (−1)
n
.
Indeed, the reader is right: now everything is clear, since the coeﬃ-
cient of z
n
in e
kz
is
k
n
n!
. The conclusion is clear: not only the identity is
true, but it has a four-line solution!!!
108
There aren’t only algebra problems that can be solved in an elegant
manner using formal series, but also some beautiful concocts of numbers
theory and combinatorics. We shall focus a little bit more on such type
of problems in the sequel.
Example 7. Let 0 = a
0
< a
1
< a
2
< . . . be a sequence of positive
integers such that the equation a
i
+2a
j
+4a
k
= n has a unique solution
i, j, k. Find a
1998
.
IMO Shortlist, 1998
Solution. Here is a very nice answer: 9817030729. Let A =
¦a
0
, a
1
, . . . ¦ and b
n
= 1 if n ∈ A and 0 otherwise. Next, consider the for-
mal series f(x) =

n≥0
b
n
x
n
, the generating function of the set A (we can
write in a more intuitive way f(x) =

n≥0
x
a
n
). The hypothesis imposed
on the set A translates into
f(x)f(x
2
)f(x
4
) =
1
1 −x
.
Replace x by x
2
k
. We obtain the recursive relation
f(x
2
k
)f(x
2
k+1
)f(x
2
k+2
) =
1
1 −x
2
k
.
Now, observe two relations:

k≥0
f(x
2
k
) =

k≥0
(f(x
2
3k
)f(x
2
3k+1
)f(x
2
3k+2
)) =

k≥0
1
1 −x
2
3k
and

k≥1
f(x
2
k
) =

k≥0
(f(x
2
3k+1
)f(x
2
3k+2
)f(x
2
3k+3
)) =

k≥0
1
1 −x
2
3k+1
.
Therefore (the reader has observed that rigor was not the strong
point in establishing these relations) we have
f(x)=

k≥0
1 −x
2
3k+1
1 −x
2
3k
=

k≥0
(1+x
8
k
)f(x) =

k≥0
1 −x
2
3k+1
1 −x
2
3k
=

k≥0
(1+x
8
k
).
109
This shows that the set A is exactly the set of nonnegative integers
that use only the digits 0 and 1 when written in base 8. A quick com-
putation based on this observation shows that the magical term asked
by the problem is 9817030729.
The following problem is an absolute classic. It appeared, under dif-
ferent forms, in Olympiads from all over the world. We will present the
latest one, given in a Putnam competition:
Example 8. Find all partitions with two classes A, B of the set
of nonnegative integers having the property that for all nonnegative
integers n the equation x + y = n with x < y has as many solutions
(x, y) ∈ AA as in B B.
Solution. Consider f, g the generating functions of A, B and write
them in explicit form
f(x) =

n≥0
a
n
x
n
, g(x) =

n≥0
b
n
x
n
(as in the previous problem, a
n
equals 1 if n ∈ A and 0 otherwise). The
fact that A, B form a partition of the set of nonnegative integers can be
also rewritten as
f(x) +g(x) =

n≥0
x
n
=
1
1 −x
.
Also, the hypothesis made on the number of solutions of the equation
x +y = n imposes that
f
2
(x) −f(x
2
) = g
2
(x) −g(x
2
).
Therefore,
f(x
2
) −g(x
2
) =
f(x) −g(x)
1 −x
,
which can also be rewritten as
f(x) −g(x)
f(x
2
) −g(x
2
)
= 1 −x.
110
Now, the idea is the same as in the previous problems: replace x by
x
2
k
and iterate the process. After multiplication, we deduce that
f(x) −g(x) =

k≥0
(1 −x
2
k
) lim
x→∞
1
f(x
2
n
) −g(x
2
n
)
.
Let us assume without loss of generality that 0 ∈ A. Then the reader
can easily verify that
lim
n→∞
f(x
2
n
) = 1 and lim
n→∞
g(x
2
n
) = 0.
This shows that actually
f(x) −g(x) =

k≥0
(1 −x
2
k
) =

k≥0
(−1)
s
2
(k)
x
k
,
where s
2
(x) is the sum of digits of binary representation of x. Taking
into account the relation
f(x) +g(x) =

n≥0
x
n
=
1
1 −x
,
we ﬁnally deduce that A, B are respectively the set of nonnegative inte-
gers having even (respectively odd) sum of digits when written in binary.
We will discuss two nice problems in which formal series and complex
numbers appear in a quite spectacular way:
Example 9. Let n, k be positive integers such that n ≥ 2
k−1
and
let S = ¦1, 2, . . . , n¦. Prove that the number of subsets A ⊂ S such that

x∈A
x ≡ m (mod 2
k
) does not depend on m ∈ ¦0, 1, . . . , 2
k
−1¦.
Solution. Let us consider the function (call it formal series, if you
want)
f(x) =
n

i=1
(1 +x
i
).
If we prove that 1 + x + + x
2
k
−1
divides f(x), then we have
certainly done the job. In order to prove this, it suﬃces of course to
111
prove that any 2
k
th root of unity, except for 1 is a root of f. But it
suﬃces to observe that for any l ∈ ¦1, 2, . . . , 2
k−1
−1¦ we have
_
cos
2lπ
2
k
+i sin
2lπ
2
k
_
2
k−2−v
2
(l)
= −1
and so
f
_
cos
2lπ
2
k
+i sin
2lπ
2
k
_
= 0,
which proves our claim and ﬁnishes the solution.
Example 10. Let m, n ≥ 2 be positive integers and a
1
, a
2
, . . . , a
n
integers, none of them divisible by m
n−1
. Prove that one can ﬁnd integers
e
1
, e
2
, . . . , e
n
, not all zero, such that [e
i
[ < m for all i and such that
m
n
[e
1
a
1
+e
2
a
2
+ +e
n
a
n
.
IMO Shortlist 2002
Solution. Look at the set A =
_
n

i=1
e
i
a
i
[ 1 ≤ e
i
≤ m
_
and observe
that we can assume that A is a complete system of residues modulo m
n
(otherwise, the conclusion is immediate). Now, consider f(x) =

p∈A
x
a
.
On one hand, we have
f(x) =
n

i=1
_
_
m−1

j=0
x
ja
i
_
_
=
n

i=1
1 −x
ma
i
1 −x
a
i
.
On the other hand, take ε = e
2iπ
m
n
. Since A is a complete system of
residues modulo m
n
, we must have f(ε) = 0. Therefore (the hypothesis
ensures that ε
a
i
,= 1) we must have
n

i=1
(1 − ε
ma
i
) = 0. But this surely
contradicts the fact that none of the numbers a
1
, a
2
, . . . , a
n
is a multiple
of m
n−1
.
Finally, it is time for a tough problem. Of course, it will be a com-
binatorial problem, whose nice solution below was found by Constantin
Tanasescu.
112
Example 11. Let A be the set of all words which can be formed
using m ≥ 2 given letters. For any c ∈ A, let l(c) be its length. Also,
let C ⊆ A be a set of words. We know that any word from A can be
obtained in at most one way by concatenating words from C. Prove the
inequality:

c∈C
1
m
l(c)
≤ 1.
Solution. Let S be the set of all words which can be obtained by
concatenating words from C. Let
f(x) =

c∈C
x
l(c)
, g(x) =

s∈S
x
l(s)
.
By the deﬁnition of S, we have that:
g(x) = 1 +f(x) +f
2
(x) + =
1
1 −f(x)
.
Therefore,
f(x)g(x) = g(x) −1. (∗)
Now, S (and C) has at most m
k
elements of length k, thus g(x) < ∞
and f(x) < ∞ for x <
1
m
. Thus, for all x ∈
_
0,
1
m
_
:
f(x)g(x) = g(x) −1 < g(x)
and so f(x) < 1 for all x ∈
_
0,
1
m
_
. All we need now is to make x tend
to
1
m
and we will obtain that f
_
1
m
_
≤ 1, which is nothing else than
the desired inequality.
Proposed problems
1. Let z
1
, z
2
, . . . , z
n
be some arbitrary complex numbers. Prove that
for any ε > 0 there are inﬁnitely many numbers n such that
k
_
[z
k
1
+z
k
2
+ +z
k
n
[ > max([z
1
[, [z
2
[, . . . , [z
n
[) −ε.
113
2. Find the general formula for the sequence (x
n
)
n≥1
given by
x
n+k
= a
1
x
n+k−1
+ +a
k
x
n
in function of x
0
, x
1
, . . . , x
k−1
. Here a
1
, . . . , a
k
are arbitrary complex
numbers.
3. Prove that if we partition the natural numbers into a ﬁnite number
of inﬁnite arithmetic progressions, then there will be two of them having
the same ratio.
4. How many polynomials P with coeﬃcients 0, 1, 2 or 3 satisfy
P(2) = n, where n is a given positive integer?
Romanian TST, 1994
5. Deﬁne A
1
= ∅, B
1
= ¦0¦ and A
n+1
= ¦1 + x[ x ∈ B
n
¦, B
n+1
=
(A
n
¸ B
n
) ∪(B
n
¸ A
n
). What are the positive integers n such that B
n
=
¦0¦?
AMM
6. In how many ways can we parenthesis a non-associative product
a
1
a
2
. . . a
n
?
Catalan’s problem
7. For which positive integers n can we ﬁnd real numbers a
1
, a
2
, . . . ,
a
n
such that
¦[a
i
−a
j
[ [ 1 ≤ i < j ≤ n¦ =
_
1, 2, . . . ,
_
n
2
__
?
China TST 2002
8. Let a
1
, a
2
, . . . , a
n
relatively prime positive integers. Find in closed
form a sequence (x
n
)
n≥1
such that if (y
n
)
n≥1
is the number of positive
integral solutions to the equation a
1
x
1
+ a
2
x
2
+ + a
n
x
n
= k, then
lim
n→∞
x
n
y
n
= 1.
114
9. Let A
1
, A
2
, . . . , A
k
∈ M
n
(C) be some complex nθn matrices such
that
|A
p
1
+A
p
2
+ +A
p
k
| ≤
C
p!
for any natural number p ≥ 1. Here C does not depend on p ≥ 1 and
|X| = max
1≤i,j≤n
[x
ij
[. Then prove that A
n
i
= 0 for all 1 ≤ i ≤ k.
Gabriel Dospinescu
10. Is there an inﬁnite set of natural numbers such that all suﬃ-
ciently large integer can be represented in the same number of ways as
the sum of two elements of the set?
D. Newman
11. Find all possibilities to color a regular polygon in the way pre-
sented in example 5.
12. Find all positive integers n with the following property: for any
real numbers a
1
, a
2
, . . . , a
n
, knowing the numbers a
i
+ a
j
, i < j, deter-
mines a
1
, a
2
, . . . , a
n
uniquely.
Erdos and Selfridge
13. Suppose that a
0
= a
1
= 1, (n + 3)a
n+1
= (2n + 3)a
n
+ 3na
n−1
.
Prove that all terms of this sequence are integers.
Komal
14. Deﬁne two sequences of integer numbers (a
n
), (b
n
) : a
1
= b
1
= 0
and
a
n
= nb
n
+a
1
b
n−1
+a
2
b
n−2
+ +a
n−1
b
1
.
Prove that for any prime number p we have p[a
p
.
Komal
15. Is it possible to partition the set of all 12-digit numbers into
groups of 4 numbers such that the numbers in each group have the same
digits in 11 places and four consecutive digits in the remaining place?
115
16. Prove the following identity
n

k=1

1≤i
1
<i
2
<<i
k
≤n

ε
1

2
,...,ε
n
∈¡−1,1¦
(−1)
k
2
k

1
a
i
1

2
a
i
2
+ +ε
k
a
i
k
)
2n
=
(−1)
n
(2n)!a
2
1
a
2
2
. . . a
2
n
2
n
for any real numbers a
1
, a
2
, . . . , a
n
.
Gabriel Dospinescu
17. A set of positive integers A has the property that for some
positive integers b
i
, c
i
, the sets b
i
A + c
i
, 1 ≤ i ≤ n are disjoint subsets
of A. Prove that
n

i=1
1
b
i
≤ 1.
IMO Shortlist 2004
116
NUMBERS AND LINEAR ALGEBRA
We have seen how analysis can help in solving number theory prob-
lems. But linear algebra has an important role as well, especially because
it makes a beautiful connection between number theory and algebra.
This discussion practically started from the following diﬃcult problem
that we solved in the chapter ”Look at the exponent!” and which ap-
peared in the American Mathematical Monthly a long time ago.
For any integers a
1
, a
2
, . . . , a
n
the number

1≤i<j≤n
a
j
−a
i
j −i
is an in-
teger.
You will see a nice and short solution to this problem. At the ap-
propriate time... But ﬁrst, we need some basic facts about matrices,
determinants, and systems of linear equations. For example, the fact
that any homogeneous linear system
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
a
11
x
1
+a
12
x
2
+ +a
1n
x
n
= 0
a
21
x
1
+a
22
x
2
+ +a
2n
x
n
= 0
a
n1
x
1
+a
n2
x
2
+ +a
nn
x
n
= 0
in which
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
a
11
a
12
. . . a
1n
a
21
a
22
. . . a
2n
. . . . . . . . . . . .
a
n1
a
n2
. . . a
nn
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
,= 0
has only the trivial solution. Secondly, we need Vandermonde’s identity
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 x
1
x
2
1
. . . x
n−1
1
1 x
2
x
2
2
. . . x
n−1
2
. . . . . . . . . . . . . . .
1 x
n
x
2
n
. . . x
n−1
n
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
=

1≤i<j≤n
(x
j
−x
i
). (1)
117
With these basic facts (of course, for a better understanding, the
reader should have some more knowledge on linear algebra), we are
sical problem. This time, we will prove a result from the theory of per-
mutations. Here is a nice solution.
Example 1. Let σ be a permutation of the numbers 1, 2, . . . , n.
Then
¸
¸
¸
¸
¸
¸

1≤i<j≤n
(σ(j) −σ(i))
¸
¸
¸
¸
¸
¸
= 1! 2! . . . (n −1)!.
Solution. The formula in the left-hand side suggests that we might
use Vandermonde’s identity (1). But we also need a small trick. Using
the fact that det A = det
t
A for any matrix A, we get
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 1 1 . . . 1
σ(1) σ(2) σ(3) . . . σ(n)
. . . . . . . . . . . . . . .
σ(1)
n−1
σ(2)
n−1
σ(3)
n−1
. . . σ(n)
n−1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
=

1≤i<j≤n
(σ(j) −σ(i)).
So, by multiplying the two determinants we ﬁnd
_
_

1≤i<j≤n
(σ(j) −σ(i))
_
_
2
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 1 1 . . . 1
σ(1) σ(2) σ(3) . . . σ(n)
. . . . . . . . . . . . . . .
σ(1)
n−1
σ(2)
n−1
σ(3)
n−1
. . . σ(n)
n−1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸

¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 σ(1) . . . σ(1)
n−1
1 σ(2) . . . σ(2)
n−1
. . . . . . . . . . . .
1 σ(n) . . . σ(n)
n−1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
S
0
S
1
. . . S
n−1
S
1
S
2
. . . S
n
. . . . . . . . . . . .
S
n−1
. . . . . . S
2n−2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
,
118
where
S
i
= σ(1)
i
+σ(2)
i
+ +σ(n)
i
.
But since σ is a permutation of the numbers 1, 2, . . . , n, we have
S
i
= 1
i
+ 2
i
+ +n
i
and repeating the arguments we conclude that
_
_

1≤i<j≤n
(σ(j) −σ(i))
_
_
2
=
_
_

1≤i<j≤n
(j −i)
_
_
2
.
Hence
¸
¸
¸
¸
¸
¸

1≤i<j≤n
(σ(j) −σ(i))
¸
¸
¸
¸
¸
¸
=

1≤i<j≤n
(j −i) = (n −1)!(n −2)! . . . 1!.
Using this result, we can answer immediately the following question:
Example 2. Given a polynomial with complex coeﬃcients, can we
decide if it has a double zero only by performing additions, multiplica-
tions, and divisions on its coeﬃcients?
Solution. Yes, we can. Let f(x) = a
0
+a
1
x + +a
n
x
n
. Then this
polynomial has a double root if and only if
_
_

1≤i<j≤n
(x
i
−x
j
)
_
_
2
= 0,
where x
1
, x
2
, . . . , x
n
are the zeros of the polynomial. But we have seen
that
_
_

1≤i<j≤n
(x
i
−x
j
)
_
_
2
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
S
0
S
1
. . . S
n−1
S
1
S
2
. . . S
n
. . . . . . . . . . . .
S
n−1
. . . . . . S
2n−2
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
,
where S
i
= x
i
1
+x
i
2
+ +x
i
n
. So, we need to express S
i
= x
i
1
+x
i
2
+ +x
i
n
in terms of the coeﬃcients of the polynomial. But this is a consequence
of Newton’s and Vieta’s formulas, which combined yield
a
n
S
i
+a
n−1
S
i−1
+ +a
n−i+1
S
1
+ia
n−i
S
i
= 0, i ∈ ¦1, 2, . . . , ¦.
119
The last formula allows us to prove by induction that S
i
can be ex-
pressed only in terms of the coeﬃcients of the polynomial (this could
have been shown much easier, since after all S
i
is a symmetric polyno-
mial in n variables, hence it can be expressed only in terms of the fun-
damental symmetric polynomials, which can also be expressed in terms
of the coeﬃcients due to Vieta’s formulas). Consequently, we can decide
whether
_
_

1≤i<j≤n
(x
i
−x
j
)
_
_
2
= 0
only by using the described operations on the coeﬃcients of the polyno-
mial, which shows that the answer to the problem is positive.
You may know the following classical problem: if a, b, c ∈ Q verify
a + b
3

2 + c
3

4 = 0, then a = b = c = 0. Have you ever thought about
the general case? This cannot be done only with simple tricks. We need
much more. Of course, a direct solution could be the following: from
Eisenstein’s criterion, the polynomial f(X) = X
n
− 2 is irreducible,
so if a
0
+ a
1
n

2 + + a
n−1
n−1

2
n−1
= 0 for some rational numbers
a
0
, a
1
, . . . , a
n−1
, then the polynomial g(x) = a
0
+ a
1
x + + a
n−1
x
n−1
is not relatively prime with f. Hence gcd(f, g) is a polynomial of degree
at most n−1 that divides an irreducible polynomial f of degree n. This
cannot happen, unless g = 0, i.e. a
0
= a
1
= = a
n−1
= 0. But here
is a beautiful proof using linear algebra. This time we will have to be
careful to work in the most appropriate ﬁeld.
Example 3. Prove that if a
0
, a
1
, . . . , a
n−1
∈ Q satisfy
a
0
+a
1
n

2 + +a
n−1
n−1

2
n−1
= 0,
then a
0
= a
1
= = a
n−1
= 0.
Solution. If a
0
+a
1
n

2 + +a
n−1
n−1

2
n−1
= 0, then
ka
0
+ka
1
n

2 + +ka
n−1
n−1

2
n−1
= 0
120
for any real number k. Thus, we may assume that a
0
, a
1
, . . . , a
n−1
∈ Z
(for example, we can choose k to be the least common multiple of all de-
nominators of the numbers a
0
, a
1
, . . . , a
n−1
). The idea is to choose n val-
ues for k so that to obtain a system of linear equations, having nontrivial
solutions. Then, the determinant of the system must be zero and this
will imply a
0
= a
1
= = a
n−1
= 0. Now, let us ﬁll in the blanks. What
are the best values for k? This can be seen by noticing that
n

2
n−1

n

2 =
2 ∈ Z. So, the values (k
1
, k
2
, . . . , k
n
) = (1,
n

2, . . . ,
n

2
n−1
) are good and
the system becomes
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
a
0
+a
1

n

2 + +a
n−1

n

2
n−1
= 0
a
0

n

2 +a
1

n

2
2
+ + 2a
n−1
= 0
. . .
a
0

n

2
n−1
+ 2a
1
+ +a
n−1

n

2
n−2
= 0
Viewing (1,
n

2, . . . ,
n

2
n−1
) as a nontrivial solution to the system,
we conclude that
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
a
0
a
1
. . . a
n−1
2a
n−1
a
0
. . . a
n−2
. . . . . . . . . . . .
2a
1
2a
2
. . . a
0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
= 0.
But what can we do now? Expanding the determinant leads nowhere.
As we said before passing to the solution, we should always work in the
most appropriate ﬁeld. This time the ﬁeld is Z
2
, since in this case the
determinant can be easily computed. It equals a
n
0
= 0. Hence a
0
must
be even, that is a
0
= 2b
0
and we have
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
b
0
a
1
. . . a
n−1
a
n−1
a
0
. . . a
n−2
. . . . . . . . . . . .
a
1
2a
2
. . . a
0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
= 0.
121
Now, we interchange the ﬁrst two lines of the determinant. Its value
remains 0, but when we expand it in Z
2
, it is a
n
1
= 0. Similarly, we
ﬁnd that all a
i
are even. Let us write a
i
= 2b
i
. Then we also have
b
0
+b
1

n

2 + +b
n−1

n−1

2
n−1
= 0 and with the same reasoning we
conclude that all b
i
are even. But of course, we can repeat this as long
as we want. By the method of inﬁnite descent, we ﬁnd that a
0
= a
1
=
= a
n−1
= 0.
The above solution might seem exaggeratedly diﬃcult compared
with the one using Eisenstein’s criterion, but the idea was too nice not
to be presented here.
The following problem can become a nightmare despite its simplicity.
Example 4. Let A = ¦a
3
+ b
3
+ c
3
− 3abc[ a, b, c ∈ Z¦. Prove that
if x, y ∈ A then xy ∈ A.
Proof. The observation that
a
3
+b
3
+c
3
−3abc =
¸
¸
¸
¸
¸
¸
¸
¸
a c b
b a c
c b a
¸
¸
¸
¸
¸
¸
¸
¸
leads to a quick solution. Indeed, it suﬃces to note that
_
_
_
_
a c b
b a c
c b a
_
_
_
_
_
_
_
_
x z y
y x z
z y x
_
_
_
_
=
=
_
_
_
_
ax +cy +bz az +by +cx ay +bx +cz
ay +bx +cz ax +cy +bz az +by +cx
az +by +cx ay +bx +cz ax +cy +bz
_
_
_
_
and thus
(a
3
+b
3
+c
3
−3abc)(x
2
+y
3
+z
3
−3xyz) = A
3
+B
3
+C
3
−3ABC,
122
where A = ax +cy +bz, B = az +by +cx, C = ax +cy +bz. You see,
identities are not so hard to ﬁnd...
We all know the famous Bezout’s theorem, stating that if a
1
, a
2
, . . . ,
a
n
are relatively prime, then one can ﬁnd integers k
1
, k
2
, . . . , k
n
such
that k
1
a
1
+ k
2
a
2
+ + k
n
a
n
= 1. The following problem claims more,
at least for n = 3.
Example 5. Prove that if a, b, c are relatively prime integers, then
there are integers x, y, z, u, v, w such that
a(yw −zv) +b(zu −xw) +c(xv −yu) = 1.
Solution. First of all, there is a crucial observation to be made: the
given condition can be also written in the form det A = 1, where
A =
_
_
_
_
a x u
b y v
c z w
_
_
_
_
.
So, let us prove a much more general result.
Theorem. Any vector v whose integer components are relatively
prime is the ﬁrst column of an integral matrix with determinant equal
to 1.
There is a simple proof of this theorem, using clever manipulations of
determinant properties and induction on the dimension n of the vector
v. Indeed, for n = 2 it is exactly Bezout’s theorem. Now, assume that
it is true for vectors in Z
n−1
and take v = (v
1
, v
2
, . . . , v
n
) such that v
i
are relatively prime. Consider the numbers
v
1
g
, . . . ,
v
n−1
g
, where g is the
greatest common divisor of v
1
, . . . , v
n−1
. They are relatively prime and
123
thus we can ﬁnd an integral matrix
_
_
_
_
_
v
1
g
a
12
. . . a
1,n−1
. . . . . . . . . . . .
v
n−1
g
a
n−1,2
. . . a
n−1,n
_
_
_
_
_
having determinant equal to 1. Now, using Bezout’s theorem, we can
ﬁnd α, β such that αg +βv
n
= 1. In this case, it is not diﬃcult to verify
that the following matrix has integral entries and determinant equal to
1:
_
_
_
_
_
_
_
_
v
1
a
12
. . . a
1,n−1
(−1)
n−1
β
v
1
g
. . . . . . . . . . . . . . .
v
n−1
a
n−1,2
. . . a
n−1,n−1
(−1)
n−1
β
v
n−1
g
v
n
0 . . . 0 (−1)
n−1
α
_
_
_
_
_
_
_
_
.
We said at the beginning that the discussion started from the diﬃcult
problem that appeared in AMM, but yet we did not present its solution
yet. It is now time to do it.
Example 6. For any integers a
1
, a
2
, . . . , a
n
then

1≤i<j≤n
a
j
−a
i
j −i
∈ Z.
Armond Spencer, AMM E 2637
Solution. With this introduction, the way to proceed is clear. What
does the expression

1≤i<j≤n
(a
j
− a
i
suggest? It is the Vandermonde’s
identity (1), associated to a
1
, a
2
, . . . , a
n
. But we have a hurdle here. We
might want to use the same formula for the expression

1≤i<j≤n
(j − i).
This is a dead end. But we have seen what is

1≤i<j≤n
(j −i) equal to in
124
the ﬁrst problem. It equals (n −1)!(n −2)! . . . 1!. Now, we can write

1≤i<j≤n
a
j
−a
i
j −i
=
1
1! 2! . . . (n −1)!
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 1 1 . . . 1
a
1
a
2
a
3
. . . a
n
. . . . . . . . . . . . . . .
a
n−1
1
a
n−1
2
a
n−1
3
a
n−1
n
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
.
As usual, the last step is the most important. The above formula
can be rewritten as

1≤i<j≤n
a
j
−a
i
j −i
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 1 1 . . . 1
a
1
1!
a
2
1!
a
3
1!
. . .
a
n
1!
. . . . . . . . . . . . . . .
a
n−1
1
(n −1)!
a
n−1
2
(n −1)!
a
n−1
2
(n −1)!
. . .
a
n−1
n
(n −1)!
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
.
And now we recognize the form

1≤i<j≤n
a
j
−a
i
j −i
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 1 . . . 1
_
a
1
1
_ _
a
2
1
_
. . .
_
a
n
1
_
_
a
1
2
_ _
a
2
2
_
. . .
_
a
n
2
_
. . . . . . . . . . . .
_
a
1
n −1
_ _
a
2
n −1
_
. . .
_
a
n
n −1
_
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
,
which can be proved easily by subtracting lines. Because each number
_
a
i
j
_
is an integer, the determinant itself is an integer and the conclusion
follows.
We end the unit with a very nice and diﬃcult problem that also
appeared in AMM in 1998. A variant of this problem was given in 2004
at a TST in Romania and turned out to be a hard problem.
125
Example 7. Consider the sequence (x
n
)
n≥0
deﬁned by x
0
= 4,
x
1
= x
2
= 0, x
3
= 3 and x
n+4
= x
n
+x
n+1
. Prove that for any prime p
the number x
p
is a multiple of p.
AMM
Solution. Let us consider the matrix
A =
_
_
_
_
_
_
_
0 0 0 1
1 0 0 1
0 1 0 0
0 0 1 0
_
_
_
_
_
_
_
and let trX be the sum of the entries of the main diagonal of the matrix
X. We will ﬁrst prove that x
n
= TrA
n
(here A
0
= I
4
). This is going to
be the easy part of the solution. Indeed, for n = 1, 2, 3 it is not diﬃcult to
verify it. Now, assume that the statement is true for all i = 1, 2, . . . , n−1
and prove that it is also true for n. This is true because
x
n
= x
n−4
+x
n−3
= TrA
n−4
+TrA
n−3
= Tr(A
n−4
(A+I
4
)) = TrA
n
.
We have used here the relation A
4
= A + I
4
, which can be easily
veriﬁed by a simple computation. Hence the claim is proved.
Now, let us prove an important result, that is TrA
p
≡ TrA (mod p)
for any integral matrix and any prime p. The proof is not trivial at
all. A possible advanced solution is to start by considering the matrix A
obtained by reducing all entries of A modulo p, then by placing ourselves
in a ﬁeld in which the characteristic polynomial of A has all its zeroes
λ
1
, λ
2
, . . . , λ
n
. This ﬁeld has clearly characteristic p (it contains Z
p
) and
so we have (using the binomial formula and the fact that all coeﬃcients
_
p
k
_
, 1 ≤ k ≤ p −1 are multiples of p)
trA
p
=
n

i=1
λ
p
i
=
_
n

i=1
λ
i
_
p
= (TrA)
p
,
126
from where the conclusion is immediate via Fermat’s little theorem.
But there is a beautiful elementary solution. Let us consider two
integral matrices A, B and write
(A+B)
p
=

A
1
,...,A
p
∈¡A,B¦
A
1
A
2
. . . A
p
.
Observe that for any A, B we have Tr(AB) = Tr(BA) and by in-
duction, for any X
1
, . . . , X
n
and any cyclic permutation σ,
Tr(X
1
. . . X
n
) = Tr(X
σ(1)
. . . X
σ(n)
).
Now, note that in the sum

A
1
,...,A
p
∈¡A,B¦
A
1
A
2
. . . A
p
we can form
2
p
−2
p
groups of p-cycles and we have two more terms A
p
and B
p
. Thus

A
1
,...,A
p
∈¡A,B¦
Tr(A
1
A
2
. . . A
p
) ≡ TrA
p
+TrB
p
already noticed that Fermat’s little theorem comes handy once again),
since the sum of Tr(A
1
A
2
. . . A
p
) is a multiple of p in any cycle. Thus
we have proved that
Tr(A+B)
p
≡ TrA
p
+TrB
p
(mod p)
and by an immediate induction we also have
Tr(A
1
+ +A
k
)
p

k

i=1
TrA
p
i
.
Next, consider the matrices E
ij
that have 1 in position (i, j) and 0
elsewhere. For these matrices we clearly have TrA
p
≡ TrA (mod p) and
by using the above result we can write (using Fermat’s little theorem
one more time):
TrA
p
= Tr
_
_

i,j
a
ij
E
ij
_
_
p

i,j
Tr(a
p
ij
E
p
ij
) ≡

i,j
a
ij
TrE
ij
= TrA (mod p).
127
The result is proved and with it the fact that x
p
is a multiple of p.
Problems for training
1. Let F
1
= 1, F
2
= 1 and F
n
= F
n−1
+ F
n−2
for all n ≥ 3 be the
Fibonacci sequence. Prove that
F
n+1
F
n−1
−F
2
n
= (−1)
n
and F
m+n
= F
n
F
m−1
+F
n+1
F
m
.
2. Let the sequence of polynomials (f
n
)
n≥1
be deﬁned by f
1
(x) = 1,
f
2
(x) = x and f
n+1
(x) = xf
n
(x) + f
n−1
(x). Prove that this sequence
satisﬁes the following Fibonacci-type relations f
m+n
= f
n
f
m−1
+f
n+1
f
m
.
3. Prove that the number

2 +

3 +

5 +

7 is irrational.
4. Compute the product

0≤i<j≤n−1

j
−ε
i
)
2
, where
ε
k
= cos
2kπ
n
+i sin
2kπ
n
for all k ∈ ¦0, 1, . . . , n −1¦.
5. Consider 2005 real numbers with the following property: whenever
we eliminate one number, the rest can be divided into two groups of 1002
numbers each and having the same sum per group. Prove that all the
2005 numbers are equal.
6. Let a, b, c be relatively prime nonzero integers. Prove that for any
relatively prime integers u, v, w satisfying au + bv + cw = 0, there are
integers m, n, p such that
a = nw −pv, b = pu −mw, c = mv −nu.
Octavian Stanasila, TST 1989, Romania
7. Let p be a prime and suppose that the real numbers a
1
, a
2
, . . . ,
a
p+1
have the property: no matter how we eliminate one of them, the
rest of the numbers can be divided into at least two nonempty classes,
128
any two of them being disjoint and each class having the same arithmetic
mean. Prove that a
1
= a
2
= = a
p+1
.
8. Let a, b, c be integers and deﬁne the sequence (x
n
)
n≥0
by x
0
= 4,
x
1
= 0, x
2
= 2c, x
3
= 3b and x
n+3
= ax
n−1
+ bx
n
+ cx
n+1
. Prove that
for any prime p and any positive integer m, the number x
p
m is divisible
by p.
Calin Popescu, TST 2004, Romania
9. Prove that for any integers a
1
, a
2
, . . . , a
n
the following number
lcm(a
1
, a
2
, . . . , a
n
)
a
1
a
2
. . . a
n

1≤i<j≤n
(a
j
−a
i
)
is an integer divisible by 1!2! . . . (n − 2)!. Moreover, we cannot replace
1!2! . . . (n −2)! by any other multiple of 1!2! . . . (n −2)!.
10. Let a
1
, a
2
, . . . , a
n
∈ R. A move is transforming the n-tuple
(x
1
, x
2
, . . . , x
n
) into the n-tuple
_
x
1
+x
2
2
,
x
2
+x
3
2
, . . . ,
x
n−1
+x
n
2
,
x
n
+x
1
2
_
.
1
, a
2
, . . . , a
n
), after
ﬁnitely many moves we obtain an n-tuple (A
1
, A
2
, . . . , A
n
) such that
max
1≤i<j≤n
[A
i
−A
j
[ <
1
2
2005
.
11. Let a
(0)
1
, a
(0)
2
, . . . , a
(0)
n
∈ R and deﬁne a
(k)
i
=
a
(k−1)
i
+a
(k−1)
i+2
2
for
all k ≥ 1 and 1 ≤ i ≤ n (the indices are taken modulo n). Prove that
n

k=0
(−2)
k
_
n
k
_
a
(k)
i
= (−1)
n
a
(0)
i
for all 1 ≤ i ≤ n.
Gabriel Dospinescu
129
ARITHMETIC PROPERTIES OF POLYNOMIALS
Another topic with old fashioned tricks... will surely say the reader
too many time on a problem just because we neglected basic and triv-
ial aspects of it? This is why we think that speaking about these ”old
fashioned tricks” is not lack of imagination, but rather an imperious
need. In this small note we joined together some classical arithmetic
properties of polynomials. Of course, as usual, the list is just a small
and insigniﬁcant introduction to this ﬁeld, but some basic things should
become reﬂex and between them there are also some problems we shall
discuss. As usual, we kept some chestnuts for the end of the note, so the
tough solver will have his own part of lecture, especially in a chapter
like this one, when extremely diﬃcult problems with extremely simple
There is one result that should be remembered, that is for any poly-
nomial f ∈ Z[X] and any diﬀerent integers a, b, a−b divides f(a)−f(b).
Practically, this is the fundamental result that we shall use continuously.
We will start with an essential result, due to Schur, and which ap-
peared in many variants in contests. Although in the topic Analysis
against number theory we proved an even more general result using a
nice analytical argument, we prefer to present here a purely arithmetic
proof.
Example 1. (Schur) Let f ∈ Z[X] be a non constant polynomial.
Then the set of prime numbers dividing at least one non-zero number
between f(1), f(2), . . . , f(n), . . . is inﬁnite.
Proof. First, suppose that f(0) = 1 and consider the numbers f(n!).
For suﬃciently large n, they are non-zero integers. Moreover, f(n!) ≡ 1
(mod n!) and so if we pick a prime divisor of each of the numbers f(n!).
we obtain the conclusion (since in particular any such prime divisor is
130
greater than n). Now, if f(0) = 0, the conclusion is obvious. Suppose thus
that f(0) ,= 0 and consider the polynomial g(x) =
f(xf(0))
f(0)
. Obviously,
g ∈ Z[X] and g(0) = 1. Applying now the ﬁrst part of the solution, we
easily get the conclusion.
This result has, as we have already said, important consequences.
Here is a nice application.
Example 2. Suppose that f, g ∈ Z[X] are monic non constant irre-
ducible polynomials such that for all suﬃciently large n, f(n) and g(n)
have the same set of prime divisors. Then f = g.
Solution. Indeed, by Gauss’s lemma, the two polynomials are ir-
reducible in Q[X]. Even more, if they are not equal, then the above
remark and the fact that they have the same leading coeﬃcient implies
they are relatively prime in Q[X]. Using Bezout’s theorem we conclude
instantly that we can ﬁnd a non zero integer N and P, Q ∈ Z[X] such
that fP +gQ = N. This shows that for all suﬃciently large n, all prime
factors of f(n) divide N. But, of course, this contradicts Schur’s result.
The result of example 2 remains true if we assume the same property
valuable for inﬁnitely many numbers n. Yet, the proof uses some highly
non elementary results of Erdos in this ﬁeld. Interested reader will ﬁnd
a rich literature on this ﬁeld.
A reﬁnement of Schur’s lemma is discussed in the following example.
The ingredient is, as usual, the Chinese remainder theorem.
Example 3. Let f ∈ Z[X] be a non constant polynomial and n, k
some positive integers. Then prove that there exists a positive integer
a such that each of the numbers f(a), f(a + 1), . . . , f(a +n −1) has at
least k distinct prime divisors.
131
Solution. Let us consider an array of diﬀerent prime numbers
(p
ij
)
i,j=1,k
such that for some positive integers x
ij
such that f(x
ij
) ≡ 0
(mod p
ij
). We know that this is possible from Schur’s theorem. Now,
using the Chinese remainder theorem we can ﬁnd a positive integer a
such that a
i
−1 ≡ x
ij
(mod p
ij
). Using the fundamental result, it follows
that each of the numbers f(a), f(a + 1), . . . , f(a +n −1) has at least k
distinct prime divisors.
Classical arithmetic ”tricks” and the fundamental result that a − b
divides f(a) −f(b) are the main ingredients of the following problem.
Example 4. Find all polynomials with integer coeﬃcients f such
that for all suﬃciently large n, f(n)[n
n−1
−1.
Gabriel Dospinescu
Solution. Since clearly f(X) = X −1 is a solution, let us consider
an arbitrary solution and write it in the form f(X) = (X−1)
r
g(X) with
r ≥ 0 and g ∈ Z[X] such that g(1) ,= 0. Thus, there exists M such that
for all n > M we have g(n)[n
n−1
−1.
We will prove that g is constant. Supposing the contrary, then, since
changing g and his opposite has no eﬀect, we may assume that the
leading coeﬃcient of g is positive. Thus one can ﬁnd k > M such that
for all n > k we have g(n) > 2 and g(n)[n
n−1
− 1. Now, since n +
g(n) −n[g(n+g(n)) −g(n), we deduce that g(n)[g(n+g(n)) for all n. In
particular, for all n > k we have g(n)[g(n+g(n))[(n+g(n))
n+g(n)−1
−1
and g(n)[n
n−1
− 1. Of course, this implies that g(n)[n
n+g(n)−1
− 1 =
(n
n−1
− 1)n
g(n)
+ n
g(n)
− 1, that is g(n)[n
g(n)
− 1 for all n > k. Now,
let us consider a prime number p > k and let us look at the smallest
prime divisor of g(p+1) > 2. We clearly have q[g(p+1)[(p+1)
g(p+1)
−1
and q[(p +1)
q−1
−1. Since gcd(g(p +1), q −1) = 1 (by minimality) and
gcd((p + 1)
g(p+1)
− 1, (p + 1)
q−1
− 1) = (p + 1)
gcd(g(p+1),q−1)
− 1 = p, it
follows that we actually have p = q. This shows that p[g(p +q) and thus
132
(again using the fundamental result) p[g(1). Since this happens for any
prime number p > k, we must have g(1) = 0. This contradiction shows
that g is indeed constant.
Let g(X) = c. Thus, c[2
n(2
n
−1)−1
−1 for all n > M. (2
n
> M). Since
gcd(2
a
−1, 2
b
−1) = 2
gcd(a,b)
−1, in order to show that [c[ = 1, it suﬃces to
exhibit k < m < n such that gcd(m(2
m
−1), n(2
n
−1)) = 1. This is very
simple to realize. Indeed, it suﬃces to take m a prime number greater
than M, k and to choose n a prime number greater than m(2
m
− 1).
A simple argument shows that gcd(m(2
m
− 1), n(2
n
− 1)) = 1 and so
[c[ = 1.
Finally, let us prove that r ≤ 2. Supposing the contrary, we deduce
that
(n −1)
3
[n
n−1
−1 ⇔ (n −1)
1
[n
n−2
+n
n−3
+ +n + 1
for all suﬃciently large n and since
n
n−2
+n
n−3
+ +n + 1 =
= n −1 + (n −1)[n
n−3
+ 2n
n−4
+ + (n −3)n + (n −2)],
we obtain that n −1[n
n−3
+2n
n−4
+ +(n −3)n +(n −2) +1 for all
suﬃciently large n, which is clearly impossible, since
n
n−3
+ 2n
n−4
+ + (n −3)n + (n −2) + 1 ≡ 1 + 2 + + (n −2) + 1

(n −1)(n −2
2
+ 1 (mod n −1).
Hence r ≤ 2. Finally, the relation
n
n−1
−1 = (n −1)
2
[n
n−3
+ 2m
n−4
+ + (n −3)n + (n −2) + 1]
shows that (n − 1)
2
[n
n−1
− 1 for all n > 1 and allows to conclude that
all solutions are the polynomials ±(X −1)
r
, with r ∈ ¦0, 1, 2¦.
have the impression the problem is very simple. Actually, it is extremely
133
diﬃcult. There are many possible approaches that fail and the time spent
for solving such a problem can very well tend to inﬁnity.
Example 5. Let f ∈ Z[X] be a non constant polynomial and k ≥ 2
a positive integer such that
k
_
f(n) ∈ Q for all positive integers n. Then
there exists a polynomial g ∈ Z[X] such that f = g
k
.
Folklore
Solution. Let us assume the contrary and let us decompose f =
p
k
1
1
. . . p
k
s
s
g
k
where 1 ≤ k
i
< k and p
i
are diﬀerent irreducible polyno-
mials in Q[X]. Suppose that s ≥ 1 (which is the same as denying the
conclusion). Since p
1
is irreducible in Q[X], it is relatively prime with
p
1
p
2
. . . p
s
and thus (using Bezout’s theorem and multiplication with in-
tegers) there exist some polynomials with integer coeﬃcients Q, R and
a positive integer c such that
Q(x)p
1
(x) +R(x)p
1
(x)p
2
(x) . . . p
s
(x) = c.
Now, using the result from example 1, we can take a prime number
q > [c[ and a number n such that q[p
1
(n) ,= 0. We shall have of course
q[p
1
(n+q) (since p
1
(n+q) ≡ p
1
(n) (mod q)). The choice q > [c[ ensures
that q does not divide p
1
(n)p
2
(n) . . . p
s
(n) and so v
q
(f(n)) = v
q
(p
1
(n))+
kv
q
(g(n)). But the hypothesis says that k[v
q
(f(n)), so we must have
v
q
(p
1
(n)) > 2. In exactly the same way we obtain v
q
(p
1
(n+q)) ≥ 2. Yet,
using the binomial formula, we can easily establish the congruency
p
1
(n +q) ≡ p
1
(n) +qp
t
1
(n) (mod q
2
).
Therefore, we must have q[p
1
(n), which contradicts q > [c[ and
Q(x)p
1
(x) +R(x)p
1
(x)p
2
(x) . . . p
s
(x) = c.
This contradiction shows that the hypothesis s ≥ 1 was wrong and
thus the result of the problem follows.
134
The next problem was given in the USA TST 2005 and uses a nice
combination of arithmetic considerations and computations using com-
plex numbers. There are many arithmetic properties of polynomials spec-
ulated in this problem, although the problem itself is not so diﬃcult, if
we ﬁnd the good way to solve it, of course...
Example 6. Let us call a polynomial f ∈ Z[X] special if for any
positive integer k > 1, in the sequence f(1), f(2), f(3), . . . one can ﬁnd
numbers which are relatively prime with k. Prove that for any n > 1, at
least 71% of all monic polynomials of degree n, with coeﬃcient in the
set ¦1, 2, . . . , n!¦ are special.
Titu Andreescu, Gabriel Dospinescu, USA TST 2005
Solution. Of course, before counting such polynomials, it would be
better to ﬁnd an easier characterization for them.
Let p
1
, p
2
, . . . , p
r
all prime numbers at most equal to n and let us
consider the sets A
i
= ¦f ∈ M[ p
i
[f(m), ∀ m ∈ N

¦, where M is
the set of monic polynomials of degree n, with coeﬃcient in the set
¦1, 2, . . . , n!¦. We shall prove that the set T of special polynomials is
exactly M ¸
r
_
i=1
A
i
. Obviously, we have T ⊂ M ¸
_
i≤r
A
i
. The converse,
however is not that easy. Let us suppose that f ∈ Z[X] belongs to
M ¸
r
_
i=1
A
i
and let p be a prime number greater than n. Since f is
monic, Lagrange’s theorem ensures that we can ﬁnd m such that p is
not a divisor of f(m). It follows then that for any prime number q at
least one of the numbers f(1), f(2), f(3), . . . is not a multiple of q. Let
now k > 1 and q
1
, q
2
, . . . , q
s
its prime divisors. Then we know we can
ﬁnd u
1
, . . . , u
s
such that q
i
does not divide f(u
i
). Using the Chinese
remainder theorem, we can ﬁnd a positive integer x such that x ≡ u
i
(mod q
i
). Then f(x) ≡ f(u
i
) (mod q
i
) and thus q
i
does not divide f(x),
thus gcd(f(x), k) = 1. The equality of the two sets is thus proved.
135
Using a brutal estimation, we obtain
[T[ = [M[ −
¸
¸
¸
¸
¸
r
_
i=1
A
i
¸
¸
¸
¸
¸
≥ [M[ −
r

i=1
[A
i
[.
Let’s compute now [A
i
[. Actually, we will show that
(n!)
n
p
p
i
i
. Let f a
monic polynomial in A
i
,
f(X) = X
n
+a
n−1
X
n−1
+ +a
1
X +a
0
.
Then for any m > 1 we have
0 ≡ f(m) ≡ a
0
+ (a
1
+a
p
+a
2p−1
+a
3p−2
+. . . )m
+(a
2
+a
p+1
+a
2p
+. . . )m
2
+ +(a
p−1
+a
2p−2
+a
3p−3
+. . . )m
p−1
(mod p),
where, for simplicity, we put p = p
i
. Using Lagrange’s theorem it follows
that p[a
0
, p[a
1
+a
p
+a
2p−1
+. . . , . . . , p[a
p−1
+a
2p−2
+. . . We are going
to use this later, but we still need a small observation. Let us count the
number of s-tuples (x
1
, x
2
, . . . , x
s
) ∈ ¦1, 2, . . . , n!¦
s
such that x
1
+x
2
+
+x
s
≡ u (mod p), where u is ﬁxed. Let
ε = cos

p
+i sin

p
.
Let us observe that
0 = (ε +ε
2
+ +ε
n!
)
s
=
p−1

k=0
ε
k
[¦(x
1
, x
2
, . . . , x
s
) ∈ ¦1, 2, . . . , n!¦
s
[ x
1
+ +x
s
≡ k (mod p)¦[.
A simple argument related to the irreducibility of the polynomial
1 + X + X
2
+ + X
p−1
shows that all cardinals that appear in the
above sum are equal and that their sum is (n!)
s
, thus each cardinal
equals
(n!)
s
p
.
We are now ready to ﬁnish the proof. Assume that among the
numbers a
1
, a
p
, a
2p−1
, . . . there are exactly v
1
numbers that among
136
a
p−1
, a
2p−2
, . . . there are v
p−1
numbers. Using the above observations,
it follows that we have
[A
i
[ =
n!
p

(n!)
v
1
p
. . .
(n!)
v
p−1
p
=
(n!)
n
p
p
.
Hence, we obtain
[T[ ≥ (n!)
n

p prime
(n!)
n
p
p
.
But
1
5
5
+
1
7
7
+ <
1
5
5
_
1 +
1
5
+
1
5
2
+. . .
_
<
1
1000
and so the percent of special polynomials is at least
100
_
1 −
1
4

1
27

1
1000
_
= 75 −
100
27

1
10
> 71.
The solution of the problem ends here.
Example 7. Suppose that the non constant polynomial f with in-
teger coeﬃcients has no double roots. Then for any positive integer r
there exists n such that f(n) has at least r distinct prime divisors, all of
them appearing with exponent 1 in the decomposition of f(n) in prime
factors.
Solution. Already for r = 1 the problem is in no way obvious, so
let’s not try to attack it directly and concentrate at ﬁrst on the case
r = 1. Suppose the contrary, that is for all n, all prime divisors of
f(n) appear with exponent at least 2. Since f has no double root, we
deduce that gcd(f, f
t
) = 1 in C[X] and thus also in Q[X] (because of
the division algorithm and Euclid’s algorithm). Using Bezout’s theorem
in Q[X], we deduce that we can ﬁnd integer polynomials P, Q such that
P(n)f(n) + Q(n)f
t
(n) = c for a certain positive integer c. Using the
result stated in the ﬁrst example, we can choose q > c a prime divisor
137
of a certain f(n). The hypothesis made ensures that q
2
[f(n). But then
we also have q[f(n + q) and so q
2
[f(n + q). Using Newton’s binomial
formula, we deduce immediately that f(n+q) ≡ f(n)+qf
t
(n) (mod q
2
).
We ﬁnally deduce that q[p
t
(n) and so q[c, impossible since our choice was
q > c. Thus the case r = 1 was proved.
Let us try to prove the property by induction and suppose it is true
for r. Of course, the existence of P, Qsuch that P(n)f(n)+Q(n)f
t
(n) = c
for a certain positive integer c did not depend on r, so we keep the above
notations. By inductive hypothesis, there is n such that at least r prime
divisors of f(n) appear with exponent 1. Let these prime factors be
p
1
, p
2
, . . . , p
r
. But it is obvious that n +kp
2
1
p
2
2
. . . p
2
r
has the same prop-
erty: all prime divisors p
1
, p
2
, . . . , p
r
have exponent 1 in the decomposi-
tion of f(n+kp
2
1
p
2
2
. . . p
2
r
). Since at most a ﬁnite number among them can
be roots of f, we may very well suppose from the beginning that n is not
a root of f. Consider now the polynomial g(X) = f(n + (p
1
. . . p
r
)
2
X),
which is obviously non constant. Thus, using again the result in exam-
ple 1, we can ﬁnd q > max¦[c[, p
1
, . . . , p
r
, [p(n)[¦ a prime number and
a number u such that q[g(u). If v
q
(g(u)) = 1, victory is ours, since
a trivial veriﬁcation shows that q, p
1
, . . . , p
r
are diﬀerent prime num-
bers whose exponents in f(n +(p
1
. . . p
r
)
2
u) are all 1. The diﬃcult case
is when v
q
(g(u)) ≥ 2. In this case, we shall consider the number N =
n+u(p
1
. . . p
r
)
2
+uq(p
1
. . . p
r
)
2
. Let us prove that in the decomposition of
f(N), all prime numbers q, p
1
, . . . , p
r
appear with exponent 1. For any p
i
,
this is obvious since f(N) ≡ f(n) (mod (p
1
. . . p
r
)
2
). Using once again
binomial formula, we easily obtain that f(N) ≡ f(n + (p
1
. . . p
r
)
2
u) +
uq(p
1
. . . p
r
)
2
f
t
(N) (mod q
2
). Now, if v
q
(f(n)) ≥ 2, then since v
q
(f(n +
(p
1
. . . p
r
)
2
u)) = v
q
(g(u)) ≥ 2, we have q[u(p
1
. . . p
r
)
2
f
t
(N). Remember
that the choice was q > max¦[c[, p
1
, . . . , p
r
, [p(n)[¦ so necessarily q[u (if
q[f
t
(N) ⇒ q[(f(N), f
t
(N))[c ⇒ q ≤ [c[, contradiction). But since q[g(u),
138
we must have q[g(0) = f(n). But hopefully, we ensured that n is not a
roof of our polynomial and also that q > max¦[c[, p
1
, . . . , p
r
, [p(n)[¦ so
that the last divisibility relation cannot hold. This allows to ﬁnish the
induction step and to solve the problem.
Example 8. Find all non constant polynomials f with integer coef-
ﬁcients and with the following property: for any relatively prime positive
integers a, b, the sequence (f(an+b))
n≥1
contains an inﬁnite number of
terms, any two of which are relatively prime.
Gabriel Dospinescu
Solution. We will prove that the only polynomials with the spec-
iﬁed property are those of the form X
n
, −X
n
with n a positive inte-
ger. Because changing f with its opposite does not modify the property
of the polynomial, we can suppose that the leading coeﬃcient of f is
positive. Thus, there exists a constant M such that for any n > M
we have f(n) > 2. From now on, we consider only n > M. Let us
prove that we have gcd(f(n), n) ,= 1 for any such n. Suppose that
there is n > M such that gcd(f(n), n) = 1. Consequently, the sequence
(f(n+kf(n))
n≥1
will contain at least two relatively prime numbers. Let
them be s, r. Since f(n)[kf(n) = kf(n) + n − n[f(kf(n) + n) − f(n),
we have f(n)[f(n + kf(n)) for any positive integer k. Hence ,we ob-
tain that s, r are multiples of f(n) > 2, impossible. We have shown
that gcd(f(n), n) ,= 1 for any n > M. Thus ,for any prime p > M we
have p[f(p) and so p[f(0). Since any integer diﬀerent from zero has a
ﬁnite number of divisors, we conclude that f(0) = 0. Thus, there is a
polynomial q with integer coeﬃcients such that f(X) = Xq(X). It is
obvious that q has positive leading coeﬃcient and the same property
as f. Repeating the above argument, we infer that if q is non-constant,
then q(0) = 0 and q(X) = Xh(X). Since f is not constant, the above
argument cannot be repeated inﬁnitely many times and thus one of the
139
polynomials g, h must be constant. Consequently, there are positive inte-
gers n, k such that f(X) = kX
n
. But since the sequence (f(2n +3))
n≥1
contains at least two relatively prime integers, we must have k = 1. We
have obtained that f must have the form X
n
. But since f is a solution
if and only if −f is a solution, we infer that any solution of the problem
is a polynomial of the form X
n
, −X
n
.
Now let us prove that the polynomials of the form X
n
, −X
n
are solu-
tions. It is enough to prove for X
n
and even for X. But this follows triv-
ially from Dirichlet’s theorem. Let us observe that there is another, more
elementary approach. Let us suppose that x
1
, x
2
, . . . , x
p
are terms of the
sequence, any two of which are relatively prime. We prove that we can
p+1
so that x
1
, x
2
, . . . , x
p+1
has the same property. It
is clear that x
1
, x
2
, . . . , x
p
are relatively prime with a, so we can apply
the Chinese remainder theorem to ﬁnd x
p+1
greater than x
1
, x
2
, . . . , x
p
,
such that x
p+1
≡ (1 −b)a
−1
i
(mod x
i
), i ∈ ¦1, 2, . . . , p¦, where a
−1
i
is a’s
inverse in Z

x
i
. Then gcd(x
p+1
, x
i
) = 1 for i ∈ ¦1, 2, . . . , p¦ and thus we
p+1
.
Here is an absolute classic, that appears in at least one Olympiad
around the world each year. Very easy, it uses only the fundamental
result.
Example. Suppose that
Fie n natural nenul. Care este gradul minim al unui polinom monic
cu coeﬁcienti intregi f astfel incat n[f(k) pentru orice k natural?
Proposed problems
1. Let (a
n
)
n≥1
be an increasing sequence of positive integers such
that for a certain polynomial f ∈ Z[X] we have a
n
≤ f(n) for all n.
Suppose also that m−n[a
m
−a
n
for all distinct positive integers m, n.
140
Prove that there exists a polynomial g ∈ Z[X] such that a
n
= g(n) for
all n.
USAMO 1995
2. We call the sequence of positive integers (a
n
)
n≥1
relatively prime
if gcd(a
m
, a
n
) = 1 for any diﬀerent positive integers m, n. Find all integer
polynomials f ∈ Z[X] such that for any positive integer c, the sequence
(f
[n]
(c))
n≥1
is relatively prime. Here f
[n]
is the composition of f with
itself n times.
Leo Mosser
3. Are there polynomials p, q, r with positive integer coeﬃcients such
that
p(x) + (x
2
−3x + 2)q(x) and q(x) =
_
x
2
20

x
15
+
1
12
_
r(x)?
4. Given is a ﬁnite family of polynomials with integer coeﬃcients.
Prove that for inﬁnitely many numbers n, if we evaluate any member of
the family in n, we obtain only composite numbers.
Folklore
5. Find all polynomials with integer coeﬃcients such that f(n)[2
n
−1
for any positive integer n.
6. Suppose that f ∈ Z[X] is a non constant polynomial. Also, sup-
pose that for some positive integers r, k, the following property is sat-
isﬁed: for any positive integer n, at most r prime factors of f(n) have
appear with exponent at most equal to k. Does it follow that any root
of this polynomial appears with multiplicity at least equal to k + 1?
7. Is it true that any polynomial f ∈ Z[X] that has a root modulo
n for any positive integer n must necessarily have a rational root?
141
8. Let f, g ∈ Z[X] some non zero polynomials. Let us consider the
set D
f,g
= ¦gcd(f(n), g(n))[ n ∈ N¦. Prove that the two polynomials are
relatively prime in Q[X] if and only if D
f,g
.
M. Andronache, Gazeta Matematica 1985
9. Prove that there are no polynomials f ∈ Z[X] with the property:
there exists n > 3 and integers x
1
, . . . , x
n
such that f(x
i
) = x
i−1
, i = 1, n
(indices are taken mod n).
10. Let f ∈ Z[X] a polynomial of degree n at least 2, with integer
coeﬃcients. Prove that the polynomial f(f(X))−X has at most n integer
roots.
Gh. Eckstein, Romanian TST
11. Find all trinomials f ∈ Z[X] with the property that for any
relatively prime integers m, n, the numbers f(M), f(n) are also relatively
prime.
12. For the die hard: ﬁnd all polynomials with the above property.
13. Let f ∈ Z[X] be a non constant polynomial. Show that the
sequence f(3
n
) (mod b) is not bounded.
14. Is there a second degree polynomial f ∈ Z[X] such that for any
positive integer n all prime factors of f(n) are of the form 4k + 3?
AMM
15. Prove that for any n there exists a polynomial f ∈ Z[X] such
that all numbers f(1) < f(2) < < f(n) are
a) prime numbers b) powers of 2.
Folklore
16. Find all integers n > 1 for which there exists a polynomial
f ∈ Z[X] such that for any integer k we have f(k) ≡ 0, 1 (mod n) and
both these congruences have solutions.
142
17. Let p be a prime number. Find the maximal degree of a poly-
nomial f ∈ Z[X] having coeﬃcients in the set ¦0, 1, . . . , p −1¦, knowing
that its degree is at most p and that if p divides f(m) − f(n) then it
also divides m−n.
18. Use example 1 and properties of the cyclotomic polynomials
φ
n
(X) =

1≤k≤n
(k,n)=1
(X −e
2iπk
n
)
to prove that there are inﬁnitely many prime numbers of the form 1+kn
for any given n ≥ 2. You may be interested to characterize those numbers
m, n for which p[φ
m
(n), but p does not divide any other number of the
form φ
d
(n), where d is a divisor of m diﬀerent from m
Classical result
143
LAGRANGE INTERPOLATION
Almost everyone knows the Chinese Remainder Theorem, which is a
remarkable tool in number theory. But does everyone know the analogous
form for polynomials? Stated like this, this question may seem impossible
to answer. Then, let us make it easier and also reformulate it: is it true
that given some pair wise distinct real numbers x
0
, x
1
, x
2
, . . . , x
n
and
some arbitrary real numbers a
0
, a
1
, a
2
, . . . , a
n
, we can ﬁnd a polynomial
f with real coeﬃcients such that f(x
i
) = a
i
for i ∈ ¦0, 1, . . . , n¦? The
answer turns out to be positive and a possible solution to this question
is based on Lagrange’s interpolation formula. It says that an example of
such polynomial is
f(x) =
n

i=0
a
i

0≤j,=i≤n
x −x
j
x
i
−x
j
(1)
Indeed, it is immediate to see that f(x
i
) = a
i
for i ∈ ¦0, 1, . . . , n¦.
Also, from the above expression we can see that this polynomial has
degree less than or equal to n. Is this the only polynomial with this
supplementary property? Yes, and the proof is not diﬃcult at all. Just
suppose we have another polynomial g of degree smaller than or equal
than n and such that g(x
i
) = a
i
for i ∈ ¦0, 1, . . . , n¦. Then the polyno-
mial g − f also has degree smaller than or equal to n and vanishes at
0, 1, . . . , n. Thus, it must be null and the uniqueness is proved.
What is Lagrange’s interpolation theorem good for? We will see in
the following problems that it helps us to ﬁnd immediately the value of a
polynomial in a certain point if we know the values in some given points.
And the reader has already noticed that this follows directly from the
formula (1), which shows that if we know the value in 1 + deg f points,
then we can ﬁnd easily the value in any other point without solving a
complicated linear system. Also, we will see that it helps in establishing
144
some inequalities and bounds for certain special polynomials and will
even help us in ﬁnding and proving some beautiful identities.
Now, let us begin the journey trough some nice examples of problems
where this idea can be used. As promised, we will see ﬁrst how we can
compute rapidly the value in a certain point for some polynomials. This
was one of the favorite’s problems in the old Olympiads, as the following
examples will show. The ﬁrst example is just an immediate application
of formula (1) and became a classical problem.
Example 1.
Let f be a polynomial of degree n such that
f(0) = 0, f(1) =
1
2
, f(2) =
2
3
, . . . , f(n) =
n
n + 1
.
Find f(n + 1).
USAMO 1975, Great Britain 1989
Solution. A ﬁrst direct approach would be to write
f(x) =
n

k=0
a
k
x
k
and to determine a
0
, a
1
, . . . , a
n
from the linear system
f(0) = 0, f(1) =
1
2
, f(2) =
2
3
, . . . , f(n) =
n
n + 1
.
But this is terrible, since the determinants that must be computed
are really complicated. This is surely a dead end. But for someone who
knows Lagrange’s Interpolation Theorem, the problem is straightfor-
ward. Indeed, we have
f(x) =
n

i=0
i
i + 1

j,=i
x −j
i −j
,
so that
f(n + 1) =
n

i=0
i
i + 1

n≤i
n + 1 −j
i −j
.
145
Now, how do we compute this? The reader might say: but we have
already found the value of f(n+1)! Well, it is tacit that the answer should
be expressed in the closest possible form. But, after all, computing the
above sum is not so diﬃcult. Indeed, we can see that

j,=i
n + 1 −j
i −j
=
(n + 1)!
(n + 1 −i) i! (n −i)!
(−1)
n−i
just by writing

j,=i
n + 1 −j
i −j
=
(n + 1)n. . . (n + 1 −(i −1))(n + 1 −(i + 1)) . . . 1
i(i −1) . . . 1 (−1) . . . (−(n −i))
.
According to these small observations, we can write
f(n + 1) =
n

i=0
i
i + 1

(n + 1)!
(n + 1 −i) i! (n −i)!
(−1)
n−i
=
n

i=1
(n + 1)!
(n + 1 −i)! (i −1)!
(−1)
n−i
= (n + 1)
n

i=1
_
n
i −1
_
(−1)
n−i
= (n + 1)
n−1

i=0
_
n
i
_
(−1)
n+1−i
.
And we have arrived at a familiar formula: the binomial theorem.
According to this,
n−1

i=0
_
n
i
_
(−1)
n+1−i
= −
_
n

i=0
_
n
i
_
(−1)
n−i
−1
_
= 1.
This shows that f(n + 1) = n + 1.
The ﬁrst example was straightforward because we didn’t ﬁnd any
diﬃculties after ﬁnding the idea. It’s not the case with the following
problem.
Example 2. Let F
1
= F
2
= 1, F
n+2
= F
n
+ F
n+1
and let f be a
polynomial of degree 990 such that f(k) = F
k
for k ∈ ¦992, . . . , 1982¦.
Show that f(1983) = F
1983
−1.
Titu Andreescu, IMO 1983 Shortlist
146
Solution. So, we have f(k + 992) = F
k+992
for k = 0, 990 and we
need to prove that f(992 + 991) = F
1983
− 1. This simple observation
shows that we don’t have to bother too much with k + 992, since we
could work as well with the polynomial g(x) = f(x+992), which also has
degree 990. Now, the problem becomes: if g(k) = F
k+992
, for k = 0, 990,
then g(991) = F
1983
−1. But we know how to compute g(991). Indeed,
looking again at the previous problem, we ﬁnd that
g(991) =
990

k=0
g(k)
_
991
k
_
(−1)
k
=
990

k=0
_
991
k
_
F
k+992
(−1)
k
which shows that we need to prove the identity
990

k=0
_
991
k
_
F
k+992
(−1)
k
= F
1983
−1.
This isn’t so easy, but with a little bit of help it can be done. The
device is: never complicate things more than necessary! Indeed, we could
try to establish a more general identity that could be proved by induc-
tion. But why, since it can be done immediately with the formula for
F
n
. Indeed, we know that
F
n
=
a
n
−b
n

5
,
where a =

5 + 1
2
and b =
1 −

5
2
. Having this in mind, we can of
course try a direct approach:
990

k=0
_
991
k
_
F
k+992
(−1)
k
=
1

5
_
990

k=0
_
991
k
_
a
k+992
(−1)
k

990

k=0
_
991
k
_
b
k+992
(−1)
k
_
.
But using the binomial theorem, the above sums vanish:
990

k=0
_
991
k
_
a
k+992
(−1)
k
= a
992
990

k=0
_
991
k
_
(−a)
k
= a
992
[(1 −a)
991
+a
991
].
147
Since a
2
= a + 1, we have
a
992
[(1 −a)
991
+a
991
] = a(a −a
2
)
991
+a
1983
= −a +a
1983
.
Since in all this argument we have used only the fact that a
2
= a+1
and since b also veriﬁes this relation, we ﬁnd that
990

k=0
_
991
k
_
F
k+992
(−1)
k
=
1

5
(a
1983
−b
1983
−a +b)
=
a
1983
−b
1983

5

a −b

5
= F
1983
−1.
And this is how with the help of a precious formula and with some
smart computations we could solve this problem and also ﬁnd a nice
property of the Fibonacci numbers.
The following example is a very nice problem proposed for IMO 1997.
Here, the following steps after using Lagrange’s Interpolation formula are
even better hidden in some congruencies. It is the typical example of a
good Olympiad problem: no matter how much the contestant knows in
that ﬁeld, it causes great diﬃculties in solving.
Example 3. Let f be a polynomial with integer coeﬃcients and let
p be a prime such that f(0) = 0, f(1) = 1 and f(k) = 0, 1 (mod p) for
all positive integer k. Show that deg f is at least p −1.
IMO Shortlist 1997
Solution. As usual, such a problem should be solved indirectly,
arguing by contradiction. So, let us suppose that deg f ≤ p − 2. Then,
using the Interpolation formula, we ﬁnd that
f(x) =
p−1

k=0
f(k)

j,=k
x −j
k −j
.
148
Now, since deg f ≤ p − 2, the coeﬃcient of x
p−1
in the right-hand
side of the identity must be zero. Consequently, we have
p−1

k=0
(−1)
p−k−1
k!(p −1 −k)!
f(k) = 0.
From here we have one more step. Indeed, let us write the above
relation in the form
p−1

k=0
(−1)
k
_
p −1
k
_
f(k) = 0
and let us take this equality modulo p. Since
k!
_
p −1
k
_
= (p −k)(p −k + 1) . . . (p −1) ≡ (−1)
k
k! (mod p)
we ﬁnd that
_
p −1
k
_
≡ (−1)
k
(mod p)
and so
p−1

k=0
(−1)
k
_
p −1
k
_
f(k) ≡
p−1

k=0
f(k) (mod p).
Thus,
p−1

k=0
f(k) ≡ 0 (mod p),
which is impossible, since f(k) ≡ 0, 1 (mod p) for all k and not all of
the numbers f(k) have the same remainder modulo p (for example, f(0)
and f(1)). This contradiction shows that our assumption was wrong and
the conclusion follows.
It’s time now for some other nice identities, where polynomials do
not appear at ﬁrst sight. We will see how some terrible identities are
simple consequences of the Lagrange Interpolation formula.
149
Example 4. Let a
1
, a
2
, . . . , a
n
be pairwise distinct positive integers.
Prove that for any positive integer k the number
n

i=1
a
k
i

j,=i
(a
i
−a
j
)
is an
integer.
Great Britain
Solution. Just by looking at the expression, we recognize the La-
grange Interpolation formula for the polynomial f(x) = x
k
. But we may
have some problems when the degree of this polynomial is greater than
or equal to n. But this can be solved by working with the remainder of
f modulo g(x) = (x − a
1
)(x − a
2
) . . . (x − a
n
). So, let us proceed, by
writing f(x) = g(x)h(x) + r(x), where r is a polynomial of degree at
most n − 1. This time we don’t have to worry, since the formula works
and we obtain
r(x) =
n

i=1
r(a
i
)

j,=i
x −a
j
a
i
−a
j
.
Now, we need three observations. The ﬁrst one is r(a
i
) = a
k
i
, the sec-
ond one is that the polynomial r has integer coeﬃcients and the third one
is that
n

i=1
a
k
i

j,=i
(a
i
−a
j
)
is just the coeﬃcient of x
n−1
in the polynomial
n

i=1
r(a
i
)

j,=i
x −a
j
a
i
−a
j
. All these observations are immediate. Combining
them, we ﬁnd that
n

i=1
a
k
i

j,=i
(a
i
−a
j
)
is the coeﬃcient of x
n−1
in r, which
is an integer. Thus, not only that we have solved the problem, but we
also found a rapid way to compute the sums of the form
n

i=1
a
k
i

j,=i
(a
i
−a
j
)
.
150
The following two problems we are going to discuss refer to combi-
natorial sums. If the ﬁrst one is relatively easy to prove using a combi-
natorial argument (it is a very good exercise for the reader to ﬁnd this
argument), the second problem is much more diﬃcult. But we will see
that both are immediate consequences of the Interpolation Formula.
Example 5. Let f(x) =
n

k=0
a
k
x
n−k
. Prove that for any non-zero
real number h and any real number A we have
n

k=0
(−1)
n−k
_
n
k
_
f(A+kh) = a
0
h
n
n!.
Alexandru Lupas
Solution. Since this polynomial has degree at most n, we have no
problems in applying the Interpolation formula
f(x) =
n

k=0
f(A
k
h)

j,=k
x −A−jh
(k −j)h
.
Now, let us identify the leading coeﬃcients in both polynomials that
appear in the equality. We ﬁnd that
a
0
=
n

k=0
f(A+kh)
1

j,=k
[(k −j)h]
=
1
n!h
n
n

k=0
(−1)
n−k
_
n
k
_
f(A+kh),
which is exactly what we had to prove. Simple and elegant! Notice that
the above problem implies the well-known combinatorial identities
n

k=0
(−1)
k
_
n
k
_
k
p
= 0
for all p ∈ ¦0, 1, 2, . . . , n −1¦ and
n

k=0
(−1)
n−k
_
n
k
_
k
n
= n!.
As we promised, we will discuss a much more diﬃcult problem. The
reader might say after reading the solution: but this is quite natural! Yes,
it is natural for someone who knows very well the Lagrange Interpolation
151
formula and especially for someone who thinks that using it could lead
to a solution. Unfortunately, this isn’t always so easy.
Example 6. Prove the identity
n

k=0
(−1)
n−k
_
n
k
_
k
n+1
=
n(n + 1)!
2
.
Solution. We take the polynomial f(x) = x
n
(why don’t we take the
polynomial f(x) = x
n+1
? Simply because (−1)
n−k
_
n
k
_
appears when
writing the formula for a polynomial of degree at most n) and we write
the Interpolation Formula
x
n
=
n

k=0
k
n
x(x −1) . . . (x −k −1)(x −k + 1) . . . (x −n)
(n −k)!k!
(−1)
n−k
Now, we identify the coeﬃcient of x
n−1
in both terms. We ﬁnd that
0 =
n

k=0
(−1)
n−k
_
n
k
_
k
n
(1 + 2 + +n −k).
And now the problem is solved, since we found that
n

k=0
(−1)
n−k
_
n
k
_
k
n+1
=
n(n + 1)
2
n

k=0
(−1)
n−k
_
n
k
_
k
n
and we also know that
n

k=0
(−1)
n−k
_
n
k
_
k
n
= n!
from the previous problem.
Were Lagrange interpolation formula good only to establish identi-
ties and to compute values of polynomials, it wouldn’t have been such
a great discovery. Of course it is not the case, it plays a fundamental
role in analysis. Yet, we are not going to enter this ﬁeld and we prefer to
concentrate on another elementary aspect of this formula and see how
it can help us establish some remarkable inequalities. And some of them
will be really tough.
152
We begin with a really diﬃcult inequality, in which the interpola-
tion formula is really well hidden. Yet, the denominators give sometimes
precious indications...
Example 7. Prove that for any real numbers x
1
, x
2
, . . . , x
n
∈ [−1, 1]
the following inequality is true:
n

i=1
1

j,=i
[x
j
−x
i
[
≥ 2
n−2
.
Solution. The presence of

j,=i
[x
j
−x
i
[ is the only hint to this prob-
lem. But even if we know it, how do we choose the polynomial? The
answer is simple: we will choose it to be arbitrary and only in the end
we will decide which is one is optimal. So, let us proceed by taking
f(x) =
n−1

k=0
a
k
x
k
an arbitrary polynomial of degree n −1. Then we have
f(x) =
n

k=1
f(x
k
)

j,=k
x −x
j
x
k
−x
j
.
Combining this with the triangular inequality , we arrive at a new
inequality
[f(x)[ ≤
n

k=1
[f(x
k
)[

j,=k
¸
¸
¸
¸
x −x
j
x
k
−x
j
¸
¸
¸
¸
.
Only now comes the beautiful idea, which is in fact the main step.
From the above inequality we ﬁnd that
¸
¸
¸
¸
f(x)
x
n−1
¸
¸
¸
¸

n

k=1
[f(x
k
)[

j,=k
[x
k
−x
j
[
¸
¸
¸
¸
¸
¸

j,=k
_
1 −
x
j
x
_
¸
¸
¸
¸
¸
¸
153
and since this is true for all non-zero real numbers x, we may take the
limit when x → ∞ and the result is pretty nice
[a
n−1
[ ≤
n

k=1
[f(x
k
)[

j,=k
[x
k
−x
j
[
.
This is the right moment to decide what polynomial to take. We
need a polynomial f such that [f(x)[ ≤ 1 for all x ∈ [−1, 1] and such
that the leading coeﬃcient is 2
n−2
. This time our mathematical culture
will decide. And it says that Chebyshev polynomials are the best, since
they are the polynomials with the minimum deviation on [−1, 1] (the
reader will wait just a few seconds and he will see a beautiful proof of
this remarkable result using Lagrange’s interpolation theorem). So, we
take the polynomial deﬁned by f(cos x) = cos(n −1)x. It is easy to see
that such a polynomial exists, has degree n − 1 and leading coeﬃcient
2
n−2
, so this choice solves our problem.
Note also that the inequality [a
n−1
[ ≤
n

k=1
[f(x
k
)[

j,=k
[x
k
−x
j
[
can be
proved by identifying the leading coeﬃcients in the identity
f(x) =
n

k=1
f(x
k
)

j,=k
x −x
j
x
k
−x
j
and then using the triangular inequality.
The following example is a ﬁne concoct of ideas. The problem is not
simple at all, since many possible approaches fail. Yet, in the frame-
work of the previous problems and with the experience of Lagrange’s
interpolation formula, it is not so hard after all.
Example 8. Let f ∈ R[X] be a polynomial of degree n with leading
coeﬃcient 1 and let x
0
< x
1
< x
2
< < x
n
be some integers. Prove
154
that there exists k ∈ ¦1, 2, . . . , n¦ such that
[f(x
k
)[ ≥
n!
2
n
.
Crux Matematicorum
Solution. Naturally (but would this be naturally without having
f(x) =
n

k=0
f(x
k
)

j,=k
x −x
j
x
k
−x
j
.
Now, repeating the argument in the previous problem and using the
fact that the leading coeﬃcient is 1, we ﬁnd that
n

k=0
[f(x
k
)[

j,=k
[x
k
−x
j
[
≥ 1.
It is time to use that we are dealing with integers. This will allow us
to ﬁnd a good inferior bound for

j,=k
[x
k
−x
j
[ ≥ 1. This is easy, since

j,=k
[x
k
−x
j
[ = (x
k
−x
0
)(x
k
−x
1
) . . . (x
k
−x
k−1
)(x
k+1
−x
k
) . . . (x
n
−x
k
)
≥ k(k −1)(k −2) . . . 1 1 2 . . . (n −k) = k!(n −k)!.
And yes, we are done, since using these inequalities, we deduce that
n

k=0
[f(x
k
)[
k!(n −k)![
≥ 1.
Now, since
n

k=0
1
k!(n −k)!
=
1
n!
n

k=0
_
n
k
_
=
2
n
n!
,
it follows trivially that
[f(x
k
)[ ≥
n!
2
n
.
We shall discuss one more problem before entering in a more detailed
study of Chebyshev polynomials and their properties, a problem given
155
in the Romanian mathematical Olympiad and which is a very nice ap-
plication of Lagrange’s interpolation formula. It is useless to say that it
follows trivially using a little bit of integration theory and Fourier series.
Example 9. Prove that for any polynomial f of degree n and with
leading coeﬃcient 1 there exists a point z such that
[z[ = 1 and [f(z)[ ≥ 1.
Solution. Of course, the idea is always the same, but this time
it is necessary to ﬁnd the good points in which we should write the
interpolation formula. As usual, we shall be blind and we shall try to
ﬁnd these points. Till then, let us call them simply x
0
, x
1
, x
2
, . . . , x
n
and
write
n

k=0
[f(x
k
)[

j,=k
[x
k
−x
j
[
≥ 1.
This inequality was already proved in the two problems above. Now,
consider the polynomial
g(x) =
n

i=0
(x −x
i
).
We have then
[g
t
(x
i
)[ =
¸
¸
¸
¸
¸
¸

j,=i
(x
i
−x
j
)
¸
¸
¸
¸
¸
¸
.
Now, of course we would like, if possible, to have [x
i
[ = 1 and also
n

k=0
1
[g
t
(x
k
)[
≤ 1. In this case it would follow from
n

k=0
[f(x
k
)[

j,=k
[x
k
−x
j
[
≥ 1
that at least one of the numbers [f(x
k
)[ is at least equal to 1 and the
problem would be solved. Thus, we should ﬁnd a monic polynomial g of
degree n +1 with all roots of modulus 1 and such that
n

k=0
1
[g
t
(x
k
)[
≤ 1.
156
This is trivial: it suﬃces of course to consider g(x) = x
n+1
− 1. The
conclusion follows.
We have an explanation to give: we said the problem follows trivially
with a little bit of integration theory tools. Indeed, if we write f(x) =
n

k=0
a
k
x
k
then one can check with a trivial computation that
a
k
=
1

_

0
f(e
it
)e
−ikt
dt
and from here the conclusion follows since we will have
2π =
¸
¸
¸
¸
_

0
f(e
it
)e
−int
dt
¸
¸
¸
¸

_

0
[f(e
it
[dt ≤ 2π max
[z[=1
[f(z)[.
given to students in 10-th grade) is not something common...
The next problems will be based on a very nice identity that will
allow us to prove some classical results about norms of polynomials,
to ﬁnd the polynomials having minimal deviation on [−1, 1] and also to
establish some new inequalities. In order to do all this, we need two quite
technical lemmas, which is not diﬃcult to establish, but very useful.
Lemma 1. If we put t
k
= cos

n
, 0 ≤ k ≤ n, then
f(x) =
n

k=0
(x −t
k
) =

x
2
−1
2
n
[(x +
_
x
2
−1)
n
−(x −
_
x
2
−1)
n
].
Proof. The proof is simple. Indeed, if we consider
g(x) =

x
2
−1
2
n
[(x +
_
x
2
−1)
n
−(x −
_
x
2
−1)
n
],
using the binomial formula we can establish immediately that it is a
polynomial. Moreover, from the obvious fact that lim
x→∞
g(x)
x
n+1
= 1, we
deduce that it is actually a monic polynomial of degree n + 1. The fact
157
that g(t
k
) = 0 for all 0 ≤ k ≤ n is easily veriﬁed using Moivre’s formula.
All this proves the ﬁrst lemma.
A little bit more computational is the second lemma.
Lemma 2. The following relations are true:
i)

j,=k
(t
k
−t
j
) =
(−1)
k
n
2
n−1
if 1 ≤ k ≤ n −1;
ii)
n

j=1
(t
0
−t
j
) =
n
2
n−2
;
iii)
n−1

j=0
(t
n
−t
j
) =
(−1)
n
n
2
n−2
.
Proof. Simple computations, left to the reader, allow us to write:
f
t
(x) =
n
2
n
[(x +
_
x
2
−1)
n
+ (x −
_
x
2
−1)
n
]
+
x
2
n

x
2
−1
[(x +
_
x
2
−1)
n
−(x −
_
x
2
−1)
n
].
Using this formula and Moivre’s formula we easily deduce i). To
prove ii) and iii) it suﬃces to compute lim
x→±1
f
t
(x), using the above for-
mula. We leave the computations to the reader.
Of course, the reader hopes that all these computations will have a
honourable purpose. He’s right, since these lemmas will allow us to prove
some very nice results. The ﬁrst one is a classical theorem of Chebyshev,
about minimal deviation of polynomials on [−1, 1].
Example 10. (Chebyshev theorem) Let f ∈ R[X] be a monic poly-
nomial of degree n. Then
max
x∈[−1,1]
[f(x)[ ≥
1
2
n−1
and this bound cannot be improved.
Solution. Using again the observation from problem 7, we obtain
the identity:
I =
n

k=0
f(t
k
)

j,=k
1
t
k
−t
j
.
158
Thus, we have
1 ≤ max
0≤k≤n
[f(t
k
)[
n

k=0
1
¸
¸
¸
¸
¸
¸

j,=k
(t
k
−t
j
)
¸
¸
¸
¸
¸
¸
.
Now, it suﬃces to apply lemma 2 to conclude that we actually have
n

k=0
1
¸
¸
¸
¸
¸
¸

j,=k
(t
k
−t
j
)
¸
¸
¸
¸
¸
¸
= 2
n−1
.
This shows that max
x∈[−1,1]
[f(x)[ ≥
1
2
n−1
and so the result is proved.
To prove that this result is optimal, it suﬃces to use the polynomial
T
n
(x) = cos(narccos(x)). It is an easy exercise to prove that this is
really a polynomial (called the nth polynomial of Chebyshev of the ﬁrst
kind) and that it has leading coeﬃcient 2
n−1
and degree n. Then the
polynomial
1
2
n−1
T
n
is monic of degree n and
max
x∈[−1,1]
¸
¸
¸
¸
1
2
n−1
T
n
(x)
¸
¸
¸
¸
=
1
2
n−1
.
There are many other proof of this result , many of them are much
easier, but we chosen this one because it shows the power of Lagrange
interpolation theory. Not to say that the use of the two lemmas allowed
us to prove that the inequality presented in example 7 is actually the
best.
Some years ago, Walther Janous presented in Crux the following
problem as open problem. It is true that it is a very diﬃcult one, but
here is a very simple solution using the results already achieved.
Example 11. Suppose that a
0
, a
1
, . . . , a
n
are real numbers such that
for all x ∈ [−1, 1] we have
[a
0
+a
1
x + +a
n
x
n
[ ≤ 1.
159
Then for all x ∈ [−1, 1] we also have
[a
n
+a
n−1
x + +a
0
x
n
[ ≤ 2
n−1
.
Walther Janous, Crux Matematicorum
Solution. Actually, we are going to prove a stronger result, that is:
Lemma. Denote
|f| = max
x∈[−1,1]
[f(x)[.
Then for any polynomial f ∈ R[X] of degree n the following inequal-
ity is satisﬁed:
[f(x)[ ≤ [T
n
(x)[ |f| for all [x[ ≥ 1.
Proof. Using Lagrange’s interpolation formula and modulus in-
equality, we deduce that for all u ∈ [−1, 1] we have:
¸
¸
¸
¸
f
_
1
u

¸
¸
¸

1
[u[
n
|f|
n

k=0

j,=k
1 −t
j
u
[t
k
−t
j
[
.
The very nice idea is to use now again Lagrange interpolation for-
mula, this time for the polynomial T
n
. We shall then have
¸
¸
¸
¸
T
n
_
1
u

¸
¸
¸
=
1
[u[
n
¸
¸
¸
¸
¸
¸
n

k=0
(−1)
k

j,=k
1 −ut
j
t
k
−t
j
¸
¸
¸
¸
¸
¸
=
1
[u[
n
n

k=0

j,=k
1 −ut
j
[t
k
−t
j
[
(the last identity being ensured by lemma 2). By combining the two
results, we obtain
¸
¸
¸
¸
f
_
1
u

¸
¸
¸

¸
¸
¸
¸
T
n
_
1
u

¸
¸
¸
|f| for all u ∈ [−1, 1]
and the conclusion follows.
Coming back to the problem and considering the polynomial f(x) =
n

k=0
a
k
x
k
, the hypothesis says that |f| ≤ 1 and so by the lemma we have
[f(x)[ ≤ [T
n
(x)[ for all [x[ ≥ 1.
160
We will then have for all x ∈ [−1, 1]:
[a
n
+a
n−1
x + +a
0
x
n
[ =
¸
¸
¸
¸
x
n
f
_
1
x

¸
¸
¸

¸
¸
¸
¸
x
n
T
n
_
1
x

¸
¸
¸
.
It suﬃces to prove that
¸
¸
¸
¸
x
n
T
n
_
1
x

¸
¸
¸
≤ 2
n−1
,
which can be also written as
(1 +
_
1 −x
2
)
n
+ (1 −
_
1 −x
2
)
n
≤ 2
n
.
But this inequality is very easy to prove: just set a =

1 −x
2
∈ [0, 1]
and observe that h(a) = (1−a)
n
+(1+a)
n
is a convex function on [0, 1],
thus its superior bound is attained in 0 or 1 and there the inequality is
trivially veriﬁed. Therefore we have
[a
n
+a
n−1
x + +a
0
x
n
[ ≤ 2
n−1
and the problem is solved.
We end this topic with a very diﬃcult problem, that reﬁnes a prob-
lem given in a Japanese mathematical Olympiad in 1994. The problem
has a nice story: given initially in an old Russian Olympiad, it asked to
prove that
max
x∈[0,2]
n

i=1
[x −a
i
[ ≤ 108
n
max
x∈[0,1]
n

i=1
[x −a
i
[
for any real numbers a
1
, a
2
, . . . , a
n
. The Japanese problems asked only
to prove the existence of a constant that could replace 108. A brutal
choice of points in Lagrange interpolation theorem gives a better bound
of approximately 12 for this constant. Recent work by Alexandru Lupas
reduces this bound to 1 +2

6. In the following, we present the optimal
bound.
161
Example 12. For any real numbers a
1
, a
2
, . . . , a
n
, the following in-
equality holds:
max
x∈[0,2]
n

i=1
[x −a
i
[ ≤
(3 + 2

2)
n
+ (3 −2

2)
n
2
max
x∈[0,1]
n

i=1
[x −a
i
[.
Gabriel Dospinescu
Solution. Let us denote
|f|
[a,b]
= max
x∈[a,b]
[f(x)[
for a polynomial f and let, for simplicity,
c
n
=
(3 + 2

2)
n
+ (3 −2

2)
n
2
.
We thus need to prove that |f|
[0,2]
≤ c
n
|f|
[0,1]
where
f(x) =
n

i=1
(x −a
i
).
We shall prove that this inequality is true for any polynomial f,
which allows us to suppose that |f|
[0,1]
= 1. We shall prove that for
all x ∈ [1, 2] we have [f(x)[ ≤ x
n
. Let us ﬁx x ∈ [1, 2] and consider the
numbers x
k
=
1 +t
k
2
. Using Lagrange interpolation formula, we deduce
that
[f(x)[ ≤
n

k=0
¸
¸
¸
¸
¸
¸

j,=k
x −x
k
x
k
−x
j
¸
¸
¸
¸
¸
¸
=
n

k=0

j,=k
x −x
j
[x
k
−x
j
[

n

k=0

j,=k
2 −x
j
[x
k
−x
j
[
=
n

k=0

j,=k
3 −t
j
[t
k
−t
j
[
.
Using lemma 2, we can write
n

k=0

j,=k
3 −t
j
[t
k
−t
j
[
=
2
n−1
n
n−1

k=1

j,=k
(3 −t
j
)
+
2
n−2
n
_
_
n−1

j=0
(3 −t
j
) +
n

j=1
(3 −t
j
)
_
_
.
162
Using again the two lemmas, we obtain:
n
2
n
[(3 + 2

2)
n
+ (3 −2

2)
n
] +
3
2
n+1

2
[(3 + 2

2)
n
−(3 −2

2)
n
]
=
n−1

k=1

j,=k
(3 −t
j
) +
n−1

j=0
(3 −t
j
) +
n

j=1
(3 −t
j
).
All we have to do now is to compute
n−1

j=0
(3 −t
j
) +
n

j=1
(3 −t
j
) = 6
n−1

j=1
(3 −t
j
).
But using lemma 1, we deduce immediately that
n−1

j=1
(3 −t
j
) =
1
2
n+1

2
[(3 + 2

2)
n
−(3 −2

2)
n
].
Putting all these observations together and making a small compu-
tation, that we let to the reader, we easily deduce that [f(x)[ ≤ c
n
. This
proves that |f|
[0,2]
≤ c
n
|f|
[0,1]
and solves the problem.
Problems for training
1. A polynomial of degree 3n takes the value 0 at 2, 5, 8, . . . , 3n −1,
the value 1 at 1, 4, 7, . . . , 3n − 2 and the value 2 at 0, 3, 6, . . . , 3n. It’s
value at 3n + 1 is 730. Find n.
USAMO 1984
2. A polynomial of degree n veriﬁes p(k) = 2
k
for all k = 1, n + 1.
Find its value at n + 2.
Vietnam 1988
3. Prove that for any real number a we have the following identity
n

k=0
(−1)
k
_
n
k
_
(a −k)
n
= n!.
Tepper’s identity
163
4. Find
n

k=0
(−1)
k
_
n
k
_
k
n+2
and
n

k=0
(−1)
k
_
n
k
_
k
n+3
.
AMM
5. Prove that
n

k=0
x
n+1
k

j,=k
(x
k
−x
j
)
=
n

k=0
x
k
and compute
n

k=0
x
n+2
k

j,=k
(x
k
−x
j
)
.
6. Prove the identity
n

k=1
(−1)
k−1
_
n
k
_
k
(n −k)
n
= n
n
n

k=2
1
k
.
Peter Ungar, AMM E 3052
7. Let a, b, c be real numbers and let f(x) = ax
2
+bx +c such that
max¦[f(±1)[, [f(0)[¦ ≤ 1. Prove that if [x[ ≤ 1 then
[f(x)[ ≤
5
4
and
¸
¸
¸
¸
x
2
f
_
1
x

¸
¸
¸
≤ 2.
Spain, 1996
8. Let f ∈ R[X] a polynomial of degree n that veriﬁes [f(x)[ ≤ 1 for
all x ∈ [0, 1], then
¸
¸
¸
¸
f
_

1
n

¸
¸
¸
≤ 2
n+1
−1.
9. Let a, b, c, d ∈ R such that [ax
3
+bx
2
+cx+d[ ≤ 1 for all x ∈ [−1, 1].
What is the maximal value of [c[? Which are the polynomials in which
the maximum is attained?
Gabriel Dospinescu
164
10. Let a ≥ 3 be a real number and p be a real polynomial of degree
n. Prove that
max
i=0,n+1
[a
i
−p(i)[ ≥ 1.
India, 2001
11. Find the maximal value of the expression a
2
+b
2
+c
2
if [ax
2
+
bx +c[ ≤ 1 for all x ∈ [−1, 1].
Laurentiu Panaitopol
12. Let a, b, c, d ∈ R such that [ax
3
+ bx
2
+ cx + d[ ≤ 1 for all
x ∈ [−1, 1]. Prove that
[a[ +[b[ +[c[ +[d[ ≤ 7.
IMO Shortlist, 1996
13. Let A =
_
p ∈ R[X][ deg p ≤ 3, [p(±1)[ ≤ 1,
¸
¸
¸
¸
p
_
±
1
2

¸
¸
¸
≤ 1
_
.
Find sup
p∈A
max
[x[≤1
[p
tt
(x)[.
IMC, 1998
14. a) Prove that for any polynomial f having degree at most n, the
following identity is satisﬁed:
xf
t
(x) =
n
2
f(x) +
1
n
n

k=1
f(xz
k
)
2z
k
(1 −z
k
)
2
,
where z
k
are the roots of the polynomial [X
n
+ 1.
b) Deduce Bernstein’s inequality: |f
t
| ≤ n|f| where
|f| = max
[x[≤1
[f(x)[.
P.J. O’Hara, AMM
15. Deﬁne F(a, b, c) = max
x∈[0,3]
[x
3
− ax
2
− bx − c[. What is the least
possible value of this function over R
3
?
China TST 2001
165
HIGHER ALGEBRA IN COMBINATORICS
Till now, we have seen numerous applications of analysis and higher
algebra in number theory and algebra. It is time to see the contribution
of this ”non-elementary mathematics” to combinatorics. It is quite hard
to imagine that behind a simple game, such as football for example
or behind a quotidian situation such as handshakes there exists such
complicated machinery, but this happens sometimes and we will prove it
in the next. For the beginning of the discussion, the reader doesn’t need
any special knowledge, just imagination and the most basic properties
of the matrices, but, as soon as we advance, things change. Anyway,
the most important fact is not the knowledge, but the ideas and, as we
will see, it is not easy to discover that ”non-elementary” fact that hides
after a completely elementary problem. Since we have clariﬁed what is
the purpose of the unit, we can begin the battle.
The ﬁrst problem we are going to discuss is not classical, but it is easy
and a very nice application of how linear-algebra can solve elementary
problems. Here it is.
Example 1. Let n ≥ 3 and let A
n
, B
n
be the sets of all even,
respectively, odd permutations of the set ¦1, 2, . . . , n¦. Prove the equality

σ∈A
n
n

i=1
[i −σ(i)[ =

σ∈B
n
n

i=1
[i −σ(i)[.
Nicolae Popescu, Gazeta Matematica
Solution. Writing the diﬀerence

σ∈A
n
n

i=1
[i −σ(i)[ −

σ∈B
n
n

i=1
[i −σ(i)[
as

σ∈S
n
ε(σ)
n

i=1
[i −σ(i)[ = 0,
166
where
ε(σ) =
_
1, if σ ∈ A
n
−1, if σ ∈ B
n
det A =

σ∈S
n
ε(σ)a
1σ(1)
a
2σ(2)
. . . a
nσ(n)
.
We have taken here S
n
= A
n
∪ B
n
. But we don’t have any product
in our sum! That is why we will take an arbitrary positive number a and
we will consider the matrix A = (a
[i−j[
)
1≤i,j≤n
. This time,
det A =

σ∈S
n
(−1)
ε(α)
a
[1−σ(1)[
. . . a
[−nσ(n)[
=

σ∈A
n
a
n

i=1
[i−σ(i)[

σ∈B
n
a
n

i=1
[i−σ(i)[
This is how we have obtained the identity
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 x x
2
. . . x
n−2
x
n−1
x 1 x . . . x
n−3
x
n−2
x
2
x 1 . . . x
n−4
x
n−3
. . . . . . . . . . . . . . . . . .
x
n−1
x
n−2
. . . . . . x 1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
=

σ∈S
n
σ par˘a
x
n

i=1
[i−σ(i)[

σ∈S
n
σ impar˘a
x
n

i=1
[i−σ(i)[
. (1)
Anyway, we still do not have the desired diﬀerence. What can we do
to obtain it? The most natural way is to derive the last relation, which is
nothing else than a polynomial identity, and then to take x = 1. Before
167
doing that, let us observe that the polynomial
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 x x
2
. . . x
n−2
x
n−1
x 1 x . . . x
n−3
x
n−2
x
2
x 1 . . . x
n−4
x
n−3
. . . . . . . . . . . . . . . . . .
x
n−1
x
n−2
. . . . . . x 1
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
is divisible by (x − 1)
2
. This can be easily seen by subtracting the ﬁrst
line from the second and the third one and taking from each of these
line x − 1 as common factor. Thus, the derivative of this polynomial is
a polynomial divisible by x − 1, which shows that after we derive the
relation (1) and take x = 1, in the left-hand side we will obtain 0. Since
in the right-hand side we obtain exactly

σ∈A
n
n

i=1
[i −σ(i)[ −

σ∈B
n
n

i=1
[i −σ(i)[
the identity is established.
Here is another nice application of this trick. We have seen how many
permutation do not have a ﬁxed point. The question that arises is how
many of them are even. Here is a direct answer to the question, using
determinants.
Example 2. Find the number of even permutations of the set
¦1, 2, . . . , n¦ that do not have ﬁxed points.
Solution. Let us consider C
n
, D
n
, respectively, the sets of even and
odd permutations of the set ¦1, 2, . . . , n¦, that do not have any ﬁxed
points. We know how to ﬁnd the sum [C
n
[ + [D
n
[. We have seen it in
the unit ”Principiul includerii si excluderii” that it is equal to
n!
_
1 −
1
1!
+
1
2!
− +
(−1)
n
n!
_
.
168
Hence if we manage to compute the diﬀerence [C
n
[ − [D
n
[, will be
able to answer to the question. If we write
[C
n
[ −[D
n
[ =

σ∈A
n
σ(i),=i
1 −

σ∈B
n
σ(i),=i
1,
we observe that this reduces to computing the determinant of the matrix
T = (t
ij
)
1≤i,j≤n
, where
t
ij
=
_
1, if i ,= j
0, if i = j
That is,
[C
n
[ −[D
n
[ =
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
0 1 1 . . . 1
1 0 1 . . . 1
. . . . . . . . . . . . . . .
1 1 1 . . . 0
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
.
But it is not diﬃcult to compute this determinant. Indeed, we add
all columns to the ﬁrst one and we give n − 1 as common factor, then
we subtract the ﬁrst column from each of the other columns. The result
is [C
n
[ −[D
n
[ = (−1)
n−1
(n −1) and the conclusion is quite surprising:
[C
n
[ =
1
2
n!
_
1 −
1
2!
+
1
3!
− +
(−1)
n−2
(n −2)!
_
+ (−1)
n−1
(n −1).
We will focus in the next problems on a very important combina-
torial tool, that is the incidence matrix (cum se spune la matricea de
incidenta?). What is this? Suppose we have a set X = ¦x
1
, x
2
, . . . , x
n
¦
and X
1
, X
2
, . . . , X
k
a family of subsets of X. Now, deﬁne the matrix
A = (a
ij
)
i=1,n
j=1,k
, where
a
ij
=
_
1, if x
i
∈ X
j
0, if x
i
,∈ X
j
169
This is the incidence matrix of the family X
1
, X
2
, . . . , X
k
and the
set X. In many situations, computing the product
t
A A helps us to
modelate algebraically the conditions and the conclusions of a certain
problem. From this point, the machinery activates and the problem is
on its way of solving.
Let us discuss ﬁrst a classical problem, though a diﬃcult one. It ap-
peared in USAMO 1979, Tournament of the Towns 1985 and in Bulgar-
ian Spring Mathematical Competition 1995. This says something about
the classical character and beauty of this problem.
Example 3. Let A
1
, A
2
, . . . , A
n+1
be distinct subsets of the set
¦1, 2, . . . , n¦, each of which having exactly three elements. Prove that
there are two distinct subsets among them that have exactly one point
in common.
Solution. Of course, we argue by contradiction and suppose that
[A
i
∩ A
j
[ ∈ ¦0, 2¦ for all i ,= j. Now, let T be the incidency matrix of
the family A
1
, A
2
, . . . , A
n+1
and compute the product
t
T T =
_
_
_
_
_
_
_
_
n

k=1
t
2
k1
n

k=1
t
k1
t
k2
. . .
n

k=1
t
k
t
kn+1
. . . . . . . . . . . .
n

k=1
t
kn+1
t
k1
n

k=1
t
kn+11
t
k2
. . .
n

k=1
t
2
kn+1
_
_
_
_
_
_
_
_
.
But we have of course
n

k=1
x
2
ki
= [A
i
[ = 3
and
n

k=1
x
ki
x
kj
= [A
i
∩ A
j
[ ∈ ¦0, 2¦.
170
Thus, considered in the ﬁeld (R
2
, +, ), we have
t
T T =
_
_
_
_
´
1
´
0 . . .
´
0
´
0
. . . . . . . . . . . . . . .
´
0
´
0 . . .
´
0
´
1
_
_
_
_
,
where X is the matrix having as elements the residues classes of the
elements of the matrix X. Since of course det X = det X, the previous
relation shows that det
t
T T is odd, hence non-zero. This means that
t
TyT is an invertible matrix of size n+1, thus rank
t
T T = n+1 which
t
T T ≤ rankT ≤ n. This shows that our
assumption was wrong and there exist indeed indices i ,= j such that
[A
i
∩ A
j
[ = 1.
The following problem is very diﬃcult to solve by elementary means,
but the solution using linear-algebra is straightforward.
Example 4. Let n be an even number and A
1
, A
2
, . . . , A
n
be distinct
subsets of the set ¦1, 2, . . . , n¦, each of them having an even number of
elements. Prove that among these subsets there are two having an even
number of elements in common.
Solution. Indeed, if T is the incidency matrix of the family
A
1
, A
2
, . . . , A
n
, we obtain as in the previous problem the following rela-
tion
t
T T =
_
_
_
_
[A
1
[ [A
1
∩ A
2
[ . . . [A
1
∩ A
n
[
. . . . . . . . . . . .
[A
n
∩ A
1
[ [A
n
∩ A
2
[ . . . [A
n
[
_
_
_
_
.
Now, let us suppose that all the numbers [A
i
∩ A
j
[ are odd and
interpret the above relation in the ﬁeld (R
2
, +, ). We ﬁnd that
t
T T =
_
_
_
_
´
0
´
1 . . .
´
1
´
1
. . . . . . . . . . . . . . .
´
1
´
1 . . .
´
1
´
0
_
_
_
_
,
171
which means again that det
t
T . . . T is odd. Indeed, if we work in
(R
2
, +, ), we obtain
¸
¸
¸
¸
¸
¸
¸
¸
´
0
´
1 . . .
´
1
´
1
. . . . . . . . . . . . . . .
´
1
´
1 . . .
´
1
´
0
¸
¸
¸
¸
¸
¸
¸
¸
=
´
1.
The technique used is exactly the same as in the second example,
only this time we work in a diﬀerent ﬁeld. Note that this is the moment
when we use the hypothesis that n is even. Now, since det
t
TT = det
2
T,
we obtain that det T is also an odd number. Hence we should try to prove
that in fact det T is an even number and the problem will be solved. Just
observe that the sum of elements of the column i of T is [A
i
[, hence an
even number. Thus, if we add all lines to the ﬁrst line, we will obtain
only even numbers on the ﬁrst line. Since the value of the determinant
doesn’t change under this operation, the conclusion is plain: det T is
an even number. Since a number cannot be both even and odd, our
assumption was wrong and the problem is solved.
Working in a simple ﬁeld such as (R
2
, +, ) can allow us to ﬁnd quite
interesting solutions. For example, we will discuss the following problem,
used for the IMO preparation of the Romanian IMO team in 2004.
Example 5. The squares of a nn table are colored with white and
black. Suppose that there exists a non-empty set of lines A such that
any column of the table has an even number of white squares that also
belong to A. Prove that there exists a non-empty set of columns B such
that any line of the table contains an even number of white squares that
also belong to B.
Gabriel Dospinescu
172
Solution. This is just the combinatorial translation of the well-
known fact that a matrix T is invertible in a ﬁeld if and only if its
transpose is also invertible in that ﬁeld. But this is not that easy to see.
Let us proceed easily. In each white square we write the number 1
and in each black square we put a 0. We thus obtain a binary matrix
T = (t
ij
)
1≤i,j≤n
. From now on, we work only in (R
2
, +, ). Suppose that
A contains the columns a
1
, a
2
, . . . , a
k
. It follows that
k

i=1
t
a,j
= 0 for all
j = 1, n. Now, let us take
x
i
=
_
1, if i ∈ A
0, if i ,∈ A
It follows that the system
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
t
11
z
1
+t
21
z
2
+ +t
n1
z
n
= 0
t
12
z
1
+t
22
z
2
+ +t
n2
z
n
= 0
. . .
t
1n
z
1
+t
2n
z
2
+ +t
nm
z
n
= 0
1
, x
2
, . . . , x
n
). Thus, det T = 0 and
consequently det
t
T = 0. But this means that the system
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
u
11
y
1
+u
12
y
2
+ +u
1n
y
n
= 0
u
21
y
1
+u
22
y
2
+ +u
2n
y
n
= 0
. . .
u
n1
y
1
+u
n2
y
2
+ +u
nn
y
n
= 0
also has a non-trivial solution in R
2
. Now, we take B = ¦i[ y
i
,= 0¦ and
we will clearly have B ,= ∅ and

x∈B
u
ix
= 0, i = 1, n. But this mean that
any line of the table contains an even number of white squares that also
belong to B and the problem is solved.
In the end of this sub-unit, we will discuss a very diﬃcult problem,
in which just knowing the trick of computing
t
A A does not suﬃce. It
173
is true that it is one of the main steps, but there are much more things
to do after we compute
t
A A. And if for these ﬁrst problems we have
used only intuitive or well-known properties of the matrices and ﬁelds,
this time we need a more sophisticated machinery: the properties of the
characteristic polynomial and eingenvalues of a matrix. It is exactly that
kind of problem that kills you just when we feel most strong.
Example 6. Let S = ¦1, 2, . . . , n¦ and A be a family of pairs of
elements from S with the following property: for any i, j ∈ S there exist
exactly m indices k ∈ S for which (i, k), (k, j) ∈ A. Find all possible
values of m, n for which this is possible.
Gabriel Carrol
Solution. This time, it is easy to see what hides after the problem.
Indeed, if we take T = (t
ij
)
1≤i,j≤n
, where
a
ij
=
_
1, if (i, j) ∈ A
0, otherwise
the existence of the family A reduces to
T
2
=
_
_
_
_
_
_
_
m m . . . m
m m . . . m
. . . . . . . . . . . .
m m . . . m
_
_
_
_
_
_
_
.
So, we must ﬁnd all values of m, n for which there exist a binary
matrix T such that
T
2
=
_
_
_
_
_
_
_
m m . . . m
m m . . . m
. . . . . . . . . . . .
m m . . . m
_
_
_
_
_
_
_
.
174
Let us consider
TB =
_
_
_
_
_
_
_
m m . . . m
m m . . . m
. . . . . . . . . . . .
m m . . . m
_
_
_
_
_
_
_
.
and ﬁnd the eigenvalues of B. This is not diﬃcult, since if x is an ein-
genvalue, then
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
m−x m . . . m
m m . . . m
. . . . . . . . . . . .
m m . . . m−x
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
.
If we add all columns to the ﬁrst one and then take the common
factor mn −x, we obtain the equivalent form
(mn −x)
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
1 m . . . m
1 m−x . . . m
. . . . . . . . . . . .
1 m . . . m−x
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
= 0.
In this ﬁnal determinant, we subtract from each column the ﬁrst
column multiplied by m and we obtain in the end the equation
x
n−1
(mn −x) = 0, which shows that the eigenvalues of B are precisely
0, 0, . . . , 0
. ¸¸ .
n−1
, mn. But these are exactly the squares of the eigenvalues of
T. Thus, T has the eingevalues 0, 0, . . . , 0
. ¸¸ .
n−1
,

mn, because the sum of the
eingenvalues is nonnegative (being equal to the sum of the elements of
the matrix situated on the main diagonal). Since TrT ∈ R, we ﬁnd that
mn must be a perfect square. Also, because TrT ≤ n, we must have
m ≤ n.
175
Now, let us prove the converse. So, suppose that m ≤ n and mn is
a perfect square and write m = du
2
, n = dv
2
. Let us take the matrices
I = (11 . . . 11
. ¸¸ .
dv
), O = (00 . . . 00
. ¸¸ .
dv
).
Now, let us deﬁne the circulant matrix
S =
_
_
_
_
_
_
_
_
_
_
_
_
111 . . . 1
. ¸¸ .
u
00 . . . 0
. ¸¸ .
v−u
0 11 . . . 1
. ¸¸ .
u
00 . . . 0
. ¸¸ .
v−u−1
. . .
111 . . . 1
. ¸¸ .
u−1
00 . . . 0
. ¸¸ .
v−u
1
_
_
_
_
_
_
_
_
_
_
_
_
∈ M
v,n
(¦0, 1¦).
Finally, we take
A =
_
_
_
_
_
_
_
S
S
. . .
S
_
_
_
_
_
_
_
∈ M
n
(¦0, 1¦).
It is not diﬃcult to see that
A
2
=
_
_
_
_
_
_
_
m m . . . m
m m . . . m
. . . . . . . . . . . .
m m . . . m
_
_
_
_
_
_
_
.
The last idea that we present here (but certainly these are not all
the methods of higher mathematics applied to combinatorics) is the use
of vector spaces. Again, we will not insist on complicated notions from
the theory of vector spaces, just the basic notions and theorems. Maybe
the most useful fact is that if V is a vector space of dimension n (that is,
V has a basis of cardinal n), then any n+1 or more vectors are linearly
dependent. As a direct application, we will discuss the following problem,
176
which is very diﬃcult to solve by means of elementary mathematics. Try
ﬁrst to solve it elementary and you will see how hard it is. The following
example is classical, too, but few people know the trick behind it.
Example 7. Let n be a positive integer and let A
1
, A
2
, . . . , A
n+1
be nonempty subsets of the set ¦1, 2, . . . , n¦. Prove that there ex-
ist nonempty and disjoint index sets I
1
= ¦i
1
, i
2
, . . . , i
k
¦ and I
2
=
¦j
1
, j
2
, . . . , j
m
¦ such that
A
i
1
∪ A
i
2
∪ ∪ A
i
k
= A
j
1
∪ A
j
2
∪ ∪ A
j
m
.
Solution. Let us associate to each subset A
i
a vector v
i
∈ R
n
, where
v
i
= (x
1
i
, x
2
i
, . . . , x
n
i
) and
x
j
i
=
_
0, if j ∈ A
i
1, if j ,∈ A
i
Since dimR
n
= n, these vectors we have just constructed must be
linearly dependent. So, we can ﬁnd a
1
, a
2
, . . . , a
n+1
∈ R, not all of them
0, such that
a
1
v
1
a
2
v
2
+ +a
n+1
v
n+1
= 0.
Now, we take I
1
= ¦i ∈ ¦1, 2, . . . , n + 1¦[ a
i
> 0¦ and I
2
= ¦i ∈
¦1, 2, . . . , n+1¦[ a
i
< 0¦. It is plain that I
1
, I
2
are nonempty and disjoint.
Now, let us prove that
_
i∈I
1
A
i
=
_
i∈I
2
A
i
and the solution will be complete.
Let us take x ∈
_
i∈I
1
A
i
and suppose that x ,∈
_
i∈I
2
A
i
. Then the vectors v
i
with i ∈ I
2
have zero on their xth component, so the xth component of
the vector a
1
v
1
+a
2
v
2
+ +a
n+1
v
n+1
is

x∈A
j
j∈I
1
a
j
> 0, which is impossible,
since a
1
v
1
+a
2
v
2
+ +a
n+1
v
n+1
= 0. This shows that
_
i∈I
1
A
i

_
i∈I
2
A
i
.
But the reversed inclusion can be proved in exactly the same way, so we
conclude that
_
i∈I
1
A
i
=
_
i∈I
2
A
i
.
177
In the end of this ”non-elementary” discussion, we solve another
problem, proposed for the TST 2004 in Romania, whose idea is also
related to vector spaces.
Example 8. 30 boys and 20 girls are preparing for the 2004 Team
Selection Test. They observed that any two boys have an even number
of common acquaintances among the girls and exactly 9 boys know an
odd number of girls. Prove that there exists a group of 16 boys such that
any girls attending the preparation is known by an even number of boys
from this group.
Gabriel Dospinescu
Solution. Let us consider the matrix A = (a
ij
) where
a
ij
=
_
1, if B
i
knows F
j
0, otherwise
We have considered here that B
1
, B
2
, . . . , B
30
are the boys and
F
1
, F
2
, . . . , F
20
are the girls. Now, consider the matrix T = A
t
A. We
observe that all the elements of the matrix T, except those from the
main diagonal are even (because t
ij
=
20

k=1
a
ik
a
jk
is the number of com-
mon acquaintances among the girls of the boys B
i
, B
j
). The elements on
the main diagonal of T are exactly the number of girls known by each
boy. Thus, if we consider the matrix T in (R
2
, +, ), it will be diagonal,
with exactly nine non-zero elements on its main diagonal. From now on,
we will work only in (R
2
, +, ). We have seen till now that rankT = 9.
Using Sylvester inequality, it follows that
9 = rankT ≥ rankA+rank
t
A−20 = 2rank
t
A−20
178
hence r = rank
t
A ≤ 14. Let us consider now the linear system in
(R
2
, +, ):
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
a
11
x
1
+a
21
x
2
+ +a
301
x
30
= 0
a
12
x
1
+a
22
x
2
+ +a
302
x
30
= 0
. . .
a
120
x
1
+a
220
x
2
+ +a
3020
x
30
= 0
The set of solutions of this system is a vector space of dimension
30 − r ≥ 16. That is why we can choose a solution (x
1
, x
2
, . . . , x
30
) of
the system, having at least 16 components equal to
´
1. Finally, consider
the set M = ¦i ∈ ¦1, 2, . . . , 30¦[ x
i
=
´
1¦. We have proved that [M[ ≥ 16
and also

j∈M
a
ji
= 0 for all i = 1, 20. But we observe that

j∈M
a
ji
is just
the number of boys B
k
with k ∈ M such that B
k
knows F
i
. Thus, if we
choose the group of those boys B
k
with k ∈ M, then each girl is known
by an even number of boys from this group and the problem is solved.
Problems for training
1. Let p > 2 be an odd prime and let n ≥ 2. For any permutation
σ ∈ S
n
, we consider
S(σ) =
n

k=1
kσ(k).
Let A
j
, B
j
, respectively, be the set of even, respectively odd permu-
tations σ for which S(σ) ≡ j (mod p). Prove that n > p if and only if
A
j
and B
j
have the same number of elements for all j ∈ ¦0, 1, . . . , p−1¦.
Gabriel Dospinescu
2. Let n ≥ 2. Find the greatest number p such that for all k ∈
¦1, 2, . . . , p¦ we have

σ∈A
n
_
n

i=1
if(i)
_
k
=

σ∈B
n
_
n

i=1
if(i)
_
k
,
179
where A
n
, B
n
are, respectively, the sets of all even, respectively, odd
permutations of the set ¦1, 2, . . . , n¦.
Gabriel Dospinescu
3. Is there in the plane a conﬁguration of 22 circles and 22 points
on their union (the union of their circumferences) such that any circle
contains at least 7 points and any point belongs to at least 7 circles?
Gabriel Dospinescu, Moldova TST 2004
4. Let A
1
, A
2
, . . . , A
m
be distinct subsets of a set A with n ≥ 2 ele-
ments. Suppose that any two of these subsets have exactly one elements
in common. Prove that m ≤ n.
5. The edges of a regular 2
n
-gon are colored red and blue. A step
consists of recoloring each edge which is the same color as both of its
neighbours in red, and recoloring each other edge in blue. Prove that
after 2
n−1
steps all of the edges will be red and show that this need not
hold after fewer steps.
6. Problema de la Vietnamezi cu cunostintele
7. n ≥ 2 teams compete in a tournament and each team plays against
any other team exactly once. In each game, 2 points are given to the
winner, 1 point for a draw and 0 points for the looser. It is known that
for any subset S of teams, one can ﬁnd a team (possibly in S) whose
total score in the games with teams in S is odd. Prove that n is even.
8. On an m n sheet of paper is drawn a grid dividing the sheet
into unit squares. The two sides of length n are taped together to form a
cylinder. Prove that it is possible to write a real number in each square,
not all zero, so that each number is the sum of the numbers in the
neighboring squares, if and only if there exist integers k, l such that
180
n + 1 does not divide k and
cos
2lπ
m
+ cos

n + 1
=
1
2
.
Ciprian Manolescu, Romanian TST 1998
9. In a contest consisting of n problems, the jury deﬁnes the diﬃculty
of each problem by assigning it a positive integral number of points (the
same number of points may be assigned to diﬀerent problems). Any
of points for the problem; any other participant receives 0 points. After
the participants submitted their answers, the jury realizes that given any
ordering of the participants (where ties are not permitted), it could have
deﬁned the problems’ diﬃculty levels to make that ordering coincide
with the participants’ ranking according to their total scores. Determine,
in terms of n, the maximum number of participants for which such a
scenario could occur.
10. Let S = ¦x
0
, x
1
, . . . , x
n
¦ ⊂ [0, 1] be a ﬁnite set of real numbers
with x
0
= 0, x
1
= 1, such that every distance between pairs of elements
occurs at least twice, except for the distance 1. Prove that S consists of
rational numbers only.
11. Let x
1
, . . . , x
n
be real numbers and suppose that the vector space
spanned by x
i
−x
j
over the rationals has dimension m. Then the vector
space spanned only by those x
i
−x
j
for which x
i
−x
j
,= x
k
−x
l
whenever
(i, j) ,= (k, l) also has dimension m.
Strauss theorem
181
12. Let A
1
, A
2
, . . . , A
m
some subsets of ¦1, 2, . . . , n¦. Then there are
disjoint sets I, J with non-empty union such that
_
i∈I
A
i
=
_
j∈J
A
j
and

i∈I
A
i
=

j∈J
A
j
.
Lindstrom theorem
13. There is no partition of the set of edges of the complete graph
on n vertices into (strictly) fewer than n −1 complete bipartite graphs.
Graham-Pollak theorem
14. Let 2n+1 real numbers with the property that no matter how we
eliminate one of them, the rest of them can be divided into two groups
of n numbers, the sum of the numbers in the two groups being the same.
Then all numbers are equal.
15. In a society, acquaintance is mutual and even more, any two
persons have exactly one friend. Then there is a person that knows all
the others.
Universal friend theorem
16. Let A
1
, . . . , A
m
and B
1
, . . . , B
p
subsets of ¦1, 2, . . . , n¦ such that
A
i
∩ B
j
is an odd number for all i, j. Then mp ≤ 2
n−1
.
Benyi Sudakov
17. Let A
1
, . . . , A
n
, B
1
, . . . , B
n
⊂ A = ¦1, 2, . . . , n¦ with the proper-
ties:
a) for any nonempty subset T of A, there is i ∈ A such that [A
i
∩T[
is odd.
b) for any i, j ∈ A, A
i
and B
j
have exactly one common element.
Then prove that B
1
= B
2
= = B
n
.
Gabriel Dospinescu
182
18. A symmetric matrix of zeros and ones has only ones on the main
diagonal. Prove that we can ﬁnd some rows in this matrix such that their
sum is a vector having all of its components odd.
183
GEOMETRY AND NUMBERS
Again an apparently paradoxical note!!! Indeed, it may look weird,
but geometry it really useful in number theory and sometimes it can
help proving diﬃcult results with some extremely simple arguments. In
the sequel we are going to show some applications of geometry in num-
ber theory, almost all of them playing around the celebrated Minkowski
theorem. We will see that this theorem gives a very simple criterion for
a nice region (we are also going to explain what we understand by nice)
to have a non-trivial lattice point and the existence of this point will
have important consequences in the theory of representation of numbers
by quadratic forms or in approximation of real numbers with rational
numbers. As usual, we will content to present only a mere introduc-
tion to this ﬁeld, extremely well developed. The reader will surely have
mentioned in the bibliographies.
First of all, let us state the conditions in which we will work and
what is a nice ﬁgure. In general, we will work in R
n
and we will call
convex body a bounded subset A of R
n
which is convex (that is for all
a, b ∈ A and all 0 ≤ t ≤ 1 we have ta+(1−t)b ∈ A), which is symmetric
about the origin (that is, for all x ∈ A we also have −x ∈ A). We will
admit that convex bodies have volumes (just think about it in the plane
or space, which will be practically always used in our applications).
Let us start by proving the celebrated Minkowski’s theorem.
Theorem. (Minkowski) Suppose that A is a convex body in R
n
hav-
ing volume strictly greater than 2
n
. Then there is a lattice point in A
diﬀerent from the origin.
The proof is surprisingly simple. Indeed, let us start by making a sort
of partition of R
n
in cubes of edge 2, having as centers the points that
have all coordinates even numbers. It is clear that any two such cubes will
184
have disjoint interiors and that they cover all space. That is why we can
say that the volume of the convex body is equal to the sum of volumes of
the intersections of the body with each cube (since the body is convex,
it is clear that the sum will be ﬁnite). But of course, one can bring any
cube into the cube centered around the origin by using a translation by
a vector all of whose coordinates are even. Since translations preserve
volume, we will have now an agglomeration of bodies in the central cube
(the one centered in the origin) and the sum of volumes of all these
bodies is strictly greater that 2
n
. Necessarily there will be two bodies
which intersect in a point X. Now, look at the cubes where these two
bodies where taken from and look at the points in these cubes that give
by translations the point X. We have found two diﬀerent points x, y
in our convex body such that x − y ∈ 2Z
n
. But since A is centrally
symmetric and convex, it follows that
x −y
2
is a lattice point diﬀerent
from the origin and belonging to A. The theorem is thus proved.
Here is a surprising result that follows directly from this theorem.
Problem 1. Suppose that in each lattice point in space except for
the origin one draws a ball of radius r > 0 (common for all the balls).
Then any line that passes through the origin will intercept a certain ball.
Solution. Let us suppose the contrary and let us consider a very long
cylinder having as axe the line and basis a circle of radius
r
2
. We choose
it suﬃciently long to ensure that it will have a volume strictly greater
than 8. This is clearly a convex body in space and using Minkowski’s
theorem we deduce the existence of a non-trivial lattice point in this
cylinder (or on the border). This means that the line will intercept the
ball centered around this point.
Actually, the theorem proved before admits a more general formula-
tion, which is even more useful.
185
Theorem 2. (Minkowski) Let A is a convex body in R
n
and
v
1
, v
2
, . . . , v
n
some linearly independent vectors in R
n
. Also consider
the fundamental parallelepiped P =
_
n

i=1
x
i
v
i
[ 0 ≤ x
i
≤ 1
_
and de-
note V ol(P) its volume. Assuming that A has a volume strictly greater
than 2
n
V ol(P), A must contain at least a point of the lattice L =
Zv
1
+ +Zv
n
diﬀerent from the origin.
With all these terms, it would seem that this is extremely diﬃcult
to prove. Actually, it follows trivially from the ﬁrst theorem. Indeed,
by considering the linear application f sending v
i
into the vector e
i
=
(0, 0, . . . , 1, 0, . . . , 0) one can easily see that P is sent into the ”normal”
cube in R
n
(that is, the set of vectors all of whose components are
between 0 and 1) and that f maps L into Z
n
. Since the transformation
is linear, it will send A into a convex body of volume
V ol(A)
V ol(P)
> 2
n
. It
suﬃces to apply the ﬁrst theorem to this convex body and to look at
the preimage of the lattice point (in Z
n
), in order to ﬁnd a non-trivial
point of A∩ L. The second theorem is thus proved.
We have already proved that any prime number of the form 4k + 1
is the sum of two squares. Let us prove it diﬀerently, using Minkowski’s
theorem.
Problem 2. Any prime number of the form 4k +1 is the sum of two
squares.
Proof. We have already proved that for any prime number of the
form 4k + 1, call it p, one can ﬁnd a such that p[a
2
+ 1. Then let us
consider v
1
= (p, 0), v
2
= (a, 1). Visibly, they are linearly independent
and moreover for any point (x, y) in the lattice L = Zv
1
+Zv
2
we have
p[x
2
+ y
2
. Indeed, there are m, n ∈ Z such that x = mp + na, y = n
and thus x
2
+ y
2
≡ n
2
(a
2
+ 1) ≡ 0 (mod p). Moreover, the area of the
fundamental parallelogram is |v
1
∧ v
2
| = p. Next, consider as convex
186
body the disc centered in the origin and having as radius

2p. Obviously,
its area is strictly greater than four times the area of the fundamental
parallelogram. Thus, there is a point (x, y) diﬀerent from the origin that
lies in this disc and also in the lattice L = Zv
1
+Zv
2
. For this point we
have p[x
2
+y
2
and x
2
+y
2
< 2p, which shows that p = x
2
+y
2
.
Proving that a certain Diophantine equation has no solution is a very
classical problem, but what can we do when we are asked to prove that
a certain equation has solutions? Minkowski’s theorem and in general
geometry of numbers allow quick responses to such problems. Here is an
example, taken from a polish Olympiad.
Problem 3. Consider positive integers such that ac = b
2
+ b + 1.
Then the equation ax
2
−(2b + 1)xy +cy
2
= 1 has integer solutions.
Solution. Here is a very quick approach: consider in R
2
the set
of points verifying ax
2
− (2b + 1)xy + cy
2
< 2. A simple computation
shows that it is an elliptical disc having as area

3
> 4. An elliptical
disc is obviously a convex body and even more this elliptical disc is
symmetric about the origin. Thus, by Minkowski’s theorem we can ﬁnd
a point of this region diﬀerent from the origin. Since ac = b
2
+b +1, we
have for all x, y not both 0 the inequality ax
2
− (2b + 1)xy + cy
2
> 0.
Thus for (x, y) ∈ Z
2
¸ ¦(0, 0)¦ a lattice point of this region, we have
ax
2
−(2b+1)xy+cy
2
= 1 and the existence of a solution of the equation
is proved.
The following problem (as the above one) has a quite diﬃcult elemen-
tary solution. The solution using geometry of numbers is more natural,
but it is not at all obvious how to proceed. Yet... the experience of the
preceding problem should ring a bell.
187
Problem 4. Suppose that n is a natural number for which the
equation x
2
+xy +y
2
= n has rational solutions. Then this equation has
integer solutions as well.
Komal
Solution. Of course, the problem reduces to: if there are integer
numbers a, b, c such that a
2
+ ab + b
2
= c
2
n, then x
2
+ xy + y
2
= n
has integer solutions. We will assume that a, b, c are non-zero (otherwise
the conclusion follows trivially). Even more, a classical argument allows
to assume that a, b are relatively prime. We will try again to ﬁnd a
couple of integers (x, y) ∈ Z
2
¸ ¦(0, 0)¦ such that x
2
+ xy + y
2
< 2n
and n divides x
2
+ xy + y
2
. In this case we will have x
2
+ xy + y
2
= n
and the conclusion will follow. First, let us look at the region deﬁned
by x
2
+ xy + y
2
< 2n. Again, simple computations show that it is an
elliptical disc of area

3
n. Next, consider the lattice formed by the
points (x, y) such that n divides ax − by. The area of the fundamental
parallelepiped is clearly at most n. By Minkowski’s theorem, we can ﬁnd
(x, y) ∈ Z
2
¸ ¦(0, 0)¦ such that x
2
+xy +y
2
< 2n and n divides ax −by.
We claim that this will give an integer solution of the equation. Observe
that ab(x
2
+xy +y
2
) = c
2
xyn+(ax−by)(bx−ay) and so n also divides
x
2
+ xy + y
2
(since n is relatively prime with a, b). This allows us to
conclude.
Before continuing with some more diﬃcult problems, let us remind
that for any symmetric real matrix A such that

1≤i,j≤n
a
ij
x
i
x
j
> 0
for all x = (x
1
, x
2
, . . . , x
n
) ∈ R
n
¸ ¦0¦ the set of points verifying

1≤i,j≤n
a
ij
x
i
x
j
≤ 1
188
has a volume equal to
V ol(B
n
)

det A
, where
V ol(B
n
) =
π
n
2
Γ
_
1 +
n
2
_
(here Γ(x) =
_

0
e
−t
t
x−1
dt is Euler’s gamma function). The proof of this
in any decent book of multivariate integral calculus. In particular, the
reader has noticed that these results can be applied to previous problems
to facilitate the computations of diﬀerent areas and volumes. With these
results (that we will admit) in mind, let’s attack some serious problems.
If we spoke about squares, why not present the beautiful classical
proof of Lagrange’s theorem on representations using 4 squares.
Problem 5. (Lagrange’s theorem) Any natural number is a sum of
4 squares.
Proof. This is going to be much more complicated, but the idea is
always the same. The main diﬃculty is ﬁnding the appropriate lattice
and convex body. First of all, let us prove the result for prime numbers.
Let thus p > 2 a prime number and consider the sets A = ¦x
2
[ x ∈ Z
p
¦,
B = ¦−y
2
− 1[ y ∈ Z
p
¦. Since there are
p + 1
2
squares in Z
p
(as we
have already seen in previous notes), these two sets cannot be disjoint.
In particular, there are x, y such that 0 ≤ x, y ≤ p −1 and p[x
2
+y
2
+1.
This is the observation that will allow us to ﬁnd a good lattice. Consider
now the vectors
v
1
= (p, 0, 0, 0), v
2
= (0, p, 0, 0), v
3
= (x, y, 1, 0), v
4
= (y, −x, 0, 1)
and the lattice L generated by these vectors. A simple computation
(using the above formulas) allows to prove that the volume of the fun-
damental parallelepiped is p
2
. Moreover, one can easily verify that for
all point (x, y, z, t) ∈ L one has p[x
2
+y
2
+z
2
+t
2
. Even more, one can
189
also prove (by employing the non-elementary results stated before this
problem) that the volume of the convex body A = ¦x = (a, b, c, d) ∈
R
4
[ a
2
+b
2
+c
2
+d
2
< 2p¦ is equal to 2π
2
p
2
> 16V ol(P), thus A∩ L is
not empty. It suﬃces then to choose a point (x, y, z, t) ∈ L ∩ A and we
will clearly have x
2
+ y
2
+ z
2
+ t
2
= p. Thus the theorem is proved for
prime numbers.
Of course, everything would be nice if the product of two sums of 4
squares is always a sum of 4 squares. Hopefully, it is the case, but the
proof is not obvious at all. It follows form the miraculous identity:
(a
2
+b
2
+c
2
+d
2
)(x
2
+y
2
+z
2
+t
2
) = (ax +by +cz +dt)
2
= (ay −bx +ct −dz)
2
+ (az −bt +dy −cx)
2
+ (at +bz −cy −dx)
2
.
Of course, very nice, but how could one think at such an identity?
The eternal question... Well, this time there is a very nice reason: instead
of thinking in eight variables, let us reason only with four. Consider the
numbers z
1
= a +bi, z
2
= c +di, z
3
= x +yi, z
4
= z +ti. Introduce the
matrices
M =
_
z
1
z
2
−z
2
z
1
_
, N =
_
z
3
z
4
−z
4
z
3
_
.
We have
det(M) = [z
1
[
2
+[z
2
[
2
= a
2
+b
2
+c
2
+d
2
and similarly
det(N) = x
2
+y
2
+z
2
+t
2
.
It is then normal to try to express (a
2
+b
2
+c
2
+d
2
)(x
2
+y
2
+z
2
+t
2
)
as det(MN). But surprise! We have
MN =
_
z
1
z
3
−z
2
z
4
z
1
z
4
+z
2
z
3
−z
1
z
4
+z
2
z
3
z
1
z
3
−z
2
z
4
_
190
and so det(MN) is again a sum of 4 squares. The identity appears thus
naturally...
Let us concentrate a little bit more on approximations of real num-
bers. We have some beautiful results of Minkowski that deserve to be
presented after this small introduction to geometry of numbers.
Problem 6. (Minkowski’s linear forms’ theorem) Let A = (a
ij
) be
a nn invertible matrix of real numbers and suppose that c
1
, c
2
, . . . , c
n
are positive real numbers such that c
1
c
2
. . . c
n
> [ det A[. Then there
are integers x
1
, x
2
, . . . , x
n
, not all 0, such that
¸
¸
¸
¸
¸
¸
n

j=1
a
ij
x
j
¸
¸
¸
¸
¸
¸
< c
i
for all
i = 1, . . . , n.
Solution. We need to prove that there exists a non-zero vector
X that also belongs to the region ¦Y ∈ R
n
[ [A
−1
Y [
i
< c
i
, i =
1, . . . , n¦ (here A
−1
Y = ([A
−1
Y [
1
, . . . , [A
−1
Y [
n
). But observe that
¦Y ∈ R
n
[ [A
−1
Y [
i
< c
i
, i = 1, . . . , n¦ is exactly the image trough
A
−1
of the parallelepiped ¦Y ∈ R
n
[ −c
i
< Y
i
< c
i
, i = 1, . . . , n¦ which
has volume 2
n
c
1
. . . c
n
, thus ¦Y ∈ R
n
[ [A
−1
Y [
i
< c
i
, i = 1, . . . , n¦ is
a convex body of volume
1
det A
2
n
c
1
. . . c
n
> 2
n
. By Minkowski’s theo-
rem, this body will contain a non-zero lattice point, which will verify
the conditions of the problem.
And here is a nice consequence of the previous theorem.
Problem 7. Suppose A = (a
ij
)
1≤i≤m
1≤j≤n
is a matrix with m, n real
numbers and a ≥ 1 is a real number. Then one can ﬁnd x
1
, x
2
, . . . , x
n
integers between −a and a, not all 0, such that
¸
¸
¸
¸
¸
¸
n

j=1
a
ij
x
j
¸
¸
¸
¸
¸
¸
< a

n
m
for all 1 ≤ i ≤ m.
Solution. All we need to do is to apply the result in problem 6 for
the invertible matrix
_
A I
m
I
n
0
_
, whose determinant equals 1 or −1
191
and make the choice c
1
= = c
m
= a

n
m
, c
m+i
= a, 1 ≤ i ≤ n.
Incredibly, but the proof ends here!!!
Proposed problems
1. Suppose that a, b, c are positive integers such that ac = b
2
+ 1.
Then there exist x, y, z, t integers such that a = x
2
+ y
2
, b = z
2
+ t
2
,
c = xz +yt.
Imo Shortlist
2. Suppose that a natural number is the sum of three squares of
rational numbers. Then prove that it is also a sum of squares of three
natural numbers (the use of three squares theorem is forbidden!).
Davenport-Cassels lemma
3. Consider a disc of radius R. At each lattice point of this disc,
except for the origin, one plants a circular tree of radius r. Suppose that
r is optimal with respect to the following property: if one regards from
the origin, he can see at least a point situated at the exterior of the disc.
Then prove that
1

R
2
+ 1
≤ r <
1
R
.
AMM
4. Suppose that a, b, c are positive integers such that a > b > c.
Prove that we can ﬁnd three integers x, y, z, not all 0, such that
ax +by +cz = 0 and max¦[x[, [y[, [z[¦ <
2

3
a + 1.
Miklos Schweitzer competition
5. Suppose that a, b, c are positive integers such that ac = b
2
+ 1.
Prove that the equation ax
2
+ 2bxy +cy
2
= 1 is solvable in integers.
6. Suppose that a
ij
(1 ≤ i, j ≤ n) are rational numbers such that
for any x = (x
1
, . . . , x
n
) ∈ R
n
¸ ¦0¦ we have

1≤i,j≤n
a
ij
x
i
x
j
> 0. Then
192
there are integers (not all zero) x
1
, . . . , x
n
such that

1≤i,j≤n
a
ij
x
i
x
j
< n
n

det A,
where A = (a
ij
).
Minkowski
7. Suppose that x
1
, x
2
, . . . , x
n
are algebraic integers such that for
any 1 ≤ i ≤ n there is at least a conjugate of x
i
which is not between
x
1
, x
2
, . . . , x
n
. Then the set of n-tuples (f(x
1
), f(x
2
), . . . , f(x
n
)) with
f ∈ Z[X] is dense in R
n
.
8. Suppose that a, b are rational numbers such that the equation
ax
2
+ by
2
= 1 has at least one rational solution. Then it has inﬁnitely
many rational solutions.
Kurschak contest
9. Let us denote A(C, r) the set of points w on the unit sphere in
R
n
with the property that [wk[ ≥
C
|k|
r
for any non-zero vector k ∈ Z
n
(here wk is the usual scalar product and |k| is the Euclidean norm of
the vector k ∈ Z
n
). Prove that if r > n −1 there exists C > 0 such that
A(C, r) is non-empty, but if r < n −1 there is no such C > 0.
Mathlinks contest (after an ENS entrance exam problem)
10. Using the non-elementary results presented in the topic, prove
that if A = (a
ij
)
1≤i,j≤n
is a symmetric integer matrix such that

1≤i,j≤n
a
ij
x
i
x
j
> 0 for all x = (x
1
, x
2
, . . . , x
n
) ∈ R
n
¸ ¦0¦, then we
can ﬁnd an integer matrix B such that A = B
t
B. Deduce the result
from problem 1.
11. Let n ≥ 5 and a
1
, . . . , a
n
, b
1
, . . . , b
n
some integers verifying that
all pairs (a
i
, b
i
) are diﬀerent and [a
i
b
i+1
− a
i+1
b
i
[ = 1, 1 ≤ i ≤ n (here
(a
n+1
, b
n+1
) = (a
1
, b
1
)). Prove that one can ﬁnd 1 < [i −j[ < n−1 such
that [a
i
b
j
−a
j
b
i
[ = 1.
193
Korea TST
12. Let a, b, c, d be positive integers such that there are 2004 pairs
(x, y) such that x, y ∈ [0, 1] and ax + by, cx + dy ∈ Z. If (a, c) = 6, ﬁnd
(b, d).
13. A polygon of area greater than n is given in a plane. Prove that
it contains n + 1 points A
i
(x
i
, y
i
) such that x
i
− x
j
, y
i
− y
j
∈ Z for all
1 ≤ i, j ≤ n + 1.
China TST 1988
194
THE SMALLER, THE BETTER
Often, a large amount of simple ideas can solve very diﬃcult prob-
lems. We have seen or will see a few such examples in our journey through
the world of numbers: clever congruences that readily solve Diophantine
equations, properties of the primes of the form 4k + 3 or even complex
numbers and analysis. All these can be called ”tricks”, but in fact they
are much more, as you are going to see soon.
In this unit, we will discuss a fundamental concept in number theory,
the order of an element. It may seem contradictory for us to talk about
simple ideas and then say ”a fundamental concept”. Well, what we are
going to discuss about is the bridge between simplicity and complexity.
The reason for which we say it is a simple idea can be easily guessed
from the deﬁnition: given are the positive integer n > 1 and the integer a
such that gcd(a, n) = 1, the smallest possible integer d for which n[a
d
−1
is called the order of a modulo n. The deﬁnition is correct, since from
Euler’s theorem we have n[a
ϕ(n)
−1 so such numbers d indeed exist. The
complexity of this concept will follow from the examples.
In what follows we will denote by o
n
(a) the order of a modulo n.
There is a simple property of o
n
(a), which has important consequences:
if k is a positive integer such that n[a
k
− 1, then d[k. Indeed, because
n[a
k
−1 and n[a
d
−1, we ﬁnd that n[a
gcd(k,d)
−1. But from the deﬁnition
of d it follows that d ≤ gcd(k, d), which cannot hold unless d[k. Nice and
easy. But could such a simple idea be good at anything? The answer
is positive and will follow from the solutions of the problems to come.
But, before that, we note a ﬁrst application of this simple observation:
o
n
(a)[ϕ(n). This is a consequence of the above property and of Euler’s
theorem.
Now, an old and nice problem, which may seem really trivial after
this introduction. But do not get excited so easily, the problem has an
195
extremely short solution, but this does not mean that it is obvious.
It appeared in Saint Petersburg Mathematical Olympiad and also in
Gazeta Matematica.
Example 1. Prove that n[ϕ(a
n
−1) for all positive integers a, n.
Solution. What is o
a
n
−1
(a)? It may seem a silly question, since
of course o
a
n
−1
(a) = n. Using the observation in the introduction, we
obtain exactly n[ϕ(a
n
−1).
Here is another beautiful application of the order of an element. It
is the ﬁrst case case of Dirichlet’s theorem that we intend to discuss and
is also a classical property.
Example 2. Prove that any prime factor of the nth Fermat number
2
2
n
+ 1 is congruent to 1 modulo 2
n+1
. Show that there are inﬁnitely
many prime numbers of the form 2
n
k + 1 for any ﬁxed n.
Solution. Let us consider a prime p such that p[2
2
n
+ 1. Then
p[2
2
n+1
− 1 and consequently o
p
(2)[2
n+1
. This ensures the existence of
a positive integer k ≤ n + 1 such that o
p
(2) + 2
k
. We will prove that in
fact k = n + 1. The proof is easy. Indeed, if this is not the case, then
o
p
(2)[2
n
and so p[2
o
p
(2)
−1[2
2
n
−1. But this is impossible, since p[2
2
n
+1.
Therefore, we have found that o
p
(2) = 2
n+1
and we have to prove that
o
p
(2)[p −1 to ﬁnish the ﬁrst part of the question. But this follows from
the introduction.
The second part is a direct consequence of the ﬁrst. Indeed, it is
enough to prove that there exists an inﬁnite set of Fermat’s numbers
(2
2
n
k
+ 1)
n
k
>a
any two relatively prime. Then we could take a prime
factor of each such Fermat’s number and apply the ﬁrst part to obtain
that each such prime is of the form 2
n
k + 1. But not only it is easy to
ﬁnd such a sequence of Fermat’s coprime numbers, but in fact any two
diﬀerent Fermat’s numbers are relatively prime. Indeed, suppose that
d[gcd(2
2
n
+1, 2
2
n+k
+1). Then d[2
2
n+1
−1 and so d[2
2
n+k
−1. Combining
196
this with d[2
2
n+k
+1, we obtain a contradiction. Hence both parts of the
problem are solved.
We continue with another special case of the well-known diﬃcult
theorem of Dirichlet on arithmetical sequences. Though classical, the
following problem is not straightforward and this explains probably its
presence on a Korean TST in 2003.
Example 3. For a prime p, let f
p
(x) = x
p−1
+x
p−2
+ +x + 1.
a) If p[m, prove that there exists a prime factor of f
p
(m) that is
relatively prime with m(m−1).
b) Prove that there are inﬁnitely many numbers n such that pn + 1
is prime.
Solution.
a) is straightforward. In fact, we will prove that any prime factor of
f
p
(m) is relatively prime with m(m− 1). Take such a prime divisor q.
Because q[1+m+ +m
p−1
, it is clear that gcd(q, m) = 1. Moreover, if
gcd(q, m−1) ,= 1, then q[m−1 and because q[1+m+ +m
p−1
, it follows
that q[p. But p[m and we ﬁnd that q[m, which is clearly impossible.
More diﬃcult is b). But we are tempted to use a) and to explore the
properties of f
p
(m), just like in the previous problem. So, let us take a
prime q[f
p
(m) for a certain positive integer m divisible by p. Then we
have of course q[m
p
−1. But this implies that o
q
(m)[q and consequently
o
q
(m) ∈ ¦1, p¦. If o
q
(m) = p, then q ≡ 1 (mod p). Otherwise, q[m− 1
and because q[f
p
(m), we deduce that q[p, hence q = p. But we have seen
while solving a) that this is not possible, so the only choice is p[q−1. Now,
we need to ﬁnd a sequence (m
k
)
k≥1
of multiples of p such that f
p
(m
k
)
are pairwise relatively prime. This is not as easy as in the ﬁrst example.
Anyway, just by trial and error, it is not diﬃcult to ﬁnd such a sequence.
There are many other approaches, but we like the following one: take
m
1
= p and m
k
= pf(m
1
)f
p
(m
2
) . . . f
p
(m
k−1
). Let us prove that f
p
(m
k
)
197
is relatively prime to f
p
(m
1
), f
p
(m
2
), . . . , f
p
(m
k−1
). Fortunately, this is
easy, since f
p
(m
1
)f
p
(m
2
) . . . f
p
(m
k−1
)[f
p
(m
k
) − f
p
(0)[f
p
(m
k
) − 1. The
solution ends here.
The following problem became classical and variants of it have been
given in contests for years. It seems to be a favorite Olympiad problem,
since it uses elementary facts and the method is more than beautiful.
Example 4. Find the smallest number n with the property that
2
2005
[17
n
−1.
Solution. The problem actually asks for o
2
2005(17). We know
that o
2
2005(17)[ϕ(2
2005
) = 2
2004
, so o
2
2005(17) = 2
k
, where k ∈
¦1, 2, . . . , 2004¦. The order of an element has done its job. Now, it is
time to work with exponents. We have 2
2005
[17
2
k
−1. Using the factor-
ing
17
2
k
−1 = (17 −1)(17 + 1)(17
2
+ 1) . . . (17
2
k−1
+ 1),
we proceed by ﬁnding the exponent of 2 in each factor of this product.
But this is not diﬃcult, because for all i ≥ 0 the number 17
2
t
+ 1 is
a multiple of 2, but not a multiple of 4. Thus, v
2
(17
2
k
− 1) = 4 + k
and the order is found by solving the equation k + 4 = 2005. Thus,
o
2
2005(17) = 2
2001
is the answer to the problem.
Another simple, but not straightforward application of the order
of an element is the following divisibility problem. Here, we also need
some properties of the prime numbers, that we have already studied in
a previous unit.
Example 5. Find all primes p, q such that p
2
+ 1[2003
q
+ 1 and
q
2
+ 1[2003
p
+ 1.
Gabriel Dospinescu
Solution. Let us suppose that p ≤ q. We discuss ﬁrst the trivial case
p = 2. In this case, 5[2003
q
+ 1 and it is easy to deduce that q is even,
198
hence q = 2, which is a solution of the problem. Now, suppose that p > 2
and let r be a prime factor of p
2
+ 1. Because r[2003
2q
− 1, it follows
that o
r
(2003)[2q. Suppose that (q, o
r
(2003)) = 1. Then o
r
(2003)[2 and
r[2003
2
−1 = 2
3
3 7 11 13 167. It seems that this is a dead end, since
there are too many possible values for r. Another simple observation
narrows the number of possible cases: because r[p
2
+ 1, must be of the
form 4k + 1 or equal to 2 and now we do not have many possibilities:
r ∈ ¦2, 13¦. The case r = 13 is also impossible, because 2003
q
+ 1 ≡ 2
(mod 13) and r[2003
q
+ 1. So, we have found that for any prime factor
r of p
2
+ 1, we have either r = 2 or q[o
r
(2003), which in turn implies
q[r−1. Because p
2
+1 is even, but not divisible by 4 and because any odd
prime factor of it is congruent to 1 modulo q, we must have p
2
+ 1 ≡ 2
(mod q). This implies that p
2
+ 1 ≡ 2 (mod q), that is q[(p −1)(p + 1).
Combining this with the assumption that p ≤ q yields q[p + 1 and in
fact q = p+1. It follows that p = 2, contradicting the assumption p > 2.
Therefore the only pair is (2,2).
More diﬃcult is the following problem, proposed by Reid Barton
for the USA TST in 2003. Anyway, using the order of an element, the
problem is not very diﬃcult and the solution follows naturally. Let us
see...
Example 6. Find all ordered triples of primes (p, q, r) such that
p[q
r
+ 1, q[r
p
+ 1, r[p
q
+ 1.
Reid Barton, TST USA 2003
Solution. It is quite clear that p, q, r are distinct. Indeed, if for
example p = q, then the relation p[q
r
+ 1 is impossible. We will prove
that we cannot have p, q, r > 2. Suppose this is the case. The ﬁrst
condition p[q
r
+ 1 implies p[q
2r
− 1 and so o
p
(q)[2r. If o
p
(q) is odd, it
follows that p[q
r
−1, which combined with p[q
r
+ 1 yields p = 2, which
199
is impossible. Thus, o
p
(q) is either 2 or 2r. Could we have o
p
(q) = 2r?
No, since this would imply that 2r[p −1 and so 0 ≡ p
q
+1 (mod r) ≡ 2
(mod r), that is r = 2, false. Therefore, the only possibility is o
p
(q) = 2
and so p[q
2
−1. We cannot have p[q−1, because p[q
r
+1 and p ,= 2. Thus,
p[q + 1 and in fact p[
q + 1
2
. In the same way, we ﬁnd that q[
r + 1
2
and
r[
p + 1
2
. This is clearly impossible, just by looking at the largest among
p, q, r. So, our assumption was wrong and indeed one of the three primes
must equal 2. Suppose without loss of generality that p = 2. Then q is
odd, q[r
2
+1 and r[2
q
+1. Similarly, o
r
(2)[2q. If q[o
r
(2), then q[r −1 and
so q[r
2
+1−(r
2
that q is odd. Thus, o
r
(2)[2 and r[3. As a matter of fact, this implies
that r = 3 and q = 5, yielding the triple (2,5,3). It is immediate to
verify that this triple satisﬁes all conditions of the problem. Moreover,
all solutions are given by cyclic permutations of the components of this
triple.
Can you ﬁnd the smallest prime factor of the number 2
2
5
+ 1. Yes,
with a large amount of work, you will probably ﬁnd it. But what about
the number 12
2
15
+1? It has more than 30000 digits, so you will probably
be bored before ﬁnding its smallest prime factor. But here is a beautiful
and short solution, which does not need a single division.
Example 7. Find the smallest prime factor of the number 12
2
15
+1.
Solution. Let p be this prime number. Because p[12
2
16
−1, we ﬁnd
that o
p
(12)[2
16
. Exactly as in the solution of the ﬁrst example, we ﬁnd
that o
p
(12) = 2
16
and so 2
16
[p −1. Therefore p ≥ 1 +2
16
. But it is well-
known that 2
16
+1 is a prime (and if you do not believe, you can check;
it is not that diﬃcult). So, we might try to see if this number divides
12
2
15
+ 1. Let q = 2
16
+ 1. Then 12
2
15
+ 1 = 2
q−1
3
q−1
2
+ 1 ≡ 3
q−1
2
+ 1
(mod q). It remains to see whether
_
3
q
_
= −1. But this is done in
200
3
q−1
2
+ 1 ≡ 0 (mod 2) and 2
16
+ 1 is the smallest prime factor of the
number 12
2
15
+ 1.
Ok, you must be already tired of this old fashioned idea that any
prime factor of 2
2
n
+ 1 is congruent to 1 modulo 2
n+1
. Yet, here is a
problem that will keep you occupied for a certain period of time, even
if it uses only this simple idea.
Example 8. Prove that for any n > 1 the largest prime factor of
2
2
n
+ 1 is at least equal to n 2
n+2
+ 1.
China TST, 2005
Solution. The reader will not imagine how simple this problem re-
ally is. If the start is correct... Indeed, let us write 2
2
n
+ 1 = p
k
1
1
. . . p
k
r
r
with p
1
< . . . , p
r
prime numbers. We know that we can ﬁnd q
i
∈ N
such that p
i
= 1 +2
n+1
q
i
. Now, reduce the relation 2
2
n
+1 = p
k
1
1
. . . p
k
r
r
modulo 2
2n+2
. It follows that 1 ≡ 1 + 2
n+1
r

i=1
k
i
q
i
(mod 2
2n+2
) and so
r

i=1
k
i
q
i
≥ 2
n+1
. But then q
r
r

i=1
k
i
≥ 2
n+1
. Now everything becomes sim-
ple, since we have 2
2
n
+1 > (1+2
n+1
)
k
1
++k
r
and so k
1
+ +k
r

2
n
n + 1
.
This shows that q
r
≤ 2(n + 1) and the proof ﬁnishes here.
Problems for training
1. Let a, n > 2 be positive integers such that n[a
n−1
−1 and n does
not divide any of the numbers a
x
− 1, where x < n − 1 and x[n − 1.
Prove that n is a prime number.
2. Let p be a nonzero polynomial with integral coeﬃcients. Prove
that there are at most ﬁnitely many numbers n for which p(n) and
2
2
n
+ 1 are not relatively prime.
3. Let p > 3 be a prime. Prove that any positive divisor of the
number
2
p
+ 1
3
is of the form 2kp + 1.
201
Fermat
4. Let a > b > 1 and n > 1 be positive integers. Prove that any
positive divisor of the number a
n
− b
n
is either of the form nk + 1 or
divides a number of the form a
d
−b
d
, with d[n, d < n.
5. Find all positive integers m, n for which n[1 +m
3
n
+m
23
n
.
Bulgaria, 1997
6. Find the smallest repunit divisible by 19.
Gazeta Matematica
7. Let p be a prime and q > 5 a prime factor of the number 2
p
+3
p
.
Prove that q > p.
Laurentiu Panaitopol, TST Romania
8. Let m > 1 be an odd number. Find the smallest number n such
that 2
1989
[m
n
−1.
IMO 1989 Shortlist
9. Let 0 < m < n be integers such that 1978
m
and 1978
n
have the
same last three digits. Find the least possible value of m+n.
IMO 1978
10. Let p be a prime number and let d a positive divisor of p − 1.
Prove that there is a positive integer n such that o
p
(n) = d.
11. Let q = k 2
m
+ 1 be a divisor of the number 2
2
n
+ 1, where k
is odd. Find o
q
(k) in terms of n and v
2
(m)
J. van de Lune
12. Let n be a positive integer such that n −1 = FR, where all the
prime factors of F are known and gcd(F, R) = 1. Suppose further that
there is an integer a such that n[a
n−1
−1 and for all primes p dividing
n −1 we have gcd(n, a
n−1
p
−1) = 1. Prove that any prime factor of n is
congruent to 1 modulo F.
202
Proth, Pocklington, Lehmer Test
13. Let a > 1 be an integer and let us deﬁne o
p
(a) = 0 if p[a. Prove
that the function f : ¦2, 3, 5, 7, 11, . . . ¦ →N, f(p) =
p −1
o
p
(a)
is unbounded.
Jon Froemke, Jerrold W Grossman, AMM
14. Let d = o
p
(n) and let k = v
p
(n
d
−1).
a) If k > 1 then o
p
j (n) = d for j ≤ k and o
p
j (n) = p
j−k
d for all
j ≥ k.
b) If k = 1 then let l = v
p
(n
pd
−1). Prove that o
p
(n) = d, o
p
j (n) = pd
for 2 ≤ j ≤ l and o
p
j (n) = p
j−l+1
d, for all j ≥ l.
15. Let A be a ﬁnite set of prime numbers and let a ≥ 2 be a positive
integer. Prove that there are only ﬁnitely many positive integers n such
that all prime factors of a
n
−1 are in A.
16. Prove that for any prime p there is a prime number q that does
not divide any of the numbers n
p
−p, with n ≥ 1.
IMO 2003
17. Let a > 1 be a positive integer. Prove that for inﬁnitely many n
the largest prime factor of a
n
−1 is greater than nlog
a
n.
Gabriel Dospinescu
203
DENSITY AND REGULAR DISTRIBUTION
Everyone knows that (¦na¦)
n≥1
is dense in [0,1] if a is an irrational
number, a classical theorem of Kronecker. Various applications of this
nice result have appeared in diﬀerent contests and will probably make
the object of Olympiad problems in the future. Yet, there are some
examples in which this result is ineﬀective. A simple one is as follows:
using Kronecker’s theorem one can easily prove that for any positive
integer a that is not a power of 10 there exists n such that a
n
begins with
2006. The natural question: what fraction of numbers between 1 and n
have this property (speaking here about large values of n) is much more
diﬃcult and to answer it we need some stronger tools. This is the reason
for which we will try to discuss some classical approximation theorems,
particularly the very eﬃcient Weil criterion and its consequences. The
proofs are non-trivial and require some heavy-duty analysis. Yet, the
consequences that will be discussed here are almost elementary.
Of course, one cannot start a topic about approximation theorems
without talking ﬁrst about Kronecker’s theorem. We skip the proof, not
only because it is very well-known, but because we will prove a much
stronger result about the sequence (¦na¦)
n≥1
two beautiful problems, consequences of this theorem.
Example 1. Prove that the sequence ([n

2003])
n≥1
contains arbi-
trarily long geometric progressions with arbitrarily large ratio.
Solution. Let us take p a very large number. We will prove that
there are arbitrarily long geometric sequences with ratio p. Given
n ≥ 3, let us prove that we can ﬁnd a positive integer m such that
[p
k
m

2003] = p
k
[m

2003] for all 1 ≤ k ≤ n. If the existence of such a
number is proved, then the conclusion is immediate. But observe that
204
[p
k
m

2003] = p
k
[m

2003] is equivalent to [p
k
¦m

2003¦] = 0, or to
¦m

2003¦ <
1
p
n
. The existence of a positive integer m with the last
property is ensured by Kronecker’s theorem.
Here is a problem that is apparently very diﬃcult, but which is again
a simple consequence of Kronecker’s theorem.
Example 2. Consider k ≥ 1 and a such that log a is irrational.
Deﬁne the sequence x
n
as the number formed by the ﬁrst k digits of
the number [a
n
] with n ≥ 1. Prove that this sequence is not eventually
periodical.
Solution. The solution is based on certain simple, but useful re-
marks. First of all, the number formed with the ﬁrst k digits of a num-
ber m is [10
k−1+¡log m¦
]. The proof of this claim is not diﬃcult. Indeed,
let us write m = x
1
x
2
. . . x
p
, with p ≥ k. Then m = x
1
. . . x
k
10
p−k
+
x
k+1
. . . x
p
, hence x
1
. . . x
k
10
p−k
≤ m < (x
1
. . . x
k
+1) 10
p−k
. It follows
that x
1
. . . x
k
=
_
m
10
p−k
_
and, since p = 1 +[log m], the claim is proved.
Another remark is the following: there is a positive integer r such
that x
rT
> 10
k−1
. Indeed, assuming the contrary, we ﬁnd that for all
r > 0 we have x
rT
= 10
k−1
. Using the ﬁrst observation, it follows that
k −1 +¦log[a
rT
]¦ < log(1 + 10
k−1
) for all r. Thus
log
_
1 +
1
10
k−1
_
> log[a
rT
] −[log[a
rT
]] > log(a
rT
−1) −[log a
rT
]
= ¦rT log a¦ −log
a
rT
a
rT
−1
.
It suﬃces now to consider a sequence of positive integers (r
n
) such
that 1 −
1
n
< ¦r
n
T log a¦ (the existence is a simple consequence of Kro-
necker’s lemma) and we will deduce that:
log
_
1 +
1
10
k−1
_
+
1
n
+ log
a
r
n
T
a
r
n
T
−1
> 1 for all n.
205
The last inequality is clearly impossible.
Finally, assume the existence of such an r. It follows that for n > r
we have x
nT
= x
rT
, thus
¦log[a
nT
]¦ ≥ log
_
1 +
1
10
k−1
_
.
This shows that
log
_
1 +
1
10
k−1
_
≤ log[a
nT
] −[log[a
nT
]] ≤ nT log a −[log a
nT
]
= ¦nT log a¦ for all n > r.
Before passing to the quantitative results stated at the beginning
of this chapter, we must speak about a simple, yet surprising result,
which turns out to be very useful when dealing with real numbers and
their properties. Sometimes, it can even help us reducing the problem
to integers, as we will see in one of the examples. But ﬁrst, let us state
and prove this result.
Example 3. (Dirichlet) Let x
1
, x
2
, . . . , x
k
be some real numbers and
let ε > 0. There exists a positive integer n and integers p
1
, p
2
, . . . , p
k
such
that [nx
i
−p
i
[ < ε for all i.
Solution. Thus we need to prove that if we have a ﬁnite set of real
numbers, we can multiply all its elements by a suitable integer such that
the elements of the new set are as close to integers as we want.
Let us choose an integer N >
1
ε
and partition the interval [0, 1) in
N intervals,
[0, 1) =
N
_
s=1
J
s
, J
s
=
_
s −1
N
,
s
N
_
.
Now, choose n = N
k
+ 1 and associate to any positive integer
q ∈ ¦1, 2, . . . , n¦ a sequence of k positive integers α
1
, α
2
, . . . , α
k
, where
206
α
i
= s if and only if ¦qx
i
¦ ∈ J
s
. We obtain at most N
k
sequences cor-
responding to these numbers and so by Dirichlet’s criterion we can ﬁnd
1 ≤ u < v ≤ n such that the same sequence is associated to u and v.
This means that for all 1 ≤ i ≤ k we have
[¦ux
i
¦ −¦vx
i
¦[ <
1
N
≤ ε.
It suﬃces thus to pick n = v −u, p
i
= [vx
i
] −[ux
i
].
And here is how we can use this result in problems where it is more
comfortable to work with integers. But don’t kid yourself, there are
not many such problems. The one we are going to discuss has had a
circuitous itinerary between world’s Olympiads: proposed for IMO long
time ago, it appeared next at the W.L. Putnam Competition and later
on in a Japanese Mathematical Olympiad.
Example 4. Let x
1
, x
2
, . . . , x
2n+1
be positive real numbers with the
property: for any 1 ≤ i ≤ 2n+1 one can make two groups of n numbers
by using x
j
, j ,= i, such that the sum of the numbers in each group is
the same. Prove that all numbers are equal.
Solution. Of course, the problem for integers is very well-known and
easy: it suﬃces to observe that in this case all numbers x
i
have the same
parity and the use of inﬁnite descent solves the problem (either they are
all even and in this case we divide each one by two and obtain a new
set with smaller sum and the same properties; otherwise, we subtract 1
from each one and then divide by 2).
Now, assume that they are real numbers, which is clearly much more
subtle. First of all, if they are all rational, it suﬃces to multiply by their
common denominator and apply the ﬁrst case. Thus assume that at least
one of the numbers is irrational. Consider ε > 0, a positive integer n,
and some integers p
1
, p
2
, . . . , p
k
such that [nx
i
− p
i
[ < ε for all i. We
claim that if ε > 0 is small enough, the corresponding p
1
, p
2
, . . . , p
k
have
207
the same property as x
1
, x
2
, . . . , x
2n+1
. Indeed, take some i and write
the partition condition in the form

j,=i
a
ij
nx
j
= 0 or

j,=i
a
ij
(nx
j
−p
j
) = −

j,=i
a
ij
p
j
(where of course a
ij
∈ ¦−1, 1¦). Then
¸
¸
¸
¸
¸
¸

j,=i
a
ij
p
j
¸
¸
¸
¸
¸
¸
=
¸
¸
¸
¸
¸
¸

j,=i
a
ij
(nx
j
−p
j
)
¸
¸
¸
¸
¸
¸
≤ 2nε.
Thus if we choose ε <
1
2n
, then

j,=i
a
ij
p
j
= 0 and so p
1
, p
2
, . . . , p
k
have the same property. Since they are all integers, they must be equal
(again, because of the ﬁrst case). Thus we have proved that for any
N > 2n there are integers n
N
, p
N
such that [n
N
x
i
−p
N
[ ≤
1
N
.
Because at least one of the numbers x
1
, x
2
, . . . , x
2n+1
is irrational, it
is not diﬃcult to prove that the sequence (n
N
)
N>2n
is unbounded. But
2
N
> [n
N
[ max
i,j
[x
i
− x
j
[, hence max
i,j
[x
i
− x
j
[ = 0 and the problem is
solved.
Now, let us turn to more quantitative results about the set of frac-
tional parts of natural multiples of diﬀerent real numbers. The following
criterion, due to Weil, is famous and deserves to be discussed because of
its beauty and apparent simplicity.
Weil criterion. Let (a
n
)
n≥1
be a sequence of real numbers from the
interval [0,1]. Then the following statements are equivalent:
a) For any real numbers 0 ≤ a ≤ b ≤ 1,
lim
n→∞
[¦i[ 1 ≤ i ≤ n, a
i
∈ [a, b]¦[
n
= b −a;
b) For any continuous function f : [0, 1] →R,
lim
n→∞
1
n
n

k=1
f(a
k
) =
_
1
0
f(x)dx;
208
c) For any positive integer p ≥ 1,
lim
n→∞
1
n
n

k=1
e
2iπpa
k
= 0.
In this case we will say that the sequence is equidistributed.
We will present just a sketch of the solution, containing yet all the
necessary ingredients.
First of all, we observe that a) says precisely that b) is true for
the characteristic function of any sub-interval of [0,1]. By linearity, this
remains true for any piecewise function. Now, there is a well-known and
easy to verify property of continuous functions: they can be uniformly
approximated with piecewise functions. That is, given ε > 0, we can ﬁnd
a piecewise function g such that [g(x) −f(x)[ < ε for all x ∈ [0, 1]. But
then if we write
¸
¸
¸
¸
¸
1
n
n

k=1
f(a
k
) −
_
1
0
f(x)dx
¸
¸
¸
¸
¸

1
n
n

k=1
[f(a
k
) −g(a
k
)[ +
_
1
0
[f(x) −g(x)[dx
+
¸
¸
¸
¸
¸
1
n
n

k=1
g(a
k
) −
_
1
0
g(x)dx
¸
¸
¸
¸
¸
and apply the result in b) for the function g, we easily deduce that b)
is true for any continuous function.
The fact that b) implies c) is immediate. More subtle is that b)
implies a). Let us consider the subinterval I = [a, b] with 0 < a < b < 1.
Next, consider two sequences of continuous functions f
k
, g
k
such that f
k
is zero on [0, a], [b, 1] and 1 on
_
a +
1
k
, b −
1
k
_
(being aﬃne otherwise),
while g
k
has ”the same” properties but is greater than or equal to λ
I
(the characteristic function of I = [a, b]). Therefore
1
n
n

j=1
f
k
(a
j
) ≤
[¦i[ 1 ≤ i ≤ n, a
i
∈ [a, b]¦[
n

1
n
n

j=1
g
k
(a
j
).
209
But from the hypothesis,
1
n
n

j=1
f
k
(a
j
) →
_
1
0
f
k
(x)dx = b −a −
1
k
and
1
n
n

j=1
g
k
(a
j
) →
_
1
0
g
k
(x)dx = b −a +
1
k
.
Now, let us take ε > 0 and k suﬃciently large. The above inequalities
show that actually for all suﬃciently large positive integer n
¸
¸
¸
¸
[¦i[ 1 ≤ i ≤ n, a
i
∈ [a, b]¦[
n
−b +a
¸
¸
¸
¸
≤ 2ε
this proof for the case a = 0 or b = 1.
Finally, let us prove that c) implies b). Of course, a linearity argu-
ment allows us to assume that b) is true for any trigonometric poly-
nomials of any degree. Because any continuous function f : [0, 1] → R
satisfying f(0) = f(1) can be uniformly approximated by trigonometric
polynomials (this is a really non-trivial result due to Weierstrass), we
deduce that b) is true for continuous functions f for which f(0) = f(1).
Now, given f : [0, 1] →R continuous, it is immediate that for any ε > 0
we can ﬁnd two continuous functions g, h, both having equal values at
0 and 1 and such that
[f(x) −g(x)[ ≤ h(x) and
_
1
0
h(x)dx ≤ ε.
Using the same arguments as those used to prove that b) implies a),
one can easily see that b) is true for any continuous function.
The ﬁrst problem that we discuss is in fact the most common result
in order to appreciate the power of Weil’s criterion. Before presenting the
second problem, we need another deﬁnition: we say that the sequence
210
(a
n
)
n≥1
is uniformly distributed mod 1 if the sequence of fractional parts
of a
n
is equidistributed. So, here is the classical example.
Example 5. Let a be an irrational number. Then the sequence
(na)
n≥1
is uniformly distributed mod 1.
Solution. Well, after so much work, we deserve a reward: this is a
simple consequence of Weil’s criterion. Indeed, it suﬃces to prove that
c) is veriﬁed, which reduces to proving that
lim
n→∞
1
n
n

k=1
e
2iπpka
= 0 (∗)
for all integers p ≥ 1. But this is just a geometric series!!! A one-line
computation shows that (∗) is trivially satisﬁed and thus we have the
desired result.
It’s probably time to solve the problem presented in the very begin-
ning of this note: how to compute the density of those numbers n for
which 2
n
begins with 2006 (for example). Well, again a reward: this is
going to be equally easy (of course, the reader needs some rest before
looking at some deeper results...).
Example 6. What is the density of the set of positive integers n for
which 2
n
begins with 2006?
Solution. Indeed, 2
n
begins with 2006 if and only if there is a
p ≥ 1 and some digits a
1
, a
2
, . . . , a
p
∈ ¦0, 1, . . . , 9¦ such that 2
n

2006a
1
a
2
. . . a
p
, which is clearly equivalent to the existence of p ≥ 1
such that
2007 10
p
> 2
n
≥ 2006 10
p
.
This can be rewritten in the form
log 2007 +p > nlog 2 ≥ log 2006 +p
This implies [nlog 2] = p +3 hence log
2007
1000
> ¦nlog 2¦ > log
2006
1000
.
Thus the density of the desired set is exactly the density of the set of
211
positive integers n satisfying
log
2007
1000
> ¦nlog 2¦ > log
2006
1000
.
From example 5, the last set has density log
2007
2006
and so this is the
We have seen a beautiful proof of the fact that if a is irrational, then
(na)
n≥1
is uniformly distributed mod 1. Actually, much more is true,
but this much more is also much more diﬃcult to prove. The next two
examples are two important theorems. The ﬁrst is due to Van der Corput
and shows how a brilliant combination of algebraic manipulations and
Weil’s criterion can yield diﬃcult and nice results.
Example 7. (Van der Corput) Let (x
n
) be a sequence of real num-
bers such that the sequences (x
n+p
− x
n
)
n≥1
are equidistributed for all
p ≥ 1. Then the sequence (x
n
) is also equidistributed.
This is not an Olympiad problem!!! But mathematics is not just
Olympiad and from time to time (in fact, from a certain time on) one
should try to discover what is behind such great results. This is the
reason for which we present a proof of this theorem, a diﬃcult proof
that uses the ”well-known” but not easy to remember inequality of Van
der Corput.
Lemma. (Van der Corput) For any complex numbers z
1
, z
2
, . . . , z
n
and any h ∈ ¦1, 2, . . . , n¦, the following inequality is true (with the con-
vention that z
i
= 0 for any integer i not in ¦1, 2, . . . , n¦):
h
2
¸
¸
¸
¸
¸
n

i=1
z
i
¸
¸
¸
¸
¸
2
≤ (n +h −1)
_
2
p−1

r=1
(h −r)Re
_
n−r

i=1
z
i
z
i+r
_
+h
n

i=1
[z
i
[
2
_
.
Unbelievable, but true! Not to mention that the proof of this in-
equality is anything but easy. We will limit to give the main idea of the
212
proof, the computations being very technical. The idea behind this fun-
damental inequality is another fundamental one. You would have never
guessed: the Cauchy Schwarz inequality!!! The simple observation that
h
n

i=1
z
i
=
n+h−1

i=1
h−1

j=0
z
i−j
allows us to write (via Cauchy Schwarz’s inequality):
h
2
¸
¸
¸
¸
¸
n

i=1
z
i
¸
¸
¸
¸
¸
2
≤ (n +h −1)
n+h−1

i=1
¸
¸
¸
¸
¸
¸
h−1

j=0
z
i−j
¸
¸
¸
¸
¸
¸
2
.
And next ? Well... this is where the readers will get some satisfac-
tion... if they have the patience to expand
n+p−1

i=1
¸
¸
¸
¸
¸
¸
p−1

j=0
z
i−j
¸
¸
¸
¸
¸
¸
2
and see that
it is nothing else than
2
p−1

r=1
(p −r)Re
_
n−r

i=1
z
i
z
i+r
_
+p
n

i=1
[z
i
[
2
.
Wishing them good luck with the computations, we will now prove
Van der Corput’s theorem, by using this lemma and Weil’s criterion.
Of course, the idea is to prove that
lim
n→∞
1
n
n

k=1
e
2iπpx
k
= 0 for all p ≥ 1.
Fix such a p ≥ 1 and take for the moment a positive real number h and
ε ∈ (0, 1) (h may depend on ε). Also, denote z
j
= e
2iπpx
j
. Using the
lemma, we have:
¸
¸
¸
¸
¸
¸
1
n
n

j=1
z
j
¸
¸
¸
¸
¸
¸
2

1
n
2

n +h −1
h
2
_
_
hn + 2
h−1

i=1
(h −i)Re
_
_
n−i

j=1
z
j
z
i+j
_
_
_
_
.
213
Now, let us regard
Re
_
_
n−i

j=1
z
j
z
i+j
_
_
= Re
_
_
n−i

j=1
e
2iπp(x
j
−x
i+j
)
_
_

¸
¸
¸
¸
¸
¸
n−i

j=1
e
2iπp(x
j
−x
i+j
)
¸
¸
¸
¸
¸
¸
.
Using Weil’s criterion for the sequences (x
n+i
− x
n
) for i =
1, 2, . . . , h −1 we deduce that for all suﬃciently large n we have
¸
¸
¸
¸
¸
¸
n−i

j=1
e
2iπp(x
j
−x
i+j
)
¸
¸
¸
¸
¸
¸
≤ εn.
Therefore
¸
¸
¸
¸
¸
¸
1
n
n

j=1
z
j
¸
¸
¸
¸
¸
¸
2

1
n
2

n +h −1
h
2
_
hn + 2εn
h−1

i=1
(h −i)
_
<
n +h −1
nh
(1 +ε) <
2(1 +ε)
h
for all suﬃciently large n. Now, by choosing h >
2(1 +ε)
ε
2
, we deduce
that for all suﬃciently large n we have
¸
¸
¸
¸
¸
¸
1
n
n

j=1
z
j
¸
¸
¸
¸
¸
¸
≤ ε.
This shows that Weil’s criterion is veriﬁed and thus (x
n
) is equidis-
tributed.
This was surely the most diﬃcult result of this unit, but why not
taking one more step once we are already here? Let us prove the following
weaker (but as the reader will probably agree, absolutely nontrivial)
version of famous theorem of Weil. It is related to the equidistribution
of the sequence (¦f(n)¦)
n≥1
where is a real polynomial having at least
one irrational coeﬃcient except for the free term. We will not prove this
here, but focus on the following result.
214
Example 8. (Weil) Let f be a polynomial with real coeﬃcients and
irrational leading coeﬃcient. Then the sequence (¦f(n)¦)
n≥1
is equidis-
tributed.
The reader has probably noticed that this is an immediate conse-
quence of Van der Corput’s theorem (but just imagine the amount of
work done to arrive at this conclusion!!!). Indeed, the proof by induction
is immediate.
Indeed, if f has degree 1, then the conclusion is immediate (see
example 5). Now, if the result holds for polynomials of degree at most k,
it suﬃces (by Van der Corput’s theorem) to prove that for all positive
integers p, the sequence (f(n + p) − (f(n)) is equidistributed. But this
is exactly the induction hypothesis applied to the polynomial (whose
leading coeﬃcient is clearly irrational) f(X + p) − f(X). The proof by
induction ﬁnishes here.
Problems for training
1. Compute sup
n≥1
_
_
min
p,q∈N
p+q=n
[p −q

3[
_
_
.
Putnam Competition
2. Prove that by using diﬀerent terms of the sequence [n
2

2006] one
can construct geometric sequences of any length.
3. Let x be an irrational number and let f(t) = min(¦t¦, ¦1 − t¦).
Prove that given any ε > 0 one can ﬁnd a positive integer n such that
f(n
2
x) < ε.
Iran 2004
4. Prove that the sequence consisting of the ﬁrst digit of 2
n
+ 3
n
is
not periodical.
215
5. Suppose that A = ¦n
1
, n
2
, . . . ¦ is a set of positive integers such
that the sequence (cos n
k
)
k≥1
is convergent. Then prove that A has zero
density.
Marian Tetiva
6. Suppose that f is a real, continuous, and periodical function such
that the sequence
_
n

k=1
[f(k)[
k
_
n≥1
is bounded. Prove that f(k) = 0 for
all positive integers k. Give a necessary and suﬃcient condition ensuring
the existence of a constant c > 0 such that
n

k=1
[f(k)[
k
> c ln n for all n.
Gabriel Dospinescu
7. Does the sequence sin(n
2
) + sin(n
3
) converge?
Gabriel Dospinescu
8. Let f be a polynomial with integral coeﬃcients and let a be an
irrational number. Can all numbers f(k), k = 1, 2, . . . be in the set
A = ¦[na][ n ≥ 1¦? Is it true that any set of positive integers with
positive density contains an inﬁnite arithmetical sequence?
9. Let a, b be positive real numbers such that ¦na¦ + ¦nb¦ < 1 for
all n. Then at least one of them is an integer.
10. Prove that for every k one can ﬁnd distinct positive inte-
gers n
1
, n
2
, . . . , n
k
such that [n
1

2], [n
2

2], . . . , [n
k

2] and [n
1

3],
[n
2

3], . . . , [n
k

3] are both geometrical sequences.
After a romanian IMO TST problem
11. A ﬂea moves in the positive direction of an axis, starting from
the origin. It can only jump over distances equal to

2 and

2005.
Prove that there exists n
0
such that the ﬂea will be able to arrive in any
interval [n, n + 1] for all n ≥ n
0
.
Romanian Contest, 2005
216
12. Let a, b, c be positive real numbers. Prove that the sets
A = ¦[na][ n ≥ 1¦, B = ¦[nb][ n ≥ 1¦, C = ¦[nc][ n ≥ 1¦
cannot form a partition of the set of positive integers.
Putnam
13. Let z
1
, z
2
, . . . , z
n
be arbitrary complex numbers. Prove that for
any ε > 0 there are inﬁnitely many positive integers n such that
ε +
k
_
[z
k
1
+z
k
2
+ +z
k
n
[ < max¦[z
1
[, [z
2
[, . . . , [z
n
[¦.
217
THE SUM OF DIGITS OF A POSITIVE INTEGER
Problems about the sum of digits of a positive integer often occur in
mathematical contests because of their diﬃculty and the lack of stan-
dard ways to tackle the problem. This is why a synthesis of the most
frequent techniques that occur in such cases would be useful. We have
selected several representative problems to show how the main results
and techniques work and why they are so important.
We will work only in base 10 and we will denote the decimal sum of
digits of the positive integer x by s(x). The following ”formula” can be
checked easily:
s(n) = n −9

k≥1
_
n
10
k
_
(1)
From (1) we can easily deduce some well-known results about s(n) such
as s(n) ≡ n(mod9) and s(m+n) ≤ s(m) +s(n). Unfortunately, (1) is a
clumsy formula, which can hardly be used in applications. On the other
hand, there are several more or less known results about sum of digits,
results which may oﬀer simple ways to tackle hard problems. This is
what we will discuss about in the following.
The easiest of these techniques is, probably, just the careful anal-
ysis of the structure of the numbers and their digits. This can work
surprisingly well, as we will see in the following examples.
1. Prove that among any 79 consecutive numbers, one can choose at
least one whose sum of digits is a multiple of 13.
Baltic, 1997
Solution. Note that among the ﬁrst 40 numbers, there are exactly
4 multiples of 10. Also, it is clear that the last but one digit of one of
them is at least 6. Let x be this number. Obviously, x, x+1,..., x+39 are
among our numbers, so s(x), s(x) +1,...,s(x) +12 occur as sum of digits
218
in some of our numbers. Obviously, one of these numbers is a multiple
of 13 and we are done.
We will continue with two harder problems, which still do not require
any special result or technique.
2. Find the greatest N such that one can ﬁnd N consecutive numbers
with the property that the sum of digits of the k-th number is divisible
by k, for k = 1, 2, ..., N.
Tournament of Towns, 2000
Solution. The answer here is not trivial at all, namely 21. The
main idea is that among s(n+2), s(n+12) and s(n+22) there are two
consecutive numbers, which is impossible since they should all be even.
In truth, we make transports at a +10 only when the last but one digit
of a is 9, but this situation can occur at most once in our case. So, for
N > 21, we have no solution. For N = 21, we can choose N+1, N+2,...,
N +21, where N = 291 10
11!
−12. For i = 1 we have nothing to prove.
For 2 ≤ i ≤ 11, s(N +i) = 2 +9 +0 +9(11! −1) +i −2 = i +11! while
for 12 ≤ i ≤ 21, s(N +i) = 2 +9 +1 +(i −12) = i, so our numbers have
the desired property.
3. How many positive integers n ≤ 10
2005
can be written as the sum
of 2 positive integers with the same sum of digits?
2005
− 9023. At ﬁrst glance, it might seem al-
most impossible to ﬁnd the exact number of positive integers with this
property. In fact, the following is true: a positive integer cannot be writ-
ten as the sum of two numbers with the same sum of digits iﬀ all of its
digits (eventually) excepting the ﬁrst are 9 and the sum of its digits is
odd.
219
Let n be such a number. Suppose there are a, b ∈ Z
+
such that
n = a+b and s(a) = s(b). The main fact is that when we add a+b = n,
there are no transports. This is clear enough. It follows that s(n) =
s(a) +s(b) = 2s(a), which is impossible since s(n) is odd.
Now we will prove that any number n which is not one of the numbers
stated above, can be written as the sum of 2 positive integers with the
Lemma. There is a ≤ n such that s(a) ≡ s(n −a)(mod2).
Proof. If the s(n) is even, take a = 0. If s(n) is odd, then n must
have a digit which is not the ﬁrst and is not equal to 9, otherwise it
would have one of the forbidden forms. Let c be the value of this digit
and p its position (from right to left). Then let us chose a = 10
p−1
(c+1).
At the adding a + (n −a) = n there is exactly one transport, so
s(a) +s(n −a) = 9 +s(n) ≡ 0(mod2) ⇒ s(a) ≡ s(n −a)(mod2)
which proves our claim.
Back to the original problem. All we have to do now is take one-by-
one a ”unity” from a number and give it to the other until the 2 numbers
have the same sum of digits. This will happen since they have the same
parity. So, let us do this rigorously. Let
a = a
1
a
2
..a
k
, n −a = b
1
b
2
...b
k
The lemma shows that the number of elements of the set I = ¦i ∈
¦1, 2, ..., k¦ : 2 does not divide a
i
+b
i
¦ is even, so it can be divided into 2
sets with the same number of elements, say I
1
and I
2
. For i = 1, 2, ..., k
deﬁne A
i
= (a
i
+b
i
)/2 if i ∈ I, (a
i
+b
i
+1)/2 if i ∈ I
1
or (a
i
+b
i
−1)/2
if i ∈ I
2
and B
i
= a
i
+b
i
−A
i
. It is clear that the numbers
A = A
1
A
2
...A
k
, B = B
1
B
2
...B
k
220
have the property that s(A) = s(B) and A + B = n. The proof is
complete.
We have previously seen that s(n) ≡ n( mod9). This is probably the
most famous property of the function s and it has a series of remarkable
applications. Sometimes it is combined with some simple inequalities
such as s(n) ≤ 9(¸lg n| + 1). Some immediate applications are the fol-
lowing:
4. Find all n for which one can ﬁnd a and b such that
s(a) = s(b) = s(a +b) = n.
Vasile Zidaru and Mircea Lascu, JBMO TST, 2002
Solution. We have a ≡ b ≡ a + b ≡ n(mod9), so 9 divides n. If
n = 9k, we can take a = b = 10
k
−1 and we are done since s(10
k
−1) =
s(2 10
k
−2) = 9k.
5. Find all the possible values of the sum of digits of a perfect square.
Iberoamerican, 1995
Solution. What does sum of digits has to do with perfect squares?
Apparently, nothing, but perfect squares do have something to do with
remainders mod 9! In fact, it is very easy to prove that the only possible
values of a perfect square mod 9 are 0, 1, 4 and 7. So, we deduce that
the sum of digits of a perfect square must be congruent to 0, 1, 4 or 7
mod 9. To prove that all such numbers work, we will use a small and
very common (but worth to remember!) trick: use numbers that consist
almost only of 9-s. We have the following identities:
99...99
. ¸¸ .
n
2
= 99...99
. ¸¸ .
n−1
8 00...00
. ¸¸ .
n−1
1 ⇒ s(99...99
. ¸¸ .
n
2
) = 9n
99...99
. ¸¸ .
n−1
1
2
= 99...99
. ¸¸ .
n−2
82 00...00
. ¸¸ .
n−2
81 ⇒ s(99..99
. ¸¸ .
n−1
1
2
) = 9n + 1
221
99...99
. ¸¸ .
n−1
2
2
= 99...99
. ¸¸ .
n−2
84 00...00
. ¸¸ .
n−2
64 ⇒ s(99..99
. ¸¸ .
n−1
2
2
) = 9n + 4
99...99
. ¸¸ .
n−1
4
2
= 99...99
. ¸¸ .
n−2
88 00...00
. ¸¸ .
n−2
36 ⇒ s(99..99
. ¸¸ .
n−1
4
2
) = 9n + 7
and since s(0) = 0, s(1) = 1, s(4) = 4 and s(25) = 7 the proof is
complete.
6. Compute s(s(s(4444
4444
))).
IMO 1975
Solution. Using the inequality s(n) ≤ 9(¸lg n|+1) several times we
have
s(4444
4444
) ≤ 9(
_
lg 4444
4444
_
+ 1) < 9 20, 000 = 180, 000;
s(s(4444
4444
)) ≤ 9(
_
lg s(4444
4444
)
_
+ 1) ≤ 9(lg 180, 000 + 1) ≤ 36,
so s(s(s(4444
4444
))) ≤ 12. On the other hand, s(s(s(n))) ≡ s(s(n)) ≡
s(n) ≡ n(mod9) and since
4444
4444
≡ 7
4444
= 7 7
3
1481
≡ 7(mod9),
the only possible answer is 7.
Finally, we present a beautiful problem which appeared in the Rus-
sian Olympiad and, later, in Kvant.
7. Prove that for any N there is n ≥ N such that s(3
n
) ≥ s(3
n+1
).
Solution. Suppose by way of contradiction that there is one N such
that s(3
n+1
) − s(3
n
) > 0, ∀n ≥ N. But, for n ≥ 2, s(3
n+1
) − s(3
n
) ≡
0(mod9), so s(3
n+1
) −s(3
n
) ≥ 9, ∀n ≥ N. It follows that
n

k=N+1
_
3
k+1
−3
k
_
≥ 9(n −N) ⇒ s(3
n+1
) ≥ 9(n −N), n ≥ N + 1.
But s(3
n+1
) ≤ 9(
_
lg 3
n+1
_
+ 1), so 9n − 9N ≤ 9 + 9(n + 1) lg 3, for all
n ≥ N + 1. This is obviously a contradiction.
If so far we have studied some remarkable properties of the function
s, which were quite well-known, it is time to present some problems and
222
results which are less familiar, but interesting and hard. The ﬁrst result
is the following:
Statement. If 1 ≤ x ≤ 10
n
, then s(x(10
n
−1)) = 9n.
Proof. The idea is very simple. All we have to do is write x =
a
1
a
2
...a
j
with a
j
,= 0 (we can ignore the ﬁnal 0-s of x) and note that
x(10
n
−1) = a
1
a
2
...a
j−1
(a
j
−1) 99...99
. ¸¸ .
n−j
(9 −a
1
)...(9 −a
j
)(10 −a
j
),
which obviously has the sum of digits equal to 9n.
The previous result is by no means hard, but we will see that it can
be the key in many situations. A ﬁrst application is:
8. Compute s(9 99 999 ... 99...99
. ¸¸ .
2
n
).
USAMO, 1992
Solution. The problem in trivial if we know the previous result. We
have
N = 9 99 999 ... 99...99
. ¸¸ .
2
n−1
< 10
1+2+...+2
n−1
< 10
2
n
−1
so s(99...99
. ¸¸ .
2
n
N) = 9 2
n
.
However, there are very hard applications of this apparently unim-
portant result, such as the following problem.
9. Prove that for any n there is a positive integer n which is divisible
by its sum of digits.
IMO Shortlist, 1998
Solution. Only to assure our readers that this problem did not ap-
pear on the ISL out of nowhere, such numbers are called Niven numbers
and they are an important research source in number theory. Now, let’s
solve it. We will see that constructing such a number is hard. First, we
223
will get rid of the case n = 3
k
, when we can take the number 11...11
. ¸¸ .
n
(it
can be easily proved by induction that 3
k+2
[10
3
k
−1).
Due to the trick that we should search numbers with many equal
digits and the last result, we decide that the required number p should
be aa...aa
. ¸¸ .
s
b (10
t
− 1), with aa...aa
. ¸¸ .
s
b ≤ 10
t
− 1. This number will have
s+t +1 digits and its sum of digits will be 9t. Therefore, we will require
s + t = n − 1 and 9t[ aa...aa
. ¸¸ .
s
b (10
t
− 1). We now use the fact that
for t a power of 3, 9t[10
t
− 1. So, let us take t = 3
k
where k is chosen
such that 3
k
< n < 3
k+1
. If we also take in account the condition
aa...aa
. ¸¸ .
s
b ≤ 10
t
− 1, the choice p = 11...11
. ¸¸ .
n−3
k
−1
2(10
3
k
− 1) when n ≤ 2 3
k
and p = 22...22
. ¸¸ .
23
k
(10
23
k
−1) otherwise becomes natural.
We continue our investigations in ﬁnding suitable techniques for
problems involving sum of digits with a very beautiful result. The follow-
ing result turned out to have several consequences, most of them being
very hard.
Statement. Any multiple of 11...11
. ¸¸ .
k
has sum of digits at least k.
Proof. We will use the extremal principle. Suppose by way of con-
tradiction that the statement is false and take M to be the smallest mul-
tiple of a such that s(M) < k, where a = 11...11
. ¸¸ .
k
. Note that s(ia) = ik
for i = 1, 2, ..., 9. So M ≥ 10a > 10
k
. Therefore, M = a
1
a
2
...a
p
, with
p ≥ k +1 and a
p
,= 0. Take the number N = M −10
p−k
a. Obviously, N
is a multiple of a. We will try to prove that s(N) < k. In this way, we
would contradict the minimality of M and the proof would be complete.
But this is not hard at all since if a
k+1
< 9, we have s(N) = s(M) < k
and if a
k+1
= 9, we have s(N) < s(M) < k.
224
We will show 3 applications of this fact, which might seem simple,
but which might be unsolvable without it.
10. Prove that for every k, we have
lim
n→∞
s(n!)
ln
k
ln n
= ∞
Gabriel Dospinescu
Solution. Due to the simple fact that 10
]lg n|
− 1 ≤ n ⇒ 10
]lg n|

1[n!, we have that s(n!) ≥ ¸lg n|, from which our conclusion follows
easily.
11. Let S be the set of positive integers whose decimal representation
contains only of at most 1988 1-s and the rest 0-s. Prove that there is a
positive integer which does not divide any member of S.
Tournament of Towns, 1988
Solution. Again, the solution follows directly from our result. We
can choose the number 10
1989
− 1, whose multiples have sum of digits
greater than 1988.
12. Prove that for any k > 0, there is an inﬁnite arithmetical se-
quence having the ratio relatively prime to 10, such that all its members
have the sum of digits greater than k.
IMO Shortlist, 1999
Solution. Let us remind that this is the last problem of ISL 1999,
so the hardest. The oﬃcial solution is indeed one for such a problem.
But, due to our ”theorem” we can chose the sequence a
n
= n(10
m
−1),
where m > k and we are done.
Now, as a ﬁnal proof of the utility of these two results, we will present
a hard, but beautiful, problem from the USAMO.
13. Let n be a ﬁxed positive integer. Denote by f(n) the smallest k
for which one can ﬁnd a set X ⊂ Z
+
of cardinality n with the property
225
that
s
_

x∈Y
x
_
= k
for all nonempty subsets Y of X. Prove that C
1
lg n < f(n) < C
2
lg n
for some constants C
1
and C
2
.
Gabriel Dospinescu and Titu Andreescu, USAMO 2005
Solution. We will prove that
¸lg(n + 1)| ≤ f(n) ≤ 9 lg
_
n(n + 1)
2
+ 1
_
,
which is enough to establish our claim. Let l be the smallest integer such
that
10
l
−1 ≥
n(n + 1)
2
.
Consider the set X = ¦j(10
l
−1) : 1 ≤ j ≤ n¦. By the previous inequality
and our ﬁrst statement, it follows that
s
_

x∈Y
x
_
= 9l
for all nonempty subsets Y of X, so f(n) ≤ 9l and the RHS is proved.
Let m be the largest integer such that n ≥ 10
m
− 1. We will use the
following well-known
Lemma. Any set M = ¦a
1
, a
2
, ..., a
m
¦ has a nonempty subset whose
sum of elements is divisible by m.
Proof. Consider the sums a
1
, a
1
+a
2
,..., a
1
+a
2
+... +a
m
. If one of
then is a multiple of m, them we are done. Otherwise, there are 2 of them
congruent mod m, say the i-th and the j-th. Then, m[a
i+1
+a
i+2
+...+a
j
so we are done.
From the lemma, it follows that any set X with n elements has a
subset, say Y , whose sum of elements is divisible by 10
m
− 1. By our
226
second statement, it follows that
s
_

x∈Y
x
_
≥ m ⇒ f(n) ≥ m,
and the proof is complete.
The last solved problem is one we consider to be very hard, and
which uses diﬀerent techniques than the ones we have mentioned so far.
14. Let a, b ∈ Z
+
such that s(an) = s(bn) for all n ∈ Z
+
. Prove that
lg a −lg b ∈ Z.
Solution. We start with an observation. If (max¦a, b¦, 10) = 1, then
the problem becomes trivial. Suppose a = max¦a, b¦. Then, by Euler’s
theorem, a[10
ϕ(a)
− 1, so there is an n such that an = 10
ϕ(a)
− 1 and
since numbers consisting only of 9-s have the sum of digits greater than
all previous numbers, it follows that an = bn, so a = b.
Let us solve now the harder problem. For any k ≥ 1, there is a n
k
such that 10
k
≤ an
k
≤ 10
k
+a −1. It follows that s(an
k
) is bounded, so
s(bn
k
) is bounded as well. On the other hand,
10
k
b
a
≤ bn
k
< 10
k
b
a
+b,
so, for suﬃciently large p, the ﬁrst (nonzero) digits of b/a are exactly
the same as the ﬁrst p digits of bn
k
for large enough k. This means the
the sum of the ﬁrst p digits of b/a is bounded, which could only happen
when this fraction has ﬁnitely many decimals. Analogously, we can prove
Let a = 2
x
5
y
m and b = 2
z
5
t
m
t
, where (m, 10) = (m
t
, 10)=1. It fol-
lows that m[m
t
and m
t
[m, so m = m
t
. Now, we can write the hypothesis
as
s(2
z
5
u
mn2
c−x
5
c−y
) = s(2
x
5
y
mn2
c−x
5
c−y
) = s(mn), ∀c ≥ max¦x, y¦
227
Now, if p = max¦z +c −x, u+c −y¦ −min¦z +c −x, u+c −y¦, we get
that there is a k ∈ ¦2, 5¦ such that s(mn) = s(mk
p
n) for all n ∈ Z
+
. It
follows that
s(m) = s(k
p
m) = s(k
2p
m) = s(k
3p
m) = ...
Let t = a
p
, so lg t ∈ R−Q unless p = 0. Now, we will use the following:
Lemma. If lg t ∈ R −Q, then for any sequence of digits, there is a
n ∈ Z
+
, such that t
n
m starts with the selected sequence of digits.
Proof. If we will prove that ¦¦lg t
n
m¦ : n ∈ Z
+
¦ is dense in (0, 1),
then we are done. But lg t
n
m = nlg t + m and by Kronecker’s theorem
¦¦nlg t¦ : n ∈ Z
+
¦ is dense in (0, 1), so the proof is complete.
The lemma implies the very important result that s(t
n
m) is un-
bounded for p ,= 0, which is a contradiction. So p = 0 and hence
z +c −x = u +c −y, so a = 10
x−z
b and the proof is complete.
This problem can be nicely extended to any base. The proof of the
general case is quite similar, although there are some very important
diﬀerences.
The upmetioned methods are just a point to start from in solving
such problems since the variety of problems involving sum of digits is
very large. The techniques are useful only when they are applied cre-
atively. Finally, we invite our readers to solve this proposed problems:
Proposed Problems
1. Prove that among any 39 there is one whose sum of digits is
divisible by 11.
USSR, 1961
2. Prove that among any 18 consecutive 2-digit numbers there is at
least one Niven number.
228
Tournament of Towns, Training, 1997
3. Are there positive integers n such that s(n) = 1, 000 and s(n
2
) =
1, 000, 000?
USSR, 1985
4. Prove that for any positive integer n there are inﬁnitely many
numbers m which do not contain any zero, such that s(n) = s(mn).
USSR, 1970
5. Find all x such that s(x) = s(2x) = s(3x) = ... = s(x
2
).
Kurschak, 1989
6. Are there arbitrarily long arithmetical sequences whose terms
have the same sum of digits? What about inﬁnite aritmetical sequences?
***
7. Prove that
lim
n→∞
s(2
n
) = ∞.
***
8. Are there p ∈ Z[X] such that
lim
n→∞
s(p(n)) = ∞?
***
9. Prove that there are arbitrarily long sequences of consecutive
numbers which do not contain any Niven number.
***
10. We start with a perfect number, diﬀerent form 6 (which is equal
to the sum of its divisors, except itself), and calculate its sum of digits.
Then, we calculate the sum of digits of the new number and so on. Prove
that we will eventually get 1.
***
229
11. Prove that there are inﬁnitely many x ∈ Z
+
such that
s(x) +s(x
2
) = s(x
3
).
Gabriel Dospinescu
12. a, b, c and d are primes such that 2 < a ≤ c and a ,= b. We now
that there is one M ∈ Z such that the numbers an +b and cn +d have
the same sum of digits for any n > M and base between 2 and a − 1.
Prove that a = c and b = d.
Gabriel Dospinescu
13. Let (a
n
)
n≥1
be a sequence such that s(a
n
) ≥ n. Prove that for
any n, we have
1
a
1
+
1
a
2
+... +
1
a
n
< 3.2
Can we replace 3.2 by 3?
Laurentiu Panaitopol
14. Prove that one can ﬁnd n
1
< n
2
< ... < n
50
such that
n
1
+s(n
1
) = n
2
+s(n
2
) = ... = n
50
+s(n
50
)
Poland, 1999
15. Study whether we can choose the numbers in the previous prob-
lem such that n
2
−n
1
= n
3
−n
2
= ... = n
50
−n
49
.
Gabriel Dospinescu
16. Deﬁne f(n) = n + s(n). A number m is called special if there
is a k such that f(k) = m. Prove that there are inﬁnitely many special
numbers 10
n
+b iﬀ b −1 is special.
Christopher D. Long
17. Find a Niven number with 100 digits.
Sankt Petersburg, 1990
230
18. Let S be a set such that for any α ∈ R−Q, there is a number n ∈
Z
+
such that ¸α
n
| ∈ S. Prove that S contains numbers with arbitrarily
large sum of digits.
Gabriel Dospinescu
19. Let a be a positive integer such that s(a
n
+ n) = 1 + s(n) for
any n > M, where M is given. Prove that a is a power of 10.
Gabriel Dospinescu
20. Let k ∈ Z
+
. Prove that there is a positive integer m such that
the equation n +s(n) = m has exactly k solutions.
Mihai Manea, Romanian IMO TST, 2003
21. Are there 19 positive integers with the same sum of digits, which
Rusia, 1999
22. Let a, b > 0. Prove that the sequence s(¸an + b|) contains a
constant subsequence.
Laurentiu Panaitopol, Romanian IMO TST, 2002
23. If s(n) = 100 and s(44n) = 800, ﬁnd s(3n).
Rusia, 1999
24. Find the smallest positive integer which can be expressed at the
same time as the sum of 2002 numbers with the same sum of digits and
as the sum of 2003 numbers with the same sum of digits.
Rusia, 2002
25. Prove that

n≥1
s(n)
n(n + 1)
=
10
9
ln 10.
O. Shallit
231
Open Questions
1. For any a > 1, we have
lim
n→∞
s(a
n
) = ∞
(proved only for a couple of values, namely 2, 4, 6, 8).
2. Is it true that
lim
n→∞
s(n!)
nln ln n
= ∞?
3. Let a, b ∈ Z
+
such that s(a
n
) = s(b
n
) for all n ∈ Z
+
. Prove that
lg a −lg b ∈ Z.
4. Prove that for any n, there are a, b ∈ Z such that lg a − lg b / ∈ Z
with the property that s(a
k
) = s(b
k
) for any k ∈ ¦1, 2, ..., n¦.
5. Is it true that
lim
n→∞
s(2
n
)
ln n
= ∞?
232
ANALYSIS AGAINST NUMBER THEORY?
”Olympiad problems can be solved without the aid of analysis or
linear algebra” is a sentence always heard when speaking about the el-
ementary problems given in contests. This is true, but the true nature
and essence of some of these problems is in analysis and this is the reason
for which such type of problems are always the highlight of a contest.
Their elementary solutions are very tricky and sometimes extremely dif-
ﬁcult, while using analysis they can be solved quickly. Well, of course,
”quickly” only if you see the sequence that hides after each problem.
Practically, our aim is to exhibit convergent sequences formed by in-
teger numbers. These sequences must become constant and from here
the problem is much easier. The diﬃculty is in ﬁnding those sequences.
Sometimes, this is easy, but most of the time this is a very diﬃcult task.
We will develop our skills in ”hunting” these sequences by solving ﬁrst
some easy problems (anyway, ”easy” is a relative concept: try to solve
them elementary and you will see if they really are easy) and after that
we will attack the chestnuts.
As usual, we begin with a classic beautiful problem, which has lots
of applications and extensions.
Example 1. Let f, g ∈ Z[X] be two non-constant polynomials such
that f(n)[g(n) for an inﬁnite natural numbers n. Prove that f divides g
in Q[X].
Solution. Indeed, we need to look at the remainder of g when di-
vided with f in Q[X]! Let us write g = fh+r, were h, r are polynomials
from Q[X] and deg r < deg f. Now, multiplying by the common denom-
inator of all coeﬃcients of polynomials h, r, the hypothesis becomes:
there exists two inﬁnite sequences (a
n
)
n≥1
, (b
n
)
n≥1
of integer numbers
and a positive integer N such that b
n
= N
r(a
n
)
f(a
n
)
(we could have some
233
problems with the roots of f, but they are in ﬁnite number and the
sequence (a
n
)
n≥1
tends to inﬁnity, so from a certain point, a
n
is not a
root of f). Since deg r < deg f, it follows that
r(a
n
)
f(a
n
)
→ 0, thus (b
n
)
n≥1
is a sequence of integer numbers that tends to 0. This implies that from
a certain point, all the terms of these sequence are 0. Well, this is the
same as r(a
n
) = 0 from a certain point n
0
, which is practically the same
thing with r = 0 (don’t forget that any non-zero polynomial has only a
ﬁnite number of roots!). But in this moment the problem is solved.
The next problem we are going to discuss is a particular case of a
much more general and classical result: if f is a polynomial with integer
coeﬃcients, k > 1 is a natural number and
k
_
f(n) ∈ Q for all natural
numbers n, then there exists a polynomial g ∈ Q[X] such that f(x) =
g
k
(x). We won’t discuss here this general result (the reader will ﬁnd a
proof in the chapter about arithmetic properties of polynomials).
Example 2. Let a ,= 0, b, c be integers such that for any natural
number n, the number an
2
+bn+c is a perfect square. Prove that there
exist x, y ∈ Z such that a = x
2
, b = 2xy, c = y
2
.
Solution. Let us begin by writing an
2
+ bn + c = x
2
n
for a certain
sequence of nonnegative integers (x
n
)
n≥1
. We could expect that x
n
−n

a
converges. And yes, it converges, but it’s not a sequence of integers, so
the convergence is useless. In fact, it’s not that useless, but we need
another sequence. The easiest way is to work with (x
n+1
−x
n
)
n≥1
, since
this sequence certainly converges to

why it wasn’t useless to ﬁnd that x
n
−n

a is convergent; we used this to
establish the convergence of (x
n+1
−x
n
)
n≥1
). This time, the sequence is
formed by integer numbers, so it is constant from a certain point. Thus,
we can ﬁnd a number M such that if n ≥ M then x
n+1
= x
n
+

a. Thus,
a must be a perfect square, let us say a = x
2
. A simple induction shows
that x
n
= x
M
+(n −M)x and so (x
M
−Mx +nx)
2
= x
2
n
2
+bn +c for
234
all n ≥ M. A simple identiﬁcation of coeﬃcients ﬁnishes the solution,
since we can take y = x
M
−Mx.
The following problem is based on the same idea, but it really doesn’t
seem to be related with mathematical analysis. In fact, as we will see,
it is closely related to the concept of convergence.
Example 3. Let a, b, c > 1 be positive integers such that for any
positive integer n there exists a positive integer k such that a
k
+b
k
= 2c
n
.
Prove that a = b.
Laurentiu Panaitopol
Solution. What does the problem say in fact? That we can ﬁnd a se-
quence of positive integers (x
n
)
n≥1
such that a
x
n
+b
x
n
= 2c
n
. What could
be the convergent sequence here? We see that (x
n
)
n≥1
is appreciatively
kn for a certain constant k. Thus, we could expect that the sequence
(x
n+1
− x
n
)
n≥1
converges. Let us see if this is true or not. From where
could we ﬁnd x
n+1
−x
n
? Certainly, by writing that a
x
n+1
+b
x
n+1
= 2c
n+1
and after that considering the value
a
x
n+1
+b
x
n+1
a
x
n
+b
x
n
= c. Now, let us sup-
pose that a > b and let us write
a
x
n+1
+b
x
n+1
a
x
n
+b
x
n
= c in the form
a
x
n+1
−x
n
1 +
_
b
a
_
x
n+1
1 +
_
b
a
_
x
n
= c,
from where it is easy to see that a
x
n+1
−x
n
converges to c. Why is it so
easy? It would be easy if we could show that x
n
→ ∞. Fortunately,
this is immediate, since 2a
x
n
> 2c
n
⇒ x
n
> nlog
a
c. So, we found
that a
x
n+1
−x
n
converges. Being a sequence of integer numbers, it must
become constant, so there exist M such that for all n ≥ M we have
235
a
x
n+1
−x
n
= c. This means that for all n ≥ M we also have
1 +
_
b
a
_
x
n+1
1 +
_
b
a
_
x
n
= 1.
But this is impossible, since a > b. Thus, our assumption was wrong
and we must have a ≤ b. Due to symmetry in a and b, we conclude that
a = b.
Another easy example is the following problem, in which ﬁnding the
right convergent sequence of integers in not diﬃcult at all. But, attention
must be paid to details!
Example 4. Let a
1
, a
2
, . . . , a
k
be positive real numbers such that
at least one of them is not an integer. Prove that there exits inﬁnitely
many natural numbers n such that n and [a
1
n] +[a
2
n] + +[a
k
n] are
relatively prime.
Gabriel Dospinescu, Arhimede Magazine
Solution. Of course, the solution of such a problem is better to be
indirect. So, let us assume that there exists a number M such that for all
n ≥ M the numbers n and [a
1
n] + [a
2
n] + + [a
k
n] are not relatively
prime. Now, what are the most eﬃcient numbers n to be used? Yes,
they are the prime numbers, since if n is prime and it is not relatively
prime with , [a
1
n] + [a
2
n] + + [a
k
n], then it must divide [a
1
n] +
[a
2
n] + + [a
k
n]. This suggests us to consider the sequence of prime
numbers (p
n
)
n≥1
. Since this sequence is inﬁnite, there is a number N
such that if n ≥ N then p
n
≥ M. According to our assumption, this
implies that for all n ≥ N there exist a natural number x
n
such that
[a
1
p
n
] +[a
2
p
n
] + +[a
k
p
n
] = x
n
p
n
. And now, you have already guessed
what is the convergent sequence! Yes, it is (x
n
)
n≥N
. This is obvious,
since
[a
1
p
n
] + [a
2
p
n
] + + [a
k
p
n
]
p
n
tends to n ≥ Na
1
+ a
2
+ + a
k
.
236
Thus, we can ﬁnd a number P such that for x
n
= a
1
+a
2
+ +a
k
for
all n ≥ P. But this is the same as ¦a
1
p
n
¦ +¦a
2
p
n
¦ + +¦a
k
p
n
¦ = 0.
Of course, this says that a
i
p
n
∈ Z for all i = 1, k and n ≥ P. Well,
the conclusion is immediate: a
i
∈ Z for all i = 1, k, which contradicts
the hypothesis. Consequently, we were wrong again and the problem
statement is right!
Step by step, we start to have some experience in ”guessing” the
sequences. Thus, it’s time to solve some more diﬃcult problems. The
next problem we are going to discuss may seem obvious after reading
the solution. In fact, it’s just that type of problem whose solution is very
short, but very hard to ﬁnd.
Example 5. Let a, b ∈ Z such that for all natural numbers n the
number a 2
n
+b is a perfect square. Prove that a = 0.
Poland TST
Solution. Again, we argue by contradiction. Suppose that a ,= 0.
Then, of course, a > 0, otherwise for large values of n the number
a 2
n
+b is negative. According to the hypothesis, there exists a sequence
of positive integers (x
n
)
n≥1
such that for all natural numbers n, x
n
=

a 2
n
+b. Then, a direct computation shows that lim
n→∞
(2x
n
−x
n+2
) =
0. This implies the existence of a natural number N such that for all
n ≥ P we have 2x
n
= x
n+2
. But 2x
n
= x
n+2
is equivalent with b = 0.
Then, a and 2a are both perfect squares, which is impossible for a ,= 0.
This shows, as usually, that our assumption was wrong and indeed a = 0.
A classical result of Schur states that for any non-constant polyno-
mial f with integer coeﬃcients, the set of prime numbers dividing at
least one of the numbers f(1), f(2), f(3), . . . is inﬁnite. The following
problem is a generalization of this result.
Example 6. Suppose that f is a polynomial with integer coeﬃcients
and (a
n
) is a strictly increasing sequence of natural numbers such that
237
a
n
≤ f(n) for all n. Then the set of prime numbers dividing at least one
term of the sequence is inﬁnite.
Solution. The idea is very nice: for any ﬁnite set of prime numbers
p
1
, p
2
, . . . , p
r
and any k > 0, we have

α
1

2
,...,α
N
∈Z
+
1
p

1
1
. . . p

N
N
< ∞.
Indeed, it suﬃces to remark that we have actually

α
1

2
,...,α
N
∈Z
+
1
p

1
1
. . . p

N
N
N

j=1

i≥0
1
p
ki
j
=
n

j=1
p
k
j
p
k
j
−1
.
On the other hand, by taking k =
1
2 deg(f)
we clearly have

n≥1
1
(f(n))
k
= ∞.
Thus, if the conclusion of the problem is not true, we can ﬁnd
p
1
, p
2
, . . . , p
r
such that any term of the sequence is of the form
p

1
1
. . . p

N
N
and thus

n≥1
1
a
k
n

α
1

2
,...,α
N
∈Z
+
1
p

1
1
. . . p

N
N
< ∞.
On the other hand, we also have

n≥1
1
a
k
n

n≥1
1
(f(n))
k
= ∞,
which is clearly impossible.
The same idea is employed in the following problem.
Example 7. Let a, b ≥ 2 be natural numbers. Prove that there is
a multiple of a which contains all digits 0, 1, . . . , b − 1 when written in
base b.
238
Solution. Let’s suppose the contrary. Then any multiple of a misses
at least a digit when written in base b. Since the sum of inverses of all
multiples of a diverges (because 1+
1
2
+
1
3
+ = ∞), it suﬃces to show
that the sum of inverses of all natural numbers missing at least one digit
in base b is convergent and we will reach a contradiction. But of course,
it suﬃces to prove it for a ﬁxed (but arbitrary) digit j. For any n ≥ 1,
there are at most (b − 1)
n
numbers which have n digits in base b, all
diﬀerent from j. Thus, since each one of them is at least equal to b
n−1
,
the sum of inverses of numbers that miss the digit j when written in base
b is at most equal to

n
b
_
b −1
b
_
n
, which converges. The conclusion
follows.
appeared in the Tournament of the Towns in 1982, in a Russian Team
Selection Test in 1997 and also in the Bulgarian Olympiad in 2003. It’s
beauty explains probably the preference for this problem.
Example 8. Let f ∈ Z[X] be a polynomial with leading coeﬃcient
1 such that for any natural number n the equation f(x) = 2
n
has at
least one natural solution. Prove that deg f = 1.
Solution. So, the problem states that there exists a sequence of
positive integers (x
n
)
n≥1
such that f(x
n
) = 2
n
. Let us suppose that
deg f = k > 1. Then, for large values of x, f(x) behaves like x
k
. So,
trying to ﬁnd the right convergent sequence, we could try ﬁrst to ”think
big”: we have x
k
n

= 2
n
, that is for large n, x
n
behaves like 2
n
k
. Then,
a good possibly convergent sequence could be x
n+k
− 2x
n
. Now, the
hard part: proving that this sequence is indeed convergent. First, we
will show that
x
n+k
x
n
converges to 2. This is easy, since the relation
239
f(x
n+k
) = 2
k
f(x
n
) implies
f(x
n+k
)
x
k
n+k
_
x
n+k
x
n
_
k
= 2
k

f(x
n
)
x
k
n
and since
lim
x→∞
f(x)
x
k
= 1 and lim
n→∞
x
n
= ∞
(do you see why?), we ﬁnd that indeed
lim
n→∞
x
n+k
x
n
= 2.
We will see that this will help us a lot. Indeed, let us write
f(x) = x
k
+
k−1

i=0
a
i
x
i
.
Then f(x
n+k
) = 2
k
f(x
n
) can be also written
x
n+k
−2x
n
=
k−1

i=0
a
i
(2
k
x
i
n
−x
i
n+k
)
k−1

i=0
(2x
n
)
i
x
k−i−1
n+k
But from the fact that lim
n→∞
x
n+k
x
n
= 2. it follows that the right-hand
side in the above relation is also convergent. So, (x
n+k
−2x
n
)
n≥1
is con-
vergent and it follows that there exist M, N such that for all n ≥ M we
have x
n+k
= 2x
n
+N. But now the problem is almost done, since the last
result combined with f(x
n+k
) = 2
k
f(x
n
) yields f(2x
n
+ N) = 2
k
f(x
n
)
for n ≥ M, that is f(2x + N) = 2
k
f(x). So, an arithmetical property
of the polynomial turned into an algebraic one using analysis. This al-
gebraic property helps us to immediately solve the problem. Indeed, we
see that if z is a complex root of f, then 2z + N, 4z + 3N, 8z + 7N, . . .
are all roots of f. Since f is non-zero, this sequence must be ﬁnite and
this can happen only for z = −N. Since −N is the only root of f, we
deduce that f(x) = (x + N)
k
. But since the equation f(x) = 2
2k+1
has
240
natural roots, we ﬁnd that 2
1
k
∈ N, which implies, contradiction. Thus,
our assumption was wrong and deg f = 1.
The idea of the following problem is so beautiful, that after reading
the solution the reader will have the impression that the problem is
trivial. Wrong! The problem is really diﬃcult and to make again an
solution. He will see the diﬃculty.
Example 9. Let π(n) be the number of prime numbers smaller than
or equal to n. Prove that there exist inﬁnitely many numbers n such that
π(n)[n.
AMM
Solution. First, let us prove the following result, which is the key
of the problem.
Lemma. For any increasing sequence of positive integers (a
n
)
n≥1
such that lim
n→∞
a
n
n
= 0, the sequence
_
n
a
n
_
n≥1
contains all natural num-
bers. In particular, for inﬁnitely many n we have that n divides a
n
.
Proof. Even if it seems unbelievable, this is true and moreover the
proof is extremely short. Let m ≥ 1 be a natural number. Consider the
set A =
_
n ≥ 1[
a
mn
mn

1
m
_
. This set contains and it is bounded, since
lim
n→∞
a
mn
mn
= 0. Thus it has a maximal element k. If
a
mk
mk
=
1
m
, then
m is in the sequence
_
n
a
n
_
n≥1
. Otherwise, we have a
m(k+1)
≥ a
mk

k + 1, which shows that k + 1 is also in the set, contradiction with the
maximality of k. The lemma is proved.
Thus, all we need to show is that lim
n→∞
π(n)
n
= 0. Fortunately, this is
well-known and not diﬃcult to prove. There are easier proofs than the
following one, but we prefer to deduce it from a famous and beautiful
result of Erdos.
241
Erdos’s theorem. We have

p≤n
p prime
p ≤ 4
n
.
The proof of this result is magniﬁcient. The proof is by induction.
For small values of n it is clear. Now, assume the inequality true for all
values smaller than n and let us prove that

p≤n
p prime
p ≤ 4
n
. If nis even,
we have nothing to prove, since

p≤n
p prime
=

p≤n−1
p prime
p ≤ 4
n−1
< 4
n
.
Now, assume that n = 2k + 1 and consider the binomial coeﬃcient
_
2k + 1
k
_
=
(k + 2) . . . (2k + 1)
k!
.
A simple application of the fact that
2
2k+1
=

i≥0
_
2k + 1
i
_
shows that
_
2k + 1
k
_
≤ 4
k
.
Thus, using the inductive hypothesis, we ﬁnd that

p≤n
p prime
p ≤

p≤k+1
p prime
p

k+2≤p≤2k+1
p prime
p ≤ 4
k+1
4
k
= 4
n
.
Now, the fact that lim
n→∞
π(n)
n
= 0 is trivial. Indeed, ﬁx k ≥ 1. We
have for all large n the inequality
nlg 4 ≥

k≤p≤n
p prime
lg p ≥ lg k(π(n) −π(k)),
which shows that
π(n) ≤
π(k)
n
+
lg 4
lg k
.
242
This shows of course that lim
n→∞
π(n)
n
= 0. The problem is solved.
It is time now for the last problem, which is, as usual, very hard. We
don’t exaggerate if we say that the following problem is exceptionally
diﬃcult.
Example 10. Let a, b > 1 be natural numbers such that for any
natural number n, a
n
−1[b
n
−1. Prove that b is a natural power of a.
Marius Cavachi, AMM
Solution. This time we will be able to ﬁnd the right convergent
sequence only after some double recurrences. Let us see. So, initially
we are given that there exists a sequence of positive integers (x
1
n
)
n≥1
such that x
1
n
=
b
n
−1
a
n
−1
Then, x
1
n

=
_
b
a
_
n
for large values of n. So, we
could expect that the sequence (x
2
n
)
n≥1
, x
2
n
= bx
1
n
−ax
2
n+1
is convergent.
Unfortunately,
x
2
n
=
b
n+1
(a −1) −a
n+1
(b −1) +a −b
(a
n
−1)(a
n+1
−1)
,
which is not necessarily convergent. But... if we look again at this
sequence, we see that for large values of n it grows like
_
b
a
2
_
n
, so
much slower. And this is the good idea: repeat this procedure until
the ﬁnal sequence behaves like
_
b
a
k+1
_
n
, where k is chosen such that
a
k
≤ b < a
k+1
. Thus, the ﬁnal sequence will converge to 0. Again,
the hard part has just begun, since we have to prove that if we deﬁne
x
i+1
n
= bx
i
n
−a
i
x
i
n+1
then lim
n→∞
x
k+1
n
= 0. This isn’t easy at all. The idea
is to compute x
3
n
and after that to prove the following statement: for
any i ≥ 1 the sequence (x
i
n
)
n≥1
has the form
c
i
b
n
+c
i−1
a
(i−1)n
+ +c
1
a
n
+c
0
(a
n+i−1
−1)(a
n+i−2
−1) . . . (a
n
−1)
for some constants c
0
, c
1
, . . . , c
i
. Proving this is not so hard, the hard part
was to think about it. How can we prove the statement otherwise than by
243
induction? And induction turns out to be quite easy. Supposing that the
statement is true for i, then the corresponding statement for i+1 follows
from x
i+1
n
= bx
i
n
−a
i
x
i
n+1
directly (note that to make the diﬀerence, we
just have to multiply the numerator c
i
b
n
+c
i−1
a
(i−1)n
+ +c
1
a
n
+c
0
with
b and a
n+i
−1. Then, we proceed in the same way with the second fraction
and the term b
n+1
a
n+i
will vanish). So, we have found a formula which
shows that as soon as a
i
> b we have lim
n→∞
x
i
n
= 0. So, we have deduced
that lim
n→∞
x
k+1
n
= 0. Another step of the solution is to take the minimal
index j such that lim
n→∞
x
j
n
= 0. Obviously, j > 1 and the recurrence
relation x
i+1
n
= bx
i
n
−a
i
x
i
n+1
shows that x
i
n
∈ Z for all n, i. Thus, there
exists M such that whenever n ≥ M we have x
j
n
= 0. This is the same
as bx
j−1
n
= a
j
x
j−1
n+1
for all n ≥ M, which implies x
j−1
n
=
_
b
a
j
_
n−M
x
j−1
M
for all n ≥ M. Let us suppose that b is not a multiple of a. Since
_
b
a
j
_
n−M
x
j−1
M
∈ Z for all n ≥ M, we must have x
j−1
M
= 0 and so
x
j−1
n
= 0 for n ≥ M, which means lim
n→∞
x
j
n
the minimality of j. Since we have reached a contradiction, we must
have a[b. Let us write b = ca. Then, the relation a
n
− 1[b
n
− 1 implies
a
n
− 1[c
n
− 1. And now are ﬁnally done. Why? We have just seen that
a
n
− 1[c
n
− 1 for all n ≥ 1. But our previous argument applied for c
instead of b shows that a[c. Thus, c = ad and we deduce again that a[d.
Since this process cannot be inﬁnite, b must be a power of a.
It worth saying that there exist an even stronger result: it is enough
to suppose that a
n
− 1[b
n
− 1 for an inﬁnite number n, but this is a
much more diﬃcult problem. It follows from a result found by Bugeaud,
Corvaja and Zannier in 2003:
If a, b > 1 are multiplicatively independent in Q

(that is log
a
b ,∈ Q),
then for any ε > 0 there exists n
0
= n
0
(a, b, ε) such that gcd(a
n
−1, b
n

244
1) < 2
εn
for all n ≥ n
0
. Unfortunately, the proof is too advanced to be
presented here.
Problems for training
1. Let f ∈ Z[X] be a polynomial of degree k such that for all n ∈ N
we have
k
_
f(n) ∈ Z. Prove that there exists integer numbers a, b such
that f(x) = (ax +b)
k
.
2. Find all arithmetic progressions of positive integers (a
n
)
n≥1
such
that for all n ≥ 1 the number a
1
+a
2
+ +a
n
is a perfect square.
3. Let p be a polynomial with integer coeﬃcients such that there
exists a sequence of pair wise distinct positive integers (a
n
)
n≥1
such
that p(a
1
) = 0, p(a
2
) = a
1
, p(a
3
) = a
2
, . . . . Find the degree of this
polynomial.
Tournament of the Towns, 2003
4. Let f, g : N

→ N

two functions such that [f(n) −n[ ≤ 2004

n
and n
2
+ g
2
(n) = 2f
2
(n). Prove that if f or g is surjective, then these
functions have inﬁnitely many ﬁxed points.
Gabriel Dospinescu, Moldova TST 2004
5. Let a, b be natural numbers such that for any natural number n,
the decimal representation of a +bn contains a sequence of consecutive
digits which form the decimal representation of n (for example, if a =
600, b = 35, n = 16 we have 600 + 16 35 = 1160). Prove that b is a
power of 10.
Tournament of the Towns, 2002
6. Let a, b > 1 be positive integers. Prove that for any given k > 0
there are inﬁnitely many numbers n such that ϕ(an +b) < kn, where ϕ
is the Euler totient function.
245
Gabriel Dospinescu
7. Let b an integer at least equal to 5 and deﬁne the number
x
n
= 11 . . . 1
. ¸¸ .
n−1
22 . . . 2
. ¸¸ .
n
5 in base b. Prove that x
n
is a perfect square for
all suﬃciently large n if and only if b = 10.
Laurentiu Panaitopol, IMO Shortlist 2004
8. Find all triplets of integer numbers a, b, c such that for any positive
integer n, a 2
n
+b is a divisor of c
n
+ 1.
Gabriel Dospinescu
11. Suppose that a is a real number such that all numbers
1
a
, 2
a
, 3
a
, . . . are integers. Then prove that a is also integer.
Putnam
12. Find all complex polynomials f having the property: there exists
a ≥ 2 a natural number such that for all suﬃciently large n, the equation
f(x) = a
n
2
has at least a positive rational solution.
Gabriel Dospinescu, Revue de Mathematiques Speciales
13. Let f be a complex polynomial having the property that for all
natural number n, the equation f(x) = n has at least a rational solution.
Then f has degree at most 1.
14. Let A be a set of natural numbers, which contains at least one
number among any 2006 consecutive natural numbers and let f a non-
constant polynomial with integer coeﬃcients. Prove that there exists a
number N such that for any n ≥ N there are at least

ln ln n diﬀerent
prime numbers dividing the number

N≤k≤n
k∈A
f(k).
Gabriel Dospinescu
246
15. Prove that in any strictly increasing sequence of positive integers
(a
n
)
n≥1
which satisﬁes a
n
< 100n for all n, one can ﬁnd inﬁnitely many
terms containing at least 1986 consecutive 1.
Kvant
16. Any inﬁnite arithmetical progression contains inﬁnitely many
terms that are not powers of integers.
17. Find all a, b, c such that for all suﬃciently large n, the number
a 4
n
+b 6
n
+c 9
n
is a perfect square.
18. Let f, g two real polynomials of degree 2 such that for any real
x, if f(x) is integer, so is g(x). Then there are integers m, n such that
g(x) = mf(x) +n for all x.
19. Try to generalize the preceding problem (this is for the die-
hards!!!).
20. Find all pairs of natural numbers a, b such that for every positive
integer n the number an +b is triangular if and only if n is triangular.
After a Putnam problem
21. Let (a
n
)
n≥1
be an inﬁnite and strictly increasing sequence of
positive integers such that for all n ≥ 2002, a
n
[a
1
+ a
2
+ + a
n−1
.
Prove that there exists n
0
such that for all n ≥ n
0
we have a
n
= a
1
+
a
2
+ +a
n−1
.
Tournament of the Towns, 2002
22. Find all real polynomials such that the image of any repunit is
also a repunit.
After a problem from Kvant
247
23. Fie doua multimi ﬁnite de numere reale pozitive cu proprietatea
ca
_

x∈A
x
n
[ n ∈ R
_

_

x∈B
x
n
[ n ∈ R
_
.
Sa se arate ca exista k ∈ R astfel incat A = ¦x
k
[ x ∈ B¦.
Gabriel Dospinescu
248
For an odd prime p, deﬁne the function
_
a
p
_
: Z → ¦−1, 1¦ by
_
a
p
_
= 1 if the equation x
2
= a has at least a solution in Z
p
and, other-
wise,
_
a
p
_
= −1. In the ﬁrst case, we say that a is a quadratic residue
modulo p, otherwise we say that it is a non quadratic residue modulo
p. This function is called Legendre’s symbol and plays a fundamental
role in number theory. Perhaps the most remarkable result involving
this symbol is Gauss’s quadratic reciprocity law. This states that for
diﬀerent odd prime numbers p, q the following equality holds:
_
p
q
__
q
p
_
= (−1)
p−1
2

q−1
2
.
This is a nontrivial result whose proof will be sketched later. Until
then, we will unfold some easier properties of Legendre’s symbol. First,
let us present an useful theoretical (but not practical at all) way of
computing
_
a
p
_
due to Euler.
Theorem. The following identity is true:
_
a
p
_
= a
p−1
2
(mod p).
We will prove this result and many other simple remarks concerning
quadratic residues in what follows. First, let us assume that
_
a
p
_
= 1
and consider x a solution of the equation x
2
= a in Z
p
. Using Fermat’s
theorem, we ﬁnd that a
p−1
2
= x
p−1
= 1 (mod p). Thus the equality
_
a
p
_
= a
p−1
2
(mod p) holds for all quadratic residues a modulo p. In
p−1
2
= 1 (mod p). Now, we
will prove that there are exactly
p −1
2
p
¸ ¦0¦.
This will enable us to conclude that quadratic residues are precisely the
249
roots of the polynomial X
p−1
2
− 1 and also that non quadratic residues
are exactly the roots of the polynomial X
p−1
2
+ 1 (from Fermat’s little
theorem). Note that Fermat’s little theorem implies that the polynomial
X
p−1
− 1 = (X
p−1
2
− 1)(X
p−1
2
+ 1) has exactly p − 1 roots in the ﬁeld
Z
p
. But in a ﬁeld, the number of diﬀerent zeros of a polynomial cannot
exceed its degree. Thus each of the polynomials X
p−1
2
−1 and X
p−1
2
+1
has at most
p −1
2
zeros in Z
p
. These two observations show that in fact
each of these polynomials has exactly
p −1
2
zeros in Z
p
. Let us observe
next that there are at least
p −1
2
all numbers i
2
(mod p) with 1 ≤ i ≤
p −1
2
they are all diﬀerent. This shows that there are exactly
p −1
2
residues in Z
p
¸ ¦0¦ and also proves Euler’s criterion.
We have said that Euler’s criterion is a very useful result. Indeed, it
allows a very quick proof of the fact that
_
a
p
_
: Z → ¦−1, 1¦ is a group
morphism. Indeed, we have
_
ab
p
_
= (ab)
p−1
2
(mod p) = a
p−1
2
b
p−1
2
(mod p) =
_
a
p
__
b
p
_
.
The relation
_
ab
p
_
=
_
a
p
__
b
p
_
shows that while studying Le-
gendre’s symbol, it suﬃces to focus on the prime numbers only. Also,
the same Euler’s criterion implies that
_
a
p
_
=
_
b
p
_
whenever a ≡ b
(mod p).
It is now time to come back to Gauss’s celebrated quadratic reci-
procity law. First of all, we will prove a lemma (due to Gauss).
Lemma. Let p be an odd prime and let a ∈ Z such that gcd(a, p) = 1.
If m is the number of positive integers x such that x <
p
2
and
p
2
< ax
(mod p) < p, then
_
a
p
_
= (−1)
m
.
250
Proof. Let x
1
, x
2
, . . . , x
m
be those numbers x for which x <
p
2
and
p
2
< ax (mod p) < p. Let k =
p −2
2
− m and y
1
, . . . , y
k
all numbers
smaller than
p
2
, diﬀerent from x
1
, x
2
, . . . , x
m
.
Observe that
p−1
2

x=1
(ax) = a
p−1
2
_
p −1
2
_
! ≡
_
a
p
__
p −1
2
_
! (mod p).
On the other hand,
p−1
2

x=1
(ax) =

ax (mod p)>
p
2
(ax) (mod p)

ax (mod p)<
p
2
(ax) (mod p).
We clearly have

ax (mod p)>
p
2
(ax) (mod p)

ax (mod p)<
p
2
(ax) (mod p)
=
m

i=1
ax
i
(mod p)
k

j=1
ay
i
(mod p).
On the other hand, the numbers p − ax
i
(mod p) and ay
i
(mod p)
give a partition of 1, 2, . . . ,
p −1
2
(mod p). Indeed, it suﬃces to prove
that p − ax
i
(mod p) ,= ay
j
(mod p), which is clearly true by the deﬁ-
nition of x
i
, y
j
<
p
2
. Hence we can write
m

i=1
ax
i
(mod p)
k

j=1
ay
i
(mod p)
= (−1)
m
m

i=1
(p −ax
i
(mod p))
k

j=1
ay
j
(mod p)
≡ (−1)
m
p−1
2

i=1
i (mod p) = (−1)
m
_
p −1
2
_
! (mod p).
Combining these facts, we ﬁnally deduce that
_
a
p
_
= (−1)
m
.
251
Using Gauss’s lemma, the reader will enjoy proving the next two
classical results.
Theorem. The identity
_
2
p
_
= (−1)
p
2
−1
8
holds for any odd prime
number p.
Theorem. (quadratic reciprocity law) For any diﬀerent odd primes
p, q, the following identity holds:
_
p
q
__
q
p
_
= (−1)
p−1
2

q−1
2
.
Using this powerful arsenal, we are now able to solve some interest-
ing problems. Most of them are merely direct applications of the above
results, but we think that they are still worthy not necessarily because
they appeared in various contests.
Example 1. Prove that the number 2
n
+ 1 does not have prime
divisors of the form 8k −1.
Vietnam TST 2004
Solution. Indeed, assume that p is a prime divisor of the form 8k−1
that divides 2
n
+ 1. Of course, if n is even, the contradiction is im-
mediate, since in this case we would have −1 ≡ (2
n
2
)
2
(mod p) and
so −1 = (−1)
p−1
2
=
_
−1
p
_
= 1. Now, assume that n is odd. Then
−2 ≡ (2
n+1
2
)
2
(mod p) and so
_
−2
p
_
= 1. This can be also written in
the form
_
−1
p
__
2
p
_
= 1, or (−1)
p−1
2
+
p
2
−1
8
= 1. Fortunately, if p is of
the form 8k −1 the later cannot hold and this is the contradiction that
solves the problem.
Based on the same idea and with a bit more work, we arrive at the
following result.
Example 2. Prove that for any positive integer n, the number 2
3
n
+1
has at least n prime divisors of the form 8k + 3.
252
Gabriel Dospinescu
Solution. Using the result of the previous problem, we deduce that
2
n
+ 1 does not have prime divisors of the form 8k + 7. We will prove
that if n is odd, then it has no prime divisors of the form 8k +5 either.
Indeed, let p be a prime divisor of 2
n
+ 1. Then 2
n
≡ −1 (mod p) and
so −2 ≡ (2
n+1
2
)
2
(mod p). Using the same argument as the one in the
previous problem, we deduce that
p
2
−1
8
+
p −1
2
is even, which cannot
happen if p is of the form 8k + 5.
Now, let us solve the proposed problem. We will assume n > 2
(otherwise the veriﬁcation is trivial). The essential observation is the
identity:
2
3
n
+ 1 = (2 + 1)(2
2
−2 + 1)(2
23
−2
3
+ 1) . . . (2
23
n−1
−2
3
n−1
+ 1)
Now, we will prove that for all 1 ≤ i < j ≤ n − 1, gcd(2
23
i
− 2
3
i
+
1, 2
23
j
−2
3
j
+1) = 3. Indeed, assume that p is a prime number dividing
gcd(2
23
i
−2
3
i
+ 1, 2
23
j
−2
3
j
+ 1) We will then have p[2
3
i+1
+ 1. Thus,
2
3
j
≡ (2
3
i+1
)
3
j−i−1
≡ (−1)
3
j−i−1
≡ −1 (mod p),
implying
0 ≡ 2
23
j
−2
3
j
+ 1 ≡ 1 −(−1) + 1 ≡ 3 (mod p).
This cannot happen unless p = 3. But since v
3
(gcd(2
23
i
− 2
3
i
+
1, 2
23
j
−2
3
j
+ 1)) = 1 (as one can immediately check), it follows that
gcd(2
23
i
−2
3
i
+ 1, 2
23
j
−2
3
j
+ 1) = 3
and the claim is proved.
It remains to show that each of the numbers 2
23
i
− 2
3
i
+ 1, with
1 ≤ i ≤ n − 1 has at least a prime divisor of the form 8k + 3 diﬀerent
from 3. It would follow in this case that 2
3
n
+1 has at least n−1 distinct
prime divisors of the form 8k+3 (from the previous remarks) and since it
253
is also divisible by 3, the conclusion would follow. Fix i ∈ ¦1, 2, . . . , n−1¦
and observe that any prime factor of 2
23
i
−2
3
i
+1 is also a prime factor
of 2
3
n
+1 and thus, from the ﬁrst remark, it must be of the form 8k +1
or 8k + 3. Because v
3
(2
23
i
− 2
3
i
+ 1) = 1, it follows that if all prime
divisors of 2
23
i
− 2
3
i
+ 1 except for 3 are of the form 8k + 1, then
2
23
i
− 2
3
i
+ 1 ≡ 8 (mod 8), which is clearly impossible. Thus at least
a prime divisor of 2
23
i
− 2
3
i
+ 1 is diﬀerent from 3 and is of the form
8k + 3 and so the claim is proved. The conclusion follows.
At ﬁrst glance, the following problem seems trivial. Far from being
true! It is actually very tricky, because brute force will take us nowhere.
In the framework of the above results, this should not be so diﬃcult
now.
Example 3. Find a number n between 100 and 1997 such that
n[2
n
+ 2.
APMO, 1997
Solution. If we search for odd numbers, then we will certainly fail
(actually, this result due to Sierpinski has been proved in the topic ”Look
at the exponent!”). So let us search for even numbers. The ﬁrst step
would be choosing n = 2p, for some prime number p. Unfortunately this
cannot work by Fermat’s little theorem. So let us try setting n = 2pq,
with p, q diﬀerent prime numbers. We need pq[2
2pq−1
+ 1 and so we
must have
_
−2
p
_
=
_
−2
q
_
= 1. Also, using Fermat’s little theorem,
p[2
2q−1
+ 1 and q[2
2p−1
+ 1. A small veriﬁcation shows that q = 3, 5, 7
are not good choices, so let us try q = 11. In this case we ﬁnd p = 43
and so it suﬃces to show that pq[2
2pq−1
+ 1 for q = 11 and p = 43.
This is immediate, since the hard work has already been completed: we
have shown that it suﬃces to have p[q
2q−1
, q[2
2p−1
+ 1, and
_
−2
p
_
=
254
_
−2
q
_
= 1 in order to have pq[2
2pq−1
+1. But as one can easily check, all
these conditions are veriﬁed and the number 2 11 43 is a valid answer.
Were we wrong when choosing to present the following example? It
apparently has no connection with quadratic reciprocity, but let us take
a closer look.
Example 4. Let f, g : Z
+
→ Z
+
functions with the properties:
i) g is surjective;
ii) 2f
2
(n) = n
2
+g
2
(n) for all positive integers n.
If, moreover, [f(n)−n[ ≤ 2004

n for all n, prove that f has inﬁnitely
many ﬁxed points.
Gabriel Dospinescu, Moldova TST, 2005
Solution. Let p
n
be the sequence of prime numbers of the form
8k + 3 (the fact that there are inﬁnitely many such numbers is a trivial
consequence of Dirichlet’s theorem, but we invite the reader to ﬁnd an
elementary proof). It is obvious that for all n we have
_
2
p
n
_
= (−1)
p
2
n
−1
8
= −1.
Using the condition i) we can ﬁnd x
n
such that g(x
n
) = p
n
for all n. It
follows that 2f
2
(x
n
) = x
2
n
+p
2
n
, which can be rewritten as 2f
2
(x
n
) ≡ x
2
n
(mod p
n
). Because
_
2
p
n
_
= −1, the last congruence shows that p
n
[x
n
and p
n
[f(x
n
). Thus there exist sequences of positive integers a
n
, b
n
such
that x
n
= a
n
p
n
, f(x
n
) = b
n
p
n
for all n. Clearly, ii) implies the relation
2b
2
n
= a
2
n
+ 1. Finally, using the property [f(n) −n[ ≤ 2004

n we infer
that
2004

x
n

¸
¸
¸
¸
f(x
n
)
x
n
−1
¸
¸
¸
¸
=
¸
¸
¸
¸
b
n
a
n
−1
¸
¸
¸
¸
,
that is
lim
n→∞
_
a
2
n
+ 1
a
n
=

2.
255
The last relation immediately implies that lim
n→∞
a
n
= 1. Therefore,
starting from a certain rank, we have a
n
= 1 = b
n
that is f(p
n
) = p
n
.
The conclusion now follows.
We continue with a diﬃcult classical result, that often proves very
useful. It characterizes numbers that are quadratic residues modulo all
suﬃciently large prime numbers. Of course, perfect square are such num-
bers, but how to prove that they are the only ones?
Example 5. Suppose that a ∈ N is not a perfect square. Then
_
a
p
_
= −1 for inﬁnitely many prime numbers p.
Solution. One may assume that a is square-free. Let us write a =
2
e
q
1
q
2
. . . q
n
, where q
i
are diﬀerent odd primes and e ∈ ¦0, 1¦. Let us
assume ﬁrst that n ≥ 1 and consider some odd distinct primes r
1
, . . . , r
k
each of them diﬀerent from q
1
, . . . , q
n
. We will show that there exists a
prime p, diﬀerent from r
1
, . . . , r
k
, such that
_
a
p
_
= −1. Let s be a non
n
.
Using the Chinese remainder theorem, we can ﬁnd a positive integer
b such that
_
¸
¸
¸
¸
¸
_
¸
¸
¸
¸
¸
_
b ≡ 1 (mod r
i
), 1 ≤ i ≤ k
b ≡ 1 (mod 8),
b ≡ 1q
i
, 1 ≤ i ≤ n −1
b ≡ s (mod q
n
)
Now, write b = p
1
. . . p
m
with p
i
odd primes, not necessarily distinct.
Using the quadratic reciprocity law, it follows immediately that
m

i=1
_
2
p
i
_
=
m

i=1
(−1)
p
2
i
−1
8
= (−1)
b
2
−1
8
= 1
and
m

j=1
_
q
i
p
j
_
=
m

j=1
(−1)
p
j
−1
2

q
i
−1
2
_
p
j
q
i
_
= (−1)
q
i
−1
2

b−1
2
_
b
q
i
_
=
_
b
q
i
_
256
for all i ∈ ¦1, 2, . . . , n¦. Hence
m

i=1
_
a
p
i
_
=
_
_
m

j=1
_
2
p
j
_
_
_
2
n

i=1
m

j=1
_
q
i
p
j
_
=
n

i=1
_
b
q
i
_
=
_
b
q
n
_
=
_
s
q
n
_
= −1.
Thus, there exists i ∈ ¦1, 2, . . . , m¦ such that
_
a
p
i
_
= −1. Because
b ≡ 1 (mod r
i
), 1 ≤ i ≤ k we also have p
i
∈ ¦1, 2, . . . ¦ ¸ ¦r
1
, . . . , r
k
¦ and
the claim is proved.
The only remaining case is a = 2. But this one is very simple, since it
suﬃces to use Dirichlet’s theorem to ﬁnd inﬁnitely many primes p such
that
p
2
−1
8
is odd.
As in other units, we will now focus on a special case. This time it is
a problem almost trivial in the above framework and almost impossible
to solve otherwise (we say almost because there is a beautiful, but very
diﬃcult solution using analytical tools, that we will not present here).
Example 6. Suppose that a
1
, a
2
, . . . , a
2004
are nonnegative integers
such that a
n
1
+a
n
2
+ +a
n
2004
is a perfect square for all positive integers
n. What is the minimal number of such integers that must equal 0?
Solution. Suppose that a
1
, a
2
, . . . , a
k
are positive integers such that
a
n
1
+ a
n
2
+ + a
n
k
is a perfect square for all n. We will show that k
is a perfect square. In order to prove this, we will use the above result
and show that
_
k
p
_
= 1 for all suﬃciently large prime p. This is not a
diﬃcult task. Indeed, consider a prime p, greater than any prime divisor
of a
1
a
2
. . . a
k
. Using Fermat’s little theorem, a
p−1
1
+a
p−1
2
+ +a
p−1
k
≡ k
(mod p), and since a
p−1
1
+a
p−1
2
+ +a
p−1
k
is a perfect square, it follows
that
_
k
p
_
= 1. Thus k is a perfect square. And now the problem becomes
257
trivial, since we must ﬁnd the greatest perfect square smaller than 2004.
A quick computation shows that this is 44
2
= 1936 and so the desired
minimal number is 68.
Here is another nice application of this idea. The following example
Example 7. Suppose that f ∈ Z[X] is a second degree polynomial
such that for any prime p there exists at least an integer n for which
p[f(n). Prove that f has rational zeros.
Solution. Let f(x) = ax
2
+ bx + c be this polynomial. It suﬃces
of course to prove that b
2
− 4ac is a perfect square. This boils down
to proving that it is a quadratic residue modulo any suﬃciently large
prime. Pick a prime number p and an integer n such that p[f(n). Then
b
2
−4ac ≡ (2an +b)
2
(mod p)
and so
_
b
2
−4ac
p
_
= 1.
This shows that our claim is true and ﬁnishes the solution.
Some of the properties of Legendre’s symbol can be also found in
the following problem.
Example 8. Let p be an odd prime and let
f(x) =
p−1

i=1
_
i
p
_
X
i−1
.
a) Prove that f is divisible by X−1 but not by (X−1)
2
if and only
if p ≡ 3 (mod 4);
b) Prove that if p ≡ 5 (mod 8) then f is divisible by (X − 1)
2
and
not by (X −1)
3
.
Romanian TST, 2004
258
Solution. The ﬁrst question is not diﬃcult at all. Observe that
f(1) =
p−1

i=1
_
i
p
_
= 0
by the simple fact that there are exactly
p −1
2
dratic residues in ¦1, 2, . . . p −1¦. Also,
f
t
(1) =
p−1

i=1
(i −1)
_
i
p
_
=
p−1

i=1
i
_
i
p
_
,
because f(1) = 0. The same idea of summing up in reversed order allows
us to write:
p−1

i=1
i
_
i
p
_
=
p−1

i=1
(p −i)
_
p −i
p
_
= (−1)
p−1
2
p−1

i=1
2(p −i)
_
i
p
_
= −(−1)
p−1
2
f
t
(1)
(we used again the fact that f(1) = 0).
Hence for p ≡ 1 (mod 4) we must also have f
t
(1) = 0. In this case
f is divisible by (X −1)
2
. On the other hand, if p ≡ 3 (mod 4), then
f
t
(1) =
p−1

i=1
i
_
i
p
_

p−1

i=1
i =
p(p −1)
2
≡ 1 (mod 2)
and so f is divisible by X −1 but not by (X −1)
2
.
The second question is much more technical, even though it uses the
same main idea. Observe that
f
tt
(1) =
p−1

i=1
(i
2
−3i + 2)
_
i
p
_
=
p−1

i=1
i
2
_
i
p
_
−3
p−1

i=1
i
_
i
p
_
(once again we use the fact that f(1) = 0). Observe that the condition
p ≡ 5 (mod 8) implies, by a), that f is divisible by (X −1)
2
so actually
f
tt
(1) =
p−1

i=1
i
2
_
i
p
_
.
259
Let us break this sum into two pieces and treat each of them inde-
pendently. Let us deal with
p−1
2

i=1
(2i)
2
_
2i
p
_
= 4
_
2
p
_
p−1
2

i=1
i
2
_
i
p
_
.
Note that
p−1
2

i=1
i
2
_
i
p
_

p−1
2

i=1
i
2

p−1
2

i=1
i =
p
2
−1
8
≡ 1 (mod 2),
so
p−1
2

i=1
(2i)
2
_
2i
p
_
≡ ± (mod 8)
(actually, using the fact that
_
2
p
_
= (−1)
p
2
−1
8
, we obtain that its value
is −4). On the other hand,
p−1
2

i=1
(2i −1)
2
_
2i −1
p
_

p−1
2

i=1
_
2i −1
p
_
(mod 8).
If we prove that the last quantity is a multiple of 8, then the problem
will be solved. But note that f(1) = 0 implies
0 =
p−1
2

i=1
_
2i
p
_
+
p−1
2

i=1
_
2i −1
p
_
.
Also,
p−1
2

i=1
_
2i
p
_
= 1 +
p−3
2

i=1
_
2i
p
_
= 1 +
p−3
2

i=1
_
_
_
_
2
_
p −1
2
−i
_
p
_
_
_
_
= 1 +
p−3
2

i=1
_
2i + 1
p
_
=
p−1
2

i=1
_
2i −1
p
_
.
Therefore
p−1
2

i=1
_
2i −1
p
_
= 0 and the problem is ﬁnally solved.
260
Finally, a diﬃcult problem.
Example 9. Find all positive integers n such that 2
n
−1[3
n
−1.
Solution. We will prove that n = 1 is the only solution to the
problem. Suppose that n > 1 is a solution. Then 2
n
− 1 cannot be a
multiple of 3, hence n is odd. Therefore, 2
n
≡ 8 (mod 12). Because any
odd prime diﬀerent from 3 is of one of the forms 12k±1, 12k±5 and since
2
n
−1 ≡ 7 (mod 12), it follows that 2
n
−1 has at least a prime divisor
of the form 12k ± 5, call it p. Obviously, we must have
_
3
p
_
= 1 and
using the quadratic reciprocity law, we ﬁnally obtain
_
p
3
_
= (−1)
p−1
2
.
On the other hand
_
p
3
_
=
_
±2
3
_
= −(±1). Consequently, −(±1) =
(−1)
p−1
2
= ±1, which is the desired contradiction. Therefore the only
solution is n = 1.
Problems for training
1. Prove that for any odd prime p, the smallest positive quadratic
non residue modulo p is smaller than 1 +

p.
2. Let p be a prime number. Prove that the following statements are
equivalent:
i) there is a positive integer n such that p[n
2
−n + 3;
ii) there is a positive integer m such that p[m
2
−m+ 25.
3. Let x
1
= 7 and let x
n+1
= 2x
2
n
− 1. Prove that 2003 does not
divide any term of the sequence.
4. Let p be a prime of the form 4k + 1. Compute
p−1

k=1
__
2k
2
p
_
−2
_
k
2
p
__
.
261
Korea TST 2000
5. Prove that the number 3
n
+2 does not have prime divisors of the
form 24k + 13.
Laurentiu Panaitopol, Gazeta Matematica
6. What is the number of solutions to the equation a
2
+ b
2
= 1 in
Z
p
Z
p
. What about the equation a
2
−b
2
= 1?
7. Suppose that p is an odd prime and that A, B are two diﬀerent
non empty subsets of ¦1, 2, . . . , p −1¦ for which
i) A∪ B = ¦1, 2, . . . , p −1¦;
ii) If a, b are in the same set among A, B, then ab (mod p) ∈ A;
iii) If a ∈ A, b ∈ B, then ab ∈ B.
Find all such subsets A, B.
India
8. Let a, b, c be positive integers such that b
2
− 4ac is not a perfect
square. Prove that for any n > 1 there are n consecutive positive integers,
none of which can be written in the form (ax
2
+ bxy + cy
2
)
z
for some
integers x, y and some positive integer z.
Gabriel Dospinescu
9. Let be integers relatively prime with an odd prime p. Prove that:
p−1

i=1
_
ai
2
+bi
p
_
= −
_
a
p
_
.
10. Compute
p−1

k=1
_
f(k)
p
_
, where f is a polynomial with integral
coeﬃcients and p is an odd prime.
11. Suppose that for a certain prime p, the values the polynomial
with integral coeﬃcients f(x) = ax
2
+bx+c takes at 2p −1 consecutive
integers are perfect squares. Prove that p[b
2
−4ac.
IMO Shortlist
262
12. Suppose that φ(5
m
− 1) = 5
n
− 1 for a pair (m, n) of positive
integers. Here φ is Euler’s totient function. Prove that gcd(m, n) > 1.
Taiwan TST
13. Let p be a prime of the form 4k+1 such that p
2
[2
p
−2. Prove that
the greatest prime divisor q of 2
p
−1 satisﬁes the inequality 2
q
> (6p)
p
.
Gabriel Dospinescu
263
SOLVING ELEMENTARY INEQUALITIES WITH
INTEGRALS
Why are integral pertinent for solving inequalities? Well, when we
say integral, we say in fact area. And area is a measurable concept, a
comparable one. That is why there are plenty of inequalities which can
be solved with integrals, some of them with a completely elementary
statement. They seem elementary, but sometimes ﬁnding elementary
solutions for them is a real challenge. Instead, there are beautiful and
short solutions using integrals. Of course, the hard part is to ﬁnd the
integral that hides after the elementary form of the inequality (and to
be sincere, the idea of using integrals to solve elementary inequalities
is practically inexistent in Olympiad books). First, let us state some
properties of integrals that we will use here.
1) For any integrable function f : [a, b] →R we have
_
b
a
f
2
(x)dx ≥ 0.
2) For any integrable functions f, g : [a, b] → R such that f ≤ g we
have
_
b
a
f(x)dx ≤
_
b
a
g(x)dx (monotony for integrals).
3) For any integrable functions f, g : [a, b] →R and any real numbers
α, β we have
_
b
a
(αf(x)+βg(x))dx = α
_
b
a
f(x)dx+β
_
b
a
g(x) (linearity of integrals).
Also, the well-known elementary inequalities of Cauchy-Schwarz,
Chebyshev, Minkowski, H¨older, Jensen, Young have corresponding in-
tegral inequalities, which are derived immediately from the algebraic
inequalities (indeed, one just have to apply the corresponding inequal-
ities for the numbers f
_
a +
k
n
(b −a)
_
, g
_
a +
k
n
(b −a)
_
, . . . with
264
k ∈ ¦1, 2, . . . , n¦ and to use the fact that
_
b
a
f(x)dx = lim
n→∞
b −a
n
n

k=1
f
_
a +
k
n
(b −a)
_
.
The reader will take a look at the glossary if he doesn’t manage to
state them.
It seems at ﬁrst glance that this is not a very intricate and diﬃcult
theory. Totally false! We will see how strong is this theory of integration
and especially how hard it is to look beneath the elementary surface
of a problem. To convince yourself of the strength of the integral, take
a look at the following beautiful proof of the AM-GM inequality using
integrals. This magniﬁcent proof was found by H. Alzer and published
in the American Mathematical Monthly.
Example 1. Prove that for any a
1
, a
2
, . . . , a
n
≥ 0 we have the in-
equality
a
1
+a
2
+ +a
n
n

n

a
1
a
2
. . . a
n
.
Solution. Let us suppose that a
1
≤ a
2
≤ ≤ a
n
and let
A =
a
1
+a
2
+ +a
n
n
, B =
n

a
1
a
2
. . . a
n
.
Of course, we can ﬁnd an index k ∈ ¦1, 2, . . . , n − 1¦ such that
a
k
≤ G ≤ a
k+1
. Then it is immediate to see that
A
G
−1 =
1
n
k

i=1
_
G
a
i
_
1
t

1
G
_
dt +
1
n
n

i=k+1
_
a
i
G
_
1
G

1
t
_
dt
and the last quantity is clearly nonnegative, since each integral is non-
negative.
Truly wonderful, isn’t it? So, after all, integrals are nice! This is also
conﬁrmed by the following problem, an absolute classic whose solution
by induction can be a real nightmare.
265
Example 2. Prove that for any real numbers a
1
, a
2
, . . . , a
n
the fol-
lowing inequality holds:
n

i=1
n

j=1
a
i
a
j
i +j
≥ 0.
Solution. Now, we will see how easy is this problem if we manage
to handle integrals and especially to see from where they come. The
essential suggestion is the observation that
a
i
a
j
i +j
=
_
1
0
a
i
a
j
t
i+j−1
dt.
And now the problem is solved. What follows are just formalities;
the hard part was translating the inequality. After that, we will decide
what is better to do. So,
n

i,j=1
a
i
a
j
i +j
≥ 0
is equivalent to
n

i,j=1
_
1
0
a
i
a
j
t
i+j−1
dt ≥ 0,
or, using the linearity of the integrals, to
_
1
0
_
_
n

i,j=1
a
i
a
j
t
i+j−1
_
_
dt ≥ 0.
This form suggests us that we should use the ﬁrst property, that is
we should ﬁnd an integrable function f such that
f
2
(t) =
n

i,j=1
a
i
a
j
t
i+j−1
dt.
This isn’t hard, because the formula
_
n

i=1
a
i
x
i
_
2
=
n

i,j=1
a
i
a
j
x
i
x
j
266
solves the task. We just have to take
f(x) =
n

i=1
a
i
x
i−
1
2
.
We continue the series of direct applications of classical integral in-
equalities with a problem proposed by Walther Janous and which may
also put serious problems if not attacked appropriately.
Example 3. Let t ≥ 0 and the sequence (x
n
)
n≥1
deﬁned by
x
n
=
1 +t + +t
n
n + 1
.
Prove that
x
1

x
2

3

x
3

4

x
4
≤ . . .
Walther Janous, Crux Mathematicorum
Solution. It is clear that for t > 1 we have
x
n
=
1
t −1
_
t
1
u
n
du
and for t < 1 we have
x
n
=
1
1 −t
_
t
1
u
n
du.
This is how the inequality to be proved reduces to the more general
inequality
k
¸
¸
¸
¸
_
_
b
a
f
k
(x)dx
b −a

k+1
¸
¸
¸
¸
_
_
b
a
f
k+1
(x)dx
b −a
for all k ≥ 1 and any nonnegative integrable function f : [a, b] → R.
And yes, this is a consequence of the Power Mean Inequality for integral
functions.
The following problem has a long and quite complicated proof by
induction. Yet, using integrals it becomes trivial.
267
Example 4. Prove that for any positive real numbers x, y and any
positive integers m, n
(n −1)(m−1)(x
m+n
+y
m+n
) + (m+n −1)(x
m
y
n
+x
n
y
m
)
≥ mn(x
m+n−1
y +y
m+n−1
x).
Austrian-Polish Competition ,1995
Solution. We transform the inequality as follows:
mn(x −y)(x
m+n−1
−y
m+n−1
) ≥ (m+n −1)(x
m
−y
m
)(x
n
−y
n
) ⇔
x
m+n−1
−y
m+n−1
(m+n −1)(x −y)

x
m
−y
m
m(x −y)

x
n
−y
n
n(x −y)
(we have assumed that x > y). The last relations can be immediately
translated with integrals in the form
(y −x)
_
x
y
t
m+n−2
dt ≥
_
x
y
t
m−1
dt
_
x
y
t
n−1
dt.
And this follows from the integral form of Chebyshev inequality.
A nice blending of arithmetic and geometric inequality as well as
integral calculus allows us to give a beautiful short proof of the following
inequality.
Example 5. Let x
1
, x
2
, . . . , x
k
be positive real numbers and m, n
positive real numbers such that n ≤ km. Prove that
m(x
n
1
+x
n
2
+ +x
n
k
−k) ≥ n(x
m
1
x
m
2
. . . x
m
k
−1).
IMO Shortlist 1985, proposed by Poland
Solution. Applying AM-GM inequality, we ﬁnd that
m(x
n
1
+ +x
n
k
−k) ≥ m(k
k
_
(x
1
x
2
. . . x
k
)
n
−k).
Let
P =
k

x
1
x
2
. . . x
k
.
268
We have to prove that
mkP
n
−mk ≥ nP
mk
−n,
which is the same as
P
n
−1
n

P
mk
−1
mk
.
This follows immediately from the fact that
P
x
−1
xln P
=
_
1
0
e
xt ln P
dt.
We have seen a rapid but diﬃcult proof for the following problem,
using the Cauchy-Schwarz inequality. Well, the problem originated by
playing around with integral inequalities and the following solution will
show how one can create diﬃcult problems starting from trivial ones.
Example 6. Prove that for any positive real numbers a, b, c such
that a +b +c = 1 we have
(ab +bc +ca)
_
a
b
2
+b
+
b
c
2
+c
+
c
a
2
+a
_

3
4
.
Gabriel Dospinescu
Solution. As in the previous problem, the most important aspect
is to translate the expression
a
b
2
+b
+
b
c
2
+c
+
c
a
2
+a
in the integral
language. Fortunately, this isn’t diﬃcult, since it is just
_
1
0
_
a
(x +b)
2
+
b
(x +c)
2
+
c
(x +a)
2
_
dx.
Now, using the Cauchy-Schwarz inequality, we infer that
a
(x +b)
2
+
b
(x +c)
2
+
c
(x +a)
2

_
a
x +b
+
b
x +c
+
c
a +x
_
2
.
Using again the same inequality, we minor
a
x +b
+
b
x +c
+
c
a +x
with
1
x +ab +bc +ca
. Consequently,
a
(x +b)
2
+
b
(x +c)
2
+
c
(x +a)
2

1
(x +ab +bc +ca)
2
269
and we can integrate this to ﬁnd that
a
b
2
+b
+
b
c
2
+c
+
c
a
2
+a

1
(ab +bc +ca)(ab +bc +ca + 1)
.
Now, all we have to do is to notice that
ab +bc +ca + 1 ≤
4
3
.
Now, another question for the interested reader: can we prove the
general case (solved in Cauchy Schwarz’s inequality topic) using integral
calculus? It seems a diﬃcult problem.
There is an important similarity between the following problem and
example 2, yet here it is much more diﬃcult to see the relation with
integral calculus.
Example 7. Let n ≥ 2 and S the set of the sequences
(a
1
, a
2
, . . . , a
n
) ⊂ [0, ∞) which verify
n

i=1
n

j=1
1 −a
i
a
j
1 +j
≥ 0.
Find the maximum value of the expression
n

i=1
n

j=1
a
i
+a
j
i +j
, over all
sequences from S.
Gabriel Dospinescu
Solution. Consider the function f : R →R, f(x) = a
1
+a
2
x+ +
a
n
x
n−1
. Let us observe that
n

i=1
n

j=1
a
i
a
j
i +j
=
n

i=1
a
i
_
_
n

j=1
a
j
i +j
_
_
=
n

i=1
a
i
_
1
0
x
i
f(x)dx
=
_
1
0
_
xf(x)
n

i=1
a
i
x
i−1
_
dx =
_
1
0
xf
2
(x)dx.
270
So, if we denote M =

1≤i,j≤n
1
i +j
, we infer that
_
1
0
xf
2
(x)dx ≤ M.
On the other hand, we have the identity
n

i=1
n

j=1
a
i
+a
j
i +j
= 2
_
a
1
2
+ +
a
n
n + 1
+ +
a
1
n + 1
+ +
a
n
2n
_
= 2
_
1
0
(x +x
2
+ +x
n
)f(x)dx.
This was the hard part: translating the properties of the sequences
in S and also the conclusion. Now, the problem becomes easy, since we
must ﬁnd the maximal value of
2
_
1
0
(x +x
2
+ +x
n
)f(x)dx
where
_
1
0
xf
2
(x)dx ≤ M.
Well, Cauchy-Schwarz inequality for integrals is the way to proceed.
Indeed, we have
__
1
0
(x +x
2
+ +x
n
)f(x)dx
_
2
=
__
1
0
_
xf
2
(x)
_
x(1 +x + +x
n−1
)
2
dx
_
2
=
_
1
0
xf
2
(x)dx
_
1
0
(1 +x + +x
n−1
)
2
dx ≤ M
2
.
This shows that
n

i=1
n

j=1
a
i
+a
j
i +j
≤ 2M and now the conclusion easily
follows: the maximal value is 2

1≤i,j≤n
1
i +j
, attained for a
1
= a
2
= =
a
n
= 1.
271
is a mathematical crime. It is time to say it again. This is why we
designed this topic, to present a new method of treating inequalities
involving fractions. Some relevant examples will be treated revealing
that bunching could be a great pain for the reader wanting to use it.
Example 8. Prove that for any positive real numbers a, b, c the
following inequality holds:
1
3a
+
1
3b
+
1
3c
+
3
a +b +c

1
2a +b
+
1
2b +a
+
1
2b +c
+
1
2c +b
+
1
2c +a
+
1
2a +c
.
Gabriel Dospinescu
Solution. Of course, the reader has noticed that this is stronger
than Popoviciu’s inequality, so it seems that classical methods will have
no chances. And what if we say that this is Schur’s inequality revisited?
Indeed, let us write Schur’s inequality in the form:
x
3
+y
3
+z
3
+ 3xyz ≥ x
2
y +y
2
x +y
2
z +z
2
y +z
2
x +x
2
z
where x = t
a−
1
3
, y = t
b−
1
3
, z = t
c−
1
3
and integrate the inequality as t
ranges between 0 and 1. And surprise... since what we get is exactly the
desired inequality.
In the same category, here is another application of this idea.
Example 9. Prove that for any positive real numbers a, b, c the
following inequality holds:
1
3a
+
1
3b
+
1
3c
+ 2
_
1
2a +b
+
1
2b +c
+
1
2c +a
_
≥ 3
_
1
a + 2b
+
1
b + 2c
+
1
c + 2a
_
.
Gabriel Dospinescu
272
Solution. If the previous problem could be solved using bunching
(or not? Anyway, we haven’t tried), this one is surely impossible to solve
in this manner. With the experience from the previous problem, we see
that the problem asks in fact to prove that
x
3
+y
3
+z
3
+ 2(x
2
y +y
2
z +z
2
x) ≥ 3(xy
2
+yz
2
+zx
2
)
for any positive real numbers x, y, z.
Let us assume that x = min(x, y, z) and write y = x +m, z = x +n
for some nonnegative real numbers m, n. Simple computations show that
the inequality is equivalent to
2x(m
2
−mn +n
2
) + (n −m)
3
+m
3
≥ (n −m)m
2
.
Therefore, it suﬃces to prove that
(n −m)
3
+m
3
≥ (n −m)m
2
,
which is the same as (via the substitution t =
n −m
m
) t
3
+ 1 ≥ t for all
t ≥ −1, which is immediate.
Starting this topic, we said that there is a deep relation between
integrals and areas, but in the sequel we seemed to neglect the last
concept. We ask the reader to accept our apologizes and bring to their
attention two mathematical gems, in which they will surely have the
occasion to play around with areas. If only this was easy to see... In fact,
these problems are discrete forms of Young and Steﬀensen inequalities
for integrals.
Example 10. Let a
1
≥ a
2
≥ ≥ a
n+1
= 0 and let b
1
, b
2
, . . . , b
n

[0, 1]. Prove that if
k =
_
n

i=1
b
i
_
+ 1,
273
then
n

i=1
a
i
b
i

k

i=1
a
i
.
blance with Steﬀensen’s inequality: for any continuous functions f, g :
[a, b] →R such that f is decreasing and 0 ≤ g ≤ 1 we have
_
a+k
a
f(x)dx ≥
_
b
a
f(x)g(x)dx,
where
k =
_
b
a
f(x)dx.
So, probably an argument using areas (this is how we avoid in-
tegrals and argue with their discrete forms, areas!!!) could lead to a
neat solution. So, let us consider a coordinate system XOY and let
us draw the rectangles R
1
, R
2
, . . . , R
n
such that the vertices of R
i
are
the points (i − 1, 0), (i, 0), (i − 1, a
i
), (i, a
i
) (we need n rectangles of
heights a
1
, a
2
, . . . , a
n
and weights 1, so that to view
k

i=1
a
i
as a sum
of areas) and the rectangles S
1
, S
2
, . . . , S
n
, where the vertices of S
i
are
the points
_
_
i−1

j=1
b
j
, 0
_
_
,
_
_
i

j=1
b
j
, 0
_
_
,
_
_
i−1

j=1
b
j
, a
i
_
_
,
_
_
i

j=1
b
j
, a
i
_
_
(where
0

j=1
b
j
= 0). We have made this choice because we need two sets of pair
wise disjoint rectangles with the same heights and areas a
1
, a
2
, . . . , a
n
and a
1
b
1
, a
2
b
2
, . . . , a
n
b
n
so that we can compare the areas of the unions
of the rectangles in the two sets. Thus, looking in a picture, we ﬁnd im-
mediately what we have to show: that the set of rectangles S
1
, S
2
, . . . , S
n
can be covered with the rectangles R
1
, R
2
, . . . , R
k+1
. Intuitively, this is
evident, by looking again at the picture. Let us make it rigorous. Since
274
the weight of the union of S
1
, S
2
, . . . , S
n
is
n

j=1
b
j
< k + 1 (and the
weight of R
1
, R
2
, . . . , R
k+1
is k + 1), it is enough to prove this for any
horizontal line. But if we consider a horizontal line y = p and an index
r such that a
r
≥ p > a
r+1
, then the corresponding weight for the set
R
1
, R
2
, . . . , R
k+1
is p, which is at least b
1
+b
2
+ +b
p
, the weight for
S
1
, S
2
, . . . , S
n
. And the problem is solved.
And now the second problem, given this time in a Balkan Mathe-
Example 11. Let (x
n
)
n≥0
be an increasing sequence of nonnegative
integers such that for all k ∈ N the number of indices i ∈ N for which
x
i
≤ k is y
k
< ∞. Prove that for any m, n ∈ N we have the inequality
m

i=0
x
i
+
n

j=0
y
j
≥ (m+ 1)(n + 1).
Solution. Again, experienced reader will see immediately a similar-
ity with Young’s inequality: for any strictly increasing one to one map
f : [0, A] → [0, B] and any a ∈ (0, A), b ∈ (0, B) we have the inequality
_
a
0
f(x)dx +
_
b
0
f
−1
(x)dx ≥ ab.
Indeed, it suﬃces to take the given sequence (x
n
)
n≥0
as the one to
one increasing function in Young’s inequality and the sequence (y
n
)
n≥0
as the inverse of f. Just view
m

i=0
x
i
and
m

j=0
y
j
as the corresponding
integrals and the similarity will be obvious.
Thus, probably again a geometrical solution is hiding behind some
rectangles. Indeed, consider the vertical rectangles with weight 1 and
heights x
0
, x
1
, . . . , x
m
and the rectangles with weight 1 and heights
y
0
, y
1
, . . . , y
n
. Then in a similar way one can prove that the set of these
275
rectangles covers the rectangle of sides m+1 and n +1. Thus, the sum
of their areas is at least the are of this rectangle.
It will be diﬃcult to solve the following beautiful problems using
integrals, since the idea is very well hidden. Yet, there is such a solution
and it is more than beautiful.
Example 12. Prove that for any a
1
, a
2
, . . . , a
n
, b
1
, b
2
, . . . , b
n
≥ 0 the
following inequality holds

1≤i<j≤n
([a
i
−a
j
[ +[b
i
−b
j
[) ≤

1≤i,j≤n
[a
i
−b
j
[.
Poland, 1999
Solution. Let us deﬁne the functions f
i
, g
i
: [0, ∞) →R,
f
i
(x) =
_
1, t ∈ [0, a
i
],
0, t > a
i
and g
i
(x) =
_
1, x ∈ [0, b
i
],
0, x > b
i
.
Also, let us deﬁne
f(x) =
n

i=1
f
i
(x), g(x) =
n

i=1
g
i
(x).
Now, let us compute
_

0
f(x)g(x)dx. We see that
_

0
f(x)g(x)dx =
_

0
_
_

1≤i,j≤n
f
i
(x)g
j
(x)
_
_
dx
=

1≤i,j≤n
_

0
f
i
(x)g
j
(x)dx =

1≤i,j≤n
min(a
i
, b
j
).
A similar computation shows that
_

0
f
2
(x)dx =

1≤i,j≤n
min(a
i
, a
j
)
and
_

0
g
2
(x)dx =

1≤i,j≤n
min(b
i
, b
j
).
276
Since
_

0
f
2
(x)dx+
_

0
g
2
(x)dx =
_

0
(f
2
(x)+g
2
(x))dx ≥ 2
_

0
f(x)g(x)dx,
we ﬁnd that

1≤i,j≤n
min(a
i
, a
j
) +

1≤i,j≤n
min(b
i
, b
j
) ≥ 2

1≤i,j≤n
min(a
i
, b
j
).
Now, remember that 2 min(x, y) = x + y − [x − y[ and the last
inequality becomes

1≤i,j≤n
[a
i
−a
j
[ +

1≤i,j≤n
[b
i
−b
j
[ ≤ 2

1≤i,j≤n
[a
i
−b
j
[
and since

1≤i,j≤n
[a
i
−a
j
[ = 2

1≤i<j≤n
[a
i
−a
j
[,
the problem is solved.
Using this idea, here is a diﬃcult problem, whose elementary solution
is awful and which has a 3-lines solution using the above idea... Of course,
this is easy to ﬁnd for the author of the problem, but in a contest things
change!
Example 13. Let a
1
, a
2
, . . . , a
n
> 0 and let x
1
, x
2
, . . . , x
n
be real
numbers such that
n

i=1
a
i
x
i
= 0.
a) Prove that the inequality

1≤i<j≤n
x
i
x
j
[a
i
−a
j
[ ≤ 0 holds;
b) Prove that we have equality in the above inequality if and only if
there exist a partition A
1
, A
2
, . . . , A
k
of the set ¦1, 2, . . . , n¦ such that
for all i ∈ ¦1, 2, . . . , k¦ we have

j∈A
i
x
j
= 0 and a
j
1
= a
j
2
if j
1
, j
2
∈ A
i
.
277
Solution. Let λ
A
be the characteristic function of the set A. Let us
consider the function
f : [0, ∞) →R, f =
n

i=1
x
i
λ
[0,a
i
]
.
Now, let us compute
_

0
f
2
(x)dx =

1≤i,j≤n
x
i
x
j
_

0
λ
[0,a
i
]
(x)λ
[0,a
a
]
(x)dx
=

1≤i,j≤n
x
i
x
j
min(a
i
, a
j
).
Hence

1≤i,j≤n
x
i
x
j
min(a
i
, a
j
) ≥ 0.
Since
min(a
i
, a
j
) =
a
i
+a
j
−[a
i
−a
j
[
2
and

1≤i,j≤n
x
i
x
j
(a
i
+a
j
) = 2
_
n

i=1
x
i
__
n

i=1
a
i
x
i
_
= 0,
we conclude that

1≤i<j≤n
x
i
x
j
[a
i
−a
j
[ ≤ 0.
Let us suppose that we have equality. We ﬁnd that
_

0
f
2
(x)dx = 0
and so f(x) = 0 almost anywhere. Now, let b
1
, b
2
, . . . , b
k
the distinct
numbers that appear among a
1
, a
2
, . . . , a
n
> 0 and let A
i
= ¦j ∈
¦1, 2, . . . , n¦[ a
j
= b
i
¦. Then A
1
, A
2
, . . . , A
k
is a partition of the set
¦1, 2, . . . , n¦ and we also have
k

i=1
_
_

j∈A
i
x
j
_
_
λ
[0,b
i
]
= 0
278
almost anywhere, from where we easily conclude that

i∈A
i
x
j
= 0 for all i ∈ ¦1, 2, . . . , k¦.
The conclusion follows.
And since we have proved the nice inequality

1≤i,j≤n
x
i
x
j
min(a
i
, a
j
) ≥ 0
for any numbers x
1
, x
2
, . . . , x
n
, a
1
, a
2
, . . . , a
n
> 0 let’s make a step fur-
ther and give the magniﬁcent proof found by Ravi B. (see mathlinks
site) for one of the most diﬃcult inequalities ever given in a contest,
solution based on this result:
Example 14. Prove the following inequality

1≤i,j≤n
min(a
i
a
j
, b
i
b
j
) ≤

1≤i,j≤n
min(a
i
b
j
, a
j
b
i
).
G. Zbaganu, USAMO, 1999
Solution. Let us deﬁne the numbers r
i
=
max(a
i
, b
i
)
min(a
i
, b
i
)
−1 and x
i
=
sgn(a
i
− b
i
) (if, by any chance, one of a
i
, b
j
= 0, we can simply put
r
i
= 0). The crucial observation is the following identity:
min(a
i
b
j
, a
j
b
i
) −min(a
i
a
j
, b
i
b
j
) = x
i
x
j
min(r
i
, r
j
).
Proving this relation can be achieved by distinguishing 4 cases, but
let us remark that actually we may assume that a
i
≥ b
i
and a
j
≥ b
j
,
which leaves us with only two cases. The ﬁrst one is when at least one of
the two inequalities a
i
≥ b
i
and a
j
≥ b
j
becomes an equality. This case
is trivial, so let us assume the contrary. Then
x
i
x
j
min(r
i
, r
j
) = b
i
b
j
min
_
a
i
b
i
−1,
a
j
b
j
−1
_
= b
i
b
j
_
min
_
a
i
b
i
,
a
j
b
j
_
−1
_
= min(a
i
b
j
, a
j
b
i
) −b
i
b
j
= min(a
i
b
j
, a
j
b
i
) −min(a
i
a
j
, b
i
b
j
).
279
Now, we can write

1≤i,j≤n
min(a
i
b
j
, a
j
b
i
)−

1≤i,j≤n
min(a
i
a
j
, b
i
b
j
) =

i,j
x
i
x
j
min(r
i
, r
j
) ≥ 0,
the last inequality being nothing else than the main ingredient of the
preceding problem.
Finally, here is a very funny problem, which is a consequence of
this last hard inequality. Consider this a hint and try to solve it, since
otherwise the problem is really extremely hard.
Example 15. Let x
1
, x
2
, . . . , x
n
some positive real numbers such
that

1≤i,j≤n
[1 −x
i
x
j
[ =

1≤i,j≤n
[x
i
−x
j
[.
Prove that
n

i=1
x
i
= n.
Gabriel Dospinescu
Solution. Consider b
i
= 1 in the inequality from example 14. We
obtain:

1≤i,j≤n
min(x
i
, x
j
) ≥

1≤i,j≤n
min(1, x
i
x
j
).
Now, use the formula min(u, v) =
u +v −[u −v[
2
and rewrite the
above inequality in the form
2n
n

i=1
x
i

1≤i,j≤n
[x
i
−x
j
[ ≥ n
2
+
_
n

i=1
x
i
_
2

1≤i,j≤n
[1 −x
i
x
j
[.
Taking into account that

1≤i,j≤n
[1 −x
i
x
j
[ =

1≤i,j≤n
[x
i
−x
j
[,
we ﬁnally obtain
2n
n

i=1
x
i
≥ n
2
+
_
n

i=1
x
i
_
2
,
280
which can be rewritten as
_
n

i=1
x
i
−n
_
2
≤ 0
Therefore
n

i=1
x
i
= n.
Problems for practice
1. Show that for all a, b ∈ N

ln
_
bn + 1
an + 1
_
<
1
an + 1
+
1
an + 2
+ +
1
bn
< ln
b
a
.
2. Prove that for any a > 0 and any positive integer n the inequality
1
a
+ 2
a
+ +n
a
<
(n + 1)
a+1
−1
a + 1
holds. Also, for a ∈ (−1, 0) we have the reversed inequality.
Folklore
3. Prove that for any real number x
n
n

k=0
x
2k
≥ (n + 1)
n

k=1
x
2k−1
.
Harris Kwong, College Math. Journal
4. Let a continuous and monotonically increasing function f :
[0, 1] →R such that f(0) = 0 and f(1) = 1. Prove that
9

k=1
f
_
k
10
_
+
10

k=1
f
−1
_
k
10
_

99
10
.
Sankt Petersburg, 1991
5. Prove the following inequality
a
n
+b
n
2
+
_
a +b
2
_
n
≥ 2
a
n
+a
n−1
b + +ab
n−1
+b
n
n + 1
for any positive integer n and any nonnegative real numbers a, b.
281
Mihai Onucu Drambe
6. Prove that if a
1
≤ a
2
≤ ≤ a
n
≤ 2a
1
the the following inequality
holds
a
n

1≤i,j≤n
min(a
i
, a
j
) ≥
_
n

i=1
a
i
_
2
+
_
2n −
n

i=1
a
i
_
2
.
Gabriel Dospinescu
7. For all positive real number x and all positive integer n we have:
_
2n
0
_
x

_
2n
1
_
x + 1
+
_
2n
2
_
x + 2
− +
_
2n
2n
_
x + 2n
> 0.
Komal
8. Prove that the function f : [0, 1) →R deﬁned by
f(x) = log
2
(1 −x) +x +x
2
+x
4
+x
8
+. . .
is bounded.
Komal
9. Prove that for any real numbers a
1
, a
2
, . . . , a
n
n

i,j=1
ij
i +j −1
a
i
a
j

_
n

i=1
a
i
_
2
.
10. Let k ∈ N, α
1
, α
2
, . . . , α
n+1
= α
1
. Prove that

1≤i≤n
1≤j≤k
α
k−j
i
α
j−1
i+1

k
n
k−2
_
n

i=1
α
i
_
k−1
.
Hassan A. Shah Ali, Crux Mathematicorum
11. Prove that for any positive real numbers a, b, c such that a+b =
c = 1 we have:
_
1 +
1
a
_
b
_
1 +
1
b
_
c
_
1 +
1
c
_
a
≥ 1 +
1
ab +bc +ca
.
282
12. Prove that for all a
1
, a
2
, . . . , a
n
, b
1
, b
2
, . . . , b
n
≥ 0 the inequality
holds
_
_

1≤i,j≤n
min(a
i
, a
j
)
_
_
_
_

1≤i,j≤n
min(b
i
, b
j
)
_
_

_
_

1≤i,j≤n
min(a
i
, b
j
)
_
_
.
Don Zagier
13. Prove that for any x
1
≥ x
2
≥ ≥ x
n
> 0 we have
n

i=1
¸
x
2
i
+x
2
i+1
+ +x
2
n
i
≤ π
n

i=1
x
i
.
Adapted after an IMC 2000 problem
14. Let ϕ be Euler’s totient function, where ϕ(1) = 1. Prove that
for any positive integer n we have
1 >
n

k=1
ϕ(k)
k
ln
2
k
2
k
−1
> 1 −
1
2
n
.
Gabriel Dospinescu
15. Let p
1
, p
2
, . . . , p
n
some positive numbers which add up to 1 and
x
1
, x
2
, . . . , x
n
some positive real numbers. Let also
A =
n

i=1
a
i
x
i
and G =
n

i=1
x
p
i
i
.
a) Let us denote
I(x, a) =
_

0
tdt
(1 +t)(x +at)
2
.
Prove that
ln
A
G
=
n

i=1
p
i
(x
i
−A)
2
I(x
i
, A).
Deduce the arithmetic-geometric inequality.
b) Suppose that x
i

1
2
and deﬁne A
t
, G
t
the corresponding means
for 1 −x
i
. Prove that
A
G

A
t
G
t
.
Oral examination ENS
283
16. Prove that for any positive real numbers x
1
, x
2
, . . . , x
n
such that
n

i=1
1
1 +x
i
=
n
2
,
we have the inequality

1≤i,j≤n
1
x
i
+x
j

n
2
2
.
Gabriel Dospinescu
17. Prove that we can ﬁnd a constant c such that for any x ≥ 1 and
any positive integer n we have
¸
¸
¸
¸
¸
n

k=1
kx
(k
2
+x)
2

1
2
¸
¸
¸
¸
¸

c
x
.
IMC, 1996
18. Let 0 = x
1
< < x
2n+1
= 1 some real numbers. Prove that if
x
i+1
−x
i
≤ h for all 1 ≤ i ≤ 2n then
1 −h
2
<
2n

i=1
x
2i
(x
2i+1
−x
2i−1
) <
1 +h
2
.
Turkey TST, 1996
19. Prove that for any a
1
, a
2
, . . . , a
n
≥ 0 we have the following in-
equality

1≤i,j≤n
a
i
a
j
i +j
≤ π
n

i=1
a
2
i
.
Hilbert’s inequality
284

TWO USEFUL SUBSTITUTIONS We know that in most inequalities with a constraint such as abc = 1 x y z the substitution a = , b = , c = simpliﬁes the solution (don’t kid y z x yourself, not all problems of this type become easier!). But have you ever thought about other similar substitutions? For example, what if we had the conditions x, y, z > 0 and xyz = x + y + z + 2? Or x, y, z > 0 and xy + yz + zx + 2xyz = 1? There are numerous problems that reduce to these conditions and to their corresponding substitutions. You will be probably surprised when ﬁnding out that the ﬁrst set of conditions implies the existence of positive real numbers a, b, c such that x= b+c , a y= c+a , b z= a+b . c

Let us explain why. The condition xyz = x + y + z + 2 can be written in the following equivalent way: 1 1 1 + + = 1. 1+x 1+y 1+z Proving this is just a matter of simple computations. Take now a= 1 , 1+x b= 1 , 1+y c= 1 . 1+z

b+c 1−a = . Of course, in the same Then a + b + c = 1 and x = a a a+b b+c c+a c+a ,z= . The converse (that is, , , way we ﬁnd y = b c a b a+b satisfy xyz = x + y + z = 2) is much easier and is settled again by c basic computations. Now, what about the second set of conditions? If you look carefully, you will see that it is closely related to the ﬁrst one. 1 1 Indeed, x, y, z > 0 satisfy xy + yz + zx + 2xyz = 1 if and only if , , x y 1 1 1 1 1 verify = + + + 2, so the substitution here is z xyz x y z x= a , b+c y= b , c+a
2

z=

c . a+b

So, let us summarize: we have seen two nice substitutions, with even nicer proofs, but we still have not seen any applications. We will see them in a moment ... and there are quite a few inequalities that can be solved by using these ”tricks”. First, an easy and classical problem, due to Nesbitt. It has so many extensions and generalizations, that we must discuss it ﬁrst. Example 1. Prove that a b c 3 + + ≥ b+c c+a a+b 2 for all a, b, c > 0. Solution. With the ”magical” substitution, it suﬃces to prove that 3 if x, y, z > 0 satisfy xy + yz + zx + 2xyz = 1, then x + y + z = . 2 3 Let us suppose that this is not the case, i.e. x + y + z < . Because 2 (x + y + z)2 3 xy + yz + zx ≤ , we must have xy + yz + zx < and 3 4 3 x+y+z 1 since xyz ≤ , we also have 2xyz < . It follows that 3 4 3 1 1 = xy + yz + zx + 2xyz < + = 1, a contradiction, so we are done. 4 4 Let us now increase the level of diﬃculty and make an experiment: imagine that you did not know about these substitutions and try to solve the following problem. Then look at the solution provided and you will see that sometimes a good substitution can solve a problem almost alone. Example 2. Let x, y, z > 0 such that xy + yz + zx + 2xyz = 1. Prove that 1 1 1 + + ≥ 4(x + y + z). x y z Mircea Lascu, Marian Tetiva Solution. With our substitution the inequality becomes b+c c+a a+b + + ≥4 a b c
3

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

.

But this follows from 4s a a ≤ + , b+c b c stitution. Here is a geometric application of the previous problem. Example 3. Prove that in any acute-angled triangle ABC the following inequality holds 1 cos2 A cos2 B+cos2 B cos2 C+cos2 C cos2 A ≤ (cos2 A+cos2 B+cos2 C). 4 Titu Andreescu Solution. We observe that the desired inequality is equivalent to cos A cos C cos A cos B cos B cos C + + ≤ cos C cos A cos B ≤ Setting x= cos B cos C , cos A y= cos A cos C , cos B z= cos A cos B , cos C 1 4 cos A cos B cos C + + cos B cos C cos C cos A cos A cos B 4b b b ≤ + , c+a c a 4c c c ≤ + . a+b a b

Simple and eﬃcient, these are the words that characterize this sub-

the inequality reduces to 4(x + y + z) ≤ 1 1 1 + + . x y z

But this is precisely the inequality in the previous example. All that remains is to show that xy + yz + zx + 2xyz = 1. This is equivalent to cos2 A + cos2 B + cos2 C + 2 cos A cos B cos C = 1, a well-known identity, proved in the chapter ”Equations and beyond”. The level of diﬃculty continues to increase. When we say this, we refer again to the proposed experiment. The reader who will try ﬁrst to solve the problems discussed without using the above substitutions will certainly understand why we consider these problems hard.
4

Example 4. Prove that if x, y, z > 0 and xyz = x + y + z + 2, then √ √ √ 2( xy + yz + zx) ≤ x + y + z + 6. Mathlinks site Solution. This is tricky, even with the substitution. There are two main ideas: using some identities that transform the inequality into an easier one and then using the substitution. Let us see. What does √ √ √ 2( xy + yz + zx) suggest? Clearly, it is related to √ √ √ ( x + y + z)2 − (x + y + z). Consequently, our inequality can be written as √ x+ √ y+ √ z≤ 2(x + y + z + 3).

The ﬁrst idea that comes to mind (that is using the Cauchy√ √ √ 3(x + y + z) ≤ Schwarz inequality in the form x + y + z ≤ 2(x + y + z + 3)) does not lead to a solution. Indeed, the last inequality is not true: setting x+y+z = s, we have 3s ≤ 2(s+3). This is because s3 s3 from the AM-GM inequality it follows that xyz ≤ , so ≥ s + 2, 27 27 2 ≥ 0, implying s ≥ 6. which is equivalent to (s − 6)(s + 3) Let us see how the substitution helps. The inequality becomes b+c + a c+a + b a+b ≤ c 2 b+c c+a a+b + + +3 a b c

The last step is probably the most important. We have to change b+c c+a a+b the expression + + + 3 a little bit. a b c We see that if we add 1 to each fraction, then a + b + c will appear as common factor, so in fact b+c c+a a+b + + + 3 = (a + b + c) a b c
5

1 1 1 + + a b c

.

we transform this inequality into (x − 1)2 (y − 1)2 (z − 1)2 1 + 2 + 2 ≥ . b. There are many proofs for this inequality. 2+2 x y +2 z +2 x + y2 + z2 + 6 6 . 2+2 x y +2 z +2 2 This last form suggests using the Cauchy-Schwarz inequality to prove that (x − 1)2 (y − 1)2 (z − 1)2 (x + y + z − 3)2 + 2 + 2 ≥ 2 . amusingly. USAMO 2003 Solution. + + a b c We continue with a 2003 USAMO problem. Prove that for any positive real numbers a. Taking our substitution into account. c the following inequality holds (2a + b + c)2 (2b + c + a)2 (2c + a + b)2 + 2 + 2 ≤ 8. it suﬃces to prove that if xyz = x + y + z + 2. 2 + x2 2 + y2 2 + z2 This is in fact the same as 2x + 1 2y + 1 2z + 1 5 + 2 + 2 ≤ .And now we have ﬁnally solved the problem. none of them easy. x2 + 2 y +2 z +2 2 Now. The desired inequality is equivalent to 1+ 2+ b+c a b+c a 2 2+ 2 + 2+ c+a b c+a b 2 1+ 2 + 2+ a+b c a+b c 2 2 ≤ 8. by employing again the Cauchy-Schwarz inequality: b+c + a c+a + b a+b ≤ c (b + c + c + a + a + b) 1 1 1 . Example 5. Zuming Feng. 2a2 + (b + c)2 2b + (c + a)2 2c + (a + b)2 Titu Andreescu. then (2 + x)2 (2 + y)2 (2 + z)2 + + ≤ 8. The following solution is again not easy. but it is natural for someone familiar with this kind of substitution.

Prove that if x. a+b But now the solution is almost over. 2 2 2 2 2 2 2 2 2 Hence there are positive real numbers a. This shows that it suﬃces to prove that 2(s−3)2 ≥ s2 −18 for all s ≥ 6. Now. Indeed. this inequality is equivalent to 2(x + y + z − 3)2 ≥ (x + y + z)2 − 2(xy + yz + zx) + 6. there is an easy solution using the substitutions described in this unit. which is equivalent to (s−3)(s−6) ≥ 0. Let us write the given condition as x y y z z x x y z · + · + · + 2 · · = 1. z ≥ 0 satisfy xy + yz + zx + xyz = 4 then x + y + z ≥ xy + yz + zx. from xyz ≥ 8 (recall who x. 1998 Solution. we are left with proving that 2(x + y + z − 3)2 ≥ x2 + y 2 + z 2 + 6. Yet. c such that x= 2a . But this is not diﬃcult. y. b+c y= 2b . we ﬁnd that xy + yz + zx ≥ 12 and x + y + z ≥ 6 (by the same AM-GM inequality). India. y. Example 6.So. And this diﬃcult problem is solved! The following problem is also hard. c+a z= 2c . the inequality becomes a(a + b)(a + c) + b(b + a)(b + c) + c(c + a)(c + b) ≥ 7 . b. z are and use the AM-GM inequality three times). We have seen a diﬃcult solution in the chapter ”Equations and beyond”. b+c c+a a+b (c + a)(c + b) (a + b)(a + c) (b + a)(b + c) After clearing denominators. clearly true. since the inequality x + y + z ≥ xy + yz + zx is equivalent to a b c 2ab 2bc 2ca + + ≥ + + .

It is not diﬃcult to ﬁnd the formula (a + b − c)(b + c − a)(c + a − b) = a2 b2 c2 . in which the substitution described plays a key role. It is readily seen that this form is stronger than Schur’s inequality (a + b − c)(b + c − a)(c + a − b) ≤ abc. First. Example 7. b+c−a. Prove that if x. we may assume that a. b z= a+b . But this is Schur’s inequality! We end the discussion with a diﬃcult problem. the desired inequality can be written as (a + b + c)R2 ≥ (a + b)(b + c)(c + a) . it reduces to a(a − b)(a − c) + b(b − a)(b − c) + c(c − a)(c − b) ≥ 0. But this time using the substitution only will not suﬃce. b. z > 0 satisfy xyz = x + y + z + 2. since otherwise the left-hand side in (1) is negative. then xyz(x − 1)(y − 1)(z − 1) ≤ 8. b. c the inequality becomes (a + b)(b + c)(c + a)(a + b − c)(b + c − a)(c + a − b) ≤ 8a2 b2 c2 (1) for any positive real numbers a. (a + b + c)R2 Consequently. 8 8 . c are the sides of a triangle ABC. After basic computations. Using the substitution x= b+c . This is true because no more than one of the numbers a+b−c. c+a−b can be negative. a y= c+a .≥ 2ab(a + b) + 2bc(b + c) + 2ca(c + a). Let R be the circumradius of the triangle ABC. y. c. Gabriel Dospinescu Solution.

b. Mircea Lascu 9 . 8 4 2. y. Gabriel Dospinescu.But we know that in any triangle ABC. a b c b+c c+a a+b 2 J. Let x. 3 The ﬁrst inequality reduces to 1 a2 + b2 + c2 ≥ (a + b + c)2 . then xyz ≤ 1 3 and xy + yz + zx ≥ . By combining these two results. z > 0 such that xy + yz + zx = 2(x + y + z). then xy + yz + zx ≥ 2(x + y + z) and √ x+ √ y+ √ z≤ 3√ xyz. Prove that if x. Problems for training 1. 2 4. Prove that if x. This inequality follows from the following ones: 8 8(a + b + c)(a2 + b2 + c2 ) ≥ (a + b + c)3 3 and 1 9(a + b)(b + c)(c + a) ≤ (a + b + c)3 . c the following inequality holds a b c 9 b+c c+a a+b + + ≥ + + + . 3 while the second is a consequence of the AM-GM inequality. z > 0 and xyz = x + y + z + 2. Hence it suﬃces to prove that 8(a + b + c)(a2 + b2 + c2 ) ≥ 9(a + b)(b + c)(c + a). the desired inequality follows. Prove that xyz ≤ x + y + z + 2. Prove that for any positive real numbers a. z > 0 satisfy xy + yz + zx + 2xyz = 1. y. y. 9R2 ≥ a2 + b2 + c2 . Nesbitt 3.

z = c + . then b c a xy + yz + zx ≥ 2(x + y + z). c > 0 and x = a + . 1997 10 . y = b + . c > 0. (c + a − b)2 (a + b − c)2 3 (b + c − a)2 + + ≥ . b.5. (cos A + cos B)2 + (cos B + cos C)2 + (cos C + cos A)2 ≤ 3. Prove that in every acute-angled triangle ABC. Prove that in any triangle ABC the following inequality holds 1 cos A + cos B + cos C ≥ (3 + cos(A − B) + cos(B − C) + cos(C − A)). Vasile Cartoaje 8. (b + c)2 + a2 (c + a)2 + b2 (a + b)2 + c2 5 Japan. b. Prove that for any a. 4 Titu Andreescu 6. Prove that if a. 1 1 1 7.

. since they are too well-known. a1 . a1 a2 an But this is just actually the deﬁnition of a geometrical progression. a1 . . Prove that the ﬁnite sequence a0 . . In this unit we will not focus on the theoretical results. .ALWAYS CAUCHY-SCHWARZ In recent years the Cauchy-Schwarz inequality has become one of the most used results in elementary mathematics. Example 1. . . Note that Lagrange’s identity allowed us to work with equivalences. . . Let us begin with a very simple problem. There are countless problems that reduce readily to this inequality and even more problems in which the CauchySchwarz inequality is the key idea of the solution. This time the inequality is hidden in a closed form. employing then gradually the Cauchy-Schwarz inequality in some of the most diﬃcult problems. . Yet. This is equivalent to the proportionality of the n-tuples (a0 . an−1 ) and (a1 . . We see that the relation given in the problem is in fact the equality case in the Cauchy-Schwarz inequality. 11 . a2 . seeing the Cauchy-Schwarz inequality at work is not so well spread out. Another easy application of the Cauchy-Schwarz inequality is the following problem. This is the reason why we will see this inequality in action in several simple examples ﬁrst. it underlines something less emphasized: the analysis of the equality case. an ). Yet. an indispensable tool of any serious problem solver. . Hence the problem is solved. . 0 1 n−1 1 2 n Solution. an of positive real numbers is a geometrical progression if and only if (a2 + a2 + · · · + a2 )(a2 + a2 + · · · + a2 ) = (a0 a1 + a1 a2 + · · · + an−1 an )2 . that is a0 a1 an−1 + = ··· = . a direct application of the inequality.

f (x) = ln p(ex ) is convex. And the problem is solved. note an interesting consequence of the problem: the function f : (0. that is why we said in the introduction to this problem that it has a solution based on calculus. p(xk ) ≥ pk (x1 x2 . Moreover. since we have of course p2 (xy) ≤ p2 (|xy|). we see that the conditions x. Russian Mathematical Olympiad Solution. xk ). Let p be a polynomial with positive real coeﬃcients. ∞) → (0. And now the Cauchy-Schwarz inequality comes into the picture: (a0 + a1 xy + · · · + an xn y n )2 √ √ = ( a0 · a0 + a1 x2 · a2 y 2 + · · · + √ an xn · √ an y n )2 ≤ (a0 + a1 x2 + · · · + an x2n )(a0 + a1 y 2 + · · · + an y 2n ). 1 2 k which follows the Jensen’s inequality for the convex function f (x) = ln p(ex ). . If we work only with the closed expression p(x2 )p(y 2 ) ≥ p2 (xy).which suggests using calculus. We will not prove this here. Prove that p(x2 )p(y 2 ) ≥ p2 (xy) for any positive real numbers x. . The idea of that solution is to prove that the second derivative of is nonnegative. . . y > 0 are useless. 12 . There exists a solution by using derivatives. but it is not as elegant as the featured one: Example 2. Additionally. y. ∞). the chances of seeing a way to proceed are small. let us write p(x) = a0 + a1 x + · · · + an xn . So. The desired inequality becomes (a0 + a1 x2 + · · · + an x2n )(a0 + a1 y 2 + · · · + an y 2n ) ≥ (a0 + a1 xy + · · · + an xn y n )2 . but we note a simple consequence: the more general inequality p(xk )p(xk ) .

b = y. In fact. This ≤ (a + b + c) encourages us to take a = x. we x y z x+y+z try another approach. the reversed inequality holds. y. Unfortunately. 13 . a b c So. though this time you might be surprised why the ”trick” fails at a ﬁrst approach: 1 1 1 Example 3. since in this case x−1 y−1 z−1 + + = 1 and a + b + c = x + y + z. Hence this approach fails. Iran. this idea works and the problem is solved. this inequality is 2 9 not true. using again the Cauchy-Schwarz inequality. We will also see during our trip through the exciting world of the Cauchy-Schwarz inequality a nice application of Aczel’s inequality. Then. the not so well-known inequality of Aczel. c = z. then x y z √ √ √ x − 1 + y − 1 + z − 1 ≤ x + y + z. 9 which is equivalent to x + y + z ≤ . 2 1 1 1 9 since 2 = + + ≥ . Prove that if x. z > 0 satisfy + + = 2. The obvious and most natural approach is to apply the Cauchy-Schwarz inequality in the form √ x−1+ y−1+ √ z−1≤ 3(x + y + z − 3) √ and then to try to prove the inequality 3(x + y + z − 3) ≤ x + y + z.Here is another application of the Cauchy-Schwarz inequality. but this time in the form √ x − 1+ √ √ y − 1+ z − 1 = a· x−1 √ + b· a y−1 √ + c· b z−1 c x−1 y−1 z−1 + + . a b c √ We would like to have the last expression equal to x + y + z. We continue with a classical result. 1998 Solution. that is x + y + z ≥ .

bn be real numbers and let A. We observe ﬁrst that we may assume that A2 > a2 + a2 + · · · + a2 and B 2 > b2 + b2 + · · · + b2 . in which the condition seems to be useless. . an . where a = a2 + a2 + · · · + a2 and b = b2 + b2 + · · · + b2 . 1 2 n 1 2 n Otherwise the left-hand side of the desired inequality is smaller than or equal to 0 and the inequality becomes trivial. . From our assumption and the Cauchy-Schwarz inequality. we infer that a1 b1 +a2 b2 +· · ·+an bn ≤ a2 + a2 + · · · + a2 · n 1 2 b2 + b2 + · · · + b2 < AB n 1 2 Hence we can rewrite the inequality in the more appropriate form a1 b1 + a2 b2 + · · · + an bn + (A2 − a)(B 2 − b) ≤ AB. we can n n 1 2 1 2 apply the Cauchy-Schwarz inequality. . B > 0 such that A2 ≥ a2 + a2 + · · · + a2 or B 2 ≥ b2 + b2 + · · · + b2 . Let a1 . As a consequence of this inequality we discuss the following problem. it is the key that suggests using Aczel’s inequality. Solution. b2 . . . . a2 .Example 4. 14 . . 1 2 n 1 2 n Then (A2 − a2 − a2 − · · · − a2 )(B 2 − b2 − b2 − · · · − b2 ) 1 2 n 1 2 n ≤ (AB − a1 b1 − a2 b2 − · · · − an bn )2 . ﬁrst in the form √ a1 b1 +a2 b2 +· · ·+an bn + (A2 − a)(B 2 − b) ≤ ab+ (A2 − a)(B 2 − b) and then in the form √ ab + (A2 − a)(B 2 − b) ≤ (a + A2 − a)(b + B 2 − b) = AB. In fact. Now. . b1 . And by combining the last two inequalities the desired inequality follows.

we want to prove that (a2 + a2 + · · · + a2 − 1)(b2 + b2 + · · · + b2 − 1) 1 2 n 1 2 n ≤ (a1 b1 + a2 b2 + · · · + an bn − 1)2 (1) in order to obtain the desired contradiction. . Of a diﬀerent kind. .Example 5. xn the following 15 . It is also a reﬁnement of the Cauchy-Schwarz inequality. Dorin Andrica. First of all. At ﬁrst glance. since they have the same sign n 1 2 (this follows immediately from the hypothesis of the problem). TST 2004. Let us take a more careful look. . with A = 1 and B = 1. n n 1 2 1 2 Titu Andreescu. But this is exactly Aczel’s inequality. . b2 . . k) such that for any real numbers x1 . Let a1 . . . Example 6. 1 2 n 1 2 n Prove that a2 + a2 + · · · + a2 > 1 and b2 + b2 + · · · + b2 > 1. the problem does not seem to be related to Aczel’s inequality. x2 . bn be real numbers such that (a2 +a2 +· · ·+a2 −1)(b2 +b2 +· · ·+b2 −1) > (a1 b1 +a2 b2 +· · ·+an bn −1)2 . while the n n 1 2 1 2 conclusion is (1). it is not diﬃcult to observe that an indirect approach is more eﬃcient. we have now the conditions 1 > a2 + a2 + · · · + a2 and 1 > b2 + b2 + · · · + b2 . USA Solution. Now. . . an . Moreover. The conclusion follows. . the following example shows that an apparently very diﬃcult inequality can become quite easy if we do not complicate things more than necessary. b1 . Indeed. And all of a sudden we arrived at the result in the previous problem. we may even assume that both numbers a2 + a2 + · · · + a2 − 1 n 1 2 and b2 + b2 + · · · + b2 − 1 are negative. . . a2 . as we can see from the solution. For given n > k > 1 ﬁnd in closed form the best constant T (n.

We also observe that in the left-hand side there are n − k variables that do not appear in the right-hand side and that the left-hand side is minimal when these variables are equal. We observe that n n 2 (xi − xj ) is nothing else than n 1≤i<j≤n k i=1 2 x2 i − i=1 k xi 2 and also (xi − xj ) = k 1≤i<j≤k i=1 2 x2 i − i=1 xi . k) k k k 2  . the inequality can be written in the equivalent form n n 2  ≥ T (n. The result is k k 2  ≥ T (n. k) − n) i=1 16 x2 i (1) . So.inequality holds: (xi − xj )2 ≥ T (n. n i=1 x2 − i i=1 xi x2 − i i=1 i=1 xi which is equivalent to k 2 k (T (n. Gabriel Dospinescu Solution. let us take them all to be zero. k) − 1) i=1 xi ≥ (kT (n. according to Lagrange’s identity. we cannot make any reasonable conjecture about T (n. n i=1 x2 − i i=1 xi x2 − i i=1 i=1 xi And now we see that it is indeed a reﬁnement of the Cauchy-Schwarz inequality. k) 1≤i<j≤n 1≤i<j≤k (xi − xj )2 . Consequently. k). k) k k k 2  . only if in the end it turns out that T (n. k) > 0. In this form. so we need an eﬃcient transformation.

we can take a k-tuple (x1 . since it is equivalent to (kA − (n − k)B)2 + k(n − k)B 2 ≥ 0. i n . k) = constant. We continue the series of diﬃcult inequalities with a very nice problem of Murray Klamkin. k) = . n then it will follow that T (n. Now. the conclusion is settled: T (n.Now. If we manage to prove this inequality. let us k 2 Hence we must have kT (n. which is clear. . n−k k n k (3) where we have taken A = i=k+1 xi and B = i=1 xi . We ﬁnd that n ≥ n−k n 2 n i=k+1 x2 i xi i=k+1 and so it suﬃces to prove that n n A2 ≥ (A + B)2 − B 2 . if kT (n. because we need i=k+1 n xi . that is showing that  2 n k n n 2 n k xi − x2 − xi ≥ i k i=1 i=1 i=1 k   (2) xi i=1 for any real numbers x1 . This time. x2 . xn . . . But (2) is of course equivalent to k n n 2 n i=k+1 n x2 i ≥ i=1 xi n − k k 2 xi i=1 . . Finally. Now. xk ) such k k that i=1 xi = 0 and i=1 x2 = 0 and we contradict the inequality (1). one part of the problem is obvious 17 n is the best k . But (3) is straight- forward. . k) − n > 0. . k) ≤ proceed with the converse. . . x2 . k) − n ≤ 0 that is T (n. we have to apply the Cauchy-Schwarz inequality.

attained when two variables are zero and the third one is 1 or −1. 3 But for the lower bound things are not so easy. Finding the upper bound does not seem to be too diﬃcult. we should try to prove the inequality a2 x2 + b2 y 2 + c2 z 2 + b2 x2 + c2 y 2 + a2 z 2 + c2 x2 + a2 y 2 + b2 z 2 ≥ a + b + c. Thus. We have used here the hypothesis x2 + y 2 + z 2 = 1. Find the extreme values of the expression a2 x2 + b2 y 2 + c2 z 2 + b2 x2 + c2 y 2 + a2 z 2 + c2 x2 + a2 y 2 + b2 z 2 where x.from the Cauchy-Schwarz inequality. b. Investigating what happens when xyz = 0. Let us see. Hence. Crux Mathematicorum Solution. Example 7. 3(a2 + b2 + c2 ) is the upper bound and this value if attained for √ 3 x=y=z= . so the new inequality cannot have a very complicated form. Why not squaring it? After all. It becomes a2 x2 + b2 y 2 + c2 z 2 · 18 b2 x2 + c2 y 2 + a2 z 2 . Murray Klamkin. we observe that a2 x2 + b2 y 2 + c2 z 2 + b2 x2 + c2 y 2 + a2 z 2 + c2 x2 + a2 y 2 + b2 z 2 = a2 + b2 + c2 . Let a. but the second one is not immediate. y. c be positive real numbers. we conclude that the minimal value should be a + b + c. since from the Cauchy-Schwarz inequality it follows that a2 x2 + b2 y 2 + c2 z 2 + ≤ b2 x2 + c2 y 2 + a2 z 2 + c2 x2 + a2 y 2 + b2 z 2 ≤ 3(a2 x2 + b2 y 2 + c2 z 2 + c2 y 2 + a2 z 2 + c2 x2 + a2 y 2 + b2 z 2 ) = 3(a2 + b2 + c2 ). z are real numbers such that x2 + y 2 + z 2 = 1.

Fasten your seat belts! Let us write the inequality in the form n i=1 ai 1 − ai 2 n ≤ i=1 a2 n(n − 1) i + . . . . We will concoct a solution using a combination between the Cauchy-Schwarz inequality and Jensen’s inequality. attained for example when (x. the following inequality holds 2 i=1 1≤i<j≤n ai aj n(n − 1) ≤ . And indeed. z) = (1. it is true and it follows from what else?. the Cauchy-Schwarz inequality: a2 x2 + b2 y 2 + c2 z 2 · b2 x2 + c2 y 2 + a2 z 2 ≥ abx2 + bxy 2 + caz 2 and the other two similar inequalities. y.+ b2 x2 + c2 y 2 + a2 z 2 · + c2 x2 + a2 y 2 + b2 z 2 · c2 x2 + a2 y 2 + b2 z 2 a2 x2 + b2 y 2 + c2 z 2 ≥ ab + bc + ca which has great chances to be true. an n 1 such that ai = . This shows that the minimal value is indeed a + b + c. Prove that for any nonnegative numbers a1 . (1 − ai )2 . We will discuss two hard inequalities and after that we will leave for the reader the pleasure of solving many other problems based on these techniques. 0). Example 8. 0. in which intuition is better than technique. a2 . (1 − ai )(1 − aj ) 2(2n − 1)2 Vasile Cartoaje Solution. but we warn the reader that such a solution cannot be invented easily. . It is now time for the champion inequalities. This is a very hard problem. 2 (1 − ai ) (2n − 1)2 We apply now the Cauchy-Schwarz inequality to ﬁnd that n i=1 ai 1 − ai 2 n n ≤ i=1 ai i=1 19 ai (1 − ai )2 n = i=1 ai 2 .

TST 2004. for a short computa−6x ≤ 0. Example 9. . 2 f (x) = x(1 − 2x) (1 − x)2 and to see if it is concave. Hence we can apply Jensen’s tion shows that f (x) = (1 − x)4 inequality to complete the solution. . . 1 → R. it remains to prove the inequality n i=1 ai 2 ≤ (1 − ai )2 n i=1 n(n − 1) a2 i + . We end this discussion with a remarkable solution. Let us deﬁne si = a1 + a2 + · · · + ai for i ≥ 1 and s0 = 0. 2 (1 − ai ) (2n − 1)2 The latter can be written of course in the following form: n i=1 ai (1 − 2ai ) 2n(n − 1) ≤ . a2 . Now. an be real numbers and let S be a nonempty subset of {1. . let us observe that 20 . This is not diﬃcult. . found by the member of the Romanian Mathematical Olympiad Committee. Now. Romania Solution. . . ik < jk . Then i∈S ai is of the form sj1 −si1 +sj2 −si2 +· · ·+sjk −sik . to the diﬃcult problem given in 2004 in one of the Romanian Team Selection Tests. 2.Thus. n}. with 0 ≤ i1 < i2 < · · · < ik ≤ n. Prove that 2 ai i∈S ≤ (ai + · · · + aj )2 . partition S into groups of consecutive numbers. . 2 (1 − ai ) (2n − 1)2 This encourages us to study the function f : 0. . j1 < j2 < · · · < jk and also i1 < j1 . . Claudiu Raicu. . 1≤i≤j≤n Gabriel Dospinescu. . Let a1 .

a2k = sjk and observe the obvious (but important) inequality (sj − si )2 ≥ 0≤i<j≤n+1 1≤i<j≤2k (ai − aj )2 .. which proves that (1) is correct. 21 . . a2 = sj1 . Hence we need to show that (sj1 − si1 + sj2 − si2 + · · · + sjk − sik )2 ≤ 0≤i<j≤n+1 (sj − si )2 . . .      2  (a − a + a − · · · + a  1 2 3 2k−1 − a2k )     ≤ k((a1 − a2k )2 + (a3 − a2 )2 + · · · + (a2k−1 − a2k−2 )2 ) and by summing up all these inequalities. The solution ends here. In the right-hand side we obtain an even smaller quantity than 1≤i<j≤2k (ai − aj )2 . . And this is how we arrived at the inequality (a1 − a2 + a3 − · · · + a2k−1 − a2k )2 ≤ 1≤i<j≤2k (ai − aj )2 (1) The latter inequality can be proved by using the Cauchy-Schwarz inequality k-times:   (a1 − a2 + a3 − · · · + a2k−1 − a2k )2       ≤ k((a1 − a2 )2 + (a3 − a4 )2 + · · · + (a2k−1 − a2k )2 )      (a1 − a2 + a3 − · · · + a2k−1 − a2k )2   ≤ k((a1 − a4 )2 + (a3 − a6 )2 + · · · + (a2k−1 − a2 )2 )    . Let us take a1 = si1 ..the left-hand side is nothing else than n s2 + i i=1 (sj − si )2 = 1≤i<j≤n 1≤i<j≤n+1 (sj − si )2 . a2k−1 = sik .

. 1 2 n China. Prove that for any positive real numbers a. . . then it is true for all x > 0. a2 . Let a1 .Problems for training 1. c be nonnegative real numbers. . b1 . Revista Matematica Timisoara 3. a + b b + c c + a 2 3 abc (a + b)(b + c)(c + a) Titu Andreescu. c ≥ 1 the following inequality holds: √ a−1+ √ b−1+ √ c−1≤ a(bc + 1). bn be real numbers such that ai aj > 0. . √ (a + b + c + 3 abc)2 1 1 1 1 ≥ + + + √ . AMM 22 . . Titu Andreescu. Let p be a polynomial with positive real coeﬃcients. Prove that 1 1 if p ≥ is true for x = 1. 1≤i<j≤n Prove the inequality  2  1≤i=j≤n  1≤i=j≤n  ai aj   1≤i=j≤n  bi bj  ai bj  ≥  Alexandru Lupas. b. b. . c. 4. Prove that (ax2 + bx + c)(cx2 + bx + a) ≥ (a + b + c)2 x2 for all nonnegative real numbers x. . Gazeta Matematica 2. For any positive integer n ﬁnd the number of ordered n-tuples of integers (a1 . an . a2 . 2002 5. . b. x p(x) Titu Andreescu. . b2 . Prove that for any real numbers a. an ) such that a1 + a2 + · · · + an ≥ n2 and a2 + a2 + · · · + a2 ≤ n3 + 1. Let a. . . MOSP 1999 6.

. n−1 Czech-Polish-Slovak Competition. 1 + x1 1 + x2 1 + xn Prove the inequality √ x1 + √ x2 + · · · + √ xn ≥ (n − 1) 1 1 1 √ + √ + ··· + √ x1 x2 xn . Let n ≥ 2 be an even integer. 2002 8. . x2 . Prove that 3(sin x1 + sin x2 + · · · + sin x10 ) ≤ cos x1 + cos x2 + · · · + cos x10 . c. Given are real numbers x1 . Vojtech Jarnik Competition. 2002 π 10. Prove that for any real numbers a. 3 Vasile Cartoaje. Titu Andreescu. Kvant 23 . x2 .7. USAMO 2002 9. We consider all polynomials of the form xn + an−1 xn−1 + · · · + a1 x + 1. 2001 11. with real coeﬃcients and having at least one real zero. such that 2 sin2 x1 + sin2 x2 + · · · + sin2 x10 = 1. . . x10 ∈ 0. z the following inequality holds ax + by + cz + 2 (a2 + b2 + c2 )(x2 + y 2 + z 2 ) ≥ (a + b + c)(x + y + z). . Show that ABC is similar to a triangle whose sides are integers and ﬁnd the smallest set of such integers. Let x1 . . b. The triangle ABC satisﬁes the relation cot A 2 2 + 2 cot B 2 2 + 3 cot C 2 2 = 6s 7r 2 . y. x. . Determine the least possible value of a2 + a2 + · · · + 1 2 a2 . . xn be positive real numbers such that 1 1 1 + + ··· + = 1. Saint Petersburg.

Gabriel Dospinescu 24 . an . xn the following inequality holds n n 2 |xi − xj | i=1 i=1 ≤ 2(n2 − 1)  3  n n  |xi − xj |2  . . . . Let n > 2 and x1 . xn such that xi xj = i≤i<j≤n n . n(n − 1) 1 1 1 + + ··· + x1 x2 xn = n2 + 1. x2 . z such that xy + yz + zx = 3. . b. x2 . a2 + · · · + an a1 + · · · + an−1 Vasile Cartoaje. Gabriel Dospinescu 15. 2 the following inequality holds an a1 (x2 + · · · + xn ) + · · · + (x1 + · · · + xn−1 ) ≥ n. . Gabriel Dospinescu 14.12. a b c (y + z) + (x + z) + (x + y) ≥ 3. . . c. Prove that for any positive real numbers a. x2 . a2 . b+c c+a a+b Titu Andreescu. . Prove that for any real numbers x1 . . . y. . x. . . x1 . . . . Prove that for any positive real numbers a1 . xn be positive real numbers such that (x1 + x2 + · · · + xn ) Prove that (x2 + x2 + · · · + x2 ) 1 2 n 1 1 1 + + ··· + 2 xn x2 x2 1 2 > n2 + 4 + 2 . i=1 j=1 IMO 2003 13.

some of them really diﬃcult. c such that abc = 1 and write 1 x=a+ . In this case. independent of a. z. Now. y. 1. x2 + y 2 + z 2 − xyz = 4 and this is the answer to the problem. without any ideas. It may seem unusual. z) = (1. b. z =c+ (1) b c x. y. c and then substitute the results in the relation abc = 1. we compute the product xyz = = a2 + 1 a2 a+ 1 a b+ 1 b2 1 b c+ 1 c 1 c2 +2 + b2 + + c2 + = (x2 − 2) + (y 2 − 2) + (z 2 − 2) + 2. Of course. another question appears: is the converse true? Obviously not (take for example the numbers (x. Consider three real numbers a. z. b. y. −1)). Thus. but this problem is in fact the introduction that leads to the other themes in this discussion. y =b+ a Find an algebraic relation between 1 1 . We begin this discussion with a problem. b. We are going to discuss two real equations and two parameterizations. To generate a relation involving x. one would solve the equations from (1) with respect to a. showing how a simple idea can generate lots of nice problems. an important task characterizing the set of solutions by using parameters. Example 1. but we will go beyond. But this is a mathematical crime! Here is a nice idea. But looking again 25 (2) . c.EQUATIONS AND BEYOND Real equations with multiple unknowns have in general inﬁnitely many solutions if they are solvable.

we ﬁnd that y 2 ≥ 4 and so there exist a non-zero real number 1 v for which y = v + . y. |z|} > 2. All the conditions and if z = + . β = ln b. z = 2ch(χ). c) = v u u v are satisﬁed and the problem is solved. Since this equation has real roots. b. uv uv 1 u u v . χ ∈ R such that x = 2ch(α). |z|} ≥ 2. then there exist α. let us regard (2) as a second degree equation with respect to z. y. We will prove the following result. |y|. 26 (3) . b. This shows that there 1 exists a nonzero real number u such that x = u + . Here. We ﬁnd two solutions: z1 = uv + 1 . z be real numbers with max{|x|. v. in 2 which this time it is clear that a. . then we take (a. Indeed. β. y = 2ch(β).at (1). Example 2. Now. But since |x| > 2. Let x. Prove that there exist real numbers a. it is better to make a choice. let us consider another one x2 + y 2 + z 2 + xyz = 4. If z = uv+ A direct consequence of the previous problem is the following: If x. c with abc = 1 satisfying (1). |z|} > 2. let us take |x| > 2. c > 0 and we take α = ln a. uv z2 = u v + v u 1 1 we take (a. ch(x) = χ = ln c. |y|. b. c) = u. |y|. (we have used here u the condition |x| > 2). we see that we must have min{|x|. How do we ﬁnd the corresponding z? Simply by v solving the second degree equation. ex + e−x . z > 0 are real numbers that verify (2). which means that (x2 − 4)(y 2 − 4) ≥ 0. v. Inspired by the previous equation. and now we are almost done. the discriminant must be nonnegative. ∞). b. we write (1). where ch : R → (0. Whenever we have a condition of the form max{|x|.

such that x = 2 cos A. let us prove that all these triples are solutions. we see ﬁrst that 0 < x. b. 2) we 27 . 2 cos C) where A.where x. but here is a nice proof employing geometry and linear algebra. Now. Solving the 2 equation with respect to z and taking into account that z ∈ (0. c) and so we must have 1 − cos C − cos C − cos B 1 − cos A 1 = 0. We will prove that the set of solutions of this equation is the set of triples (2 cos A. For the converse. B. First. y. B ∈ 0. z < 2. z > 0. y = 2 cos B. that is (a. it follows that this system has a nontrivial solution. C are the angles of an acute triangle. This identity can be proved readily by using the sum-to-product formulas. 2 cos B. This reduces to the identity cos2 A + cos2 B + cos2 C + 2 cos A cos B cos C = 1. hence there are π numbers A. let us consider the system   x − y cos C − z cos B = 0   −x cos C + y − z cos A = 0    −x cos B + y cos A − z = 0 From the above observation. y. We know that in any triangle we have the relations   a = c cos B + b cos C   b = a cos C + c cos A    c = b cos A + a cos B which are simple consequences of the Law of Cosines. − cos B − cos A which expanded gives cos2 A + cos2 B + cos2 C + 2 cos A cos B cos C = 1.

c2 = z 2 − 2. Example 3. Let us write x = a+ . bn−1 cn+1 = c2 + z 2 − 4 n . cn ) = an + 1 1 1 n . b2 = b2 + 2 . c1 = z and a2 = x2 − 2. y. z = c+ . y = 2 cos B.obtain z = −2 cos(A + B). cn + n n a b c 28 . bn . z = 2 cos C. an−1 bn+1 = b2 + y 2 − 4 n . z) = (2 cos A. this follows by induction from an + 1 an 2 + a2 + 1 an−1 1 −2 a2 = an+1 + 1 an+1 an−1 + and two similar identities. With the introduction and the easy problems over it is now time to see some nice applications of the above results. Thus we can take C = π − A − B and we will have (x. The positive real numbers x. a b c So. (cn )n≥1 by an+1 = a2 + x2 − 4 n . cn + n an b c . cn−1 with a1 = x. Indeed. (bn )n≥1 . cn ) also satisﬁes (2). a b c Then 1 1 1 a2 = a2 + 2 . b + n . b2 = y 2 − 2. Let x. a reasonable conjecture is that (an . 2 cos B. c2 = c2 + 2 . 1 1 1 Solution. bn . b1 = y. Example 4. All in all we have solved the following problem. with abc = 1. y. We deﬁne the sequences (an )n≥1 . cn ) = an + 1 n 1 1 . b + n . We have established that (an . 2 cos C). z satisfy (3) if and only if there exists an acute-angled triangle ABC such that x = 2 cos A. Prove that for all n ≥ 1 the triple (an . y = b+ . bn . z > 2 satisfying (2). y.

then certainly an bn cn = 1. Now we take two numbers x. we deduce that 0= a+ 1 a 1 1 h b + . z) + xh(y. y. a So. z) + k(y.c + b c 1 1 1 a + . h(y. cn ) satisﬁes (2). Not exactly. z). |y|} > 2 and we write 1 1 x + x2 − 4 y + y2 − 4 x = b + . y such√ that min{|x|. y.b + . which shows that indeed the triple (an . bn . y) + k(x. it is now clear that the polynomials divisible by x2 + y 2 + z 2 − xyz − 4 are solutions to the problem. Well.c= . And now? The last relation suggests that we √ should prove that for each y with |y| > 2. y) = 0 whenever min{|x|. we have found that (xy + (x2 − 4)(y 2 − 4))h(x.c + a b c =0 whenever abc = 1. From the introduction. y. To show this. z) = (x2 + y 2 + z 2 − xyz − 4)g(x. There are polynomials g(x. Gabriel Dospinescu Solution. Example 5. k(y.c + b c 1 1 + k b + . It is exactly xy+ (x2 − 4)(y 2 − 4). we use the classical polynomial long division. y. the function x → x2 − 4 is 29 . The following problem is a nice characterization of the equation (2) by polynomials and also teaches us some things about polynomials in two or three variables. z) Using the hypothesis. b c 2 2 1 Then it is easy to compute a+ . it seems that this is a dead end. Find all polynomials f (x. z). But it is not obvious why any desired polynomial should be of this form. |y|} > 2. z) with real coeﬃcients such that f (x. y = c + with b = . z) with real coeﬃcients such that f whenever abc = 1.But if abc = 1.

IMO Shortlist. This equation can be written as a2 b2 c2 abc + + + =4 yz zx xy xyz and we already recognize the relation u2 + v 2 + w2 + uvw = 4 c b a where u = √ . Find all triples (x. which is not the case. v = √ . We have made use of the second condition. y) = 0 for all x. q such that x2 − 4 = p(x) . that is our polynomial is divisible with x2 + y 2 + z 2 − xyz − 4. z) of positive real numbers such that x+y+z =a+b+c a2 x + b2 y + c2 z + abc = 4xyz Titu Andreescu. y) = k(x. c > 0. Consequently. y. for each y with |y| > 2 we have h(x. But this means that h(x. 1995 Solution. y. w = √ . b. that is. 30 .√ not rational. we can yz xy zx ﬁnd an acute-angled triangle ABC such that u = 2 cos A. y) = k(x. Example 6. there aren’t polynomials p. so we use the ﬁrst one to deduce that √ √ √ x + y + z = 2 xy cos C + 2 yz cos A + 2 zx cos B. According to example 3. the following problem and the featured solution prove that sometimes an eﬃcient substitution can help more than ten complicated ideas. y) = 0 for all x. But this is easy because if such polynomials existed. Let a. O a diﬀerent kind. v = 2 cos B. w = 2 cos C. We try to use the information given by the second equation. than each q(x) zero of x2 − 4 should have even multiplicity.

Example 7. We will see that example 3 helps us ﬁnd a nice geometric solution to this inequality. √ x. 2 y= c+a . y. India. we ﬁnd that √ √ √ y x z = = sin A sin B sin C Now we square these relations and we use the fact that a cos A = √ . We have discussed the following very diﬃcult problem in the chapter ”An useful substitution”. 2 z= a+b 2 b cos B = √ . z satisfy xy + yz + zx + xyz = 4. 1998 31 . 2 zx c cos C = √ . This could be another possible solution of the problem. Notice that the condition √ √ √ x + y + z = 2 xy cos C + 2 yz cos A + 2 zx cos B is the equality case in the lemma stated in the solution of the following problem.Trying to solve this as a second degree equation in discriminant √ √ −4( y sin C − z sin B)2 . Combining the last two relations. we ﬁnd the Because this discriminant is nonnegative. Prove that if the positive real numbers x. 2 yz The conclusion is: x= b+c . we infer that √ y sin C = √ z sin B and √ x= √ y cos C + √ z cos B. then x + y + z ≥ xy + yz + zx. Hence there is a unique triple that is solution to the given system. 2 xy and it is immediate to see that this triple satisﬁes both conditions.

cos A cos B cos C Hence we need to prove that 2 cos B cos C 2 cos A cos C 2 cos A cos B + + ≥ 2(cos2 A+cos2 B+cos2 C). Let us write the condition xy + yz + zx + xyz = 4 in the form √ xy 2 + √ yz 2 + √ zx + 2 √ xy · √ yz · √ zx = 4. Q.Solution. y. 32 . Proof of the lemma. R on the lines AB. BC. It is not diﬃcult to observe that at ﬁrst glance. then x2 + y 2 + z 2 ≥ 2yz cos A + 2zx cos B + 2xy cos C. y. which is obviously true. z are arbitrary real numbers. z) = 2 cos B cos C 2 cos A cos C 2 cos A cos B . Q. we can use the result from example 3 and we deduce the existence of an acute-angled triangle ABC such that  √  yz = 2 cos A   √ zx = 2 cos B  √   xy = 2 cos C We solve the system and we ﬁnd the triplet (x. CA. respectively. Now. R and do not lie on the sides of the triangle. such that AP = BQ = CR = 1 and P. Lemma. If ABC is a triangle and x. Let us consider points P. the condition xy + yz + zx + xyz = 4 it’s not the same as the equation (3). . Then we see that the inequality is equivalent to − → −→ − − → (x · AP + y · BQ + z · CR)2 ≥ 0. cos A cos B cos C This one is a hard inequality and it follows from a more general result.

Indeed. since f (x2 ) = f 2 (x) − 2 implies that f (x) > 2 for all x. √ f 2 (x) = f (x2 ) + 2 > 2 + 2. = 2 (the last equality being immediate for a beginner in analysis). Thus. it suﬃces to write x x2 = xy . Example 8. ∞) satisfying f (x)f (y) = f (xy) + f x y . till now nothing related to our theme. Indeed. But of course. this type of identities does not appear only in inequalities. These relations should now ring a bell! It seems that we are searching for something like f (x) = 1 xk + k . this is going to be √ easy. by taking x = y = 1 y x x we obtain f (1) = 2 and then f (x2 ) = f 2 (x) − 2. observe that by symmetry in x. y xy y2 = x .. Let’s make another x small step: proving that f (x) ≥ 2 for all x. Yet. We are going to discuss two problems in which the identity is very well masked. Find all continuous functions f : (0. but still far from the solution. we just have to take x= 2 cos B cos C cos A y= 2 cos A cos C .. cos B z= 2 cos A cos B cos C in the above lemma and the problem will be solved. We are right. Sankt Petersburg Solution. we ﬁnd that for all x we have f (x) > 2+ 2+ √ 2 + . y. y x y 33 . y we must x y 1 have f =f and so f (x) = f .The lemma being proved. ∞) → (0. Wrong! Let’s observe that f (x2 ) + f (y 2 ) = f (xy)f for all x. Next. Repeating this argument. First of all.

Example 9. Let (an )n≥0 be a non-decreasing sequence of positive integers such that a0 = a1 = 47 and a2 + a2 + a2 − an−1 an an+1 = 4 for all n ≥ 1. And ﬁnally. Now. The g(x) above relation implies of course that g(xy) = g(x)g(y). Fn+1 = Fn + Fn−1 . Titu Andreescu Solution. Using also the fact that f (x) ≥ 2. let us make one more step and write f 2 (x) + f 2 (y) − 4 = f (x2 ) + f (y 2 ) = f (xy)(f (x)f (y) − f (xy)). we deduce the existence of a con1 tinuous function g : (0.With this information. We are now on the right track. these are all solutions of the equation (the veriﬁcation of the identity is immediate for this class of functions). which shows that (ln xn )n≥0 is a Fibonacci-type sequence. ∞) such that f (x) = g(x) + . n−1 n n+1 Prove that 2 + an and 2 + √ 2 + an are perfect squares for all n ≥ 0. we 1 write an = xn + . we deduce that xn = xFn . Since x0 = x1 . thus h(x) = kx for a certain k. an apparently inextricable recursive relation. do more: who is x0 ? And the answer x0 = 2 Let us remark that 2 √ 1 x0 + √ = 49 x0 34 . ∞) → [1. with xn > 1. we have to 0 √ 47 + 472 − 1 won’t suﬃces. The the given condition becomes xn xn+1 = xn xn−1 (we have used here explicitly that xn > 1). By considering h(x) = ln g(ex ). since we ﬁnd that f 2 (x) + f 2 (y) + f 2 (xy) = f (x)f (y)f (xy) + 4. Using the idea from the chapter with real equations. where F0 = F1 = 1. This shows that g(x) = xk and that our thoughts were right. we obtain that h is a continuous solution of Cauchy’s functional equation f (x + y) = f (x) + f (y).

λ2 λ4 + 1 ∈R λ4 1 ∈ R for all k ∈ R.from where we ﬁnd that √ Similarly. since √ an + 2 = (λ4Fn + λ−4Fn )2 and 2 + 2 + an = (λ2Fn + λ−2Fn )2 . Find all triples x. Romanian Inter-county Contest 35 . we obtain √ 4 x0 = = λ2 that is x0 = λ8 . (2 + c)(2 + a) Cristinel Mortici. z > 0 such that x2 + y 2 + z 2 + xyz = 4. y. All we are left to prove is that λ2k + isn’t diﬃcult. solutions to the system: x2 + y 2 + z 2 = xyz + 4 xy + yz + zx = 2(x + y + z) 2. y. Let x. we obtain that √ 4 1 x0 + √ = 3. z of positive real numbers. 1 ∈ R. And so we have found the general formula an = λ8Fn + λ−8Fn . 4 x 0 √ 1+ 5 2 2 1 x0 + √ = 7. since λ2 + and λ2(k+1) + 1 λ2(k+1) = λ2 + 1 λ2 λ2k + 1 λ2k − λ2(k−1) + 1 λ2(k−1) . But this λ2k Problems for training 1. x0 Solving the equation. And now the problem becomes easy. Prove that (2 − a)(2 − b) + (2 + a)(2 + b) (2 − b)(2 − c) + (2 + b)(2 + c) (2 − c)(2 − a) = 1.

Prove that 3 1 1 1 √ +√ +√ y x z 2 ≥ (x + 2)(y + 2)(z + 2). Titu Andreescu. then (a − 2)(b − 2) + (b − 2)(c − 2) + (c − 2)(a − 2) ≥ 0. Let x. Gabriel Dospinescu 7. Solve in positive integers the equation (x + 2)(y + 2)(z + 2) = (x + y + z + 2)2 . Titu Andreescu 9. Romanian Inter-county Contest 5. c ≥ 0 satisfy the condition |a2 +b2 +c2 −4| = abc. c) of positive real numbers. b. Titu Andreescu. Prove that in any acute-angled triangle the following inequality holds cos A cos B 2 + cos B cos C 2 + cos C cos A 2 + 8 cos A cos B cos C ≥ 4. z > 0 such that xy + yz + zx + xyz = 4. Prove that if a. solutions to the system a2 + b2 + c2 + abc = 4 a+b+c=3 Cristinel Mortici. y. Let n > 4 be a given positive integer. n 36 . Prove that in any triangle the following inequality holds sin A B C + sin + sin 2 2 2 2 ≤ cos2 A B C + cos2 + cos2 . Gazeta Matematica 4. MOSP 2000 8. 2 2 2 6. b. Find all triples (a. y) such that xy − (x + y)2 = n − 4.3. Find all pairs of positive integers (x.

∞) → (0. y. b. Titu Andreescu 11. b. m) with sum 2002 and for which the system  x y  + =k  y  x        y z + =l  z y       z    +x =m x y Titu Andreescu. c ≥ 0 satisfy a2 + b2 + c2 + abc = 4 then 0 ≤ ab + bc + ca − abc ≤ 2. USAMO 2001 37 . Prove that if a.Titu Andreescu 10. c ≥ 2 satisfy the condition a2 +b2 +c2 = abc+4. Let the sequence (an )n≥0 . l. Titu Andreescu. ∞) with the following properties: √ √ √ a) f (x)+f (y)+f (z)+f (xyz) = f ( xy)f ( yz)f ( zx) for all x. Marian Tetiva 14. IMO Shortlist 2004 13. then a + b + c + ac + bc ≥ 2 (a + b + c + 3)(a2 + b2 + c2 − 3). Find all functions f : (0. Prove that 2 + 2 + 2an is n n−1 a perfect square for all n ≥ 0. z. Hojoo Lee. has real solutions. Find all triplets of positive integers (k. Prove that if a. proposed for IMO 2002 12. b) if 1 ≤ x < y then f (x) < f (y). where a0 = a1 = 97 and an+1 = √ an−1 an + (a2 − 1)(a2 − 1) for all n ≥ 1.

. it is easy to prove the following properties of vp (a): 1) min{vp (a). . c)·lcm(c. a2 . Here. let us repeat the above idea in terms of vp (a): we have a|b if and only if for any prime number p we have vp (a) ≤ vp (b) and we have a = b if and only if for any prime number p. Euler. c? Then one thing is sure: the above methods fail. p − 1} 38 . or Wilson. vp (b)} ≤ vp (a + b) ≤ max{vp (a). Of course. a2 . vp (an )}. . an )) = max{vp (a1 ). . such as Fermat. This follows from the fact that there are multiples of p. vp (a2 ). vp (a2 ). . . The other equality is not diﬃcult. Indeed. . p p p p−1 sp (n) is the sum of digits of n when written in base b. c)2 |lcm(a. . . a) for any positive integers a. . then it is enough to prove that the exponent of any prime number in the decomposition of a is at least the exponent of that prime number in the decomposition of b. b. But what do we do when we have to prove for example that lcm(a. Less straightforward is n n 5). 2 are p p multiples of p2 and so on. . b. . . b)·lcm(b. . Also. then vp (a) = 0. . Now. proving divisibility reduces to congruences and the famous theorems from this ﬁeld. a1 . another smart idea appears: if we have to prove that a|b. if p doesn’t divide a. let us write n = a0 + a1 p + · · · + ak pk . Yet. . an )) = min{vp (a1 ). . . . 1. Observe that 3) and 4) are simple consequences of the deﬁnitions. Some other useful properties of vp (a) are: 3) vp (gcd(a1 . vp (an )} and n − sp (n) n n n 5) vp (n!) = + 2 + 3 +· · · = whenever p|n. b.LOOK AT THE EXPONENT! Most of the times. . 4) vp (lcm(a1 . . For simplicity. vp (a) = vp (b). . vp (b)} 2) vp (ab) = vp (a) + vp (b) for any positive integer numbers a. . let us denote by vp (a) the exponent of the prime number p in the decomposition of a. . ak ∈ {0. where a0 .

(3a + 3b)!(2a)!(3b)!(2b)! Example 1. as well as the ﬁrst observations. let us take a prime p and let us apply formula 5). p−1 we ﬁnd exactly 5).and ak = 0. AMM 6514 Solution. When we write n n n + 2 + 3 + . Richard Askey.. Prove that ∈ Z for (2a + 3b)!(a + 2b)!(a + b)!a!(b!)2 any positive integers a. Now. We have chosen with intention the ﬁrst problem (the classical one) a very nasty one. so that the reader doesn’t think that all the above formulas were for nothing and because it oﬀers us the opportunity to prove a very nice inequality. b. p p p we write in fact k≥1 n pk and this sum has clearly a ﬁnite number of non-zero terms. There are hundreds of variants of it in all contests around the world and in all elementary magazines.. let’s see some concrete results. We ﬁnd that vp ((3a + 3b)!(2a)!(3b)!(2b)!) = k≥1 3a + 3b 2a 3b 2b + k + k + k pk p p p and also vp = ((2a + 3b)!(a + 2b)!(a + b)!a!(b!)2 ) = k≥1 a + 2b a+b a b 2a + 3b + + + k +2 k k k k p p p p p 39 .. Enough with the introduction. Let us see. Then n n + 2 +· · · = a1 +a2 p+· · ·+ak pk−1 +a2 +a3 p+· · ·+ak pk−2 +· · ·+ak p p and now using the formula 1 + p + · · · + pi = pi+1 − 1 . let us clearify something. First.

Prove that a = b. 1 or 2. b be positive integers such that a|b2 . y < 1. With the substitution x = k . b7 |a8 . the condition b4n+3 |a4n+4 implies vp (a) ≥ vp (b) and so vp (a) = vp (b). a5 |b6 . a5 |b6 . . Let a. since another beautiful problem is waiting. 40 . we see that it suﬃces to prove the inequality only for 0 ≤ x. Why is the new inequality easy? Because we can easily compute all terms. We won’t continue studying these cases. After this operation. b3 |a4 . This isn’t easy. . but with another useful idea the inequality will become easy. is the same as a4n+1 |b4n+2 and b4n+3 |a4n+4 for all natural number n.Of course. . we have to prove that for any nonnegative real numbers x. Solution. We see that the hypothesis a|b2 . after splitting in some cases. p b y = k . it is enough to prove that for each k ≥ 1 the term corresponding to k in the ﬁrst sum is greater than or equal to the term a corresponding to k in the second sum. so that to see when [2{x}]. The idea is that [3x + 3y] = 3[x] + 3[y] + [3{x} + 3{y}] and similar relations for the other terms of the inequality. . . 4n + 1 Similarly. [2{y}] are 0. that is vp (a) ≤ lim n→∞ 4n + 2 vp (b) = vp (b). y we p have [3x + 3y] + [2x] + [3y] + [2y] ≥ [2x + 3y] + [x + 2y] + [x + y] + [x] + 2[y]. b3 |a4 . . . The conclusion follows: a = b. b7 |a8 . [3{y}]. Example 2. Let us take a prime p and try to prove that vp (a) = vp (b). But the relation a4n+1 |b4n+2 can be interpreted as (4n + 1)vp (a) ≤ (4n + 2)vp (b) for all n.

y. Solution. x}. Example 4. c)2 ) = 2 max{x. we need another observation: the sum of digits of a number A when written in binary is at most the number of digits of A in base 2. which is 1 + [log2 A] (this follows from the fact that 2k−1 ≤ A < 2k . obviously true. c. so we should better see what happens if a! · b!|n!. So. since we may assume that x ≥ y ≥ z (the symmetry allows us this supposition) and the inequality becomes 2x + y ≥ 2x. so probably it’s time to solve it. y} + max{y. y. y. we need to prove that max{x. b. b) · lcm(b. which are all based on the observations from the beginning of the discussion. y = vp (b). b. Prove that there exists a constant c such that for any positive integers a. c) · lcm(c. Now. b. z} and vp (lcm(a. a)) = max{x. which can be translated as a − s2 (a) + b − s2 (b) ≤ n − s2 (n) < n. z} + max{z. c)2 |lcm(a.We have mentioned in the beginning of the discussion a nice and easy problem. We have vp (lcm(a. It is time for some diﬃcult problems. Then v2 (a!) + v2 (b!) ≤ v2 (n!). n that verify a! · b!|n! we have a + b < n + c ln n. b) · lcm(b. z} for any nonnegative integers x. Prove that lcm(a. Paul Erdos Solution. z} + max{z. But this is easy. Example 3. although for sure the reader has already done this. we have found almost exactly what we needed: a + b < n + s2 (a) + s2 (b). a) for any positive integers a. This time the other formula for vp (n!) is useful. z. x} ≥ 2 max{x. where x = vp (a). where 41 . y} + max{y. Of course. there is no reasonable estimation of this constant. So. c) · lcm(c. b. z = vp (c). Let p an arbitrary prime number.

Prove that for any natural number n. Of course. k=0 Solution. It combines diverse ideas and techniques and the result is at least beautiful. n! is a divisor of n−1 (2n − 2k ). let us take a prime number p. that appeared in lots of mathematics competitions. 6) = 1. Thus this set is actually a good choice. The following problem appeared in Kvant as a hard problem. even easier. Is there an inﬁnite set of positive integers such that no matter how we choose some elements of this set. we take p ≤ n (otherwise doesn’t divide n!). let us see what happens with p = 2. but another one. we have the estimations a + b < n + s2 (a) + s2 (b) ≤ n + 2 + log2 ab ≤ n + 2 + 2 log2 n (since we have of course a.k is the number of digits of A in base 2). where (y. This is surely not a power of exponent at least 2. So. their sum is not an integer power of exponent at least 2? Kvant Solution. It took quite a long time before an olympic found an extraordinary solution. b ≤ n). First. Let us take A = {2n · 3n+1 |n ≥ 1} If we consider some diﬀerent numbers from this set. their sum will be of the form 2x · 3x+1 · y. for the argument to be non-trivial. So. We shall not present his solution. Example 5. The following problem shows the beauty of elementary numbertheory. And now the conclusion is immediate. since otherwise the exponent should divide both x and x + 1. This one is also a classic problem. We have v2 (n!) = n − s2 (n) ≤ n − 1 42 . Example 6.

43 . We have p|2p−1 − 1 from Fermat’s theorem. we conclude that n−1 v2 k=0 (2n − 2k ) ≥ v2 (n!) and now the problem is solved. let us assume that p > 2. p−1 From these two inequalities. Now. p−1 But we know that v2 (n!) = n − sp (n) n−1 n ≤ < p−1 p−1 p−1 and since v2 (n!) ∈ R. from the above remarks we infer that n−1 n v2 k=0 (2n − 2k ) = k=1 v2 (2k − 1) ≥ 1≤k(p−1)≤n v2 (2k(p−1) − 1) ≥ card{k|1 ≤ k(p − 1) ≤ n} Since card{k|1 ≤ k(p − 1) ≤ n} = we have found that n−1 n . p−1 v2 k=0 (2n − 2k ) ≥ n . so we are done with this case. so we also have p|2k(p−1) − 1 for all k ≥ 1. we must have v2 (n!) ≤ n . Now. n−1 (2n − 2k ) = 2 k=0 n(n−1) 2 n (2k − 1) k=1 and so.and also n−1 n−1 v2 k=0 (2 − 2 ) n k = k=0 v2 (2n − 2k ) ≥ n − 1 (since 2 n − 2k is even for k ≥ 1).

. . . we deduce that (n1 + 1) . . (nk − 1)nk + 1 is congruent to 2 modulo 4 and thus v2 (nk !) = n+1 ≤ v5 (nk !)+1. suppose that y = 0. . (nk − 1)nk + · · · + (nk−1 + 1) .Diophantine equations can also be solved using the methods employed in this topic. nk !. (nk − 1)nk + 1 is relatively prime with 10 and it follows nk nk = j for all that v2 (nk !) = v5 (nk !). We have 10n ((n1 + 1) . A crucial observation is that if nk > nk−1 + 1. Prove that the equation 1 1 1 1 = + + ··· + 10n n1 ! n2 ! nk ! does not have integer solutions such that 1 ≤ n1 < n2 < · · · < nk . But then. Thus. First of all. Let us write nk = 2x · 5y . So. . Again this implies v2 (nk !) = v5 (nk !) and we have seen that this gives no solution. . actually x = 0. Here is a diﬃcult one. impossible. . . y are positive. . . A veriﬁcation 2j 5 by hand shows that there is no solution in this case. Suppose we have found a solution of the equation and let us consider P = n1 !n2 ! . then (n1 + 1) . suppose that x. . Example 7. (n1 + 1) . Then (n1 +1) . This implies of course that j 2 5 nk nk j (because we clearly have > j ) and so nk ≤ 3. . we 2 5 44 . . Since nk is a power of 5. . . nk = nk−1 + 1. Next. It follows nk nk that ≤ 1+ and thus nk ≤ 6. (nk − 1)nk + · · · + (nk−1 + 1) . . (nk − 1)nk + 1 is again odd and thus we ﬁnd again that v2 (nk !) = n ≤ v5 (nk !). (nk − 1)nk + · · · + (nk−1 + 1) . given in a russian olympiad. . (nk − 1)nk + · · · + (nk−1 + 1) . . . (nk − 1)nk + 1 is odd and thus v2 (nk !) = n ≤ v5 (nk !). So. taking into account that nk is a power of 5. . (nk −1)nk +· · ·+(nk−1 + 1) . Tuymaada Olimpiad Solution. (nk − 1)nk + 1) = nk ! which shows that nk divides 10n . .

the proof is tricky. . it will be proved that 2 11 43 is a solution of the problem. We will let to the reader the job to verify that 2 · 11 · 43 is a solution (and especially the job to ﬁnd how we arrived at this number) and also the exercise to prove that there are actually inﬁnitely many such numbers. This problem seemed to appear for the ﬁrst time in AMM . Choose that pi which minimizes this quantity and write pi = 1 + 2ri mi with mi odd. Spencer. i−j 1≤i<j≤n Armond Spencer. which is clearly impossible. Then of course we have n ≡ 1 (mod 2mi ). an the number ai − aj is an integer. it appeared in various contests. We have 22 −1 ≡ 2 2m tmi mt ≡ −1 (mod pi ) thus we surely have ≡ 2(pi −1)t ≡ 1 (mod pi ) (the last congruence being derived from Fermat’s theorem). Example 8. . Yet. in which the idea of looking at the exponents really saves us. nk−1 = and a quick research of all possibilities shows that there are no solutions. Hence we can write n − 1 = 2m t. Thus. Prove that for any n > 1 we cannot have n|2n−1 + 1. Example 9. Prove that for any integers a1 .. Thus pi = 2. Although very short. . . Note. Let n = i=1 pk i i where p1 < · · · < ps are prime numbers. AMM E 2637 45 .. After the quadratic reciprocity law topic. proposed by Armond E. A tricky APMO problem asked once upon a time to prove there is a number 2 < n < 2000 such that n|2n + 2.ﬁnd that nk = 5. We continue with a very nice and hard problem. Proving this turns out to be a diﬃcult problem and this was proved for the ﬁrst time by Sierpinski. In the last years. a2 . s Solution. The idea is to look at v2 (pi − 1). small veriﬁcations show that all such numbers are even. the given equation has no solution in natural numbers.

Then we have   pk −1 bi  = Npk (ai − aj ) . of course.Solution. . 1. for each i ∈ {0. Now. . Since  vp   (here Nx  y∈A 1≤i<j≤n  (ai − aj ) = k≥1  Npk  1≤i<j≤n  (ai − aj )  y  is the number of terms from the sequence A that are multiples of x) and  vp  1≤i<j≤n  (i − j) = k≥1  Npk  1≤i<j≤n  (i − j) . So. if i = 0 we have 1 ≤ j ≤ ). 2. pk − 1}. . we consider a prime number p and we prove that for each k ≥ 1. 2 1≤i<j≤n i=0 n+i n+i (there are numbers k p pk congruent with i (mod p) between 1 and n. n} such that aj ≡ i (mod pk ). . . any of them is of the form n−i n i + jp. the problem will be solved if we prove our claim. . . let us ﬁx k ≥ 1 and let us suppose that there are exactly bi indices j ∈ {1. p p     k We see that if ai = i. with 0 ≤ j ≤ . then bi = p −1 Npk  and it suﬃces to prove that pk −1 (i − j) = i=0 ni pk  1≤i<j≤n 2  i=0 bi 2 pk −1   ≥ i=0 ni pk  . 2 46 . This time. there are more numbers divisible by pk in the sequence of diﬀerences (ai −aj )1≤i<j≤n than in the sequence (i−j)1≤i<j≤n . .

x2 . . Let us start by writing a = . . . . . so it is true. . b are also integers. pk − 1} such that x0 = x1 = · · · = xj and xj+1 = xj+2 = · · · = xpk −1 = x0 + 1. it is time for a challenge. This easily implies that the minimal n+i n-tuple is in fact and the problem is solved. we must have xpk −1 ≤ x0 + 1 and from here it follows that xi ∈ {x0 . . but for which x1 x0 + 1 + 2 2 < x0 2 + x1 2 + ··· + xpk −2 2 xpk −2 2 + xpk −1 − 1 2 xpk −1 . 1. Then prove that a. x1 . . Mathlinks Contest x y Solution.pk −1 Finally. b two diﬀerent positive rational numbers such that for inﬁnitely many numbers n. pk i=0. where x. an − bn is integer. z are z z diﬀerent natural numbers relatively prime. . 1. since it is equivalent with xpk −1 > x0 +1. Let a. . y. . Thus. x0 + 1} for all i ∈ {0. . .Now. pk − 1}. 2 + ··· + + The last inequality is true. when 2 pk −1 pk −1 xi = n (it is clear that i=0 i=0 bi = n = ni from the deﬁnition of bi ). Example 10. xpk −2 . But this contradicts the minimality of (x0 . there is j ∈ {0. So. xpk −1 − 1) which has the sum of the components n. 2. let us suppose that pk i=0 x1 ≤ x2 ≤ · · · ≤ xpk −1 is the n-tuple which attains the minimal pk −1 value (such a n-tuple exists since the equation i=0 xi = n has a ﬁnite number of solutions). . . For this. observe that we are practically asked to ﬁnd the minimal pk −1 value of i=0 pk −1 xi . . We know thus that z n |xn −y n 47 . . xpk −1 ). . then we consider the n-tuple (x0 + 1. Gabriel Dospinescu. . . 2. If xpk −1 > x0 + 1. b = . x1 .

a simple reason being the inequality 2un ≤ v2 (x + y) + v2 (x − y) + un − 1. then let n such that 2n |xn − y n and write n = 2un vn . it obviously follows that it can’t divide y. Let x = tu. assume that z > 1 and take p a prime divisor of z. Similarly. Assuming that p does not divide x. From the identity x2 un v n − y2 un v n = (xvn − y vn )(xvn + y vn ) . Then for any n in M we will have p|xn −y n . ii) If p is odd. Hence (un )n∈M takes only a ﬁnite number of values and from (∗) it follows that (vn )n∈M also takes a ﬁnite number of values. where vn is odd. 48 . we can prove that v2 (xvn + y vn ) = v2 (x + y). x. then let d the smallest positive integer k such that p|xk −y k . (2un )n∈M is bounded. But xvn −1 + xvn −2 + · · · + xy vn −2 + y vn −1 is obviously odd (since vn . (x2 un −1 v n − y2 un −1 v n ) it follows that un −1 v2 (x − y ) = v2 (x n n vn −y )+ k=0 vn v2 (x2 kv n + y2 kv n ). Now. Let M be the set of those numbers n. hence v2 (xvn − y vn ) = v2 (x − y). y = tv. . . y are odd). that is M is ﬁnite. Since for k > 0 we have x2 we ﬁnally deduce that 2un vn ≤ v2 (xn − y n ) ≤ v2 (x + y) + v2 (x − y) + un − 1 (∗) kv n + y2 kv n ≡ 2 (mod 4).for inﬁnitely many numbers n. We have thus two cases: i) If p = 2. Consequently.

false.d) and by the choice of d. that is p|x. Let R the inﬁnite set of those numbers m. Now. Obviously. so we should have p|A. B = y d . where j is relatively prime with p.where (u. since the ﬁnal conclusion will be obvious in any other case). un −v n ) = u(n. v) = 1. From p2 |Aj − B j we have Ajp k−1 (p−1) + Ajp k−1 (p−2) B jp k−1 + · · · + B jp k−1 (p−1) ≡ pAjp k−1 (p−1) (mod p2 ). . Let A = xd .d) |x(n. Moreover. Apj − B pj . For any m in R we have m ≤ vp (Am − B m ). . An essential observation is that we cannot have jpk − B jpk 2| A p for a certain k > 1. let us write m = pi j. Take now n in M and write it in the form n = md. we can write Aj = B j + wp and then a simple Let us prove now that we cannot have p2 | computation using Newton’s binomial formula shows that Apj − B pj = Aj(p−1) + Aj(p−2) + · · · + B j(p−1) Aj − B j ≡ pB j(p−1) + p − 1 j(p−2) 2 B p ≡ pB j(p−1) 2 (mod p2 ) and thus it would follow that p|B. Indeed. false. jpi−1 A − Bj A − B jpi−1 j j i i (we have assumed that i > 1. Then pm |pn |xn − y n = Am − B m and this happens for inﬁnitely many numbers m. othAj − B j erwise (since p|A − B). Indeed. that is p|y. tuv is not a multiple of p. We clearly have A −B m m Apj − B pj Ajp − B jp = (A − B ) j . we must have d|n. with m natural. p|A − B. otherwise we would have Ajpk−1 − B jpk−1 k k p2 |Ajp − B jp ⇒ p2 |Apj − B pj (Euler’s theorem). Yet. It follows then that p|(ud −v d .d) −v (n. we also have p2 |Ajp k−1 (p−1) + Ajp k−1 (p−2) B jp k−1 + · · · + bjp k−1 (p−1) .d) −y (n. 49 .

. d). . Prove that gcd(a. . a) for any positive integers a. b1 . Prove the identity lcm(a. c. Let m − lcm[b1 . Prove that gcd a1 m1 a2 m2 ak mk . c)2 gcd(a. y. a2 . Polish Mathematical Olympiad 3. . . . . . USAMO. . we infer that Aj−1 + Aj−2 B + · · · + B j−1 ≡ jAp−1 ≡ j which shows that vp (Aj − B j ) = vp (A − B). b are integers. . . . a2 . IMO Shortlist 1974 50 (mod p). . . This shows that z = 1 and a. bi ) = 1 for all i ∈ {1.. d be positive integers such that ab = cd. a) gcd(a.. k}. b. Problems for training 1. c) · gcd(c. . Thus. b) · lcm(b. c. b1 b2 bk = gcd(a1 . Let a1 . ak . . which is clearly impossible. c) · lcm(c. we have shown that in this case we must have m ≤ vp (Am − B m ) ≤ vp (Aj − B j ) + i. b. Let a. contradiction with the fact that x. we must have p|x and p|y. c. for inﬁnitely many numbers m we have m ≤ vp (A − B) + [log2 m]. Thus. b. . ak ). b2 . . . Using again the fact that A ≡ B (mod p). b..After all. c) · gcd(a. d) = a · gcd(a. 1972 2. 2. b2 . c)2 = lcm(a. . bk ]. b) · gcd(b. . .. bk be positive integers such that gcd(ai . z are relatively prime. b.

then n! divides a(a + b)(a + 2b) . 1991 11. Let 0 < a1 < · · · < an be integers. . n!n n n ∈ R. Porter. TST 1990 Romania 9. .M Grassl. ... Prove the identity (n + 1)lcmk=0.n for any positive integer n.. 1 2 n if and only if n + 1 is a prime. Prove that the product of the numbers between 21917 + 1 and 21991 − 1 is not a perfect square. R. . AMM E 3123 6. Peter L Montgomery. 2. Prove that for any natural number n we have (n2 )! n n n+1 2n − 1 . n + 1) 2 k=1 ak = k−1 bk and k=1 (2 )! = k=1 ak bk !. n n n n equals the least common multiple of the numbers . Show that if n is a positive integer and a and b are integers.. 51 . Let n such that 2n−2005 |n!. Komal 10. .4. T. AMM E 2686 7. Find the maximal value of the number m for which we can ﬁnd the integers 0 < b1 < · · · < bm such that n m n m n k = lcm(1. Prove that the least common multiple of the numbers 1.. . (a + (n − 1)b)bn − 1. . Tournament of the Towns.. 5. . . Prove that this number has at most 2005 non-zero digits when written in base 2. Prove that for any n ∈ N we have n!(n + 1)!(n + 2)!|(3n)!. Laurentiu Panaitopol. . Gabriel Dospinescu 8. . 2.

prove that x2n ≥ 2n − n + 1 (hint: try to prove ﬁrst the identity n−1 2 22 2n 2n 1 ). . Prove that k!k 2 +k+1 divides (k 3 )!. . we have a1 k · · · + an < 1 + [ai ]. . . 16. Prove that the product of at most 25 consecutive integers is not a square. 2n−1 AMM 17. . goes to inﬁnity as n → ∞. Narumi’s theorem + ··· + an k < 1. an > 0 such that whenever k is a prime number of a power of a prime number.IMO Shortlist. + + ··· + = n−1 1 2 n n k=0 k Adapted after a Kvant problem 18. Prove that there is a unique index i ∈ {1. y be relatively prime diﬀerent natural numbers. Find the exponent of 2 in the decomposition of the number 2n+1 2n − 2n . 1985 12. . Poland Olympiad 13. . . AMM 14. Prove that (xn )n≥1 the exponent of 2 in the decomposition of 22 2n 2 + · · · + . Let a1 . 2. Let x. Even the numerator of + 1 2 n more. n} such that a1 + 52 . Prove that for inﬁnitely many primes p the exponent of p in xp−1 − y p−1 is odd.

we will try to present a unitary view over the properties of some classes of primes and also some classical results related to representations as sum of two squares. These things are not new. respectively. Prove that A is a subset of C. which says that if n is a positive integer and a is relatively prime with √ n. So. we prefer to make some notations before discussing the problems. then there exist integers 0 < x. let C be the set of all numbers which can be written as the sum of two perfect squares.PRIMES AND SQUARES The study of the properties of the prime numbers is so well developed (yet. Also. Example 1. we will discuss as usual some non-classical and surprising problems. This is not a trivial property and we will see a beautiful proof for this theorem of Fermat. So. we need to prove that any prime number of the form 4k+1 is the sum of two squares. √ √ with 0 ≤ x. we will consider A. Thus. in the end of the unit. we have a list of ([ n] + 1)2 > n numbers and it follows that two numbers among them give the same remainder when divided by n. y ≤ [ n]. which is far from easy. B. many old conjectures and open questions wait their solution). y ≤ n such that xa ≡ ±y (mod n) for a suitable choice of the signs + and −. So. C. that some properties have become classical and need to be known. Our purpose is to present some classical things related to A. It is not diﬃcult to see 53 . but they must be included in the mathematical culture of a serious problem-solver. but the theorem itself is a diamond. B the sets of all prime numbers of the form 4k + 1 and 4k + 3. Indeed. The most spectacular property of the set A is surely the fact that any element is the sum of two squares of positive integers. let us consider all the pairs xa − y. Solution. Yet. The proof is simple. In this unit. let them be ax1 − y1 and ax2 − y2 . don’t skip this unit! Since we will use some facts several times in this paper. We will use a very nice theorem of Thue.

If we take x = x1 − x2 . we will discuss them all in a single example. we can write x2 ≡ −y 2 (mod p). it is obvious that xy is not a multiple of p. . Because p|x2 + y 2 . it is time now to study some properties of the set B. which is clearly impossible. it is clear that p and n are relatively prime. y < p. p) = (y. y < p (since p ∈ R) such that p|n2 x2 − y 2 . any number of the form n2 + 1 has only prime factors that belong to A or are equal to 2. Let p ∈ B and suppose that x. . This settles the ﬁrst question.that we may assume that x1 > x2 (we certainly cannot have x1 = x2 or y1 = y2 ). . Let us focus on the ﬁrst question. Hence we can apply Thue’s theorem and we ﬁnd √ √ the existence of positive integers 0 < x. p) = 1 and with Fermat’s theorem. we ﬁnd that 1 ≡ xp−1 ≡ (−1) p−1 2 ≡ (−1) p−1 2 ≡ −1 (mod p). Consequently. Example 2. it remains to prove the third assertion. we ﬁnd that p|x2 + y 2 and because √ 0 < x. all the conditions are satisﬁed and the theorem is proved. y) is not true. we conclude that we have in fact p = x2 + y 2 . Combining this with the observation that (x. since p|n2 + 1. . p−1 2 p−1 2 p− p−1 2 . Now. Why? Because from Wilson’s theorem we have −1 ≡ (p − 1)! (mod p) ≡ 1 · 2 . Now. Suppose that p|(x. Now. The second one follows clearly from the ﬁrst one. y). 54 . (p − 1) ≡ (−1) p−1 2 ! ≡ (2k)!2 2 (mod p) and the claim is proved. The theorem is proved. let us write p = 4k + 1 and observe that we can take n = (2k)!. Indeed. Because p|n2 + 1. Then. y = |y1 − y2 |. We will use now Wilson’s theorem to ﬁnd an integer n such that p|n2 + 1. y are integers such that p|x2 + y 2 . Solution. Conclude that A is inﬁnite and then that B is inﬁnite. Since they are easier. Then p|(x.

qm )2 + 1. p2 . . pn are all the elements of B greater than 3 and consider the odd number N = 4p1 p2 . The conclusion is plain. 1998 Solution. . . Example 3. A number is a sum of two squares if and only if any prime factor of it that also belongs to B appears at an even exponent in the decomposition of that number. India. But since pi is not a divisor of N for any i = 1. As a simple application of the ﬁrst example. which in turn becomes (2x + 1)2 + 1 ≡ 0 (mod 1997). . . since x2 + (x + 1)2 ≡ 0 (mod 1997) is of course equivalent to 2x2 + 2x + 1 ≡ 0 (mod 1997). n the contradiction is reached and thus B is inﬁnite. qm are all the elements of A and then simply apply the result from the second question. The proof is just a consequence of the ﬁrst examples and we will not insist. .Proving that B is inﬁnite is almost identical with the proof that there exist inﬁnitely many primes. Because N ≡ 3 (mod 4). the congruence n2 ≡ −1 (mod 1997) surely has at 55 . . . we consider the following problem. pn + 3. So. . where q1 . Indeed. N must have a prime factor that belongs to B. . we consider this time the number M = (q1 q2 . . let us proceed and reduce the given congruence to this special form. which is surely easy for someone who knows Fermat’s theorem regarding the elements of A and very diﬃcult otherwise. Find the number of integers x ∈ {−1997. . Having presented some basic results that we will use in this unit. . 1997} for which 1997|x2 + (x + 1)2 . . it is time to see how many applications these two examples have. . q2 . . An easy consequence of the previous observations is the following. Indeed. In the same manner we can prove that A is inﬁnite. but this time we must use the second question. It is not diﬃcult to characterize the elements of the set C. Since 1997 ∈ A. This is not diﬃcult. We know that any congruence of the second degree reduces to the congruence x2 ≡ a (mod p). suppose that p1 .

.least a solution. . pn (we can take as many as we need. . . From a previous observation. More precisely. we ﬁnd that there are exactly two solutions with x ∈ {−1997. −1996. Example 4. . Hence we know what we have to do. each of them having a prime factor that belongs to B and has exponent 1. This conclusion can be translated in the following nice problem. . This suggests that the set X is quite RARA. but here the main idea is to use the complete characterization of the set C. . 1997) = 1. . . we take diﬀerent elements of B. we know that the condition that a number is a sum of two squares is quite restrictive. More precisely. The statement of the problem suggests using the Chinese Remainder Theorem. Prove that C doesn’t have bounded gaps. 1996}. no term of which can be written as the sum of two perfect squares. . let them p1 . 1997} such that 1997|x2 + (x + 1)2 . . that we have just discussed: C = {n ∈ R| if p|n and p ∈ B. We will take long sequences of consecutive integers. so is 1997 − n0 and it is clear that it has at most two non-congruent solutions mod 1997. there are exactly two solutions that belong to {1. . then vp (n) ∈ 2R}. Because (2. . 2. the function x → 2x + 1 is a permutation of R1997 and so the initial congruence has exactly two solutions with x ∈ {1. that is there are arbitrarily long sequences of integers. . since B is inﬁnite) and then we look for a 56 . . . . 1996} because if n0 is a solution. −1}. AMM Solution. . In a similar way. . 2. p2 . Therefore there are exactly four numbers x ∈ {−1997. .

Y . but p2 does not divide x + i. from the hypothesis it follows that there exists a positive integer n such that (x + 1)(x + 2) . p − 1.. . Hence n2 ≡ 1 · 2(p − 1) ≡ −1 (mod p). Also. each having the same product of the elements. Prove that a set of p − 1 consecutive positive integers. which is clearly impossible. . . with a nice ﬂavor of elementary mathematics. Solution. . cannot be partitioned into two subsets. But from the second example we know that 57 . then there must be another one divisible by p (since in this case both the products of elements from X and Y must be multiples of p).solution of the system of congruences   x ≡ p1 − 1 (mod p2 )  1     x ≡ p − 2 (mod p2 ) 2 2  . . 2. . . . (x + p − 1) = n2 . the last congruence being true by Wilson’s theorem. (x + n) = y k has been extensively studied by many mathematicians and great results have been obtained. which means that the set of remainders of these numbers when divided by p is exactly 1. Thus. Thus. since pi |xi . . Example 5. x + 2.      x ≡ p − n (mod p2 ) n n The existence of such a solution follows from the Chinese Remainder Theorem. none of these numbers is a multiple of p. . where p ∈ B. . . Let us suppose that the positive integers x + 1. the numbers x + 1. each of them having the same product of the elements. x + p − 1 have been partitioned into two classes X. The Diophantine equation x(x + 1)(x + 2) . If at least one of them is a multiple of p. But these results are very diﬃcult to prove and we prefer to present a related problem. .. the conclusion follows. x + n cannot be written as the sum of two perfect squares. . i Since n is as large as we want. . x + 2. .

Then. The ﬁrst case is when x is odd. Practically. we must also have 4k 4 − 1 ∈ C. the details are not so obvious and. It has become a classical problem due to its beauty and diﬃculty. Prove that the equation x4 = y 2 + z 2 + 4 does not have integer solutions. The technique is better explained in the following two examples. 1999 Solution. we conclude that vp (x4 − 4) is odd and so x4 − 4 does not belong to C. Yet. Using the factorization x4 −4 = (x2 −2)(x2 +2) and the observation that x2 +2 ≡ 3 (mod 4).the congruence n2 ≡ −1 (mod p) is impossible for p ∈ B and this is the needed contradiction. Prove that at least one of the numbers x. y. we have to show that x4 − 4 does not belong to C. y. it is not clear how to begin. Hence we need to ﬁnd an element of B that has an odd exponent in the decomposition of x4 − 4. The results stated in the second example are an useful tool in solving non-standard Diophantine equations. Rookie Contest. Example 7. let us say x = 2k. Let p ∈ B and suppose that x. Moreover. which is not the case). we deduce that there exists p ∈ B such that vp (x2 + 2) is odd. Example 6. We leave to the reader the details. which are trivial. t is a multiple of p. most important. but the basic idea is the same. But since p cannot divide x2 − 2 (otherwise p|x2 + 2 − (x2 − 2). 58 . by working modulo 4. t are integers such that x2p +y 2p +z 2p = t2p . z. The following problem is much more diﬃcult. which is clearly impossible since 4k 4 − 1 ≡ 3 (mod 4) and thus 4k 4 − 1 has a prime factor that belongs to B and has odd exponent. z. Reid Barton. We have thus shown that in any solution of the equation x is even. it worth noting that the equation x2 +y 2 = 4k+3 can be solved directly.

Find the smallest nonnegative integer n for which there exists a non-constant function f : Z → [0. z is odd. In particular q|t2 −z 2 and. Example 8. we deduce that q|pt2(p−1) . 1}. This follows from the hypothesis p ∈ B and the fact that a2 ≡ 1 (mod 4) for any odd number a. ∞) with the following properties: a) f (xy) = f (x)f (y). This implies that at least one of the numbers x. 59 . Because p ∈ B. The best candidate for this factor seems to be t2p − z 2p = (t2 )p−1 + (t2 )p−2 z 2 + · · · + (z 2 )p−1 . t2 − z 2 which is congruent to 3 (mod 4). y. z. If q = p. Suppose that it is z. Because q ∈ B. Thus. Since a2 (mod 4) ∈ {0. hence q|z and also q|x2p + y 2p .Barry Powel. then q|t. we infer that q|(x. We write the equation in the form x2p + y 2p = t2p − z 2p 2 (t − z 2 ) t2 − z 2 and we look for a prime number q ∈ B with an odd exponent in the decomposition of a factor that appears in the right-hand side. z. Since x2p + y 2p ∈ C. y. contradicting the assumption that (x. the latter implies that x. we prove that t is odd. AMM Solution. we obtain that x2p + y 2p + z 2p ≡ 0 (mod 4). we ﬁnd that p|x and p|y. t are relatively prime. t) = 1. another step is required. it t2 − z 2 follows that vq (x2p + y 2p ) is even and so vq (t2 − z 2 ) is odd. y. y. there t2p − z 2p exists q ∈ B such that vq is odd. Now. we observe that it is enough to assume that x. z. Next. y. which is clearly impossible. t) = 1. The conclusion follows. Therefore q = p and so p|x2p + y 2p . because q|(t2 )p−1 +(t2 )p−2 z 2 +· · ·+(z 2 )p−1 . z are even. Supposing the contrary. It’s time for a hard problem. First of all. Hence t is odd.

b) 2f (x2 + y 2 ) − f (x) − f (y) ∈ {0. then f 2 (x2 ) = 2f (x2 ). . 1}. if p|x 1. Thus. . We will prove now that any non-constant function f which veriﬁes a) and b) must be of the form fp . we must have 2f 2 (x) = f (x) for every integer x. 1. Suppose not. We prove that there is 60 . one can easily verify that fp veriﬁes the restrictions of the problem. contradiction. Then. For this n. it is enough to ﬁnd f (p) for any prime p. 2f 2 (x) − f (x) = 2f (x2 + 02 ) − f (x) − f (0) ∈ {0. For any p ∈ B we deﬁne: fp : Z → Z. f (x) ∈ {0. ﬁnd all the functions with the above properties. Since f is non-constant. otherwise Using properties of sets A and B. Solution. we must have f (0) = 0. We will prove now that if f is non-constant and veriﬁes the conditions of the problem. ∞). we must have f (−1) = 1 and f (−x) = f (−1)f (x) = f (x) for any integer x. It is clear that we have f 2 (0) = f (0). then n > 0. respectively. Then 2f (x2 +y 2 ) = f (x)+f (y) and hence 2f 2 (x) = 2f (x2 + 02 ) = f (x) + f (0). We have already seen that f (0) = 0. n = 1 is the smallest number for which there are non-constant functions which verify a) and b). fp (x) = 0. We will use all results proved in the beginning of the note. But if 1 there exists x such that f (x) = . 2 Thus. . since f (xy) = f (x)f (y). Hence fp is a solution of the problem for any p ∈ B. Since f 2 (1) = f (1) and f is non-constant. contradiction. We remind that A and B are the sets of all primes of the form 4k + 1 and 4k + 3. we must have f (1) = 1. 1} for every integer x. . n} for all integers x and y. Since f 2 (−1) = f (1) = 1 and f (−1) ∈ [0. we will prove that for n = 1 there are functions which verify a) and b). f (x) = 0 for any integer x and f is constant. Consequently. Also. So. First.

exactly one prime number p for which f (p) = 0. Then 2f (p2 + q 2 ) ∈ {0. Suppose there is another prime q for which f (q) = 0. 61 . Suppose p ∈ A. 2. This implies that f (x) = f (y) = 0 for any odd numbers x and y and thus f is constant. Then we have 1 = f (1) = f (aq − bp) = 0. is impossible. 1}. Prove that if p ∈ A. we must have f (ap + bq) = f (aq − bp) = 0. we would have f (max{a. Since p and q are relatively prime. there are integers a and b such that aq − bp = 1. there are positive integers a and b such that p = a2 + b2 . contradiction. Therefore p ∈ A ∪ B. According to proposition 2. Consequently. which means f (p2 + q 2 ) = 0. Since 0 ≤ f (x)+f (y) ≤ 2f (x2 + y 2 ) for any x and y. Then f (x) = 0 for any even x and 2f (x2 +y 2 ) = 0 for any odd numbers x and y. But max{a. b} = 1). But it is clear that q < p and thus we have found two distinct primes p and q such that f (p) = f (q) = 0. Hence. there exists a prime number p for which f (p) = 0. So . which. b} and f (q) = 0 (otherwise. except for the order of the terms. b} > 1 and there is a prime number q such that q| max{a. p ∈ B and we have f (x) = 0 for any x divisible by p and f (x) = 1 for any x which is not divisible by p.there is exactly one prime number p for which f (p) = 0. f must be fp and the conclusion follows. Then we must have f (a) = f (b) = 0. Since f is non-constant. contradiction. Then for any integers a and b we must have: 0 = 2f (a2 +b2 )f (p2 +q 2 ) = 2f ((ap+bq)2 +(aq−bp)2 ). Problems for training 1. Let us suppose that p = 2. then it can be represented in exactly one way as the sum of the squares of two integers. as we have already seen. Prove that a positive integer can be written as the sum of two perfect squares if and only if it can be written as the sum of the squares of two rational numbers.

. a2 . Find all positive integers n with the property that the equation n = x2 + y 2 . n are positive integers. y ≡ (2k)!x (mod p). y) = 1 has exactly one solution. 2xε (mod p). Gabriel Dospinescu 6. Find all pairs of positive integers (x. . b have the property that the numbers 15a + 16b and 16a − 15b are both perfect squares. .Euler 3. . . x−y Bulgaria. . Here is another proof of the theorem from example 1. Find all n-tuples of positive integers (a1 . Gabriel Dospinescu 9. The positive integers a. Find all pairs of positive integers (m. Prove that the number 4mn − m − n cannot be a perfect square if m. n) such that m2 − 1|3m + (n! − 1)m . IMO 1984 Shortlist 8. |y|} < and 2 2k 2 + y2. 4. What is the least possible value that can be taken on by the smallest of the two squares? IMO CE AN? 7. y) such that the number x2 + y 2 is a divisor of 1995. 1995 62 . y ∈ Z such that max{|x|. Suppose p that p = 4k + 1 ∈ A and let x. Prove that p = x k Gauss 5. an ) such that (a1 ! − 1)(a2 ! − 1) . (an ! − 1) − 16 is a perfect square. where 0 ≤ x ≤ y and (x.

Moshe Laub. ROMOP. 14. Prove that there exists inﬁnitely many pairs of consecutive numbers. for any natural numbers a. Gabriel Dospinescu. Find all positive integers n such that the number 2n − 1 has a multiple of the form m2 + 9. 2001 12. then a ∈ C. if a|b then f (a) ≥ f (b) 2. Prove that if n2 + a ∈ C for any positive integer n. Let T the set of the positive integers n for which the equation n2 1. q such that |qx − p| ≤ N +1 b) Suppose that a ∈ Z is a divisor of a number of the form n2 + 1. no two of them having any prime factor that belongs to B. 1998 11. 1999 13. one can ﬁnd integer numbers p. 17. Prove that T has density . a) Prove that for any real number x and any natural number N 1 . Balkan. 6583 16. Then prove that a ∈ C. Find all functions f : N → Z with the properties: 1. b we have f (ab) + f (a2 + b2 ) = f (a) + f (b). IMO Shortlist.10. Solve in integer numbers the equation x2 = y 7 + 7. Gabriel Dospinescu 15. Prove that the equation y 2 = x5 − 4 does not have integer solutions. Mathlinks Contest 63 = a2 + b2 has solutions in positive integers.

Cohn’s theorem 64 . Then the only n > 1 such that Ln is a perfect square is n = 3.18. (for the die hards) Let L0 = 2. L1 = 1 and Ln+2 = Ln+1 + Ln be the famous Lucas’s sequence.

Anyway. which is immediate. an and any positive real numbers x1 . this particular lemma has become very popular among the American students who attended the training of the USA IMO team. . The inductive step is reduced practically to the case n = 2. . xn the inequality a2 a2 a2 (a1 + a2 + · · · + an )2 1 + 2 + ··· + n ≥ x1 x2 xn x1 + x2 + · · · + xn (1) holds. by induction. Indeed. . Some will say that it is actually the Cauchy-Schwarz inequality and they are not wrong. . . 2001. it boils down to (a1 x2 − a2 x1 )2 ≥ 0 and the equality occurs if and only a2 a1 = . With this brief introduction. x2 . xn a2 √ 2 · x2 + · · · + x2 But there is another nice proof of (1). And now we see why calling it also the Cauchy-Schwarz inequality is natural. Applying this result twice it follows that if x1 x2 a2 a2 a2 (a1 + a2 )2 a2 (a1 + a2 + a3 )2 1 + 2+ 3 ≥ + 3 ≥ x1 x2 x3 x1 + x2 x3 x1 + x2 + x3 and we see that a simple inductive argument ﬁnishes the proof. a2 . But what exactly does this lemma say? It says that for any real numbers a1 . let us discuss some problems. And there are plenty 65 . . This happened after a lecture delivered by the ﬁrst author at the Mathematical Olympiad Summer Program (MOSP) held at Georgetown University in June. since it is practically an equivalent form of this inequality: a2 a2 a2 1 + 2 + ··· + n x1 x2 xn  ≥ a2 √ 1 · x1 + x1 (x1 + x2 + · · · + xn ) 2 a2 √  n · xn . . .T2 ’S LEMMA T2 ’s lemma is clearly a direct application of the Cauchy-Schwarz inequality.

1998 Solution. a2 + ab + b2 b + bc + c2 c + ca + a2 a+b+c The second example also became representative for a whole class of problems. that became classical. so we ﬁnd it necessary to discuss it at this point. a(a2 + ab + b2 ) b(b2 + bc + c2 ) c(c2 + ca + a2 ) It follows that the left-hand side is greater than or equal to (a2 + b2 + c2 )2 a3 + b3 + c3 + ab(a + b) + bc(b + c) + ca(c + a) But we can easily observe that a3 + b3 + c3 + ab(a + b) + bc(b + c) + ca(c + a) = (a + b + c)(a2 + b2 + c2 ). 66 . we will obtain a reﬁnement of the inequality. that is a3 b3 c3 a2 + b2 + c2 + 2 + 2 ≥ . Example 1. We will change the left-hand side of the inequality so that we could apply T2 ’s lemma.of them given in mathematical contests or proposed in mathematical magazines! First. There are countless examples of this type in numerous contests and mathematical magazines. so we have proved an even stronger inequality. We will see that with T2 ’s lemma it becomes straightforward and even more. 2 + ab + b2 2 2 a b + bc + c c + ca + a 3 Tournament of the Towns. This is not diﬃcult: we just have to write it in the form b4 c4 a4 + + . an old problem. c a3 b3 c3 a+b+c + 2 + 2 ≥ . Prove that for any positive real numbers a. b.

a(b + 2c + 3d) b(c + 2d + 3a) c(d + 2a + 3b) d(a + 2b + 3c) then the way to continue is clear. 67 . which is just the Cauchy-Schwarz inequality for four variables. implies 8(ab + bc + cd + da + ac + bd) = 4(a + b + c + d)2 − 4(a2 + b2 + c2 + d2 ). But it is not diﬃcult to see that (a + b + c + d)2 = a2 + b2 + c2 + d2 + 2(ab + bc + cd + da + ac + bd). since from the lemma we obtain a b c d + + + b + 2c + 3d c + 2d + 3a d + 2a + 3b a + 2b + 3c ≥ (a + b + c + d)2 . b + 2c + 3d c + 2d + 3a d + 2a + 3b a + 2b + 3c 3 Titu Andreescu. b. For arbitrary positive real numbers a. It could be also solved by using the above lemma. Consequently. given at the IMO 1995. was discussed extensively in many publications. IMO 1993 Shortlist Solution. The problem below. If we write the left-hand side in the form a2 b2 c2 d2 + + + . c. we are left with the inequality 4(a2 + b2 + c2 + d2 ) ≥ (a + b + c + d)2 . 4(ab + bc + cd + da + ac + bd) Hence it suﬃces to prove the inequality 3(a + b + c + d)2 ≥ 8(ab + bc + cd + da + ac + bd).Example 2. d prove the inequality a b c d 2 + + + ≥ .

Let n ≥ 2. Usually. in such problems the minimal value is attained when the variables are equal. Find the minimal value of the expression x5 x5 x5 n 2 1 + + ··· + . So. we conjecture that the minimal value 1 1 is attained when x1 = x2 = · · · = xn = √ . by using n(n − 1) n the lemma. xn are positive real numbers satisfying x2 + x2 + · · · + 1 2 x2 = 1. the inequality that remains to be proved after using the lemma) is not diﬃcult. The following problem is also not diﬃcult. Example 4. x2 . + c) b (c + a) c (a + b) 2 Solution. 2(ab + bc + ca) 2(ab + bc + ca) 2 2 the last inequality following from the AM-GM inequality.Example 3. xi (x1 + · · · + xi−1 + xi+1 + · · · + xn ) i=1 68 . n Turkey. but it uses a nice combination between this lemma and the Power-Mean inequality. . 1997 Solution. Let a. b. . Indeed. . c be positive real numbers such that abc = 1. x2 + x3 + · · · + xn x1 + x3 + · · · + xn x1 + x2 + · · · + xn−1 where x1 . It is another example in which proving the intermediate inequality (that is. We have: 1 1 1 1 1 1 a2 b2 c2 + + = + + a3 (b + c) b3 (c + a) c3 (a + b) a(b + c) b(c + a) c(c + a) a3 (b 1 1 1 2 + + (ab + bc + ca)2 ab + bc + ca 3 a b c ≥ = = ≥ . . it follows that the left-hand side is greater than or equal to n 2 x3 i i=1 n . Prove that 1 1 1 3 + 3 + 3 ≥ .

The answer is really unexpected: one can prove that the inequality is true for all n = 3. 1 x3 ≥ √ and i n n i=1 √ xi ≤ √ n. .S. x3 i i=1 n(n − 1) But this is a simple consequence of the Power-Mean inequality. Indeed. we have  n 1 3 x3  i  n 1 2 x2  i   ≥   n   i=1   n  implying n i=1     ≥  i=1  n    xi i=1 n . proving that x5 x5 x5 n 1 2 + + ··· + x2 + x3 + · · · + xn x1 + x3 + · · · + xn x1 + x2 + · · · + xn−1 1 n(n − 1) reduces to proving the inequality ≥ n n 2 2 xi ≥ i=1 −1 .Shapiro asked whether the following inequality is true for any positive real numbers a1 . . . a2 .But it is not diﬃcult to observe that n n 2 xi (x1 + · · · + xi−1 + xi+1 + · · · + xn ) = i=1 i=1 xi − 1. an : a1 a2 an n + + ··· + ≥ . 7 (and for all small values of n the shortest proof is based on 69 . . 4. The conclusion follows. 5. a2 + a3 a3 + a4 a1 + a2 2 The question turned out to be extremely diﬃcult. So. 6. In 1954. H.

an and any n ≥ 3 the following inequality holds: a1 a2 an + + ··· + ≥ cn. a2 + a3 a3 + a4 a1 + a2 This time. but ﬁnding the best such constant is an extremely diﬃcult task. Then we observe that a1 (a2 + a3 ) + a2 (a3 + a4 ) + a3 (a4 + a5 ) + a4 (a + 5 + a1 ) + a5 (a1 + a2 ) = a1 (S − a1 ) + a2 (S − a2 ) + a3 (S − a3 ) + a4 (S − a4 ) + a5 (S − a5 ) 2 S 2 − a2 − a2 − a2 − a2 − a2 1 2 3 4 5 . a3 . Another question arises: is there a positive real number such that for any positive real numbers a1 . Prove that for any positive real numbers a1 . which is nothing else then 1 2 3 4 5 the Cauchy-Schwarz inequality. Example 5. 1 2 3 4 5 4 equivalent to 5(a2 + a2 + a2 + a2 + a2 ) ≥ S 2 . a1 a2 a3 a4 a5 5 + + + + ≥ . . It was ﬁrst solved by Drinﬁeld (who. a2 . . a5 . by the way. we apply the lemma and we conclude that it sufﬁces to prove the inequality (a1 + a2 + a3 + a4 + a5 )2 5 ≥ [a1 (a2 + a3 ) + a2 (a3 + a4 ) + a3 (a4 + a5 ) + a4 (a5 + a1 ) + a5 (a1 + a2 )] 2 Let us denote a1 + a2 + a3 + a4 + a5 = S. a4 . a2 . = 2 With this identity. a2 + a3 a3 + a4 a4 + a5 a5 + a1 a1 + a2 2 Solution.this lemma). . but it is false for all even numbers n ≥ 14 as well as for suﬃciently large odd numbers n. . a problem proposed for MOSP 2001. Let us examine the case n = 5. Again. we infer that the intermediate inequality is in fact 5 (a1 + a2 + a3 + a4 + a5 )2 ≥ (S 2 − a2 − a2 − a2 − a2 − a2 ). the answer is positive. The answer is quite complicated and we 70 . is a Fields’ medalist).

. . √ shows that c = 2 − 1 is such a constant (not optimal). given at the Moldavian TST in 2005. ≥ ( 2)n . a2 . yet very diﬃcult to ﬁnd. n n n (ai + ai+1 + ai+2 ) ≥ i=1 i=1 2 (2ai + ai+1 ) i=1 (2ai+2 + ai+1 ). Thus. An ingenious argument using the arithmetic-geometric means inequality does the job: let us write the inequality in the form a1 + a2 + a3 a2 + a3 + a4 an + a1 + a2 √ + + ··· + ≥ 2 · n. the real trick is to rewrite appropriately the last products. an and any n ≥ 3 the following inequality holds: √ a1 a2 an + + ··· + ≥ ( 2 − 1)n. . Now. . a2 + a3 a3 + a4 a1 + a2 The proof is completely elementary. a2 + a3 a3 + a4 a1 + a2 Observe that (ai + ai+1 + ai+2 )2 = ai + ≥ 4 ai + ai+1 2 ai+1 ai+1 + + ai+2 2 2 ai+1 + ai+2 2 2 (the last inequality being again a consequence of the AM-GM inequality).. . a2 + a3 a3 + a4 a1 + a2 Now. The following problem. Let us observe that n n (2ai+2 + ai+1 ) = i=1 71 i=1 (2ai+1 + ai ). using the AM-GM inequality.will not discuss it here (for a detailed presentation of Drinﬁeld’s method the interested reader can consult the written examination given at ENS in 1997). For any a1 .. we see that it suﬃces to prove the stronger inequality: √ an + a1 + a2 a1 + a2 + a3 a2 + a3 + a4 · .

This lemma came handy even at the IMO 2005 (problem 3). During the coordination. a student from Ireland applied this result and called it ”SQ Lemma”. We begin with a diﬃcult geometric inequality for which we have found a direct solution using T2 ’s lemma.escu”. Prove that in any triangle ABC the following inequality holds ra rb rb rc rc ra + + ≥ 3. y. ma mb mb mc mc ma Ji Chen. Here it is. the Irish deputy leader explained what ”SQ” stood for: ”. 2 + y2 + z2 x xyz(x2 + y 2 + z 2 ) It is now time for the champions.so n n n (2ai + ai+1 ) i=1 n i=1 (2ai+2 + ai+1 ) = i=1 [(2ai + ai+1 )(ai + 2ai+1 )] n 2 ≥ i=1 (2(ai + ai+1 ) ) = 2 2 n i=1 (ai + ai+1 ) . Crux Mathematicorum 72 . For example. 1 1 y z 2 2 +y +z x x hence x2 + y 2 + z 2 ≤ x5 + y 2 + z 2 1 + y2 + z2 xy + yz + zx x =2+ ≤ 3. The conclusion now follows. x5 + y 2 + z 2 a few students successfully used the above mentioned lemma. A typical solution using this lemma is as follows: x5 + y 2 + z 2 = x4 y 4 z 4 (x2 + y 2 + z 2 )2 + 2+ 2 ≥ . Example 6. z such that xyz ≥ 1 the following inequality holds x2 + y 2 + z 2 ≤ 3. In order to prove that for any positive real numbers x...

let us try to prove the stronger inequality 2(a + b + c)(c + b − a) 2(a + b + c)(c + a − b) + 4a2 + b2 + c2 4b2 + c2 + a2 2(a + b + c)(a + b − c) ≥ 3. Of course. + 2b2 − a2 · 2c2 + 2a2 − b2 In this form. by applying the AM-GM inequality to each denominator. the inequality is more that monstrous. s−a ma = √ 2b2 + 2c2 − a2 2 and the likes the desired inequality takes the equivalent form √ (a + b + c)(b + c − a) (a + b + c)(c + a − b) √ √ +√ 2a2 + 2b2 − c2 · 2a2 + 2c2 − b2 2b2 + 2a2 − c2 · 2b2 + 2c2 − a2 (a + b + c)(a + b − c) √ ≥ 3. so we try to +√ 2c2 see if a weaker form holds. 73 . this is not diﬃcult.Solution. 4c2 + a2 + b2 Written in the more appropriate form + c+b−a c+a−b a+b−c 3 + 2 + 2 ≥ 2 + b2 + c2 2 + a2 2 + b2 4a 4b + c 4c + a 2(a + b + c) we see that by T2 ’s lemma the left-hand side is at least (a + b + c)2 . Fortunately. (b + c − a)(4a2 + b2 + c2 ) + (c + a − b)(4b2 + a2 + c2 ) + (a + b − c)(4c2 + a2 + b2 ) Basic computations show that the denominator of the last expression is equal to 4a2 (b + c) + 4b2 (c + a) + 4c2 (a + b) − 2(a3 + b3 + c3 ) and consequently the intermediate inequality reduces to the simpler form 3(a3 + b3 + c3 ) + (a + b + c)3 ≥ 6[a2 (b + c) + b2 (c + a) + c2 (a + b)]. So. we start by translating the inequality into an algebraic one. since using Heron’s relation and the formulas K ra = .

we expand (a + b + c)3 and obtain the equivalent inequality 4(a3 + b3 + c3 ) + 6abc ≥ 3[a2 (b + c) + b2 (c + a) + c2 (a + b)]. we have managed to prove the intermediate inequality. Of course.Again. a y= c+a . c . since any such approach is doomed. Example 7. b. But with the substitution x= b+c . Prove that for any positive real numbers a. hence the problem is solved. b 74 z= a+b . The ﬁrst one uses a nice combination between this lemma and the substitution discussed in the unit ”Two useful substitutions”. The journey continues with a very diﬃcult problem. 2 + (b + c)2 2 2 a b + (c + a) c + (a + b) 5 Japan. from the introduction to this problem. We will present two solutions for this inequality. Indeed. which is not diﬃcult at all. it follows from the inequalities 4(a3 + b3 + c3 ) ≥ 4[a2 (b + c) + b2 (c + a) + c2 (a + b)] − 12abc and a2 (b + c) + b2 (c + a) + c2 (a + b) ≥ 6abc. After all. The ﬁrst one is just an equivalent form of Schur’s inequality. 1997 Solution. given at the Japanese Mathematical Olympiad in 1997 and which became famous due to its diﬃculty. while the second follows immediately from the identity a2 (b + c) + b2 (c + a) + c2 (a + b) − 6abc = a(b − c)2 + b(c − a)2 + c(a − b)2 . the reader has already noticed that it is useless to try a direct application of the lemma. c the following inequality holds (b + c − a)2 (c + a − b)2 (a + b − c)2 3 + 2 + 2 ≥ .

which follows from the inequality x + y + z ≥ 6. Alternative solution. x2 + 1 y +1 z +1 x + y2 + z2 + 3 Hence it is enough to prove the inequality (x + y + z − 3)2 3 ≥ . And the problem is solved. This is not an easy inequality. a2 (b + c)2 + b2 (c + a)2 + c2 (a + b)2 + (a + b)4 + (b + c)4 + (c + a)4 75 ≥ . But here is another original solution. z satisfying xyz = x + y + z + 2. We will use the proposed problem 3 from the unit ”Two useful substitutions” to reduce the above inequality to the form (x + y + z)2 − 9(x + y + z) + 18 ≥ 0. x2 + y 2 + z 2 + 3 5 But this is equivalent to (x + y + z)2 − 15(x + y + z) + 3(xy + yz + zx) + 18 ≥ 0.we have to prove that for any positive real numbers x. It is now time to use T2 ’s lemma in the form (x − 1)2 (y − 1)2 (z − 1)2 (x + y + z − 3)2 + 2 + 2 ≥ 2 . y. Let us apply T2 ’s lemma in the following form: (b + c − a)2 (c + a − b)2 (a + b − c)2 + 2 + 2 a2 + (b + c)2 b + (c + a)2 c + (a + b)2 = ((b + c)2 − a(b + c))2 ((c + a)2 − b(c + a))2 ((a + b)2 − c(a + b))2 + 2 + 2 a2 (b + c)2 + (b + c)4 b (c + a)2 + (c + a)4 c (a + b)2 + (a + b)4 4(a2 + b2 + c2 )2 . the inequality (x − 1)2 (y − 1)2 (z − 1)2 3 + 2 + 2 ≥ 2+1 x y +1 z +1 5 holds.

Hence 4ab(a2 + b2 ) + 4bc(b2 + c2 ) + 4ca(c2 + a2 ) ≤ 2(a2 + b2 + c2 )2 8 +2(a2 b2 + b2 c2 + c2 a2 ) ≤ (a2 + b2 + c2 )2 3 and so the problem is solved. Indeed. We end this discussion (which remains probably permanently open) with a diﬃcult problem. we can readily see that this solution works without the assumption that a. Because (a + b)2 + (b + c)2 + (c + a)2 ≤ 4(a2 + b2 + c2 ). 3 But this inequality is not so diﬃcult. With minor changes. it suﬃces to prove that the last quantity is greater 3 than or equal to . c are positive. b. This can be done by expanding everything. based on two hidden applications of T2 ’s lemma. we observe that the desired inequality reduces to 8 2ab(a + b)2 + 2bc(b + c)2 + 2ca(c + a)2 ≤ (a2 + b2 + c2 )2 . since (a − b)4 ≥ 0. Then.Consequently. we also ﬁnd that 4ab(a2 + b2 ) ≤ a4 + b4 + 6a2 b2 . 76 . but here 5 is an elegant proof using the observation that a2 (b + c)2 + b2 (c + a)2 + c2 (a + b)2 + (a + b)4 + (b + c)4 + (c + a)4 = [(a + b)2 + (b + c)2 + (c + a)2 ](a2 + b2 + c2 ) +2ab(a + b)2 + 2bc(b + c)2 + 2ca(c + a)2 . ﬁrst we observe that 2ab(a + b)2 + 2bc(b + c)2 + 2ca(c + a)2 ≤ 4ab(a2 + b2 ) + 4bc(b2 + c2 ) + 4ca(c2 + a2 ).

Example 8. Let a1 , a2 , . . . , an > 0 such that a1 + a2 + · · · + an = 1. Prove that: (a1 a2 +a2 a3 +· · ·+an a1 ) a2 2 a1 a2 an + 2 + ··· + 2 + a2 a3 + a3 a1 + a1 ≥ n . n+1

Gabriel Dospinescu Solution. How can we get to a1 a2 + a2 a3 + · · · + an a1 ? Probably a2 a2 a2 from 1 + 2 + · · · + n after we use the lemma. So, let us try a1 a2 a2 a3 an a1 this the following estimation: a1 a2 an a2 a2 a2 1 + +· · ·+ = 1 + 2 +· · ·+ n ≥ . a2 a3 a1 a1 a2 a2 a3 an a1 a1 a2 + a2 a3 + · · · + an a1 The new problem, proving that a1 a2 an n + + ··· + 2 ≥ n+1 a2 + a2 a2 + a3 a1 + a1 2 3 a1 a2 an + + ··· + a2 a3 a1

seems even more diﬃcult, but we will see that we have to make one more step in order to solve it. Again , we look at the right-hand side and we an a1 a2 + + ··· + as write a2 a3 a1 a1 a2 an + + ··· + a2 a3 a1 a1 a2 an + + ··· + a2 a3 a1 After applying T2 ’s lemma, we ﬁnd that a1 a2
2 2

.

a2 2

a1 a2 an + 2 +···+ 2 = a1 + a2 + · · · + an + a2 a3 + a3 a1 + a1 a1 + a2 + an + a2 a3 a1 an 2 a1 a2 + + ··· + a2 a3 a1 ≥ a1 a2 an . 1+ + + ··· + a2 a3 a1
77

a2 a3

2

an a1

2

And we are left with an easy problem: if t =

t2 nt ≥ , or t ≥ n. But this follows immediately from the AM-GM 1+t n+1 inequality. Problems for training

a1 an + ··· + , then a2 a1

1. Let a, b, c, d be positive real numbers such that a + b + c + d = 1. Prove that a2 b2 c2 d2 1 + + + ≥ . a+b b+c c+d d+a 2 Ireland, 1999 2. Let a, b, c, be positive real numbers satisfying a2 + b2 + c2 = 3abc. Prove that a b2 c2 + b c2 a2 + c a2 b2 ≥ 9 . a+b+c India 3. Let x1 , x2 , . . . , xn , y1 , y2 , . . . , yn be positive real numbers such that x1 + x2 + · · · + xn ≥ x1 y1 + x2 y2 + · · · + xn yn . Prove that x1 + x2 + · · · + xn ≤ xn x1 x2 + + ··· + . y1 y2 yn

Romeo Ilie, Romanian Olympiad, 1999 4. For arbitrary positive real numbers a, b, c prove the inequality a b c + + ≥ 1. b + 2c c + 2a a + 2b Czech-Slovak Competition, 1999 5. Prove that for any positive real numbers a, b, c satisfying a+b+c = 1, a b c 9 + + ≥ . 1 + bc 1 + ca 1 + ab 10 India
78

6. Prove that for any positive real numbers a, b, c, d satisfying ab + bc + cd + da = 1 the following inequality is true a3 b3 c3 d3 1 + + + ≥ . b+c+d c+d+a d+a+b a+b+c 3 IMO 1990 Shortlist 7. Prove that if the positive real numbers a, b, c satisfy abc = 1, then b c a + + ≥ 1. b+c+1 c+a+1 a+b+1 Vasile Cartoaje, Gazeta Matematica 8. Prove that for any positive real numbers a, b, c the following inequality holds a2 + bc b2 + ca c2 + ab + + ≥ a + b + c. b+c c+a a+b Cristinel Mortici, Gazeta Matematica 9. Prove that for any nonnegative real numbers x1 , x2 , . . . , xn , x1 x2 xn + + ··· + ≥ 2. xn + x2 x1 + x3 xn−1 + x1 Tournament of the Towns, 1982 10. Prove that for any positive real numbers a, b, c, d, e satisfying abcde = 1, a + abc b + bcd c + cde + + 1 + ab + abcd 1 + bc + bcde 1 + cd + cdea + d + dea e + eab 10 + ≥ . 1 + de + deab 1 + ea + eabc 3 Waldemar Pompe, Crux Mathematicorum

11. Prove that for any positive real numbers a, b, c the following inequality holds a b+c
2 2 2

+

b c+a

+

c a+b

3 a2 + b2 + c2 · . 4 ab + bc + ca Gabriel Dospinescu

79

12. Let n ≥ 4 an integer and let a1 , a2 , . . . , an be positive real numbers such that a2 + a2 + · · · + a2 = 1. Prove that n 1 2 a2 2 √ √ a1 a2 an 4 √ + 2 + ··· + 2 ≥ (a1 a1 + a2 a2 + · · · + an an )2 . 5 + 1 a3 + 1 a1 + 1 Mircea Becheanu, Bogdan Enescu, TST 2002, Romania 13. Find the best constant k(n) such that for any positive real numbers a1 , a2 , . . . , an satisfying a1 a2 . . . an = 1 the following inequality holds (a2 1 a1 a2 a2 a3 an a1 2 + a ) + (a2 + a )(a2 + a ) + · · · + (a2 + a )(a2 + a ) ≤ kn . + a2 )(a2 1 3 2 1 2 n 2 3 1 Gabriel Dospinescu, Mircea Lascu 14. Prove that for any positive real numbers a, b, c, (2a + b + c)2 (2b + c + a)2 (2c + a + b)2 + 2 + 2 ≤ 8. 2a2 + (b + c)2 2b + (c + a)2 2c + (a + b)2 Titu Andreescu, Zuming Feng, USAMO 2003

80

ONLY GRAPHS, NO SUBGRAPHS! There were so many strategies and useful ideas till now, that the reader might say: enough with this game of tricks! When shall we go to serious facts? Not only that we will ”dissapoint” him again, but we will try also to convince him that these are more than simple tools and tricks. They help to create a good base, which is absolutely indispensable for someone who enjoys mathematics and moreover, they are the ﬁrst step to some really beautiful and diﬃcult theorems or problems. And the reader must admit that the last problems discussed in the previous units are quite serious facts. It is worth mentioning that they are not panacea. This assertion is proved by the fact that each year problems that are based on well-known ”tricks” prove to be very diﬃcult in contests. We will focus in this unit on a very familiar theme: graphs without complete subgraphs. Why do we say familiar? Because there are hundreds of problems proposed to diﬀerent contests around the world and in mathematical magazines that deal with this subject and each one seems to add something. Before passing to the ﬁrst problem, we will assume that the basic knowledge about graphs is known and we will denote by d(A) and C(A) the number, respectively the set of vertices adjacent to A. Also, we will say that a graph does not have a complete k subgraph if there aren’t k vertices any two of them connected. For simplicity, we will say that G is k-free. First, we will discuss probably the ﬁrst classical result about triangles-free graphs, the famous Turan’ theorem. But before that, an useful lemma, which is also known as Zarankiewicz lemma and which is the main idea in Turan’ theorem’ proof. Example 1. If G is a k-free graph, then there exists a vertex having k−2 n . degree at most k−1 Zarankiewicz
81

Solution. Suppose not and take an arbitrary vertex A1 . Then |C(A1 )| > k−2 n , k−1

so there exists A2 ∈ C(A1 ). Moreover, |C(A1 ) ∩ C(A2 )| = d(A1 ) + d(A2 ) − |C(A1 ∪ A2 )| k−2 n − n > 0. k−1 Pick a vertex A3 ∈ C(A1 ) ∩ C(A2 ). A similar argument shows that ≥2 1+ |C(A1 ) ∩ C(A2 ) ∩ C(A3 )| ≥ 3 1 + Repeating this argument, we ﬁnd
k−2

k−2 n k−1

− 2n.

A4 ∈ C(A1 ) ∩ C(A2 ) ∩ C(A3 ), . . . , Ak−1 ∈
i=1

C(Ai ).

Also, we have
j

C(Ai ) ≥ j 1 +
i=1

k−2 n k−1

− (j − 1)n.

This can be proved easily by induction. Thus,
k−1

C(Ai ) ≥ (k − 1) 1 +
i=1

k−2 n k−1

− (k − 2)n > 0

and consequently we can choose
k−1

Ak ∈
i=1

C(Ai ).

But it is clear that A1 , A2 , . . . , Ak form a complete k graph, which contradicts the assumption that G is k-free. We are now ready to prove Turan’ theorem. Example 2. The maximal number of edges of a k-free graph with vertices is r k − 2 n2 − r 2 · + , 2 k−1 2
82

where r = n (mod k − 1). Turan’ theorem Solution. The theorem will be proved by induction on n. Since the ﬁrst case is trivial, let us suppose the theorem true for all k-free graphs having n − 1 vertices and let G a k-free graph with n vertices. Using Zarankiewicz’ lemma, we can ﬁnd a vertex A such that d(A) ≤ k−2 n . k−1

Since the subgraph determined by the other n−1 vertices is obviously k-free, using the inductive hypothesis we ﬁnd that G has at most
2 k − 2 (n − 1)2 − r1 r1 k−2 n + · + k−1 k−1 2 2

edges, where r1 = n − 1 (mod k − 1). Let n = q(k − 1) + r = q1 (k − 1) + r1 + 1. Then r1 ∈ {r − 1, r + k − 2} (this is because r − r1 ≡ 1 (mod k − 1)) and it is easy to check that
2 k−2 k − 2 (n − 1)2 − r1 r1 n + · + k−1 k−1 2 2

=

k − 2 n2 − r 2 r · + 2 k−1 2

and the inductive step is proved. Now, it remains to construct a kk − 2 n2 − r 2 r free graph with n vertices and · + edges. This is 2 k−1 2 not diﬃcult. Just consider k − 1 classes of vertices, r of them having q + 1 elements and the rest q elements and join the vertices situated in diﬀerent groups. It is immediate to prove that this graph is k-free, has k − 2 n2 − r 2 r · + edges and also the minimal degree of the vertices 2 2 k−1 k−2 is n . This graph is called Turan’ graph and it is denoted by k−1 T (n, k). These two examples generate lots of beautiful and diﬃcult problems. For example, knowing them means a straightforward solution for the following bulgarian problem.
83

Example 4. But Zarankiewicz’ lemma implies that if G is n-free. So. We will take of course T (2001. Tournament of the Towns. Given are 21 points on a circle. It suﬃces to construct a graph with all degrees of the vertices at least 1600. Thus. Therefore. which 84 . 2001 Solution. There are 2001 towns in a country. if n = 5 then any such graph G is not n-free. the answer 4 is n = 5. Here. every one of which is connected with at least 1600 towns by a direct bus line. Spring Mathematics Tournament. we need to prove that this graph has at least 100 edges. It is immediate to see that n−1 it is n = 5. In such problems. 6). Practically. we need the at least one vertex has degree at most n−1 n−2 maximal n for which 2001 < 1600. Here is a beautiful application of Turan’ theorem in combinatorial geometry. Find the largest n for which it is always possible to ﬁnd n towns. it is more important to choose the right graph than to apply the theorem.Example 3. Show that at least 100 pairs of points subtend an angle smaller than or equal to 120 at the center. the problem asks to ﬁnd the maximal n such that any graph G with 2001 vertices and minimum degree at least 1600 is not n-free. the solution is more or less straightforward. which is 6-free. because as soon as the graph is appropriately chosen. 1986 Solution. It seems that this is a reversed form of Turan’ theorem. then n−2 2001 . we will consider the graph with vertices in the points and we will connect two points if they subtend an angle smaller than or equal to 120 at the center. any two of which are connected by a direct bus line. whose minimal 4 degree is 2001 = 1600 and which is of course 6-free. Thus.

it is not connected with at least k + 1 other vertices.maximizes the number of edges in a k-free graph. But this is immediate. there exists a vertex whose degree is at most n = k + 1. this is trickier than the contest problem. We must show 21 that it has at most −100 = 110 edges. each of them knowing at most k languages. First. We will prove that n = 2k + 3. Ak+1 that can communicate with A. at least two speak a common language.then the conclusion of the problem holds. It remains to prove now that we can create a situation in which there are 2k + 2 delegates. . since 2 it is clear that this new graph does not have triangles and so. 4 At ﬁrst glance. Since A knows at most k languages. the complementary one. the reversed form of a reversed form is the natural one. The condition ”among any three of them there are at least two who can communicate” suggests us to take the 3-free graph with vertices in the persons and whose edges join persons that cannot communicate. we prove that if there are 2k + 3 delegates. there exists a person A and k + 1 persons A1 . This problem is an adaptation of a USAMO 1978 problem. but. Anyway. Anyway. But that language is known by at least three delegates and we are done. by Turan’ 212 − 1 theorem it has at most = 110 edges. Find the smallest number n (in terms of k) such that it is always possible to ﬁnd a language spoken by at least three delegates. Solution. Example 5. Yet. Ap that know a language also known by A. There are n delegates at a conference. Thus. . it is a simple consequence of Zarankiewicz’ lemma. . there are two persons among A1 . the following problem seem to have no relation with the previously examples. as we will see. 2 Therefore. . . From Zarankiewicz’ lemma. by creating two groups of k + 1 85 . let us look at the ”reversed” graph. A2 . We use again Turan’ graph. among any three delegates. . And the problem is solved. A2 . In the aim of this principle. but no language is known by more than two delegates. . .

any two 51 of them having common elements. Suppose that the union of any 50 subsets has more than 50 n elements. but to ﬁnd much more vertices with small degrees. all of them having degrees at most 50. we should take a graph with vertices in the subsets. 2. Prove that there are three subsets among them. . . any language is spoken by at most two delegates and there are no triangles. we can ﬁnd Ai1 . . . . . we will prove that there are at least 51 vertices whose degree are smaller than or equal to 50. Of course. we obtain that |Aij | ≤ 51 51 51 86 . we infer Since the union of these subsets has more than 51 50 n n that |Ai1 | < n − n = . Suppose this is not the case. . Ai1 is disjoint from at least 50 subsets. The idea is even easier than in the previous problems. connecting two subsets if they have common elements. 50 n elements. Gabriel Dospinescu. Let us pick such a vertex A. so we have a triangle. It is connected with at least 51 vertices. n}. Since A and B are each connected with at least 51 vertices. Ai51 . The main idea is not to use Zarankiewicz’ lemma. that is not so obvious. In a similar way.delegates. Let A1 . Therefore. . thus there are at least 51 vertices whose degrees are greater than 51. . . A101 be diﬀerent subsets of the set {1. Consequently. there is a vertex adjacent to both. Let us assume that this graph is 3-free. contradicting our assumption. The following problem turned out to be a surprise at one of the Team Selection Tests for 2004 IMO. whose degree is at least 51. A2 . In fact. Of course. but this time we need a little observation. as the conclusion suggests. . . In each group a person will have a common language with each other person from the group and will not have common languages with the members of the other group. being solved by 4 contestants. thus it must be adjacent to a vertex B. Example 6. TST 2004 Romania Solution.

Now. y: once for x and once for y. And the solution ends here. z) in which y is connected with both x. . the number of triangles from the complementary graph can be expressed only in terms of the degrees of the vertices of the graph. consider all triples (x. z) in which are connected. x∈X m− n2 − n 4 2 where X is the set of vertices of G. We end the discussion with an adaptation of a very nice and quite challenging problem from the American Mathematical Monthly. y. Example 7. .W Goodman. z) that do not form a triangle in the complementary graph (here we have used the fact that G is 3-free). . AMM Solution. But it also counts twice every triple (x.for all j ∈ {1. then the number of triangles from the complementary graph is n 1 − 2 3 d(x)(n − 1 − d(x)). y. It counts twice every triple (x. d(x)(n − 1 − d(x)) is exactly the number of 2 x∈X triples (x. Indeed. . Believe it or not. Therefore. 2. while z is not adjacent to any of x. More precisely. Indeed. A. it is enough to prove 87 . z) of vertices of G. if G is the graph. We will count the triples that do not form a triangle in the complementary graph G. 51 which contradicts the hypothesis. y. consider the sum x∈X d(x)(n−1−d(x)). y. Prove that the complementary of any 3-free graph with n vertices and m edges has at least n(n − 1)(n − 5) 2 + 24 n triangles. z: once for x and 1 once for z. 51} and so |Ai1 ∪ Ai2 ∪ · · · ∪ Ai50 | ≤ |Ai1 | + · · · + |Ai50 | < 50 n.

. . . after a few computations we ﬁnd the equivalent form of the inequality d2 (x) ≥ x∈X 4m2 . n} such that 1 < |xi − xj | < 2. . 2. Poland. 1998 2. at least two are not directly connected. if A and B are not joined by an edge. Prove that all vertices have the same degree. n But this is exactly the Cauchy-Schwarz inequality combined with the observation that d(x) = 2m. If n points lie on a circle. . . . CE AN? n62 3. . then there exists a vertex C such that AC and BC are edges. x2 . xn be real numbers. Let G be a graph with no triangles and such that no point is adjacent to all the other vertices. Let x1 . j) ∈ {1. Also. x∈X Using the observation that x∈X d(x) = 2m. In a country there are 1998 cities. . In each group of three cities. What is the maximal number of direct ﬂights? Japan. then at most segments connecting 3 √ them have length greater than 2. . . n} × {1.that n 1 − 3 2 n(n − 1)(n − 5) 2 d(x)(n − 1 − d(x)) ≥ + 24 n n2 − n m− 4 2 . Prove that there are at most n2 pairs (i. . APMO 1990 88 . 1997 4. 4 MOSP. x∈X Problems for training 1. 2.

89 . 1000000} having exactly 101 elements. 9. A graph with n2 + 1 edges and 2n vertices is given. Show that a graph with n vertices and k edges has at least n2 ) triangles. 1999 7. Let A be a subset of the set S = {1. t100 ∈ S such that the sets Aj = {x + tj |x ∈ A} are pairwise disjoint. 2. We are given 5n points in a plane and we connect some of them. . Prove that for every n one can construct a graph with no triangles and whose chromatic number is at least n.5. Prove that we can ﬁnd at least 41 people who each know at most 1958 other people. A graph with n vertices and k edges is 3-free. Prove that we can ﬁnd a monochromatic triangle. . . Prove that there exist t1 . Taiwan. 1987 10. We color these segments in 2 colours. There are 1999 people participating in an exhibition. . . Prove that we can choose a vertex such that the subgraph induced by the remaining 4k vertices has at most k 1 − 2 vertices. IMO 2003 6. so that 10n2 + 1 segments are drawn. k (4k− 3n APMO 1989 6. . t2 . Out of any 50 people. n USAMO 1995 8. . at least 2 do not know each other. China TST. Prove that it contains two triangles sharing a common edge. .

After that. There are also numerous applications in combinatorial number theory. so our purpose is to present a little bit from each of these situations. which will show what a complex ﬁeld is combinatorics.COMPLEX COMBINATORICS When reading the title. the reader will surely have the pleasure of solving the proposed problems using this technique. one will surely expect a hard unit. In this moment. 1 + x + x2 + · · · + xp−1 must divide a0 + a1 x + a2 x2 + · · · + ap−1 xp−1 . but we will continue our idea and say that complex numbers can play a very important role in counting problems and also in problems related to tilings. We ”just” want to discuss some combinatorial problems that can be solved elegantly using complex numbers. where ε = cos then a0 = a1 = · · · = ap−1 . which can only happen if a0 = a1 = · · · = ap−1 . By convention m(∅) = 0. For fear of useless repetition. a1 . It is enough to observe that the polynomials a0 +a1 x+a2 x2 +· · ·+ap−1 xp−1 and 1 + x + x2 + · · · + xp−1 cannot be relatively prime-because they share a common root-and since 1 + x + x2 + · · · + xp−1 is irreducible over Q. the lemma is proved and it is time to solve some nice problems. If p is a prime number and a0 . p p . Not before saying that in the following examples m(A) will denote the sum of the elements of the set A. the reader will probably say we are crazy. Unfortunately. which is not diﬃcult. we will present in the beginning of the discussion a useful result Lemma. this was not our intention. Therefore. 90 2π 2π + i sin . . ap−1 ∈ Q satisfy the relation a0 + a1 ε + a2 ε2 + · · · + ap−1 εp−1 = 0. We will say just a few words about the proof. . . .

. It is clear that 6 (k) a(k) εk = n k=0 x1 . Of course.4.this n is because exactly 6n numbers have n digits. 9 are divisible by 7? Solution. 6. 7. 4. all equal to 1.xn ∈{1. The same simple. Let 6 (0) (1) (6) k be the common value. 7. but tricky idea can oﬀer probably the most beautiful solution for the diﬃcult IMO 1995 problem 6. 91 . Let an be the number of n-digits numbers. Then 6 a(k) εk = (−1)n n k=0 and from the lemma we infer that an − (−1)n = an = · · · = an .6. Example 1.9} εx1 +x2 +···+xn = (ε + ε3 + ε4 + ε6 + ε7 + ε9 )n .x2 .. 7. It worth saying that Nikolai Nikolov won a special prize for the following magniﬁcent solution. formed using only the digits 1. 4. there is a solution using recurrent sequences. 6. 4. 5. all equal to 1.7. Then 7k = k=0 a(k) − (−1)n = 6n − (−1)n . Let us assume that n is divisible by 7.3. 3. 3. 3. 4. 6n − (−1)n (0) . 6 (mod 7). The remark that 1 + ε + ε2 + · · · + ε6 = 0 7 7 helps us to bring (ε + ε3 + ε4 + ε6 + ε7 + ε9 )n to the simpler form (−ε5 )n . 9 and which are congruent to k modulo 7. We leave to the Thus. in this case we have an = (−1)n + 7 reader the study of the other cases: n ≡ 12.. but it is by far less elegant than the following one.. 9. for example (the other cases can be discussed similarly). where ε = cos 2π 2π + i sin . 6.The ﬁrst example is an adaptation from a problem given in the InterCounty Contest ”Traian Lalescu”. How many numbers with n digits. 3.

. (X − εp−1 ). . . 2. Example 3. (X + p−1 ε2p ) = (X p + 1)2 . . . 2. . am be natural numbers and let f (k) the number of m-tuples (c1 . . . . . . . where n > 1 is a natural number. . 2p}. Find the number of subsets of A. . 2p . Prove that f (0) = f (1) = · · · = f (n − 1) if and only if there exists an index i ∈ {1. But 1≤c1 <c2 <···<cp ≤2p εc1 +c2 +···+cp is exactly the coeﬃcient of xp in the expansion (X + ε)(X + ε2 ) . . . j=0 xj εj = 2 and lemma implies the equality 2p p subsets with p elements.Example 2. IMO 1995 2π 2π Solution. . Since there are we have x0 + x1 + · · · + xp−1 = Therefore. Reid Burton. c2 . . i = 1. m and c1 + c2 + · · · + cm ≡ k (mod n). Let a1 . .|B|=p εm(B) = 1≤c1 <c2 <···<cp ≤2p εc1 +c2 +···+cp . m} such that n|ai . cm ) such that 1 ≤ ci ≤ ai . Thus. Consider ε = cos + i sin and let xj the number of p p subsets x ⊂ A such that |X| = p and m(X) ≡ j (mod p). p x0 − 2 = x1 = · · · = xp−1 . a2 . . but closely related idea we can solve the x0 = 2 + following nice problem. Rookie Contest .1999 92 . Marcin Kuczma. each having p elements and the sum of the elements divisible by p. . (X + ε2p ). Then it is clear that p−1 xj εj = j=0 B⊂A. Let p > 2 be a prime number and A = {1. Since X p − 1 = (X − 1)(X − ε) . p p With a somewhat diﬀerent. . 1 2p −2 . we easily ﬁnd that (X + ε)(X + ε2 ) .

. f (m). . . . . where we have chosen here a primitive root of the unity ε. The enthusiasm determined by the above solutions will surely be ATENUAT by the following problem. Now. . 2. n be multiples of p such that n is odd. It is not diﬃcult to observe that n−1 m f (k)εk = k=0 1≤ci ≤ai εc1 +c2 +···+cm = i=1 (ε + ε2 + · · · + εai ) for any complex number ε such that εn−1 + εn−2 + · · · + ε + 1 = 0. . 2. . . n} that satisﬁes p|f (1) + f (2) + · · · + f (m). m} → {1. Let p > 2 be a prime number and let m. . . . Prove that the sum of these products is divisible by . This is because the polynomial k=0 X k has only simple roots. m} such that ε + ε2 + · · · + εai = 0. 2. . The solution ends here. . . . since if f (0) = f (1) = · · · = f (n − 1) then of course we can ﬁnd i ∈ {1. after all functions f : {1. p Gabriel Dospinescu 2π 2π + i sin and xk be the sum of all p p the numbers f (1)f (2) . suppose there exists an index i ∈ {1. It is 93 . . . consider the product f (1)f (2) · n m f (m). Example 4. this is possible only if f (0) = f (1) = · · · = f (n − 1). m} such that n|ai . 2. . . Then for any root ε of the polynomial n−1 n−1 n−1 X k we have k=0 n−1 k=0 f (k)εk and so the polynomial X k divides the k=0 n−1 polynomial k=0 f (k)X k . 2. Let ε = cos {1. Thus. in which e need some tricky manipulations. 2. . . m} → Solution.Solution. . We infer that εai = 1 and so n|ai . . . For any function f : {1. By a simple degree consideration. n} that satisfy f (1) + f (2) + · · · + f (m) ≡ k (mod p). one part of the problem is already proved. .

.n} c1 c2 .cm ∈{1.c2 . bj . . (ε − 1)m xk εk = k=0 On the other hand. (x − 1)2 Plugging ε in the previous identity. ε−1 p Thus. p−1 nε nεn+2 − (n + 1)εn+1 + ε = . 94 . Recall the identity 1 + 2x + 3x2 + · · · + nxn−1 = nxn+1 − (n + 1)xn + 1 .. .. then we have the relation x0 − rc0 + (x1 rc1 )ε + · · · + (xp−1 − rcp−1 )εp−1 = 0. If r = n − p m ... (ε − 1)2 ε−1 nm . it is not diﬃcult to deduce the relations εp−1 + εp−2 + · · · + ε + 1 = 0 ⇔ 1 1 = − (εp−2 + 2εp−3 + · · · + (p − 2)ε + p − 1). if we consider (X p−2 +2X p−3 +· · ·+(p−2)X +p−1)m = b0 +b1 X +· · ·+bm(p−2) X m(p−2) . then we have nm = (ε − 1)m where ck = k≡k (mod p) − n p m (c0 + c1 ε + · · · + cp−1 εp−1 ). cm εc1 +c2 +···+cm = (ε + 2ε2 + · · · + nεn )m .2. we ﬁnd that ε + 2ε2 + · · · + nεn = Consequently..clear that: p−1 xk εk = k=0 c1 ...

where 1 2 ε1 = cos 2π 2π 2π 2π + i sin . j). (a. The idea is to put a complex number in each square of a table and then to translate the hypothesis and the conclusion in terms of complex numbers.2000 Solution. . put the number εi εj in the square whose label is (i. . Prove that it is possible to tile this rectangle using only rectangles 1 × m or only with rectangles n × 1. n are natural numbers. First. . . it is clear that r|k. Gabriel Carrol . 1). b. (1. . n n m m 95 . some easy problems. But we will better see how this technique works by solving a few problems. it remains to prove that r|k. cp−1 ∈ R. (a. . . .From the lemma. Since pk = x0 + x1 + · · · + xp−1 − r(c0 + c1 + · · · + cp−1 ) = (1 + 2 + · · · + n)m − r(b0 + b1 + · · · + bm(p−2) ) = n(n + 1) 2 m −r p(p − 1) 2 m . Consider a rectangle which can be tiled with a ﬁnite combination of 1×m or n×1 rectangles. . it follows that x0 − rc0 = x1 − rc1 = · · · = xp−1 − rcp−1 = k. where m. Next. (a. . Because clearly c0 . let them be a. 2). . Example 5. b). Here we have used the hypothesis. It is obvious that the rectangle has natural dimensions. . ε2 = cos + i sin . Now. (a.BMC Contest. It is time now to leave this kind of problems and to speak a little bit about some nice applications of complex numbers in tilings. . . 1). 1). c1 . let us partition the rectangle into 1 × 1 squares and denote this squares (1. . 2). The problem is solved. .

Can we tile a 13 × 13 table using only 1 × 4. i−1 i −1 which clearly cannot hold. 4 × 1 rectangles. given in Baltic Contest in 1998. Next. But this 2 means that n|a or m|b. 2 a b Thus. This is immediate. 96 . Indeed. it is clear that we can tile the rectangle using only horizontal or vertical rectangles. The idea in the previous problem is quite useful. it follows that the sum of the numbers from all the squares is 0 and so a b 0= 1≤i≤a 1≤j≤b εi εj 1 2 = i=1 εi 1 j=1 εj . i21 = (i + i2 + · · · + i13 )(i2 + i4 + · · · + i26 ) = i · i13 − 1 2 i26 − 1 ·i · 2 = i3 . many tilings problems having straightforward solutions by using it.1998 Solution. such that only the center of the table does not belong to any rectangle? Baltic Contest. associate to square (k. Thus. j) the number ik+2j . Obviously.The main observation is that the sum of the numbers in any 1 × m or n × 1 rectangle is 0. the assumption was wrong and such a tiling does not exist. Example 6. the sum of the numbers from each 1 × 4. 4 × 1 rectangle is 0. An example is the following problem. Thus. Therefore. at least one of the numbers i=1 εi and 1 j−1 εj is 0. In any of these cases. the sum of all numbers from the squares of the table is equal to the number in the square situated at the center of the table. Suppose such a tiling is possible and label the squares of the table as in the previous problem. but the consequence of this simple observation is really surprising.

The sum of the numbers from any ﬁgure or rectangle is 0. z2 = ib2 −1 εa2 . From the above argument. 4. Therefore. then surely they belong to S. 97 . This in turn imz1 z2 plies the equalities i3(b1 −1) = i3(b2 −1) = 1. The conclusion is immediate. On a 8 × 9 table we put rectangles 3 × 1 and ﬁgures formed by rectangles 1 × 3 by cutting the median 1 × 1 square. 1). 9. . a2 when divided by 3 are 1 and 2. The rectangles and the ﬁgures do not intersect and cannot be rotated. 8 with the columns 1. Then we have of course ib1 εa1 + ib2 εa2 = −i. b2 ) are the only uncovered squares. we can take S the set of squares that lie at the intersection of the lines 1. (a2 . 7. then they belong to S. but here complex numbers are more involved. (8. we label the squares of the table (1. 2). Gabriel Dospinescu Solution. Thus. In the square labeled (k. We have |z1 | = |z2 | = 1 and z1 + z2 = −1. which is possible if and only if the remainders of the numbers a1 . 5. (1. if two squares remain uncovered. .The following example we are going to discuss is based on the same idea. where i2 = −1 and ε2 + ε + 1 = 0. Example 7. b1 ). Again. from where we conclude that b1 ≡ b2 ≡ 1 (mod 4). 1 1 3 3 It follows that + = −1 and so z1 = z2 = 1. k=1 Let us suppose that the squares (a1 . . 2. The sum of the numbers from the table is 8  εk  9 j=1  ij  = −i. Prove that there exists a set S of 18 squares of the table such that if there are exactly two uncovered squares. . Let z1 = ib1 −1 εa1 . 9) by starting from the up-left corner. the relation z1 + z2 = −1 becomes εa1 + εa2 = −1. 5. j) we will put the number ij · εk .

1. 2000} have the sum of their elements divisible by 5? Qihong Xie. Let ak .Problems for training 1. . . 1999 3. 2. C play the following game: a subset with k elements of the set {1. The winner is A. according to the case when the sum of the elements of the selected subset is congruent to 0. How many subsets with 100 elements of the set {1. green and red balls. The faces of a die are labeled with the numbers 1. we can replace two balls in the box with one or two balls according 98 . We assume that xi = 0 and f (0) = f (1) = f (2). At each step. . There are 2000 white balls in a box. B. 2. We throw the die n times. . k = 1. . Imo Shortlist. ck ∈ R. Recreatii Matematice ¸ 4. 1986} is selected randomly. initially outside the box. n} whose union is M and for which ai + i∈M \A i∈M \B bi + i∈M \C ci ≡ 3 (mod p). . B. 2. B or C. 1987 2. n. or 2 modulo 3. Find all values of k for which A. 4. B. Let f (p) be the number of ordered triples (A. C have equal chances of winning. . 5. . 6. all selections having the same probability. . Three persons A. bk . i∈∅ Prove that there exists i ∈ M such that 3|ai + bi + ci . What is the probability that the sum of the numbers shown by the die is a multiple of 5? IMC. 3. Gabriel Dospinescu. . 2. C) of subsets (not necessarily non-empty) of the set M = {1. . There are also unlimited supplies of white. High School-Mathematics 5. .

b) of natural numbers such that any residue modulo n appears the same number of times on the frontier of the rectangle of vertices (0. 2000 6. b) and also any residue modulo n appears the same number of times in the interior of the same rectangle. 2001 9. 0). 0). a) After some ﬁnite number of steps. Gabriel Dospinescu 8. What is the value of the sum of all the obtained numbers? Gabriel Dospinescu 99 . Bulgaria. b). 1984 7. (0. a white and green with a red or a green and red with a white. A 7 × 7 table is tiled with 16 rectangles 1 × 3 such that only one square remains uncovered. What is the position of this square? Tournament of the Towns. . Prove that at least one of them is green. . in the box there are exactly three balls. Find all pairs (a. Let F be the family of the subsets of the set A = {1. 3n} which have the sum of their elements divisible by 3. . b) Is it possible that after a ﬁnite number of steps there is just one ball in the box? Bulgaria. (a. two greens with a white and red. Let k > 2 be an integer. j) having integer coordinates we write the number i + j (mod n). For each element of F . . In each point (i. For which odd natural numbers n can we tile a n × n table with 1 × k or k × 1 rectangles such that only the square in the center of the table does not belong to any rectangle? Arhimede Magazine. compute the square of sum of its elements.to the following rules: two whites or two reds with a green. (a. 2. Let n ≥ 2 be an integer.

. 3}p−1 such that p−1 p| j=0 jaj . 2}p−1 such that p| j=0 jaj . Let p > 3 be a prime number and let h be the number of sep−1 quences (a1 . a2 . . Prove that h ≤ k and that the equality appears only for p = 5. 1. . . Also. a2 .10. ap−1 ) ⊂ {0. let k be the number of sequences (a1 . ap−1 ) ⊂ {0. . 1. . . IMO 1999 Shortlist 100 . .

. . we can deﬁne another ring. r1 r2 rn 2 2. combinatorics. . rn and ﬁrst term a1 . an be some complex numbers such that for any 1 ≤ k ≤ n we have ak + ak + · · · + ak = 0. an . called the ring of formal series with coeﬃcients in A and denoted A[X]. well. The vertices of a regular polygon are colored in some fashion so that each set of vertices having the same colour is the set of vertices of a regular polygon. An element of A[X] is of the form n≥0 an X n . Vasiliev. Then there are two congruent polygons among them. The set of natural numbers (including 0) is partitioned into a ﬁnite number n ≥ 2 of inﬁnite arithmetic progressions having ratios r1 . number theory. . r2 . given a commutative ring A. where an ∈ A. As for the second one. . a2 . these formal series have some very nice applications in diﬀerent ﬁelds: algebra. What is that? Well. . . 101 . reminding that the reader is supposed to be familiar with some basic analysis tools: Example 1. . The ﬁrst problem was discussed during the preparation for IMO of the USA team. then let’s give him one small hint: the methods used to solve both problems are very similar and can be included into a larger ﬁeld. . that of formal series. As we are going to see in what follows. .FORMAL SERIES REVISITED We start with a riddle and a challenge for the reader: what is the connection between the following problems: 1. If the reader has no clue. Let a1 . a2 . but it seems it is a classical result. . Then all numbers are n 1 2 equal to 0. But let’s start working now. Then the following relation is satisﬁed: a1 a2 an n−1 + + ··· + = . it is a famous problem given in a Russian olympiad and proposed by N. .

i i i Then it suﬃces to add these relations and to prove the statement by strong induction. as (s ≥ 1) to be the collection of numbers of maximal modulus among the n numbers. let f (X) = X n + bn−1 X n−1 + · · · + b1 X + b0 n the polynomial i−1 (x − ai ). experienced reader has already noticed that this problem is a trivial consequence of Newton’s relations. . Of course. . while the second one is bounded). it suﬃces to observe 1 − zp ai i=1 that the left-hand side is unbounded. we obtain a contraBy taking a sequence zp → such that r r n 1 diction with the relation = n (indeed. But what can we do if don’t know them? Here is a nice way to solve the problem (and a way to prove Newton’s relations too). 1 − zai Developing it by using 1 = 1 + x + x2 + . Assume that not all numbers are zero and take a1 . 102 . zp 1 < 1. . . (for |x| < 1). First of all. . This shows that all numbers are equal to 0. observe that the given condition implies that ak + ak + · · · + ak = 0 1 2 n for all positive integer k.Solution. . Let the common value of the modulus be r. let us consider the function n f (z) = i=1 1 . Then for all k ≥ n + 1 we have ak + bn−1 ak−1 + · · · + b0 ak−n = 0. 1−x we obtain that f (z) = n for all suﬃciently small z (which means that |z| max(|ai |) < 1). Indeed. Now.

aq . Now. . Example 2. 1 − zxi By using the same formula as in the ﬁrst problem. . . . there are some details that make the problem more diﬃcult. consider St (i) the i-th fundamental symmetric sum in xj (j = i). Then q 1 2 q m|a1 i=2 (x1 − xi ). . Gabriel Dospinescu Solution. . x2 . . x1 .We are going to discuss a nice number theory problem. whose solution is practically based on the same idea. (i) This can also be rewritten in the nicer form q m| i=1 q−2 ai (xq−1 − x1 S1 + · · · + (−1)q−1 Sq−1 ). (1 − xq z) also has all coeﬃcients multiples of m. (1 − xq z) are multiples of m. . . .. xq and m some integers such that m|a1 xk + a2 xk + · · · + aq xk for all k ≥ 0. 1 (i) (i) 103 . Obviously. it follows that the formal series a1 (1 − x2 z) .. . a2 . Let a1 . which shows that all coeﬃcients of this formal series are integers multiples of m. . we deduce immediately that q q f (z) = i=1 ai + i=1 ai xi z + . Consider this time the formal series q f (z) = i=1 ai . a simple computation shows that we have the divisibility relation: q q q m|xq−1 1 i=1 ai − xq−2 i i=1 ai S1 + · · · + (−1)q−1 i=1 (i) ai Sq−1 .. Yet. Since all coeﬃcients of a1 (1 − x2 z) .

. . . .Now. a2 . q 1 In order to solve the problem announced in the very beginning of the presentation. Therefore we can conclude. . . since xq−1 − xq−2 S1 + · · · + (−1)q−1 Sq−1 = (x1 − x2 ) . Let us observe that for any |x| < 1 we have the identity: xa1 +kr1 + k≥0 k≥0 (1) (1) (i) (i) xa2 +kr2 + · · · + k≥0 xan +krn = k≥0 xk . an . . The above relation becomes of course the very useful relation: xa1 xa2 xan 1 + + ··· + = r1 r2 rn 1−x 1−x 1−x 1−x (1) Let us multiply the relation (1) with 1 − x and use the fact that 1 − xa lim = a. Then the following relation is satisﬁed: 1 1 1 + + ··· + = 1. all we did was to write the fact that each natural number is exactly in one of the arithmetic progressions. . r2 . . . . (x1 − xn ) = 0 gives us the not-so obvious relation xq−1 − xq−2 S1 + · · · + (−1)q−1 Sq−1 = 0 1 1 for i ≥ 2. r1 r2 rn 104 . . rn and ﬁrst term a1 . . (x1 − xi−1 (x1 − xi+1 ) . We ﬁnd of course the desired relation x→1 1 − x 1 1 1 + + ··· + = 1. we need a little lemma. r1 r2 rn Solution. . Suppose that the set of natural numbers (including 0) is partitioned into a ﬁnite number of inﬁnite arithmetic progressions of ratios r1 . which is interesting itself and which we prefer to present as a separate problem: Example 3. Indeed. (x1 − xn ). the trivial identity (x1 − x1 ) .

r1 r2 rn 2 MOSP Solution. using a beautiful and hard result due to Erdos. . . this remarkable theo- rem due to Erdos asserts that if x1 . . Let us write the relation (1) in the more appropriate form: xan xa1 + ··· + =1 1 + x + · · · + xr1 −1 1 + x + · · · + xrn −1 (2) Now. . r2 . . xk are natural numbers whose 105 . .It’s time to solve the ﬁrst problem. First of all. The set of natural numbers (including 0) is partitioned into a ﬁnite number n ≥ 2 of inﬁnite arithmetic progressions having ratios r1 . . . . So: Example 4. we can say that the relation 1 1 1 + + ··· + =1 r1 r2 rn implies that max(r1 . + + ··· + = r1 r2 rn 2 Some commentaries about these two relations are necessary. . Indeed. a2 . . We will just a small. but not obvious step and we’ll be done. Then the following relation is satisﬁed: a1 a2 an n−1 + + ··· + = . . The fundamental relation is again (1). . A small computation let to the reader will show that n i=1 ai ri − ri (ri − 1) 2 = 0. rn ) < 22 n−1 . let us derive the relation (2) and then make x → 1 in the resulting expression. an . . 2 ri But it suﬃces to use the result proved in example 3 in order to conclude that we must have a1 a2 an n−1 . . x2 . . r2 . rn and ﬁrst term a1 .

. . . r2 . u k (it is trivial by induction). . ε. Let us assume that the initial polygon (which we will call big from now on) has n edges and that it is inscribed in the unit circle.sum of inverses is strictly smaller than 1. But the reader can verify immediately n that 1 1 1 1 + + ··· + =1− u1 u2 uk u1 u2 . . Using the relation proved in example 4. . . rn can be taken as rn ). . r2 . an ) < (n−1)·22 . the vertices having as aﬃxes the numbers 1. . Example 5. . The vertices of a regular polygon are colored such that vertices having the same colour form regular polygons. the solution becomes natural. Prove that there are at least two congruent polygons among them. but we also have some explicit (even though huge) bounds on ratios and ﬁrst terms. . ε2 . . . . Russian Olympiad Solution. It is now time to solve the remarkable problem discussed in the beginning of this note. a2 . ≤ x1 x2 xk u1 u2 uk where u1 = 2. . . Thus we can write 1− 1 1 ≤1− . N. . rn ) < 2 2n−1 n−1 . rn u1 u2 . it is not at all true. then 1 1 1 1 1 1 + + ··· + + + ··· + . un+1 = u2 −un +1. un−1 = un − 1 (the last relation being again a simple induction). the reader will do a short induction to prove that un ≤ 22 that max(r1 . where ε = 106 . εn−1 . . Once again. We will see that in the framework of the previous results proved here. the problem is really diﬃcult. Vasiliev. This shows that for ﬁxed n not only there is a ﬁnite number of ways to partition the set of natural numbers into n arithmetic progressions. we also deduce that max(a1 . However. . . . And here is how we can prove (since of course any number among n−1 −1 r1 . . un−1 or even better rn ≤ u1 u2 .

Let 2iπ εj = e nj . if ζ = e identity 1 1 1 p + + ··· + = . For any complex number z. Now. . Or. . n2 . it suﬃces to observe that z. Consider now n1 . zj εj j n −1 . 1 − (zz1 )n1 1 − (zz2 )n2 1 − (zzk )nk 1 − zn (2) 107 . nk the number of edges of each monochromatic polygon and let us assume that all these numbers are diﬀerent. . . we know that P (x) 1 1 = + ··· + . . . (1) 1 − (zz1 )n1 1 − (zz2 )n2 1 − (zzk )nk 1 − zn Let us assume now that n1 < min(n2 . nk ) and let us divide (1) by z n1 . . P (X) X −z X − zζ p−1 By taking X = 1. . . p−1 1 − z 1 − zζ 1 − zζ 1 − zp then we have the Proving this lemma is a very simple task. First. zζ. the simple observation that n1 + n2 + · · · + nk = n allows the new identity n n n n k zk k n 1 z1 1 nz n n 2 z2 2 n1 n2 z + z + ··· + z nk = . a technical result. . Indeed. . we will not loose generality with all these restrictions). we obtain exactly the desired identity. . Lemma 1. zζ p−1 are exactly the roots of P (X) = X p − z p . + ··· + 1 − (zz1 )n1 1 − (zzk )nk 1 − zn Also. observe that the aﬃxes of the vertices of each monochromatic polygon are zj . for some zj complex numbers on the 2iπ p unit circle. zj εj . . .e 2iπ n (of course. the hypothesis of the problem and lemma 1 allow to write n1 n nk = . . . It follows that for non-zero z we have n n n n k zk k n 2 z2 2 nz n−n1 n 1 z1 1 + z n2 −n1 + · · · + z nk −n1 = .

we are done: it suﬃces to observe that if we make z → 0 n (by non-zero values) in (2) . 2! n thus we can say that the coeﬃcient of z n in this formal series is i=1 ai . the following identity holds: n n n   j=i n aj  n n ai i=1 − i=1  + 1≤i<j≤n n ak  − · · · + (−1)n−1  k=i. On the other hand. . Example 6. . we obtain that z1 1 = 0. but it has a four-line solution!!! 108 . a2 . Indeed. the reader is right: now everything is clear. Consider the formal series n f (z) = i=1 (ezai − 1).Well. The proof ends here.. . . The following problem that we are going to discuss appeared in various contests under diﬀerent forms. an ∈ C. First of all. The conclusion is clear: not only the identity is n! true.. It is a very nice identity that can be proved elementary in a quite messy way. Solution. Here is a magical proof using formal series. it is clear that f (z) = i=1 n zai + z 2 a2 i + .j an = n! i i=1 i=1 ai . We are going to compute it in two diﬀerent ways. we can write n z f (z) = e ai i=1 n z aj j=i n − i=1 e + · · · + (−1)n−1 i=1 ezai + (−1)n . which is surely impossible since |z1 | = 1. For any complex numbers a1 . since the coeﬃkn cient of z n in ekz is .

Find a1998 . j. the generating function of the set A (we can xan ). Example 7. but also some beautiful concocts of numbers theory and combinatorics. IMO Shortlist. Next. 1998 Solution. . observe two relations: f (x2 ) = k≥0 k≥0 k (f (x2 )f (x2 3k 3k+1 )f (x2 3k+2 )) = and f (x2 ) = k≥1 k≥0 k (f (x2 3k+1 )f (x2 3k+2 )f (x2 3k+3 )) = Therefore (the reader has observed that rigor was not the strong point in establishing these relations) we have 1 − x2 1 − x2 k k f (x) = = (1+x8 )f (x) = = (1+x8 ). Here is a very nice answer: 9817030729. We shall focus a little bit more on such type of problems in the sequel. The hypothesis imposed n≥0 write in a more intuitive way f (x) = on the set A translates into f (x)f (x2 )f (x4 ) = k 1 . 23k 23k 1−x 1−x k≥0 k≥0 k≥0 k≥0 109 3k+1 3k+1 . Let A = {a0 . . 1−x Replace x by x2 . a1 . Let 0 = a0 < a1 < a2 < . 1 − x23k+1 k≥0 Now. . be a sequence of positive integers such that the equation ai + 2aj + 4ak = n has a unique solution i. . k. } and bn = 1 if n ∈ A and 0 otherwise.There aren’t only algebra problems that can be solved in an elegant manner using formal series. . 1 − x2k 1 1 − x23k k≥0 1 . consider the formal series f (x) = n≥0 bn xn . We obtain the recursive relation f (x2 )f (x2 k k+1 )f (x2 k+2 )= 1 .

This shows that the set A is exactly the set of nonnegative integers that use only the digits 0 and 1 when written in base 8. f (x2 ) − g(x2 ) = which can also be rewritten as f (x) − g(x) = 1 − x. in Olympiads from all over the world. Therefore. B and write them in explicit form f (x) = n≥0 an xn . y) ∈ A × A as in B × B. g(x) = n≥0 bn xn (as in the previous problem. the hypothesis made on the number of solutions of the equation x + y = n imposes that f 2 (x) − f (x2 ) = g 2 (x) − g(x2 ). f (x2 ) − g(x2 ) 110 f (x) − g(x) . B of the set of nonnegative integers having the property that for all nonnegative integers n the equation x + y = n with x < y has as many solutions (x. The fact that A. given in a Putnam competition: Example 8. an equals 1 if n ∈ A and 0 otherwise). We will present the latest one. 1−x Also. g the generating functions of A. 1−x . Consider f. under different forms. The following problem is an absolute classic. A quick computation based on this observation shows that the magical term asked by the problem is 9817030729. B form a partition of the set of nonnegative integers can be also rewritten as f (x) + g(x) = n≥0 xn = 1 . Solution. Find all partitions with two classes A. It appeared.

1. B are respectively the set of nonnegative integers having even (respectively odd) sum of digits when written in binary. if you want) n f (x) = i=1 (1 + xi ). it suﬃces of course to 111 . the idea is the same as in the previous problems: replace x by x 2k and iterate the process. we deduce that f (x) − g(x) = k≥0 (1 − x2 ) lim k n x→∞ f (x2 ) 1 . k be positive integers such that n ≥ 2k−1 and let S = {1. Prove that the number of subsets A ⊂ S such that x ≡ m (mod 2k ) does not depend on m ∈ {0. Then the reader can easily verify that n→∞ lim f (x2 ) = 1 and lim g(x2 ) = 0.Now. . We will discuss two nice problems in which formal series and complex numbers appear in a quite spectacular way: Example 9. x∈A Balkan Olympiad Shortlist 2005 Solution. . . After multiplication. Taking into account the relation f (x) + g(x) = n≥0 xn = 1 . . then we have certainly done the job. Let us consider the function (call it formal series. 2k − 1}. − g(x2n ) Let us assume without loss of generality that 0 ∈ A. . n→∞ n n This shows that actually f (x) − g(x) = k≥0 (1 − x2 ) = k≥0 k (−1)s2 (k) xk . 1−x we ﬁnally deduce that A. . where s2 (x) is the sum of digits of binary representation of x. . n}. . In order to prove this. Let n. k −1 If we prove that 1 + x + · · · + x2 divides f (x). 2.

we have n   m−1 j=0  xjai  = n i=1 f (x) = i=1 1 − xmai . Now.prove that any 2k th root of unity. Example 10. 2k−1 − 1} we have 2lπ 2lπ cos k + i sin k 2 2 and so f cos 2lπ 2lπ + i sin k k 2 2 = 0. . . Look at the set A = i=1 ei ai | 1 ≤ ei ≤ m and observe that we can assume that A is a complete system of residues modulo mn (otherwise. Of course. Therefore (the hypothesis n 2iπ ensures that εai = 1) we must have i=1 (1 − εmai ) = 0. e2 . none of them divisible by mn−1 . . n ≥ 2 be positive integers and a1 . take ε = e mn . it will be a combinatorial problem. it is time for a tough problem. . we must have f (ε) = 0. . a2 . But this surely contradicts the fact that none of the numbers a1 . whose nice solution below was found by Constantin Tanasescu. 1 − xai On the other hand. such that |ei | < m for all i and such that mn |e1 a1 + e2 a2 + · · · + en an . . Finally. . 2k−2−v2 (l) = −1 which proves our claim and ﬁnishes the solution. Since A is a complete system of residues modulo mn . . . Prove that one can ﬁnd integers e1 . IMO Shortlist 2002 n Solution. . en . . . 112 . consider f (x) = p∈A xa . But it suﬃces to observe that for any l ∈ {1. . On one hand. an integers. . except for 1 is a root of f . . an is a multiple of mn−1 . a2 . the conclusion is immediate). 2. Let m. not all zero. .

ml(c) Adrian Zahariuc Solution. Proposed problems 1. Let f (x) = c∈C xl(c) . Prove the inequality: c∈C 1 ≤ 1. For any c ∈ A. which is nothing else than m m the desired inequality. thus g(x) < ∞ 1 1 and f (x) < ∞ for x < . All we need now is to make x tend m 1 1 to and we will obtain that f ≤ 1. for all x ∈ 0. zn be some arbitrary complex numbers. |zn |) − ε. f (x)g(x) = g(x) − 1. Let A be the set of all words which can be formed using m ≥ 2 given letters. |z2 |. Prove that for any ε > 0 there are inﬁnitely many numbers n such that k k k k |z1 + z2 + · · · + zn | > max(|z1 |. let C ⊆ A be a set of words. Let S be the set of all words which can be obtained by concatenating words from C. . z2 . . .Example 11. S (and C) has at most mk elements of length k. We know that any word from A can be obtained in at most one way by concatenating words from C. : m m f (x)g(x) = g(x) − 1 < g(x) 1 . . let l(c) be its length. g(x) = s∈S xl(s) . . . 113 . . By the deﬁnition of S. 1 − f (x) Now. Let z1 . Also. Thus. we have that: g(x) = 1 + f (x) + f 2 (x) + · · · = Therefore. (∗) 1 . . and so f (x) < 1 for all x ∈ 0.

lim n→∞ yn 114 . . Prove that if we partition the natural numbers into a ﬁnite number of inﬁnite arithmetic progressions. 1994 5. an relatively prime positive integers. . . then xn = 1. . 4. 2. an such that {|ai − aj | | 1 ≤ i < j ≤ n} = 1. . . .2. For which positive integers n can we ﬁnd real numbers a1 . . a2 . Let a1 . then there will be two of them having the same ratio. Find the general formula for the sequence (xn )n≥1 given by xn+k = a1 xn+k−1 + · · · + ak xn in function of x0 . xk−1 . In how many ways can we parenthesis a non-associative product a1 a2 . where n is a given positive integer? Romanian TST. . Here a1 . ak are arbitrary complex numbers. an ? Catalan’s problem 7. Deﬁne A1 = ∅. 1. . . B1 = {0} and An+1 = {1 + x| x ∈ Bn }. . . What are the positive integers n such that Bn = {0}? AMM 6. 3. Bn+1 = (An \ Bn ) ∪ (Bn \ An ). x1 . . . . Find in closed form a sequence (xn )n≥1 such that if (yn )n≥1 is the number of positive integral solutions to the equation a1 x1 + a2 x2 + · · · + an xn = k. n 2 ? China TST 2002 8. How many polynomials P with coeﬃcients 0. . . . . 2 or 3 satisfy P (2) = n. . a2 . .

Komal 15. Ak ∈ Mn (C) be some complex nθn matrices such that C p! for any natural number p ≥ 1. (bn ) : a1 = b1 = 0 and an = nbn + a1 bn−1 + a2 bn−2 + · · · + an−1 b1 . Prove that for any prime number p we have p|ap . (n + 3)an+1 = (2n + 3)an + 3nan−1 . i < j. Find all positive integers n with the following property: for any real numbers a1 .9. . Komal 14. . knowing the numbers ai + aj . . 12. Let A1 . Deﬁne two sequences of integer numbers (an ). Then prove that An = 0 for all 1 ≤ i ≤ k. Is there an inﬁnite set of natural numbers such that all suﬃciently large integer can be represented in the same number of ways as the sum of two elements of the set? D. Here C does not depend on p ≥ 1 and Ap + Ap + · · · + Ap ≤ 1 2 k X = max |xij |. . . Find all possibilities to color a regular polygon in the way presented in example 5. an uniquely. Erdos and Selfridge 13. . Newman 11. determines a1 . an . . . Prove that all terms of this sequence are integers. . Is it possible to partition the set of all 12-digit numbers into groups of 4 numbers such that the numbers in each group have the same digits in 11 places and four consecutive digits in the remaining place? 115 . . i 1≤i. a2 . Suppose that a0 = a1 = 1. .j≤n Gabriel Dospinescu 10. A2 . . a2 .

Saint Petersburg Olympiad 16.εn ∈{−1. .. = Gabriel Dospinescu 17. bi IMO Shortlist 2004 116 . ci . A set of positive integers A has the property that for some positive integers bi . an . . Prove that n i=1 1 ≤ 1...ε2 . . . . a2 . a2 n 1 2 2n for any real numbers a1 . the sets bi A + ci .. Prove the following identity n k=1 1≤i1 <i2 <···<ik ≤n ε1 . 1 ≤ i ≤ n are disjoint subsets of A.1} (−1)k (ε1 ai1 + ε2 ai2 + · · · + εk aik )2n 2k (−1)n (2n)!a2 a2 . .

1≤i<j≤n (1) . But ﬁrst. . aj − ai For any integers a1 ... a12 a22 .. xn−1 n 117 = (xj − xi ). has only the trivial solution. . determinants.... . a1n a2n .... especially because it makes a beautiful connection between number theory and algebra. But linear algebra has an important role as well... For example. Secondly. the fact that any homogeneous linear system   a11 x1 + a12 x2 + · · · + a1n xn = 0      a x + a x + ··· + a x = 0 21 1 22 2 2n n       a x + a x + ··· + a x = 0 n1 1 n2 2 nn n in which a11 a21 ..... we need some basic facts about matrices. we need Vandermonde’s identity 1 1 .. x2 n .. xn−1 1 xn−1 2 . At the appropriate time. ... ann =0 an1 an2 . ... an the number is an inj−i 1≤i<j≤n teger.. . . .... xn x2 1 x2 2 .. . and systems of linear equations. . This discussion practically started from the following diﬃcult problem that we solved in the chapter ”Look at the exponent!” and which appeared in the American Mathematical Monthly a long time ago. . a2 .... 1 x1 x2 . . You will see a nice and short solution to this problem..NUMBERS AND LINEAR ALGEBRA We have seen how analysis can help in solving number theory problems.

....... . .. . σ(3) n−1 ... (n − 1)!.. for a better understanding. This time. The formula in the left-hand side suggests that we might use Vandermonde’s identity (1). we will prove a result from the theory of permutations. ... But we also need a small trick. 1 σ(2) . As usual.. 2.. σ(n) n−1 1 · 1 . Using the fact that det A = det t A for any matrix A.... ... σ(1)n−1 σ(2)n−1 .. σ(n)n−1 σ(n) ..... . . Sn−1 Sn .. the reader should have some more knowledge on linear algebra).. 1 σ(1) σ(2) .. by multiplying the two determinants we ﬁnd  2  (σ(j) − σ(i)) = 1≤i<j≤n 1 σ(1) . Then (σ(j) − σ(i)) = 1! · 2! . we start with an easy and classical problem..With these basic facts (of course. S2n−2 . ... .. . So.. σ(2) n−1 1 σ(3) . 1 σ(n) ... .. S0 = S1 .. .. . . .. . .. n. Example 1.. Let σ be a permutation of the numbers 1. . 118 . .. Here is a nice solution. 1≤i<j≤n σ(1)n−1 σ(2)n−1 σ(3)n−1 . σ(1) n−1 1 σ(2) ..... we get 1 σ(1) . Sn−1 .. .. σ(n)n−1 = (σ(j) − σ(i)).. 1≤i<j≤n Solution. ... 1 σ(n) .. .. 1 σ(3) . S1 S2 ... .. we are ready to begin the discussion.. .

.. Given a polynomial with complex coeﬃcients. we need to express Si = xi +xi +· · ·+xi n 1 2 in terms of the coeﬃcients of the polynomial. . we can answer immediately the following question: Example 2... . ... . Sn−1 Sn . we have Si = 1i + 2i + · · · + ni and repeating the arguments we conclude that 2 2    Hence (σ(j) − σ(i)) = 1≤i<j≤n (σ(j) − σ(i)) =  1≤i<j≤n (j − i) . . n 1 2 So. and divisions on its coeﬃcients? Solution. . 1≤i<j≤n Using this result. where Si = xi +xi +· · ·+xi . . . xn are the zeros of the polynomial... . .. which combined yield an Si + an−1 Si−1 + · · · + an−i+1 S1 + ian−i Si = 0... multiplications. 119 i ∈ {1. . S1 S2 .where Si = σ(1)i + σ(2)i + · · · + σ(n)i . Then this polynomial has a double root if and only if  2  (xi − xj ) = 0. But since σ is a permutation of the numbers 1. . 1!. n. }. x2 . . .. S2n−2 . . . Sn−1 .. 2. But we have seen that   1≤i<j≤n 2 (xi − xj ) = S0 S1 . 1≤i<j≤n where x1 . But this is a consequence of Newton’s and Vieta’s formulas. . . Let f (x) = a0 + a1 x + · · · + an xn .. . we can. 2. . can we decide if it has a double zero only by performing additions. 1≤i<j≤n (j − i) = (n − 1)!(n − 2)! . . Yes.

You may know the following classical problem: if a. then a = b = c = 0. g) is a polynomial of degree at most n − 1 that divides an irreducible polynomial f of degree n. a direct solution could be the following: from Eisenstein’s criterion. . a1 . This cannot happen. since after all Si is a symmetric polynomial in n variables. Of course. hence it can be expressed only in terms of the fundamental symmetric polynomials. a0 = a1 = · · · = an−1 = 0. We need much more. then the polynomial g(x) = a0 + a1 x + · · · + an−1 xn−1 is not relatively prime with f . . Hence gcd(f.The last formula allows us to prove by induction that Si can be expressed only in terms of the coeﬃcients of the polynomial (this could have been shown much easier. . . Consequently. Prove that if a0 . a0 + a1 2 + · · · + an−1 then a0 = a1 = · · · = an−1 = 0. an−1 ∈ Q satisfy √ √ n−1 n 2n−1 = 0. If a0 + a1 n 2 + · · · + an−1 2 = 0. which can also be expressed in terms of the coeﬃcients due to Vieta’s formulas). . √ √ n−1 n−1 so if a0 + a1 n 2 + · · · + an−1 2 = 0 for some rational numbers a0 . the polynomial f (X) = X n − 2 is irreducible. c ∈ Q verify √ √ a + b 3 2 + c 3 4 = 0. i. which shows that the answer to the problem is positive. √ √ n−1 n−1 Solution. . then √ √ n−1 n ka0 + ka1 2 + · · · + kan−1 2n−1 = 0 120 . Have you ever thought about the general case? This cannot be done only with simple tricks. b. Example 3.e. unless g = 0. . This time we will have to be careful to work in the most appropriate ﬁeld. But here is a beautiful proof using linear algebra. an−1 . a1 . . we can decide whether   1≤i<j≤n 2 (xi − xj ) = 0 only by using the described operations on the coeﬃcients of the polynomial.

. an−1 an−2 . . 2n−1 ) as a nontrivial solution to the system.. . 2a1 a1 a0 . .... ... .. .. But what can we do now? Expanding the determinant leads nowhere. let us ﬁll in the blanks.. .. we can choose k to be the least common multiple of all denominators of the numbers a0 ... an−1 an−2 . the determinant of the system must be zero and this will imply a0 = a1 = · · · = an−1 = 0. the values (k1 . n 2.. . .. an−1 ∈ Z (for example. a1 . . k2 . . What √ √ n are the best values for k? This can be seen by noticing that 2n−1 · n 2 = √ √ n 2 ∈ Z. . having nontrivial solutions. 2a2 ... n 2. .. . . a1 . This time the ﬁeld is Z2 . . kn ) = (1. It equals an = 0. a1 a1 a0 . . ..    √  n n n−2  a · √2n−1 + 2a + · · · + a 2 =0 0 1 n−1 · √ √ n Viewing (1. .. . . a0 = 0. . 121 . .. As we said before passing to the solution... since in this case the determinant can be easily computed. Thus.for any real number k. Now. So. Hence a0 must 0 be even. Then. an−1 ). . a0 = 0... we should always work in the most appropriate ﬁeld. The idea is to choose n values for k so that to obtain a system of linear equations. . that is a0 = 2b0 and we have b0 an−1 . . 2n−1 ) are good and the system becomes  √ √  a0 + a1 · n 2 + · · · + an−1 · n 2n−1 = 0     n n  a · √2 + a · √22 + · · · + 2a =0 0 1 n−1  . . ... we may assume that a0 . 2a2 . . we conclude that a0 2an−1 .

Similarly. The following problem can become a nightmare despite its simplicity. we interchange the ﬁrst two lines of the determinant. c ∈ Z}. it is an = 0. y ∈ A then xy ∈ A. Prove that if x. Indeed. b. but the idea was too nice not to be presented here. we can repeat this as long as we want.Now. 122 . By the method of inﬁnite descent. Proof. The observation that a c a + b + c − 3abc = 3 3 3 b b a c c b a leads to a quick solution. Example 4. but when we expand it in Z2 . But of course. The above solution might seem exaggeratedly diﬃcult compared with the one using Eisenstein’s criterion. we 1 ﬁnd that all ai are even. Its value remains 0. Then we also have √ √ n−1 n−1 2 = 0 and with the same reasoning we b0 + b1 · n 2 + · · · + bn−1 · conclude that all bi are even. Let A = {a3 + b3 + c3 − 3abc| a. we ﬁnd that a0 = a1 = · · · = an−1 = 0. it suﬃces to note that  a c b  x z y      b a c  y x z  =    c b a z y x   ax + cy + bz az + by + cx ay + bx + cz   =  ay + bx + cz ax + cy + bz az + by + cx    az + by + cx ay + bx + cz ax + cy + bz and thus (a3 + b3 + c3 − 3abc)(x2 + y 3 + z 3 − 3xyz) = A3 + B 3 + C 3 − 3ABC. Let us write ai = 2bi .

b. there is a crucial observation to be made: the given condition can be also written in the form det A = 1. at least for n = 3. then there are integers x. Solution. They are relatively prime and 123 . . . c are relatively prime integers. . . . u. v. C = ax + cy + bz. . using clever manipulations of determinant properties and induction on the dimension n of the vector v. g g greatest common divisor of v1 .. v2 . We all know the famous Bezout’s theorem. Prove that if a. . . . for n = 2 it is exactly Bezout’s theorem. kn such that k1 a1 + k2 a2 + · · · + kn an = 1. First of all. . where  a x u   A= b y   v . an are relatively prime. . w such that a(yw − zv) + b(zu − xw) + c(xv − yu) = 1.. stating that if a1 . There is a simple proof of this theorem. Any vector v whose integer components are relatively prime is the ﬁrst column of an integral matrix with determinant equal to 1. k2 . z. You see. where g is the are relatively prime. Example 5. identities are not so hard to ﬁnd. Consider the numbers . Indeed. let us prove a much more general result. The following problem claims more. . . . .  c z w So. then one can ﬁnd integers k1 . .where A = ax + cy + bz. B = az + by + cx. y. . a2 . vn ) such that vi vn−1 v1 . vn−1 . . . Now. . assume that it is true for vectors in Z n−1 and take v = (v1 . Theorem.

But we have a hurdle here. but yet we did not present its solution yet. a1..2 . What does the expression (aj − ai suggest? It is the Vandermonde’s 1≤i<j≤n 1≤i<j≤n identity (1)... an then aj − ai ∈ Z. .2 . .. a1. an−1.. We might want to use the same formula for the expression This is a dead end. . an .. vn−1 g a12 . (−1)n−1 β . . It is now time to do it. .n−1 ..n−1 (−1)n−1 β 0 We said at the beginning that the discussion started from the diﬃcult problem that appeared in AMM.. having determinant equal to 1. j−i Armond Spencer. vn−1 g (−1)n−1 α    .. In this case... . . . v1 g     .. . . a2 .. we can ﬁnd α. Example 6. For any integers a1 . 1≤i<j≤n (j − i) equal to in 1≤i<j≤n 124 .    an−1. Now.    vn−1 an−1.thus we can ﬁnd an integral matrix      v1 g . AMM E 2637 Solution.. With this introduction.  vn 0 .. . . associated to a1 ... .. the way to proceed is clear. But we have seen what is (j − i). . ..n      an−1. β such that αg + βvn = 1.... it is not diﬃcult to verify that the following matrix has integral entries and determinant equal to 1:  v1 a12 . . a2 . using Bezout’s theorem...n−1 .

. ...... (n − 1)! a1 .. . 1≤i<j≤n . . Because each number ai is an integer. 1!... 1 an ...... A variant of this problem was given in 2004 at a TST in Romania and turned out to be a hard problem.. the last step is the most important. 1 a3 . 1 an 1! .. which can be proved easily by subtracting lines.. 1≤i<j≤n n−1 n−1 n−1 an−1 a2 a3 an 1 As usual... . ..... 1 a2 .. It equals (n − 1)!(n − 2)! . a2 n−1 . 1 an 1 an 2 . The above formula can be rewritten as 1 a1 1! .the ﬁrst problem. n−1 a2 (n − 1)! .... the determinant itself is an integer and the conclusion j follows. Now.... And now we recognize the form 1 a1 1 aj − ai = j−i a1 2 .. an n−1 . .. . an−1 2 (n − 1)! 1 a3 1! .. . 125 .. ..... ..... ..... an−1 n (n − 1)! 1≤i<j≤n aj − ai = j−i . We end the unit with a very nice and diﬃcult problem that also appeared in AMM in 1998. . . we can write 1 aj − ai 1 = j−i 1! · 2! . . an−1 1 (n − 1)! 1 a2 1! . a1 n−1 1 a2 1 a2 2 .

x3 = 3 and xn+4 = xn + xn+1 . . . . 2. that is T rAp ≡ T rA (mod p) for any integral matrix and any prime p. . The proof is not trivial at all. Indeed. This is true because xn = xn−4 + xn−3 = T rAn−4 + T rAn−3 = T r(An−4 (A + I4 )) = T rAn . 1 ≤ k ≤ p − 1 are multiples of p) k n n p trA = i=1 p λp i = i=1 126 λi = (T rA)p . then by placing ourselves in a ﬁeld in which the characteristic polynomial of A has all its zeroes λ1 . 3 it is not diﬃcult to verify it. λ2 . .Example 7. Consider the sequence (xn )n≥0 deﬁned by x0 = 4. λn . . Prove that for any prime p the number xp is a multiple of p. This is going to be the easy part of the solution. n−1 and prove that it is also true for n. . x1 = x2 = 0. Now. Now. 2. This ﬁeld has clearly characteristic p (it contains Zp ) and so we have (using the binomial formula and the fact that all coeﬃcients p . Hence the claim is proved. assume that the statement is true for all i = 1. which can be easily veriﬁed by a simple computation. A possible advanced solution is to start by considering the matrix A obtained by reducing all entries of A modulo p. let us prove an important result. AMM Solution. We have used here the relation A4 = A + I4 . Let us consider the  0   1  A=  0  0 matrix 0 0 1   0 0 1    1 0 0   0 1 0 and let trX be the sum of the entries of the main diagonal of the matrix X. . for n = 1. . We will ﬁrst prove that xn = T rAn (here A0 = I4 ).

Xn and any cyclic permutation σ. For these matrices we clearly have T rAp ≡ T rA (mod p) and by using the above result we can write (using Fermat’s little theorem one more time):  T rAp = T r  i. . . But there is a beautiful elementary solution.. j) and 0 elsewhere. . . .B} A1 A2 .. Thus we have proved that T r(A + B)p ≡ T rAp + T rB p and by an immediate induction we also have k (mod p) T r(A1 + · · · + Ak ) ≡ i=1 p T rAp .B} already noticed that Fermat’s little theorem comes handy once again).. ≡ i. since the sum of T r(A1 A2 .B} A1 A2 . consider the matrices Eij that have 1 in position (i. Ap ) is a multiple of p in any cycle. . B we have T r(AB) = T r(BA) and by induction. .. Now.. Let us consider two integral matrices A..Ap ∈{A. . Xn ) = T r(Xσ(1) . . note that in the sum A1 . . Ap ) ≡ T rAp + T rB p modulo p (the reader has A1 . A p . .j aij T rEij = T rA 127 .j p T r(ap Eij ) ≡ ij i..Ap ∈{A. Observe that for any A. T r(X1 .. . i Next. Thus p T r(A1 A2 . for any X1 . .. Xσ(n) ). .j p aij Eij  (mod p). . B and write (A + B)p = A1 ..from where the conclusion is immediate via Fermat’s little theorem..Ap ∈{A. Ap we can form 2p − 2 groups of p-cycles and we have two more terms Ap and B p . ..

Consider 2005 real numbers with the following property: whenever we eliminate one number. n − 1}. p such that a = nw − pv. 2. Compute the product 0≤i<j≤n−1 (εj − εi )2 . the rest can be divided into two groups of 1002 numbers each and having the same sum per group. Prove that 2 Fn+1 Fn−1 − Fn = (−1)n and Fm+n = Fn Fm−1 + Fn+1 Fm . Prove that the number 2 + 3 + 5 + 7 is irrational. f2 (x) = x and fn+1 (x) = xfn (x) + fn−1 (x). v. 6. b = pu − mw. ap+1 have the property: no matter how we eliminate one of them. Let p be a prime and suppose that the real numbers a1 . . . Let a. n. Let F1 = 1. Prove that all the 2005 numbers are equal. 1. 128 . . Prove that this sequence satisﬁes the following Fibonacci-type relations fm+n = fn fm−1 +fn+1 fm . where 2kπ 2kπ + i sin n n εk = cos for all k ∈ {0. the rest of the numbers can be divided into at least two nonempty classes.The result is proved and with it the fact that xp is a multiple of p. . 4. w satisfying au + bv + cw = 0. Romania 7. Problems for training 1. √ √ √ √ 3. TST 1989. F2 = 1 and Fn = Fn−1 + Fn−2 for all n ≥ 3 be the Fibonacci sequence. 5. . Let the sequence of polynomials (fn )n≥1 be deﬁned by f1 (x) = 1. b. Prove that for any relatively prime integers u. . a2 . c = mv − nu. Octavian Stanasila. . c be relatively prime nonzero integers. there are integers m. .

An ) such that 1≤i<j≤n (0) (0) (0) max |Ai − Aj | < 1 22005 . Prove that for any prime p and any positive integer m. a2 . . a2 . . a2 . . . (n − 2)! by any other multiple of 1!2! . an ∈ R and deﬁne ai n (k) = (−2)k k=0 n (k) (0) ai = (−1)n ai k for all 1 ≤ i ≤ n.Romania 8. . Prove that if we start with an arbitrary n-tuple (a1 . . (n − 2)!. ai (k−1) (k−1) + ai+2 for 2 all k ≥ 1 and 1 ≤ i ≤ n (the indices are taken modulo n). . . . a2 . . Let a1 . . Calin Popescu. . Prove that 11. an ). Gabriel Dospinescu 129 . x2 = 2c. . . x1 = 0. . . . Moreover. Romania 9. an (aj − ai ) 1≤i<j≤n is an integer divisible by 1!2! .. . a2 . after ﬁnitely many moves we obtain an n-tuple (A1 . . .. c be integers and deﬁne the sequence (xn )n≥0 by x0 = 4. . x3 = 3b and xn+3 = axn−1 + bxn + cxn+1 .. . an the following number lcm(a1 . A2 . . TST 2004. . an ∈ R. 2 2 2 2 . .. (n − 2)!. the number xpm is divisible by p. . Prove that for any integers a1 . . . we cannot replace 1!2! . . x2 . A move is transforming the n-tuple (x1 . an ) a1 a2 . . Let a.any two of them being disjoint and each class having the same arithmetic mean. Prove that a1 = a2 = · · · = ap+1 . Marius Radulescu. 10. . . . Let a1 . . xn ) into the n-tuple x1 + x2 x2 + x3 xn−1 + xn xn + x1 . . . b. . TST 1994 . .

f (n!) ≡ 1 (mod n!) and so if we pick a prime divisor of each of the numbers f (n!). so the tough solver will have his own part of lecture. .. For suﬃciently large n.. Although in the topic Analysis against number theory we proved an even more general result using a nice analytical argument. . Moreover. due to Schur. We will start with an essential result. (Schur) Let f ∈ Z[X] be a non constant polynomial. . f (n). . the list is just a small and insigniﬁcant introduction to this ﬁeld. they are non-zero integers. f (2). There is one result that should be remembered. as usual. when extremely diﬃcult problems with extremely simple statements can be asked. we prefer to present here a purely arithmetic proof.. b. In this small note we joined together some classical arithmetic properties of polynomials. . that is for any polynomial f ∈ Z[X] and any diﬀerent integers a. a − b divides f (a) − f (b).. we obtain the conclusion (since in particular any such prime divisor is 130 . suppose that f (0) = 1 and consider the numbers f (n!). Proof. especially in a chapter like this one. As usual. how many times happened to pass too many time on a problem just because we neglected basic and trivial aspects of it? This is why we think that speaking about these ”old fashioned tricks” is not lack of imagination. and which appeared in many variants in contests. Yet. . Then the set of prime numbers dividing at least one non-zero number between f (1). .ARITHMETIC PROPERTIES OF POLYNOMIALS Another topic with old fashioned tricks. but rather an imperious need. but some basic things should become reﬂex and between them there are also some problems we shall discuss. Example 1. we kept some chestnuts for the end of the note. will surely say the reader at ﬁrst about this small note. First. is inﬁnite. this is the fundamental result that we shall use continuously. Of course. Practically.

g ∈ Z[X] are monic non constant irreducible polynomials such that for all suﬃciently large n. the Chinese remainder theorem. Then f = g. of course. Suppose that f. This shows that for all suﬃciently large n. all prime factors of f (n) divide N . the two polynomials are irreducible in Q[X]. . k some positive integers. Using Bezout’s theorem we conclude instantly that we can ﬁnd a non zero integer N and P. Then prove that there exists a positive integer a such that each of the numbers f (a). we easily get the conclusion. f (a + 1). Bulgarian Olympiad 131 . if f (0) = 0. by Gauss’s lemma. this contradicts Schur’s result. the conclusion is obvious. Suppose thus f (xf (0)) that f (0) = 0 and consider the polynomial g(x) = . The result of example 2 remains true if we assume the same property valuable for inﬁnitely many numbers n. Even more. as usual. as we have already said. Solution. Obviously. A reﬁnement of Schur’s lemma is discussed in the following example. f (a + n − 1) has at least k distinct prime divisors. f (0) g ∈ Z[X] and g(0) = 1. Yet.greater than n). Interested reader will ﬁnd a rich literature on this ﬁeld. then the above remark and the fact that they have the same leading coeﬃcient implies they are relatively prime in Q[X]. Example 2. important consequences. f (n) and g(n) have the same set of prime divisors. . But. This result has. Let f ∈ Z[X] be a non constant polynomial and n. Example 3. Here is a nice application. The ingredient is. the proof uses some highly non elementary results of Erdos in this ﬁeld. if they are not equal. . Applying now the ﬁrst part of the solution. Q ∈ Z[X] such that f P + gQ = N . . Now. Indeed.

. Of course. for all n > k we have g(n)|g(n + g(n))|(n + g(n))n+g(n)−1 − 1 and g(n)|nn−1 − 1. This shows that p|g(p + q) and thus 132 .j=1. (p + 1)q−1 − 1) = (p + 1)gcd(g(p+1). Now. then. there exists M such that for all n > M we have g(n)|nn−1 − 1. Supposing the contrary. In particular. let us consider an arbitrary solution and write it in the form f (X) = (X −1)r g(X) with r ≥ 0 and g ∈ Z[X] such that g(1) = 0. Gabriel Dospinescu Solution. We clearly have q|g(p + 1)|(p + 1)g(p+1) − 1 and q|(p + 1)q−1 − 1. . it follows that each of the numbers f (a). Now. it follows that we actually have p = q. Let us consider an array of diﬀerent prime numbers (pij )i. let us consider a prime number p > k and let us look at the smallest prime divisor of g(p + 1) > 2. Since gcd(g(p + 1). we may assume that the leading coeﬃcient of g is positive. . Example 4. f (a + 1). Thus. Now. this implies that g(n)|nn+g(n)−1 − 1 = (nn−1 − 1)ng(n) + ng(n) − 1. f (n)|nn−1 − 1. f (a + n − 1) has at least k distinct prime divisors. that is g(n)|ng(n) − 1 for all n > k. since changing g and his opposite has no eﬀect. . We know that this is possible from Schur’s theorem. Find all polynomials with integer coeﬃcients f such that for all suﬃciently large n. q − 1) = 1 (by minimality) and gcd((p + 1)g(p+1) − 1. using the Chinese remainder theorem we can ﬁnd a positive integer a such that ai −1 ≡ xij (mod pij ). Thus one can ﬁnd k > M such that for all n > k we have g(n) > 2 and g(n)|nn−1 − 1. we deduce that g(n)|g(n + g(n)) for all n.Solution. We will prove that g is constant. Using the fundamental result. Since clearly f (X) = X − 1 is a solution.q−1) − 1 = p. Classical arithmetic ”tricks” and the fundamental result that a − b divides f (a) − f (b) are the main ingredients of the following problem. since n + g(n) − n|g(n + g(n)) − g(n).k such that for some positive integers xij such that f (xij ) ≡ 0 (mod pij ).

After reading the solution of the following problem. 1.b) −1. Finally. A simple argument shows that gcd(m(2m − 1). (2n > M ). in order to show that |c| = 1.(again using the fundamental result) p|g(1). This is very simple to realize. we obtain that n − 1|nn−3 + 2nn−4 + · · · + (n − 3)n + (n − 2) + 1 for all suﬃciently large n. it suﬃces to exhibit k < m < n such that gcd(m(2m − 1). c|2n(2 n −1)−1 − 1 for all n > M . Let g(X) = c. This contradiction shows that g is indeed constant. 2 Hence r ≤ 2. 2b −1) = 2gcd(a. since nn−3 + 2nn−4 + · · · + (n − 3)n + (n − 2) + 1 ≡ 1 + 2 + · · · + (n − 2) + 1 (n − 1)(n − 2 + 1 (mod n − 1). let us prove that r ≤ 2. with r ∈ {0. the relation ≡ nn−1 − 1 = (n − 1)2 [nn−3 + 2mn−4 + · · · + (n − 3)n + (n − 2) + 1] shows that (n − 1)2 |nn−1 − 1 for all n > 1 and allows to conclude that all solutions are the polynomials ±(X − 1)r . n(2n − 1)) = 1 and so |c| = 1. we must have g(1) = 0. Finally. Thus. Since gcd(2a −1. k and to choose n a prime number greater than m(2m − 1). we deduce that (n − 1)3 |nn−1 − 1 ⇔ (n − 1)1 |nn−2 + nn−3 + · · · + n + 1 for all suﬃciently large n and since nn−2 + nn−3 + · · · + n + 1 = = n − 1 + (n − 1)[nn−3 + 2nn−4 + · · · + (n − 3)n + (n − 2)]. 2}. Since this happens for any prime number p > k. it is extremely 133 . which is clearly impossible. n(2n − 1)) = 1. Actually. it suﬃces to take m a prime number greater than M. Indeed. the reader will have the impression the problem is very simple. Supposing the contrary.

we can take a prime number q > |c| and a number n such that q|p1 (n) = 0. This contradiction shows that the hypothesis s ≥ 1 was wrong and thus the result of the problem follows. Therefore. . so we must have vq (p1 (n)) > 2. we can easily establish the congruency p1 (n + q) ≡ p1 (n) + qp1 (n) (mod q 2 ). . . which contradicts q > |c| and Q(x)p1 (x) + R(x)p1 (x)p2 (x) . . pks g k where 1 ≤ ki < k and pi are diﬀerent irreducible polynos mials in Q[X]. The choice q > |c| ensures that q does not divide p1 (n)p2 (n) . Yet. Since p1 is irreducible in Q[X]. ps (x) = c. using the result from example 1. using the binomial formula. 134 . But the hypothesis says that k|vq (f (n)). ps (n) and so vq (f (n)) = vq (p1 (n))+ kvq (g(n)). R and a positive integer c such that Q(x)p1 (x) + R(x)p1 (x)p2 (x) . . ps (x) = c. Suppose that s ≥ 1 (which is the same as denying the conclusion). Let us assume the contrary and let us decompose f = pk1 1 . Now. Folklore Solution. . Example 5. Let f ∈ Z[X] be a non constant polynomial and k ≥ 2 a positive integer such that k f (n) ∈ Q for all positive integers n. . . ps and thus (using Bezout’s theorem and multiplication with integers) there exist some polynomials with integer coeﬃcients Q. We shall have of course q|p1 (n + q) (since p1 (n + q) ≡ p1 (n) (mod q)). . we must have q|p1 (n). it is relatively prime with p1 p2 . In exactly the same way we obtain vq (p1 (n + q)) ≥ 2. There are many possible approaches that fail and the time spent for solving such a problem can very well tend to inﬁnity.diﬃcult. . Then there exists a polynomial g ∈ Z[X] such that f = g k .

before counting such polynomials. of course. . . . The equality of the two sets is thus proved. . . . qs its prime divisors. we can ﬁnd a positive integer x such that x ≡ ui (mod qi ). . . . 2. if we ﬁnd the good way to solve it. p2 . q2 . ∀ m ∈ N∗ }. we have T ⊂ M \ i≤r Ai . where M is the set of monic polynomials of degree n. although the problem itself is not so diﬃcult. . There are many arithmetic properties of polynomials speculated in this problem. . . f (3). k) = 1. The converse. USA TST 2005 Solution. f (3). It follows then that for any prime number q at least one of the numbers f (1). Since f is monic. . n!}.. . 2. . however is not that easy. f (2). is not a multiple of q. at least 71% of all monic polynomials of degree n. in the sequence f (1). n!} are special. pr all prime numbers at most equal to n and let us consider the sets Ai = {f ∈ M | pi |f (m). . Lagrange’s theorem ensures that we can ﬁnd m such that p is not a divisor of f (m). Let now k > 1 and q1 . thus gcd(f (x). Let us suppose that f ∈ Z[X] belongs to M \ i=1 Ai and let p be a prime number greater than n. Gabriel Dospinescu.. . . Titu Andreescu. 135 .The next problem was given in the USA TST 2005 and uses a nice combination of arithmetic considerations and computations using complex numbers. with coeﬃcient in the set {1. us such that qi does not divide f (ui ). . Obviously. Example 6. . Then f (x) ≡ f (ui ) (mod qi ) and thus qi does not divide f (x). Let p1 . . We shall prove that the set T of special polynomials is r exactly M \ i=1 r Ai . Prove that for any n > 1. with coeﬃcient in the set {1. . Of course. Let us call a polynomial f ∈ Z[X] special if for any positive integer k > 1. it would be better to ﬁnd an easier characterization for them. . . Then we know we can ﬁnd u1 . . . one can ﬁnd numbers which are relatively prime with k. Using the Chinese remainder theorem. f (2).

Then for any m > 1 we have 0 ≡ f (m) ≡ a0 + (a1 + ap + a2p−1 + a3p−2 + . where u is ﬁxed. . a2p−1 . . we obtain r r |T | = |M | − i=1 Ai ≥ |M | − i=1 |Ai |. 2. we put p = pi . Let us count the number of s-tuples (x1 . . . . Actually. . . . for simplicity. n!}s | x1 + · · · + xs ≡ k (mod p)}|. . . but we still need a small observation. . . xs ) ∈ {1. where. . p|ap−1 + a2p−2 + . )m +(a2 +ap+1 +a2p +. Assume that among the numbers a1 . . n!}s such that x1 + x2 + · · · + xs ≡ u (mod p). . . . A simple argument related to the irreducibility of the polynomial 1 + X + X 2 + · · · + X p−1 shows that all cardinals that appear in the above sum are equal and that their sum is (n!)s . . . p|a1 + ap + a2p−1 + . . )mp−1 (mod p). p p = k=0 εk |{(x1 . p We are now ready to ﬁnish the proof. . 2. )m2 +· · ·+(ap−1 +a2p−2 +a3p−3 +. x2 . we will show that monic polynomial in Ai . xs ) ∈ {1. . . Let ε = cos Let us observe that 0 = (ε + ε2 + · · · + εn! )s p−1 2π 2π + i sin . Using Lagrange’s theorem it follows that p|a0 . . (n!)n . . . there are exactly v1 numbers that among 136 . . . . Let f a ppi i Let’s compute now |Ai |. . f (X) = X n + an−1 X n−1 + · · · + a1 X + a0 . .Using a brutal estimation. . thus each cardinal (n!)s equals . We are going to use this later. . . x2 . ap .

we deduce that we can ﬁnd integer polynomials P. Then for any positive integer r there exists n such that f (n) has at least r distinct prime divisors. Since f has no double root.. . Q such that P (n)f (n) + Q(n)f (n) = c for a certain positive integer c. all of them appearing with exponent 1 in the decomposition of f (n) in prime factors. 27 10 The solution of the problem ends here. Example 7. Using the result stated in the ﬁrst example. Already for r = 1 the problem is in no way obvious. we deduce that gcd(f.. there are vp−1 numbers. Using Bezout’s theorem in Q[X]. Suppose the contrary. Using the above observations. a2p−2 . so let’s not try to attack it directly and concentrate at ﬁrst on the case r = 1. all prime divisors of f (n) appear with exponent at least 2. f ) = 1 in C[X] and thus also in Q[X] (because of the division algorithm and Euclid’s algorithm). 5 5 < 1 1000 and so the percent of special polynomials is at least 100 1 − 1 1 1 − − 4 27 1000 = 75 − 100 1 − > 71. . we can choose q > c a prime divisor 137 .ap−1 . Suppose that the non constant polynomial f with integer coeﬃcients has no double roots. .. Iran Olympiad Solution. · . pp But 1 1 1 + 7 + ··· < 5 5 5 7 5 1+ 1 1 + 2 + . that is for all n. we obtain |T | ≥ (n!)n − p prime n! (n!)v1 (n!)vp−1 (n!)n .. = p p p pp (n!)n . it follows that we have |Ai | = Hence.

. . . . . p2 has the same propr 1 2 erty: all prime divisors p1 . 138 . We ﬁnally deduce that q|p (n) and so q|c. p1 . . . . . Since at most a ﬁnite number among them can r 1 2 be roots of f . we have q|u(p1 . pr . . The hypothesis made ensures that q 2 |f (n). By inductive hypothesis. If vq (g(u)) = 1. Let these prime factors be p1 . Using Newton’s binomial formula. . In this case. . Thus. Of course. f (N ))|c ⇒ q ≤ |c|. pr are diﬀerent prime numbers whose exponents in f (n + (p1 . . . . . . we can ﬁnd q > max{|c|. |p(n)|} a prime number and a number u such that q|g(u). . using again the result in example 1. pr )2 f (N ). . Thus the case r = 1 was proved. Let us prove that in the decomposition of f (N ). |p(n)|} so necessarily q|u (if q|f (N ) ⇒ q|(f (N ). p1 . pr )2 ). . the existence of P. pr have exponent 1 in the decomposition of f (n+kp2 p2 . since a trivial veriﬁcation shows that q. then since vq (f (n + (p1 . But then we also have q|f (n + q) and so q 2 |f (n + q). there is n such that at least r prime divisors of f (n) appear with exponent 1. victory is ours. pr )2 u) + uq(p1 . . pr )2 +uq(p1 . . . . . if vq (f (n)) ≥ 2. p2 . pr )2 . Now. we shall consider the number N = n+u(p1 . . But it is obvious that n + kp2 p2 . p2 ). . But since q|g(u). p1 . which is obviously non constant. . pr )2 f (N ) (mod q 2 ). Let us try to prove the property by induction and suppose it is true for r. . impossible since our choice was q > c. . The diﬃcult case is when vq (g(u)) ≥ 2. Remember that the choice was q > max{|c|. we deduce immediately that f (n+q) ≡ f (n)+qf (n) (mod q 2 ). so we keep the above notations. we easily obtain that f (N ) ≡ f (n + (p1 . Consider now the polynomial g(X) = f (n + (p1 . .of a certain f (n). For any pi . p1 . . pr )2 u) are all 1. pr appear with exponent 1. . . . . pr )2 X). . pr . Using once again binomial formula. all prime numbers q. p2 . . . contradiction). . . pr )2 u)) = vq (g(u)) ≥ 2. . pr . Q such that P (n)f (n)+Q(n)f (n) = c for a certain positive integer c did not depend on r. . . . this is obvious since f (N ) ≡ f (n) (mod (p1 . we may very well suppose from the beginning that n is not a root of f . .

It is obvious that q has positive leading coeﬃcient and the same property as f . there is a polynomial q with integer coeﬃcients such that f (X) = Xq(X). we infer that if q is non-constant. Since f (n)|kf (n) = kf (n) + n − n|f (kf (n) + n) − f (n). Because changing f with its opposite does not modify the property of the polynomial. we consider only n > M . impossible. the sequence (f (n + kf (n))n≥1 will contain at least two relatively prime numbers. p1 . we ensured that n is not a roof of our polynomial and also that q > max{|c|. then q(0) = 0 and q(X) = Xh(X).for any prime p > M we have p|f (p) and so p|f (0). Example 8. Gabriel Dospinescu Solution. r are multiples of f (n) > 2. This allows to ﬁnish the induction step and to solve the problem. From now on. Repeating the above argument.we must have q|g(0) = f (n).we obtain that s. Let us prove that we have gcd(f (n). any two of which are relatively prime. we conclude that f (0) = 0. the sequence (f (an + b))n≥1 contains an inﬁnite number of terms. Consequently. Hence . Let them be s. n) = 1 for any such n. r. We have shown that gcd(f (n). pr . But hopefully. there exists a constant M such that for any n > M we have f (n) > 2. . . n) = 1. |p(n)|} so that the last divisibility relation cannot hold. −X n with n a positive integer. Since any integer diﬀerent from zero has a ﬁnite number of divisors. Since f is not constant. the above argument cannot be repeated inﬁnitely many times and thus one of the 139 . . Thus. b. Suppose that there is n > M such that gcd(f (n). Find all non constant polynomials f with integer coefﬁcients and with the following property: for any relatively prime positive integers a. we have f (n)|f (n + kf (n)) for any positive integer k. . Thus . n) = 1 for any n > M . We will prove that the only polynomials with the speciﬁed property are those of the form X n . we can suppose that the leading coeﬃcient of f is positive. Thus.

. Let us observe that there is another. . . . It is clear that x1 . . x2 . . . . Very easy. We prove that we can add another term xp+1 so that x1 . . 140 . Then gcd(xp+1 . i ∈ {1. . We have obtained that f must have the form X n . . 2. xi ) = 1 for i ∈ {1. . . so we can apply the Chinese remainder theorem to ﬁnd xp+1 greater than x1 . . . . −X n are solutions. But this follows trivially from Dirichlet’s theorem. But since the sequence (f (2n + 3))n≥1 contains at least two relatively prime integers. −X n . that appears in at least one Olympiad around the world each year. there are positive integers n. Example. Suppose also that m − n|am − an for all distinct positive integers m. p}. .polynomials g. xp are relatively prime with a. x2 . Suppose that Fie n natural nenul. any two of which are relatively prime. we must have k = 1. . Let (an )n≥1 be an increasing sequence of positive integers such that for a certain polynomial f ∈ Z[X] we have an ≤ f (n) for all n. more elementary approach. Let us suppose that x1 . . It is enough to prove for X n and even for X. p} and thus we can add xp+1 . Care este gradul minim al unui polinom monic cu coeﬁcienti intregi f astfel incat n|f (k) pentru orice k natural? Proposed problems 1. . where a−1 is a’s i i ∗ inverse in Zxi . such that xp+1 ≡ (1 − b)a−1 (mod xi ). . we infer that any solution of the problem is a polynomial of the form X n . . x2 . n. Consequently. 2. h must be constant. Now let us prove that the polynomials of the form X n . xp+1 has the same property. . Here is an absolute classic. . But since f is a solution if and only if −f is a solution. it uses only the fundamental result. xp . xp are terms of the sequence. x2 . k such that f (X) = kX n .

Here f [n] is the composition of f with itself n times. Are there polynomials p. an ) = 1 for any diﬀerent positive integers m. the following property is satisﬁed: for any positive integer n. Suppose that f ∈ Z[X] is a non constant polynomial. Find all polynomials with integer coeﬃcients such that f (n)|2n −1 for any positive integer n. we obtain only composite numbers. Given is a ﬁnite family of polynomials with integer coeﬃcients.Prove that there exists a polynomial g ∈ Z[X] such that an = g(n) for all n. Prove that for inﬁnitely many numbers n. k. Poland Olympiad 6. Find all integer polynomials f ∈ Z[X] such that for any positive integer c. Folklore 5. r with positive integer coeﬃcients such that p(x) + (x2 − 3x + 2)q(x) and q(x) = x2 x 1 − + 20 15 12 r(x)? Vietnam Olympiad 4. suppose that for some positive integers r. We call the sequence of positive integers (an )n≥1 relatively prime if gcd(am . q. Is it true that any polynomial f ∈ Z[X] that has a root modulo n for any positive integer n must necessarily have a rational root? 141 . the sequence (f [n] (c))n≥1 is relatively prime. n. at most r prime factors of f (n) have appear with exponent at most equal to k. if we evaluate any member of the family in n. Leo Mosser 3. USAMO 1995 2. Also. Does it follow that any root of this polynomial appears with multiplicity at least equal to k + 1? 7.

g . . n. Is there a second degree polynomial f ∈ Z[X] such that for any positive integer n all prime factors of f (n) are of the form 4k + 3? AMM 15. Prove that the two polynomials are relatively prime in Q[X] if and only if Df. . Prove that the polynomial f (f (X))−X has at most n integer roots. . Andronache. Eckstein. M. xn such that f (xi ) = xi−1 . Sankt Petersburg Olympiad 12. with integer coeﬃcients. Let us consider the set Df. 142 . Show that the sequence f (3n ) (mod b) is not bounded.g = {gcd(f (n). Gh. 13. Romanian TST 11. g(n))| n ∈ N}.8. Prove that there are no polynomials f ∈ Z[X] with the property: there exists n > 3 and integers x1 . 10. Gazeta Matematica 1985 9. n (indices are taken mod n). 1 (mod n) and both these congruences have solutions. Prove that for any n there exists a polynomial f ∈ Z[X] such that all numbers f (1) < f (2) < · · · < f (n) are a) prime numbers b) powers of 2. i = 1. Folklore 16. Find all integers n > 1 for which there exists a polynomial f ∈ Z[X] such that for any integer k we have f (k) ≡ 0. g ∈ Z[X] some non zero polynomials. . Find all trinomials f ∈ Z[X] with the property that for any relatively prime integers m. 14. the numbers f (M ). Let f ∈ Z[X] be a non constant polynomial. Let f. f (n) are also relatively prime. For the die hard: ﬁnd all polynomials with the above property. Let f ∈ Z[X] a polynomial of degree n at least 2.

n for which p|φm (n). but p does not divide any other number of the form φd (n). You may be interested to characterize those numbers m.17. . Use example 1 and properties of the cyclotomic polynomials φn (X) = (X − e 1≤k≤n (k. . Find the maximal degree of a polynomial f ∈ Z[X] having coeﬃcients in the set {0. where d is a divisor of m diﬀerent from m Classical result 143 . knowing that its degree is at most p and that if p divides f (m) − f (n) then it also divides m − n. p − 1}. 1. . Let p be a prime number. .n)=1 2iπk n ) to prove that there are inﬁnitely many prime numbers of the form 1+kn for any given n ≥ 2. 18.

It says that an example of such polynomial is n f (x) = i=0 ai 0≤j=i≤n x − xj xi − xj (1) Indeed. 1. . then we can ﬁnd easily the value in any other point without solving a complicated linear system. . 1. which shows that if we know the value in 1 + deg f points. . . . . n. . . . Then the polynomial g − f also has degree smaller than or equal to n and vanishes at 0. and the proof is not diﬃcult at all. . . . . x1 . . . from the above expression we can see that this polynomial has degree less than or equal to n. n}? The answer turns out to be positive and a possible solution to this question is based on Lagrange’s interpolation formula. What is Lagrange’s interpolation theorem good for? We will see in the following problems that it helps us to ﬁnd immediately the value of a polynomial in a certain point if we know the values in some given points. let us make it easier and also reformulate it: is it true that given some pair wise distinct real numbers x0 . which is a remarkable tool in number theory. . . . it must be null and the uniqueness is proved. it is immediate to see that f (xi ) = ai for i ∈ {0. we can ﬁnd a polynomial f with real coeﬃcients such that f (xi ) = ai for i ∈ {0. Then. we will see that it helps in establishing 144 . 1. . But does everyone know the analogous form for polynomials? Stated like this. Also. Thus. a2 . xn and some arbitrary real numbers a0 . Is this the only polynomial with this supplementary property? Yes.LAGRANGE INTERPOLATION Almost everyone knows the Chinese Remainder Theorem. . Just suppose we have another polynomial g of degree smaller than or equal than n and such that g(xi ) = ai for i ∈ {0. 1. . an . And the reader has already noticed that this follows directly from the formula (1). a1 . . x2 . . n}. . n}. this question may seem impossible to answer. Also.

a1 . But for someone who knows Lagrange’s Interpolation Theorem. USAMO 1975. . . This is surely a dead end.some inequalities and bounds for certain special polynomials and will even help us in ﬁnding and proving some beautiful identities. Now. f (2) = . f (n) = . Example 1. i−j . . i−j so that f (n + 1) = n i=0 i i+1 145 n≤i n+1−j . Indeed. an from the linear system 1 2 n f (0) = 0. . . As promised. f (2) = . f (n) = . . as the following examples will show. 2 3 n+1 Find f (n + 1). A ﬁrst direct approach would be to write n f (x) = k=0 ak xk and to determine a0 . we will see ﬁrst how we can compute rapidly the value in a certain point for some polynomials. let us begin the journey trough some nice examples of problems where this idea can be used. Let f be a polynomial of degree n such that 1 2 n f (0) = 0. . The ﬁrst example is just an immediate application of formula (1) and became a classical problem. . . we have n f (x) = i=0 i i+1 j=i x−j . Great Britain 1989 Solution. This was one of the favorite’s problems in the old Olympiads. . f (1) = . the problem is straightforward. f (1) = . 2 3 n+1 But this is terrible. since the determinants that must be computed are really complicated. . .

we can see that n+1−j (n + 1)! = (−1)n−i i−j (n + 1 − i) · i! · (n − i)! j=i just by writing (n + 1)n . . IMO 1983 Shortlist 146 . . . i−j i(i − 1) . Fn+2 = Fn + Fn+1 and let f be a polynomial of degree 990 such that f (k) = Fk for k ∈ {992. The ﬁrst example was straightforward because we didn’t ﬁnd any diﬃculties after ﬁnding the idea. . According to this. (−(n − i)) j=i According to these small observations.Now. Titu Andreescu. 1982}. . it is tacit that the answer should be expressed in the closest possible form. . Let F1 = F2 = 1. . . But. Example 2. i And we have arrived at a familiar formula: the binomial theorem. . after all. . computing the above sum is not so diﬃcult. (n + 1 − (i − 1))(n + 1 − (i + 1)) . 1 · (−1) . how do we compute this? The reader might say: but we have already found the value of f (n+1)! Well. we can write n f (n + 1) = i=0 n i (n + 1)! · (−1)n−i i + 1 (n + 1 − i) · i! · (n − i)! (n + 1)! (−1)n−i (n + 1 − i)! · (i − 1)! n−1 i=0 = i=1 n = (n + 1) i=1 n (−1)n−i = (n + 1) i−1 n (−1)n+1−i . . n−1 i=0 n i=0 n (−1)n+1−i = − i n (−1)n−i − 1 i = 1. Indeed. . 1 n+1−j = . Show that f (1983) = F1983 − 1. This shows that f (n + 1) = n + 1. It’s not the case with the following problem.

Indeed. Now. then g(991) = F1983 − 1. we can of 2 2 course try a direct approach: Fn = 990 k=0 991 Fk+992 (−1)k k 990 k=0 1 =√ 5 990 k=0 990 k=0 991 k+992 a (−1)k − k 990 k=0 991 k+992 b (−1)k . But we know how to compute g(991). 990. the above sums vanish: 991 k+992 a (−1)k = a992 k 991 (−a)k = a992 [(1 − a)991 + a991 ]. since we could work as well with the polynomial g(x) = f (x+992). k But using the binomial theorem. the problem becomes: if g(k) = Fk+992 . for k = 0. we have f (k + 992) = Fk+992 for k = 0. but with a little bit of help it can be done. But why. looking again at the previous problem. which also has degree 990. This simple observation shows that we don’t have to bother too much with k + 992. since it can be done immediately with the formula for Fn . we know that an − bn √ . k 147 . The device is: never complicate things more than necessary! Indeed. So.Solution. 5 √ √ 5+1 1− 5 where a = and b = . Indeed. we ﬁnd that 990 990 k=0 g(991) = k=0 991 g(k) (−1)k = k 991 Fk+992 (−1)k k which shows that we need to prove the identity 990 k=0 991 Fk+992 (−1)k = F1983 − 1. Having this in mind. k This isn’t so easy. we could try to establish a more general identity that could be proved by induction. 990 and we need to prove that f (992 + 991) = F1983 − 1.

we ﬁnd that 990 k=0 991 1 Fk+992 (−1)k = √ (a1983 − b1983 − a + b) k 5 a1983 − b1983 a − b √ − √ = F1983 − 1. we ﬁnd that p−1 f (x) = k=0 f (k) j=k 148 x−j . The following example is a very nice problem proposed for IMO 1997.Since a2 = a + 1. Here. it causes great diﬃculties in solving. Then. Show that deg f is at least p − 1. we have a992 [(1 − a)991 + a991 ] = a(a − a2 )991 + a1983 = −a + a1983 . arguing by contradiction. Since in all this argument we have used only the fact that a2 = a + 1 and since b also veriﬁes this relation. So. 1 (mod p) for all positive integer k. Example 3. IMO Shortlist 1997 Solution. It is the typical example of a good Olympiad problem: no matter how much the contestant knows in that ﬁeld. Let f be a polynomial with integer coeﬃcients and let p be a prime such that f (0) = 0. k−j . let us suppose that deg f ≤ p − 2. such a problem should be solved indirectly. f (1) = 1 and f (k) = 0. using the Interpolation formula. As usual. 5 5 = And this is how with the help of a precious formula and with some smart computations we could solve this problem and also ﬁnd a nice property of the Fibonacci numbers. the following steps after using Lagrange’s Interpolation formula are even better hidden in some congruencies.

It’s time now for some other nice identities. . f (0) and f (1)). let us write the above relation in the form p−1 (−1)k k=0 p−1 f (k) = 0 k and let us take this equality modulo p. 1 (mod p) for all k and not all of the numbers f (k) have the same remainder modulo p (for example.Now. We will see how some terrible identities are simple consequences of the Lagrange Interpolation formula. since f (k) ≡ 0. k!(p − 1 − k)! From here we have one more step. the coeﬃcient of xp−1 in the right-hand side of the identity must be zero. Indeed. we have p−1 k=0 (−1)p−k−1 f (k) = 0. k=0 which is impossible. k=0 Thus. 149 . p−1 f (k) ≡ 0 (mod p). since deg f ≤ p − 2. (p − 1) ≡ (−1)k k! (mod p) we ﬁnd that p−1 k and so p−1 ≡ (−1)k (mod p) (−1)k k=0 p−1 f (k) ≡ k p−1 f (k) (mod p). Since k! p−1 k = (p − k)(p − k + 1) . where polynomials do not appear at ﬁrst sight. Consequently. . This contradiction shows that our assumption was wrong and the conclusion follows.

ai − aj Now. So. Let a1 . (x − an ). (ai − aj ) i=1 j=i 150 . the seci ond one is that the polynomial r has integer coeﬃcients and the third one n ak i is that is just the coeﬃcient of xn−1 in the polynomial (ai − aj ) i=1 j=i n r(ai ) i=1 j=i x − aj . where r is a polynomial of degree at most n − 1. we need three observations.Example 4. . since the formula works and we obtain n r(x) = i=1 r(ai ) j=i x − aj . This time we don’t have to worry. we recognize the Lagrange Interpolation formula for the polynomial f (x) = xk . The ﬁrst one is r(ai ) = ak . let us proceed. But this can be solved by working with the remainder of f modulo g(x) = (x − a1 )(x − a2 ) . Just by looking at the expression. . not only that we have solved the problem. a2 . . but we n ak i also found a rapid way to compute the sums of the form . by writing f (x) = g(x)h(x) + r(x). But we may have some problems when the degree of this polynomial is greater than or equal to n. n ak i Prove that for any positive integer k the number is an (ai − aj ) i=1 j=i integer. . . Great Britain Solution. . an be pairwise distinct positive integers. Thus. All these observations are immediate. we ﬁnd that i=1 j=i ak i (ai − aj ) is the coeﬃcient of xn−1 in r. Combining ai − aj n them. which is an integer.

it is natural for someone who knows very well the Lagrange Interpolation 151 . we will discuss a much more diﬃcult problem. . k k=0 As we promised. the second problem is much more diﬃcult. Let f (x) = k=0 ak xn−k . We ﬁnd that n n a0 = k=0 f (A + kh) j=k 1 [(k − j)h] = 1 n!hn (−1)n−k k=0 n f (A + kh). But we will see that both are immediate consequences of the Interpolation Formula. The reader might say after reading the solution: but this is quite natural! Yes. . 2. Since this polynomial has degree at most n. we have no problems in applying the Interpolation formula n f (x) = k=0 f (Ak h) j=k x − A − jh . . k which is exactly what we had to prove. If the ﬁrst one is relatively easy to prove using a combinatorial argument (it is a very good exercise for the reader to ﬁnd this argument). n Example 5. Simple and elegant! Notice that the above problem implies the well-known combinatorial identities n (−1)k k=0 n n p k =0 k (−1)n−k for all p ∈ {0. 1. (k − j)h Now. . k Alexandru Lupas Solution. n − 1} and n n k = n!. Prove that for any non-zero real number h and any real number A we have n (−1)n−k k=0 n f (A + kh) = a0 · hn · n!.The following two problems we are going to discuss refer to combinatorial sums. let us identify the leading coeﬃcients in both polynomials that appear in the equality.

formula and especially for someone who thinks that using it could lead to a solution. Yet. We ﬁnd that n 0= k=0 (−1)n−k n n k (1 + 2 + · · · + n − k). k And now the problem is solved. . . We take the polynomial f (x) = xn (why don’t we take the n polynomial f (x) = xn+1 ? Simply because (−1)n−k appears when k writing the formula for a polynomial of degree at most n) and we write the Interpolation Formula n x = k=0 n kn x(x − 1) . Unfortunately. we are not going to enter this ﬁeld and we prefer to concentrate on another elementary aspect of this formula and see how it can help us establish some remarkable inequalities. Prove the identity n (−1)n−k k=0 n n+1 n(n + 1)! k = . (x − n) (−1)n−k (n − k)!k! Now. Were Lagrange interpolation formula good only to establish identities and to compute values of polynomials. . we identify the coeﬃcient of xn−1 in both terms. And some of them will be really tough. since we found that n (−1)n−k k=0 n n+1 n(n + 1) k = 2 k n (−1)n−k k=0 n n k k and we also know that n (−1)n−k k=0 n n k = n! k from the previous problem. this isn’t always so easy. (x − k − 1)(x − k + 1) . it wouldn’t have been such a great discovery. 152 . Of course it is not the case. . it plays a fundamental role in analysis. k 2 Solution. Example 6.

We begin with a really diﬃcult inequality. xn ∈ [−1. Iran Olympiad Solution. . Then we have n f (x) = k=1 f (xk ) j=k x − xj . xk − xj Only now comes the beautiful idea. xk − xj Combining this with the triangular inequality . in which the interpolation formula is really well hidden. Prove that for any real numbers x1 . From the above inequality we ﬁnd that f (x) ≤ xn−1 n k=1 j=k 153 |f (xk )| |xk − xj | j=k 1− xj x . But even if we know it. 1] the following inequality is true: n i=1 j=i 1 |xj − xi | ≥ 2n−2 .. we arrive at a new inequality n |f (x)| ≤ k=1 |f (xk )| j=k x − xj .. Yet. . let us proceed by taking n−1 f (x) = k=0 ak xk an arbitrary polynomial of degree n − 1. The presence of j=i |xj − xi | is the only hint to this prob- lem. So. . . which is in fact the main step. the denominators give sometimes precious indications. Example 7. x2 . how do we choose the polynomial? The answer is simple: we will choose it to be arbitrary and only in the end we will decide which is one is optimal.

since many possible approaches fail. it is not so hard after all. 1] and such that the leading coeﬃcient is 2n−2 . It is easy to see that such a polynomial exists. Let f ∈ R[X] be a polynomial of degree n with leading coeﬃcient 1 and let x0 < x1 < x2 < · · · < xn be some integers. Example 8. we take the polynomial deﬁned by f (cos x) = cos(n − 1)x. so this choice solves our problem. We need a polynomial f such that |f (x)| ≤ 1 for all x ∈ [−1. Yet. Note also that the inequality |an−1 | ≤ k=1 j=k n |f (xk )| |xk − xj | can be proved by identifying the leading coeﬃcients in the identity n f (x) = k=1 f (xk ) j=k x − xj xk − xj and then using the triangular inequality. And it says that Chebyshev polynomials are the best. in the framework of the previous problems and with the experience of Lagrange’s interpolation formula.and since this is true for all non-zero real numbers x. This time our mathematical culture will decide. since they are the polynomials with the minimum deviation on [−1. This is the right moment to decide what polynomial to take. has degree n − 1 and leading coeﬃcient 2n−2 . The problem is not simple at all. So. Prove 154 . The following example is a ﬁne concoct of ideas. 1] (the reader will wait just a few seconds and he will see a beautiful proof of this remarkable result using Lagrange’s interpolation theorem). we may take the limit when x → ∞ and the result is pretty nice n |an−1 | ≤ k=1 |f (xk )| |xk − xj | j=k .

we start with the identity n f (x) = k=0 f (xk ) j=k x − xj . . 2. 2n Crux Matematicorum Solution. . . . we deduce that n k=0 |f (xk )| ≥ 1. xk − xj Now. . (xn − xk ) j=k ≥ k(k − 1)(k − 2) . . (n − k) = k!(n − k)!. (xk − xk−1 )(xk+1 − xk ) . n} such that |f (xk )| ≥ n! . This will allow us to ﬁnd a good inferior bound for j=k |xk − xj | ≥ 1. repeating the argument in the previous problem and using the fact that the leading coeﬃcient is 1. . we ﬁnd that n k=0 j=k |f (xk )| |xk − xj | ≥ 1. a problem given 155 . . we are done. And yes. . It is time to use that we are dealing with integers. since |xk − xj | = (xk − x0 )(xk − x1 ) . since n k=0 1 1 = k!(n − k)! n! n k=0 n k = 2n . Naturally (but would this be naturally without having discussed so many related problems before?).that there exists k ∈ {1. since using these inequalities. k!(n − k)!| Now. 2n We shall discuss one more problem before entering in a more detailed |f (xk )| ≥ study of Chebyshev polynomials and their properties. n! it follows trivially that n! . This is easy. . . 1 · 1 · 2 . .

we shall be blind and we shall try to ﬁnd these points. Example 9. We have then |g (xi )| = j=i (xi − xj ) . |g (xk )| k=0 156 . to have |xi | = 1 and also n |f (xk )| 1 ≤ 1. Thus. Romanian Olympiad Solution. let us call them simply x0 . Now. of course we would like. xn and write n k=0 j=k |f (xk )| |xk − xj | ≥ 1. . we should ﬁnd a monic polynomial g of n 1 degree n + 1 with all roots of modulus 1 and such that ≤ 1. if possible. As usual.in the Romanian mathematical Olympiad and which is a very nice application of Lagrange’s interpolation formula. It is useless to say that it follows trivially using a little bit of integration theory and Fourier series. . . This inequality was already proved in the two problems above. Of course. but this time it is necessary to ﬁnd the good points in which we should write the interpolation formula. Prove that for any polynomial f of degree n and with leading coeﬃcient 1 there exists a point z such that |z| = 1 and |f (z)| ≥ 1. Now. consider the polynomial n g(x) = i=0 (x − xi ). Marius Cavachi. . In this case it would follow from ≥1 |g (xk )| |xk − xj | k=0 k=0 n j=k that at least one of the numbers |f (xk )| is at least equal to 1 and the problem would be solved. x2 . Till then. the idea is always the same. x1 .

|z|=1 Of course. The fact 157 . to ﬁnd the polynomials having minimal deviation on [−1. kπ Lemma 1. If we put tk = cos . from the obvious fact that lim n+1 = 1.This is trivial: it suﬃces of course to consider g(x) = xn+1 − 1.. Moreover. The proof is simple. The next problems will be based on a very nice identity that will allow us to prove some classical results about norms of polynomials. 2n using the binomial formula we can establish immediately that it is a g(x) polynomial. if we consider √ x2 − 1 g(x) = [(x + x2 − 1)n − (x − x2 − 1)n ]. We have an explanation to give: we said the problem follows trivially with a little bit of integration theory tools. Indeed.. 2n k=0 Proof. Indeed. 1] and also to establish some new inequalities. In order to do all this. The conclusion follows. then n √ n x2 − 1 f (x) = (x − tk ) = [(x + x2 − 1)n − (x − x2 − 1)n ]. 0 ≤ k ≤ n. but very useful. knowing already this in 10-th grade (the problem was given to students in 10-th grade) is not something common. we need two quite technical lemmas. we x→∞ x deduce that it is actually a monic polynomial of degree n + 1. if we write f (x) = n ak xk then one can check with a trivial computation that k=0 ak = 1 2π 2π 0 f (eit )e−ikt dt and from here the conclusion follows since we will have 2π 2π = 0 f (eit )e−int dt ≤ 0 2π |f (eit |dt ≤ 2π max |f (z)|. which is not diﬃcult to establish.

allow us to write: f (x) = + √ n [(x + 2n x2 − 1)n + (x − x2 − 1)n ] x [(x + x2 − 1)n − (x − x2 − 1)n ]. the reader hopes that all these computations will have a honourable purpose. 2n−2 (−1)n n . Using again the observation from problem 7. The following relations are true: (−1)k n i) (tk − tj ) = n−1 if 1 ≤ k ≤ n − 1. Example 10. Of course. since these lemmas will allow us to prove some very nice results. The ﬁrst one is a classical theorem of Chebyshev. We leave the computations to the reader. 1]. Solution. using the above forx→±1 mula. 2 j=k n ii) iii) (t0 − tj ) = j=1 n−1 j=0 n . left to the reader. tk − tj . Then max |f (x)| ≥ x∈[−1. about minimal deviation of polynomials on [−1. To prove ii) and iii) it suﬃces to compute lim f (x). Lemma 2. He’s right. All this proves the ﬁrst lemma.1] 1 2n−1 and this bound cannot be improved. A little bit more computational is the second lemma. 2n x2 − 1 Using this formula and Moivre’s formula we easily deduce i). we obtain the identity: n I= k=0 f (tk ) j=k 158 1 .that g(tk ) = 0 for all 0 ≤ k ≤ n is easily veriﬁed using Moivre’s formula. (Chebyshev theorem) Let f ∈ R[X] be a monic polynomial of degree n. Simple computations. 2n−2 (tn − tj ) = Proof.

Example 11. Suppose that a0 . 1] we have |a0 + a1 x + · · · + an xn | ≤ 1. To prove that this result is optimal. It is true that it is a very diﬃcult one.1] 2n−1 max 1 Tn (x) = 1 2n−1 .Thus. . many of them are much easier. a1 . There are many other proof of this result . but we chosen this one because it shows the power of Lagrange interpolation theory. an are real numbers such that for all x ∈ [−1. it suﬃces to apply lemma 2 to conclude that we actually have n k=0 j=k 1 (tk − tj ) 1 = 2n−1 . This shows that max |f (x)| ≥ x∈[−1. Then the 1 polynomial n−1 Tn is monic of degree n and 2 x∈[−1. . Now. Walther Janous presented in Crux the following problem as open problem. Not to say that the use of the two lemmas allowed us to prove that the inequality presented in example 7 is actually the best. but here is a very simple solution using the results already achieved. 159 . . It is an easy exercise to prove that this is really a polynomial (called the nth polynomial of Chebyshev of the ﬁrst kind) and that it has leading coeﬃcient 2n−1 and degree n.1] 2n−1 and so the result is proved. it suﬃces to use the polynomial Tn (x) = cos(n arccos(x)). we have n 1 ≤ max |f (tk )| 0≤k≤n k=0 j=k 1 (tk − tj ) . Some years ago. .

Using Lagrange’s interpolation formula and modulus inequality. we deduce that for all u ∈ [−1. x∈[−1. By combining the two results.Then for all x ∈ [−1. this time for the polynomial Tn . the hypothesis says that f ≤ 1 and so by the lemma we have k=0 |f (x)| ≤ |Tn (x)| for all |x| ≥ 1. 1] and the conclusion follows. Coming back to the problem and considering the polynomial f (x) = n ak xk . We shall then have Tn 1 u 1 = |u|n n (−1) k=0 k j=k 1 − utj 1 = t k − tj |u|n n k=0 j=k 1 − utj |tk − tj | (the last identity being ensured by lemma 2).1] Then for any polynomial f ∈ R[X] of degree n the following inequality is satisﬁed: |f (x)| ≤ |Tn (x)| · f for all |x| ≥ 1. |tk − tj | The very nice idea is to use now again Lagrange interpolation formula. Walther Janous. 1] we have: f 1 u 1 ≤ f |u|n n k=0 j=k 1 − tj u . 160 . 1] we also have |an + an−1 x + · · · + a0 xn | ≤ 2n−1 . we are going to prove a stronger result. that is: Lemma. Denote f = max |f (x)|. Proof. Crux Matematicorum Solution. we obtain f 1 u ≤ Tn 1 u f for all u ∈ [−1. Actually.

it asked to prove that n n max x∈[0. . . 1]. The Japanese problems asked only to prove the existence of a constant that could replace 108. 161 . A brutal choice of points in Lagrange interpolation theorem gives a better bound of approximately 12 for this constant. Therefore we have |an + an−1 x + · · · + a0 xn | ≤ 2n−1 and the problem is solved. . thus its superior bound is attained in 0 or 1 and there the inequality is trivially veriﬁed.1] |x − ai | i=1 for any real numbers a1 . 1] 1 x ≤ 2n−1 . 1 x ≤ xn Tn 1 x . In the following. We end this topic with a very diﬃcult problem. an . The problem has a nice story: given initially in an old Russian Olympiad. But this inequality is very easy to prove: just set a = and observe that h(a) = (1 − a)n + (1 + a)n is a convex function on [0. 1]: |an + an−1 x + · · · + a0 xn | = xn f It suﬃces to prove that xn Tn which can be also written as (1 + 1 − x2 )n + (1 − 1 − x2 )n ≤ 2n .We will then have for all x ∈ [−1.2] i=1 |x − ai | ≤ 108n max x∈[0. √ 1 − x2 ∈ [0. that reﬁnes a problem given in a Japanese mathematical Olympiad in 1994. we present the optimal bound. . a2 . Recent work by Alexandru Lupas √ reduces this bound to 1 + 2 6.

Using Lagrange interpolation formula. Let us denote f [a. . (3 − tj ) + j=1 162 . which allows us to suppose that f [0. i=1 Gabriel Dospinescu Solution.b] = max |f (x)| x∈[a. . We shall prove that for all x ∈ [1. we deduce numbers xk = 2 that n |f (x)| ≤ k=0 j=k n x − xk = xk − xj n n k=0 j=k x − xj |xk − xj | ≤ k=0 j=k 2 − xj = |xk − xj | k=0 j=k 3 − tj .2] n ≤ cn f [0. |tk − tj | Using lemma 2.b] for a polynomial f and let.1] = 1. . for simplicity.2] i=1 √ √ (3 + 2 2)n + (3 − 2 2)n |x − ai | ≤ max 2 x∈[0. Let us ﬁx x ∈ [1.1] n |x − ai |. 2] we have |f (x)| ≤ xn . . a2 . an . 2] and consider the 1 + tk . cn = 2 We thus need to prove that f f (x) = i=1 [0.1] where (x − ai ). we can write n k=0 j=k 3 − tj 2n−1 = |tk − tj | n   j=0 n−1 n−1 (3 − tj ) k=1 j=k n + 2n−2 n  (3 − tj ) . the following inequality holds: n max x∈[0. For any real numbers a1 . We shall prove that this inequality is true for any polynomial f . √ √ (3 + 2 2)n + (3 − 2 2)n .Example 12.

Using again the two lemmas. 5. 7. This proves that f [0. 4. we deduce immediately that n−1 (3 − tj ) = j=1 √ √ √ [(3 + 2 2)n − (3 − 2 2)n ]. . Find n. we obtain: √ √ √ √ n 3 √ [(3 + 2 2)n − (3 − 2 2)n ] [(3 + 2 2)n + (3 − 2 2)n ] + 2n 2n+1 2 n−1 n−1 n = k=1 j=k (3 − tj ) + j=0 (3 − tj ) + j=1 (3 − tj ). the value 1 at 1. All we have to do now is to compute n−1 n n−1 (3 − tj ) + j=0 j=1 (3 − tj ) = 6 j=1 (3 − tj ). . . . 2n+1 2 1 Putting all these observations together and making a small computation. It’s value at 3n + 1 is 730.1] and solves the problem. Find its value at n + 2. 6. . we easily deduce that |f (x)| ≤ cn . Vietnam 1988 3. 3n − 1. USAMO 1984 2. But using lemma 1. 3n. . 8. A polynomial of degree n veriﬁes p(k) = 2k for all k = 1. that we let to the reader. Problems for training 1. . 3n − 2 and the value 2 at 0. . A polynomial of degree 3n takes the value 0 at 2. n + 1. . . k Tepper’s identity 163 . Prove that for any real number a we have the following identity n (−1)k k=0 n (a − k)n = n!.2] ≤ cn f [0. 3. . .

c be real numbers and let f (x) = ax2 + bx + c such that max{|f (±1)|. 1]. d ∈ R such that |ax3 +bx2 +cx+d| ≤ 1 for all x ∈ [−1. AMM E 3052 7. then f − 1 n ≤ 2n+1 − 1. c. b. Find k=0 (−1)k n n+2 k and k n (−1)k k=0 n n+3 k . b. Let a. 1996 8. Let a. 9. Spain. Prove that n k=0 j=k xn+1 k (xk − xj ) n = k=0 xk and compute n k=0 j=k xn+2 k (xk − xj ) . 1]. k AMM 5. What is the maximal value of |c|? Which are the polynomials in which the maximum is attained? Gabriel Dospinescu 164 . 6. k Peter Ungar. Prove the identity n (−1)k−1 k=1 n k (n − k)n = nn k n k=2 1 . |f (0)|} ≤ 1. Let f ∈ R[X] a polynomial of degree n that veriﬁes |f (x)| ≤ 1 for all x ∈ [0.n 4. Prove that if |x| ≤ 1 then |f (x)| ≤ 5 and x2 f 4 1 x ≤ 2.

J.n+1 India. b) Deduce Bernstein’s inequality: f |x|≤1 ≤ n f where f = max |f (x)|. p∈A |x|≤1 IMC. Prove that |a| + |b| + |c| + |d| ≤ 7. Prove that max |ai − p(i)| ≥ 1. 1].10. Let A = p ∈ R[X]| deg p ≤ 3. Deﬁne F (a. 2 Find sup max |p (x)|. a) Prove that for any polynomial f having degree at most n.3] possible value of this function over R3 ? China TST 2001 165 . i=0. What is the least x∈[0. (1 − zk )2 where zk are the roots of the polynomial |X n + 1. P. 1996 1 13. IMO Shortlist. 1]. |p(±1)| ≤ 1. c) = max |x3 − ax2 − bx − c|. p ± ≤1 . Find the maximal value of the expression a2 + b2 + c2 if |ax2 + bx + c| ≤ 1 for all x ∈ [−1. b. c. O’Hara. d ∈ R such that |ax3 + bx2 + cx + d| ≤ 1 for all x ∈ [−1. 2001 11. the following identity is satisﬁed: xf (x) = n 1 f (x) + 2 n n f (xzk ) k=1 2zk . Laurentiu Panaitopol 12. b. Let a ≥ 3 be a real number and p be a real polynomial of degree n. 1998 14. Let a. AMM 15.

Nicolae Popescu. as we will see. Since we have clariﬁed what is the purpose of the unit. the reader doesn’t need any special knowledge. just imagination and the most basic properties of the matrices. it is not easy to discover that ”non-elementary” fact that hides after a completely elementary problem. as soon as we advance. Let n ≥ 3 and let An . Here it is. For the beginning of the discussion. but the ideas and. the most important fact is not the knowledge. odd permutations of the set {1. Writing the diﬀerence n n |i − σ(i)| − σ∈An i=1 σ∈Bn i=1 n |i − σ(i)| as ε(σ) σ∈Sn i=1 166 |i − σ(i)| = 0. . but it is easy and a very nice application of how linear-algebra can solve elementary problems. The ﬁrst problem we are going to discuss is not classical. n}. things change. It is time to see the contribution of this ”non-elementary mathematics” to combinatorics. respectively. It is quite hard to imagine that behind a simple game. 2. Example 1. such as football for example or behind a quotidian situation such as handshakes there exists such complicated machinery. we have seen numerous applications of analysis and higher algebra in number theory and algebra. . . Prove the equality n n |i − σ(i)| = σ∈An i=1 σ∈Bn i=1 |i − σ(i)|. Gazeta Matematica Solution. but this happens sometimes and we will prove it in the next. but.HIGHER ALGEBRA IN COMBINATORICS Till now. . Bn be the sets of all even. we can begin the battle. Anyway. .

. det A = σ∈Sn n n (−1)ε(α) a|1−σ(1)| .. ... Before 167 . . . We have taken here Sn = An ∪ Bn . .. . which is nothing else than a polynomial identity. if σ ∈ An −1. if σ ∈ Bn ε(σ)a1σ(1) a2σ(2) . (1) Anyway..... What can we do to obtain it? The most natural way is to derive the last relation..j≤n . .... . x 1 x ... .. ..where ε(σ) = reminds us about the formula det A = σ∈Sn 1. xn−2 xn−1 xn−3 xn−2 xn−4 xn−3 . x2 x 1 . n |i−σ(i)| = σ∈Sn σ par˘ a xi=1 − σ impar˘ a σ∈Sn |i−σ(i)| xi=1 .. 1 xn−1 xn−2 . we still do not have the desired diﬀerence. . anσ(n) . a|−nσ(n)| |i−σ(i)| = σ∈An ai=1 − σ∈Bn |i−σ(i)| ai=1 This is how we have obtained the identity 1 x x2 . But we don’t have any product in our sum! That is why we will take an arbitrary positive number a and we will consider the matrix A = (a|i−j| )1≤i. x n . . This time... and then to take x = 1.

. . We have seen how many permutation do not have a ﬁxed point. is divisible by (x − 1)2 . .. .doing that. the sets of even and odd permutations of the set {1. x 1 x . Solution. .... Let us consider Cn . x2 x 1 . . . . The question that arises is how many of them are even. respectively. . n}. .. .. that do not have any ﬁxed points. x . We have seen it in the unit ”Principiul includerii si excluderii” that it is equal to n! 1 − 1 (−1)n 1 + − ··· + 1! 2! n! 168 . Find the number of even permutations of the set {1. 2. Thus. let us observe that the polynomial 1 x x2 . . Since in the right-hand side we obtain exactly n n |i − σ(i)| − σ∈An i=1 σ∈Bn i=1 |i − σ(i)| the identity is established. Here is a direct answer to the question. the derivative of this polynomial is a polynomial divisible by x − 1.. Dn .. Example 2... . using determinants. 2. .... which shows that after we derive the relation (1) and take x = 1. n} that do not have ﬁxed points... This can be easily seen by subtracting the ﬁrst line from the second and the third one and taking from each of these line x − 1 as common factor.. xn−2 xn−1 xn−3 xn−2 xn−4 xn−3 . We know how to ﬁnd the sum |Cn | + |Dn |. Here is another nice application of this trick. 1 xn−1 xn−2 . in the left-hand side we will obtain 0.. .. . ..

0 .Hence if we manage to compute the diﬀerence |Cn | − |Dn |. Now..j≤n . if xi ∈ Xj 169 . Xk a family of subsets of X. 1.. 1 1 0 ..... We will focus in the next problems on a very important combinatorial tool. If we write |Cn | − |Dn | = σ∈An σ(i)=i 1− σ∈Bn σ(i)=i 1. . . 0 |Cn | − |Dn | = 1 . Indeed. 1 ... 1 1 1 ... x2 . . we observe that this reduces to computing the determinant of the matrix T = (tij )1≤i. where j=1. then we subtract the ﬁrst column from each of the other columns... that is the incidence matrix (cum se spune la matricea de incidenta?). What is this? Suppose we have a set X = {x1 .. we add all columns to the ﬁrst one and we give n − 1 as common factor.. . . . The result is |Cn | − |Dn | = (−1)n−1 (n − 1) and the conclusion is quite surprising: 1 1 1 (−1)n−2 |Cn | = n! 1 − + − · · · + 2 2! 3! (n − 2)! + (−1)n−1 (n − 1). if xi ∈ Xj 0. will be able to answer to the question. . . if i = j But it is not diﬃcult to compute this determinant.n . xn } and X1 ... . . 1 1 . where tij = That is.k aij = 1. X2 . . deﬁne the matrix A = (aij )i=1. if i = j 0.

An+1 be distinct subsets of the set {1. .. .. Prove that there are two distinct subsets among them that have exactly one point in common. . From this point... though a diﬃcult one. we argue by contradiction and suppose that |Ai ∩ Aj | ∈ {0.. . tk tkn+1    . 2}.. In many situations. computing the product t A · A helps us to modelate algebraically the conditions and the conclusions of a certain problem. 2. Xk and the set X. This says something about the classical character and beauty of this problem. n tkn+1 tk1 k=1 k=1 tkn+11 tk2 . k=1 170 . the machinery activates and the problem is on its way of solving. . . n}. . It appeared in USAMO 1979. . A2 . . . Example 3. Let A1 .. k=1 . n . . . Solution. . . A2 .. Tournament of the Towns 1985 and in Bulgarian Spring Mathematical Competition 1995.   n  t2 kn+1 k=1 But we have of course n x2 = |Ai | = 3 ki k=1 and n xki xkj = |Ai ∩ Aj | ∈ {0.. An+1 and compute the product     t T ·T =    n n n t2 k1 k=1 k=1  tk1 tk2 . Let us discuss ﬁrst a classical problem.This is the incidence matrix of the family X1 . X2 . . . . 2} for all i = j. let T be the incidency matrix of the family A1 . . . each of which having exactly three elements.. Of course. Now. .

.. ... each of them having an even number of elements.  1 where X is the matrix having as elements the residues classes of the elements of the matrix X.... . .. we have 0 . A2 . . 0 0   . . . An ... Let n be an even number and A1 . A2 . 0 (R2 .. . but the solution using linear-algebra is straightforward.   T ·T =  |An ∩ A1 | |An ∩ A2 | . the previous relation shows that det t T · T is odd.. let us suppose that all the numbers |Ai ∩ Aj | are odd and interpret the above relation in the ﬁeld (R2 .. we obtain as in the previous problem the following relation  t |A1 | .. considered in the ﬁeld  1  tT · T =   . . This means that t T yT is an invertible matrix of size n + 1. ·).. ·).. hence non-zero. . .. Prove that among these subsets there are two having an even number of elements in common. Indeed. . . if T is the incidency matrix of the family A1 . We ﬁnd that   0 1 . . . Since of course det X = det X... An be distinct subsets of the set {1. +. Now.... .. thus rank t T · T = n + 1 which contradicts the inequality rank t T ·T ≤ rankT ≤ n. 1 0 171 .... |An |   . . +. . n}. .   1 1 .. .. 0 . Example 4. 2.. The following problem is very diﬃcult to solve by elementary means... 1 1   tT · T =  ... |A1 ∩ A2 | . . 0 . . . . . ..Thus. Solution... |A1 ∩ An | . . This shows that our assumption was wrong and there exist indeed indices i = j such that |Ai ∩ Aj | = 1.

Since a number cannot be both even and odd. if we work in (R2 . For example.which means again that det t T . Just observe that the sum of elements of the column i of T is |Ai |... Since the value of the determinant doesn’t change under this operation. used for the IMO preparation of the Romanian IMO team in 2004. we obtain that det T is also an odd number. 0 = 1. Working in a simple ﬁeld such as (R2 .. ... Thus.. +... if we add all lines to the ﬁrst line. The squares of a n×n table are colored with white and black. we will discuss the following problem. . 1 . Prove that there exists a non-empty set of columns B such that any line of the table contains an even number of white squares that also belong to B. 1 1 .. ·).. 1 1 . Gabriel Dospinescu 172 .. only this time we work in a diﬀerent ﬁeld. . hence an even number.. the conclusion is plain: det T is an even number. we will obtain only even numbers on the ﬁrst line.. The technique used is exactly the same as in the second example. T is odd. . ·) can allow us to ﬁnd quite interesting solutions.. we obtain 0 . 1 . Now. Hence we should try to prove that in fact det T is an even number and the problem will be solved. since det t T ·T = det2 T . Example 5. our assumption was wrong and the problem is solved. Note that this is the moment when we use the hypothesis that n is even. Suppose that there exists a non-empty set of lines A such that any column of the table has an even number of white squares that also belong to A. +. Indeed.

. we work only in (R2 .j = 0 for all j = 1.      u y + u y + ··· + u y = 0 n1 1 n2 2 nn n also has a non-trivial solution in R2 . ak . . We thus obtain a binary matrix T = (tij )1≤i.j≤n . Now. det T = 0 and consequently det t T = 0. in which just knowing the trick of computing t A · A does not suﬃce. n.      t z + t z + ··· + t z = 0 1n 1 2n 2 nm n admits the non-trivial solution (x1 . In each white square we write the number 1 and in each black square we put a 0.. +. n. This is just the combinatorial translation of the wellknown fact that a matrix T is invertible in a ﬁeld if and only if its transpose is also invertible in that ﬁeld.. we take B = {i| yi = 0} and we will clearly have B = ∅ and x∈B uix = 0. . From now on.Solution. . if i ∈ A 0. It 173 . i = 1. .. we will discuss a very diﬃcult problem. . It follows that i=1 ta. if i ∈ A It follows that the system   t11 z1 + t21 z2 + · · · + tn1 zn = 0      t z + t z + ··· + t z = 0 12 1 22 2 n2 n  . a2 . Thus. But this mean that any line of the table contains an even number of white squares that also belong to B and the problem is solved. xn ). ·). Suppose that k A contains the columns a1 . Let us proceed easily. . But this means that the system   u11 y1 + u12 y2 + · · · + u1n yn = 0      u y + u y + ··· + u y = 0 21 1 22 2 2n n  . But this is not that easy to see. Now. In the end of this sub-unit. .. let us take xi = 1. x2 .

is true that it is one of the main steps, but there are much more things to do after we compute t A · A. And if for these ﬁrst problems we have used only intuitive or well-known properties of the matrices and ﬁelds, this time we need a more sophisticated machinery: the properties of the characteristic polynomial and eingenvalues of a matrix. It is exactly that kind of problem that kills you just when we feel most strong. Example 6. Let S = {1, 2, . . . , n} and A be a family of pairs of elements from S with the following property: for any i, j ∈ S there exist exactly m indices k ∈ S for which (i, k), (k, j) ∈ A. Find all possible values of m, n for which this is possible. Gabriel Carrol Solution. This time, it is easy to see what hides after the problem. Indeed, if we take T = (tij )1≤i,j≤n , where 1, if (i, j) ∈ A 0, otherwise

aij =

the existence of the family A reduces to  m m m ... m ... ... ... ... m 

  m  T2 =   ...  m

 m   . ...   m

So, we must ﬁnd all values of m, n for which there exist a binary matrix T such that 
2

m

m m ... m
174

... ... ... ...

m

  m  T =  ...  m

 m   . ...   m

Let us consider  m m m ... m ... ... ... ... m 

  m  TB =   ...  m

 m   . ...   m

and ﬁnd the eigenvalues of B. This is not diﬃcult, since if x is an eingenvalue, then m−x m ... m m m ... m ... ... ... ... m m ... m−x .

If we add all columns to the ﬁrst one and then take the common factor mn − x, we obtain the equivalent form 1 (mn − x) 1 ... 1 m ... m ... ... ... m m ... m−x = 0.

m − x ...

In this ﬁnal determinant, we subtract from each column the ﬁrst column multiplied by m and we obtain in the end the equation xn−1 (mn − x) = 0, which shows that the eigenvalues of B are precisely 0, 0, . . . , 0, mn. But these are exactly the squares of the eigenvalues of
n−1 √ T . Thus, T has the eingevalues 0, 0, . . . , 0, mn, because the sum of the n−1

eingenvalues is nonnegative (being equal to the sum of the elements of the matrix situated on the main diagonal). Since T rT ∈ R, we ﬁnd that mn must be a perfect square. Also, because T rT ≤ n, we must have m ≤ n.
175

Now, let us prove the converse. So, suppose that m ≤ n and mn is a perfect square and write m = du2 , n = dv 2 . Let us take the matrices I = (11 . . . 11),
dv

O = (00 . . . 00).
dv

Now, let us deﬁne the circulant matrix   111 . . . 1 00 . . . 0     u v−u    0 11 . . . 1 00 . . . 0    S= u v−u−1  ∈ Mv,n ({0, 1}).     ...      111 . . . 1 00 . . . 0 1 
u−1 v−u

Finally, we take  S 

   S    A=  ∈ Mn ({0, 1}).  ...    S It is not diﬃcult to see that  m   m  A2 =   ...  m

m m ... m

... ... ... ...

m

 m   . ...   m

The last idea that we present here (but certainly these are not all the methods of higher mathematics applied to combinatorics) is the use of vector spaces. Again, we will not insist on complicated notions from the theory of vector spaces, just the basic notions and theorems. Maybe the most useful fact is that if V is a vector space of dimension n (that is, V has a basis of cardinal n), then any n + 1 or more vectors are linearly dependent. As a direct application, we will discuss the following problem,
176

which is very diﬃcult to solve by means of elementary mathematics. Try ﬁrst to solve it elementary and you will see how hard it is. The following example is classical, too, but few people know the trick behind it. Example 7. Let n be a positive integer and let A1 , A2 , . . . , An+1 be nonempty subsets of the set {1, 2, . . . , n}. Prove that there exist nonempty and disjoint index sets I1 = {i1 , i2 , . . . , ik } and I2 = {j1 , j2 , . . . , jm } such that Ai1 ∪ Ai2 ∪ · · · ∪ Aik = Aj1 ∪ Aj2 ∪ · · · ∪ Ajm . Solution. Let us associate to each subset Ai a vector vi ∈ Rn , where vi = (x1 , x2 , . . . , xn ) and i i i xj = i 0, if j ∈ Ai 1, if j ∈ Ai

Since dim Rn = n, these vectors we have just constructed must be linearly dependent. So, we can ﬁnd a1 , a2 , . . . , an+1 ∈ R, not all of them 0, such that a1 v1 a2 v2 + · · · + an+1 vn+1 = 0. Now, we take I1 = {i ∈ {1, 2, . . . , n + 1}| ai > 0} and I2 = {i ∈ {1, 2, . . . , n+1}| ai < 0}. It is plain that I1 , I2 are nonempty and disjoint. Now, let us prove that
i∈I1

Ai =
i∈I2

Ai and the solution will be complete. Ai . Then the vectors vi
i∈I2

Let us take x ∈
i∈I1

Ai and suppose that x ∈

with i ∈ I2 have zero on their xth component, so the xth component of the vector a1 v1 +a2 v2 +· · ·+an+1 vn+1 is
x∈Aj j∈I1

aj > 0, which is impossible, Ai ⊂
i∈I1 i∈I2

since a1 v1 + a2 v2 + · · · + an+1 vn+1 = 0. This shows that

Ai .

But the reversed inclusion can be proved in exactly the same way, so we conclude that
i∈I1

Ai =
i∈I2

Ai .

177

In the end of this ”non-elementary” discussion, we solve another problem, proposed for the TST 2004 in Romania, whose idea is also related to vector spaces. Example 8. 30 boys and 20 girls are preparing for the 2004 Team Selection Test. They observed that any two boys have an even number of common acquaintances among the girls and exactly 9 boys know an odd number of girls. Prove that there exists a group of 16 boys such that any girls attending the preparation is known by an even number of boys from this group. Gabriel Dospinescu Solution. Let us consider the matrix A = (aij ) where

aij =

1, if Bi knows Fj 0, otherwise

We have considered here that B1 , B2 , . . . , B30 are the boys and F1 , F2 , . . . , F20 are the girls. Now, consider the matrix T = A · t A. We observe that all the elements of the matrix T , except those from the
20

main diagonal are even (because tij =
k=1

aik ajk is the number of com-

mon acquaintances among the girls of the boys Bi , Bj ). The elements on the main diagonal of T are exactly the number of girls known by each boy. Thus, if we consider the matrix T in (R2 , +, ·), it will be diagonal, with exactly nine non-zero elements on its main diagonal. From now on, we will work only in (R2 , +, ·). We have seen till now that rankT = 9. Using Sylvester inequality, it follows that

9 = rankT ≥ rankA + rank t A − 20 = 2rank t A − 20
178

. .. . then each girl is known by an even number of boys from this group and the problem is solved. . if we choose the group of those boys Bk with k ∈ M . . Finally. 2. 30}| xi = 1}. . That is why we can choose a solution (x1 . Let Aj . Let us consider now the linear system in (R2 . 1. 2. p−1}. we consider n S(σ) = k=1 kσ(k). respectively. . Bj . Find the greatest number p such that for all k ∈ {1. Let p > 2 be an odd prime and let n ≥ 2. But we observe that j∈M aji is just the number of boys Bk with k ∈ M such that Bk knows Fi .hence r = rank t A ≤ 14. be the set of even. . . We have proved that |M | ≥ 16 and also j∈M aji = 0 for all i = 1. Problems for training 1.. Thus. respectively odd permutations σ for which S(σ) ≡ j (mod p). consider the set M = {i ∈ {1. . Prove that n > p if and only if Aj and Bj have the same number of elements for all j ∈ {0. x2 . . . For any permutation σ ∈ Sn . 20. Gabriel Dospinescu 2. . having at least 16 components equal to 1. p} we have n k n k if (i) σ∈An i=1 = σ∈Bn 179 i=1 if (i) . ·):   a11 x1 + a21 x2 + · · · + a301 x30 = 0      a x + a x + ··· + a x = 0 12 1 22 2 302 30  . . x30 ) of the system.      a x + a x + ··· + a x =0 120 1 220 2 3020 30 The set of solutions of this system is a vector space of dimension 30 − r ≥ 16. Let n ≥ 2. +. . . .

. It is known that for any subset S of teams. l such that 180 . The two sides of length n are taped together to form a cylinder. Moldova TST 2004 4. Gabriel Dospinescu 3. 2 points are given to the winner. not all zero. 2. and recoloring each other edge in blue. Karpov. Prove that it is possible to write a real number in each square. one can ﬁnd a team (possibly in S) whose total score in the games with teams in S is odd. On an m × n sheet of paper is drawn a grid dividing the sheet into unit squares. Problema de la Vietnamezi cu cunostintele 7. 1 point for a draw and 0 points for the looser.where An . .1972 8. 5. . Prove that m ≤ n. Let A1 . Is there in the plane a conﬁguration of 22 circles and 22 points on their union (the union of their circumferences) such that any circle contains at least 7 points and any point belongs to at least 7 circles? Gabriel Dospinescu. . if and only if there exist integers k. D. n ≥ 2 teams compete in a tournament and each team plays against any other team exactly once. n}. respectively. Suppose that any two of these subsets have exactly one elements in common. the sets of all even. . . respectively. The edges of a regular 2n -gon are colored red and blue. A2 . . Prove that after 2n−1 steps all of the edges will be red and show that this need not hold after fewer steps. Bn are. so that each number is the sum of the numbers in the neighboring squares. odd permutations of the set {1. Iran Olympiad. In each game. Prove that n is even. . Am be distinct subsets of a set A with n ≥ 2 elements. A step consists of recoloring each edge which is the same color as both of its neighbours in red. Russian Olympiad. 1998 6.

Let x1 . . j) = (k. the jury deﬁnes the diﬃculty of each problem by assigning it a positive integral number of points (the same number of points may be assigned to diﬀerent problems). After the participants submitted their answers. Let S = {x0 . Then the vector space spanned only by those xi −xj for which xi −xj = xk −xl whenever (i. 2001 10. in terms of n. the maximum number of participants for which such a scenario could occur. it could have deﬁned the problems’ diﬃculty levels to make that ordering coincide with the participants’ ranking according to their total scores. any other participant receives 0 points. . . Strauss theorem 181 cos . Prove that S consists of rational numbers only. such that every distance between pairs of elements occurs at least twice.n + 1 does not divide k and 2lπ kπ 1 + cos = . Russian Olympiad. Romanian TST 1998 9. Determine. . x1 = 1. except for the distance 1. xn be real numbers and suppose that the vector space spanned by xi − xj over the rationals has dimension m. xn } ⊂ [0. l) also has dimension m. x1 . In a contest consisting of n problems. . m n+1 2 Ciprian Manolescu. . . 1] be a ﬁnite set of real numbers with x0 = 0. Any participant who answers the problem correctly receives that number of points for the problem. the jury realizes that given any ordering of the participants (where ties are not permitted). Iran Olympiad 11. .

2. 2. j. Let 2n+1 real numbers with the property that no matter how we eliminate one of them. Am and B1 . . . . . Then all numbers are equal. There is no partition of the set of edges of the complete graph on n vertices into (strictly) fewer than n − 1 complete bipartite graphs. b) for any i. Graham-Pollak theorem 14. . A2 . Then there is a person that knows all the others. Then mp ≤ 2n−1 . Ai and Bj have exactly one common element. . there is i ∈ A such that |Ai ∩ T | is odd. n} with the properties: a) for any nonempty subset T of A. any two persons have exactly one friend. the sum of the numbers in the two groups being the same. Then there are disjoint sets I. . Universal friend theorem 16. acquaintance is mutual and even more. B1 . Then prove that B1 = B2 = · · · = Bn . . . n}. . . Benyi Sudakov 17. . Lindstrom theorem 13. In a society. . . . . .12. . . . . . the rest of them can be divided into two groups of n numbers. . . . J with non-empty union such that i∈I Ai = j∈J Aj and Ai = i∈I j∈J Aj . j ∈ A. Gabriel Dospinescu 182 . Bn ⊂ A = {1. 2. Am some subsets of {1. Let A1 . Let A1 . . . An . n} such that Ai ∩ Bj is an odd number for all i. . . Bp subsets of {1. . . 15. . Let A1 .

Prove that we can ﬁnd some rows in this matrix such that their sum is a vector having all of its components odd.18. A symmetric matrix of zeros and ones has only ones on the main diagonal. Iran Olympiad 183 .

The reader will surely have the pleasure to read some reference books about this fascinating ﬁeld. it may look weird. we will content to present only a mere introduction to this ﬁeld. Let us start by proving the celebrated Minkowski’s theorem.GEOMETRY AND NUMBERS Again an apparently paradoxical note!!! Indeed. which is symmetric about the origin (that is. let us start by making a sort of partition of Rn in cubes of edge 2. (Minkowski) Suppose that A is a convex body in Rn having volume strictly greater than 2n . First of all. As usual. It is clear that any two such cubes will 184 . Theorem. Indeed. which will be practically always used in our applications). almost all of them playing around the celebrated Minkowski theorem. We will see that this theorem gives a very simple criterion for a nice region (we are also going to explain what we understand by nice) to have a non-trivial lattice point and the existence of this point will have important consequences in the theory of representation of numbers by quadratic forms or in approximation of real numbers with rational numbers. We will admit that convex bodies have volumes (just think about it in the plane or space. we will work in Rn and we will call convex body a bounded subset A of Rn which is convex (that is for all a. The proof is surprisingly simple. for all x ∈ A we also have −x ∈ A). extremely well developed. In general. Then there is a lattice point in A diﬀerent from the origin. In the sequel we are going to show some applications of geometry in number theory. b ∈ A and all 0 ≤ t ≤ 1 we have ta + (1 − t)b ∈ A). but geometry it really useful in number theory and sometimes it can help proving diﬃcult results with some extremely simple arguments. mentioned in the bibliographies. let us state the conditions in which we will work and what is a nice ﬁgure. having as centers the points that have all coordinates even numbers.

Suppose that in each lattice point in space except for the origin one draws a ball of radius r > 0 (common for all the balls). the theorem proved before admits a more general formulation. one can bring any cube into the cube centered around the origin by using a translation by a vector all of whose coordinates are even. 185 . Now. Since translations preserve volume. Necessarily there will be two bodies which intersect in a point X. Problem 1. we will have now an agglomeration of bodies in the central cube (the one centered in the origin) and the sum of volumes of all these bodies is strictly greater that 2n . But since A is centrally x−y symmetric and convex. This means that the line will intercept the ball centered around this point. it is clear that the sum will be ﬁnite). Let us suppose the contrary and let us consider a very long r cylinder having as axe the line and basis a circle of radius . Then any line that passes through the origin will intercept a certain ball. We choose 2 it suﬃciently long to ensure that it will have a volume strictly greater than 8. Actually. it follows that is a lattice point diﬀerent 2 from the origin and belonging to A. Solution. This is clearly a convex body in space and using Minkowski’s theorem we deduce the existence of a non-trivial lattice point in this cylinder (or on the border). The theorem is thus proved. We have found two diﬀerent points x.have disjoint interiors and that they cover all space. But of course. which is even more useful. Here is a surprising result that follows directly from this theorem. y in our convex body such that x − y ∈ 2Zn . look at the cubes where these two bodies where taken from and look at the points in these cubes that give by translations the point X. That is why we can say that the volume of the convex body is equal to the sum of volumes of the intersections of the body with each cube (since the body is convex.

. it follows trivially from the ﬁrst theorem. . they are linearly independent and moreover for any point (x. With all these terms. A must contain at least a point of the lattice L = Zv1 + · · · + Zvn diﬀerent from the origin. using Minkowski’s theorem. . Actually. Any prime number of the form 4k + 1 is the sum of two squares. in order to ﬁnd a non-trivial point of A ∩ L. 0). We have already proved that any prime number of the form 4k + 1 is the sum of two squares. y) in the lattice L = Zv1 + Zv2 we have p|x2 + y 2 . We have already proved that for any prime number of the form 4k + 1. Next. y = n and thus x2 + y 2 ≡ n2 (a2 + 1) ≡ 0 (mod p). Since the transformation V ol(A) is linear. Problem 2. n ∈ Z such that x = mp + na. It V ol(P ) suﬃces to apply the ﬁrst theorem to this convex body and to look at the preimage of the lattice point (in Zn ). 0. Also consider n the fundamental parallelepiped P = i=1 xi vi | 0 ≤ xi ≤ 1 and de- note V ol(P ) its volume. call it p. the area of the fundamental parallelogram is v1 ∧ v2 = p. Proof. . Visibly. . the set of vectors all of whose components are between 0 and 1) and that f maps L into Zn . . Let us prove it diﬀerently. . one can ﬁnd a such that p|a2 + 1. 1. Indeed.Theorem 2. . there are m. it will send A into a convex body of volume > 2n . it would seem that this is extremely diﬃcult to prove. 0. 0) one can easily see that P is sent into the ”normal” cube in Rn (that is. The second theorem is thus proved. . . v2 . vn some linearly independent vectors in Rn . Indeed. . . v2 = (a. by considering the linear application f sending vi into the vector ei = (0. Then let us consider v1 = (p. Moreover. 1). consider as convex 186 . (Minkowski) Let A is a convex body in Rn and v1 . Assuming that A has a volume strictly greater than 2n · V ol(P ).

y) diﬀerent from the origin that lies in this disc and also in the lattice L = Zv1 + Zv2 . Since ac = b2 + b + 1. we have ax2 − (2b + 1)xy + cy 2 = 1 and the existence of a solution of the equation is proved. Then the equation ax2 − (2b + 1)xy + cy 2 = 1 has integer solutions. 0)} a lattice point of this region. The following problem (as the above one) has a quite diﬃcult elementary solution. but what can we do when we are asked to prove that a certain equation has solutions? Minkowski’s theorem and in general geometry of numbers allow quick responses to such problems. Thus. Problem 3. Yet. A simple computation 4π shows that it is an elliptical disc having as area √ > 4. Poland Olympiad Solution. by Minkowski’s theorem we can ﬁnd a point of this region diﬀerent from the origin.. Thus for (x. Here is an example. we have for all x. The solution using geometry of numbers is more natural.body the disc centered in the origin and having as radius √ 2p. there is a point (x. y) ∈ Z2 \ {(0. y not both 0 the inequality ax2 − (2b + 1)xy + cy 2 > 0. Obviously. but it is not at all obvious how to proceed. which shows that p = x2 + y 2 . 187 . For this point we have p|x2 + y 2 and x2 + y 2 < 2p. the experience of the preceding problem should ring a bell. Consider positive integers such that ac = b2 + b + 1. taken from a polish Olympiad. Here is a very quick approach: consider in R2 the set of points verifying ax2 − (2b + 1)xy + cy 2 < 2. its area is strictly greater than four times the area of the fundamental parallelogram. Proving that a certain Diophantine equation has no solution is a very classical problem. Thus. An elliptical 3 disc is obviously a convex body and even more this elliptical disc is symmetric about the origin..

The area of the fundamental parallelepiped is clearly at most n. consider the lattice formed by the 3 points (x. y) such that n divides ax − by. Then this equation has integer solutions as well.Problem 4. we can ﬁnd (x. c are non-zero (otherwise the conclusion follows trivially). b). 0)} such that x2 + xy + y 2 < 2n and n divides x2 + xy + y 2 . Of course. Even more. b. . We claim that this will give an integer solution of the equation. First. Suppose that n is a natural number for which the equation x2 + xy + y 2 = n has rational solutions. b are relatively prime. y) ∈ Z2 \ {(0. Komal Solution. Again. We will try again to ﬁnd a couple of integers (x.j≤n for all x = (x1 . let us remind that for any symmetric real matrix A such that aij xi xj > 0 1≤i. b. 0)} such that x2 + xy + y 2 < 2n and n divides ax − by. . y) ∈ Z2 \ {(0. . x2 . a classical argument allows to assume that a. xn ) ∈ Rn \ {0} the set of points verifying aij xi xj ≤ 1 1≤i. This allows us to conclude. Observe that ab(x2 + xy + y 2 ) = c2 xyn + (ax − by)(bx − ay) and so n also divides x2 + xy + y 2 (since n is relatively prime with a. In this case we will have x2 + xy + y 2 = n and the conclusion will follow. c such that a2 + ab + b2 = c2 n. . let us look at the region deﬁned by x2 + xy + y 2 < 2n. then x2 + xy + y 2 = n has integer solutions. the problem reduces to: if there are integer numbers a. We will assume that a. Next. Before continuing with some more diﬃcult problems. By Minkowski’s theorem. simple computations show that it is an 4π elliptical disc of area √ n.j≤n 188 .

In particular. let’s attack some serious problems. Moreover. v2 = (0. these two sets cannot be disjoint. If we spoke about squares. Consider now the vectors v1 = (p. Since there are squares in Zp (as we 2 have already seen in previous notes). This is going to be much more complicated. 0). v3 = (x. 0. p+1 B = {−y 2 − 1| y ∈ Zp }. The main diﬃculty is ﬁnding the appropriate lattice and convex body. −x. Let thus p > 2 a prime number and consider the sets A = {x2 | x ∈ Zp }. 0). let us prove the result for prime numbers. p. Problem 5. (Lagrange’s theorem) Any natural number is a sum of 4 squares. With these results (that we will admit) in mind. there are x. the reader has noticed that these results can be applied to previous problems to facilitate the computations of diﬀerent areas and volumes. v4 = (y. y. y ≤ p − 1 and p|x2 + y 2 + 1. 0. First of all. A simple computation (using the above formulas) allows to prove that the volume of the fundamental parallelepiped is p2 . one can easily verify that for all point (x. where has a volume equal to √ det A V ol(Bn ) = ∞ π2 n Γ 1+ (here Γ(x) = 0 n 2 e−t tx−1 dt is Euler’s gamma function). one can 189 .V ol(Bn ) . The proof of this result is non elementary and we invite the reader to read more about it in any decent book of multivariate integral calculus. Proof. 0. Even more. 1) and the lattice L generated by these vectors. t) ∈ L one has p|x2 + y 2 + z 2 + t2 . 1. In particular. y. 0). why not present the beautiful classical proof of Lagrange’s theorem on representations using 4 squares. 0. y such that 0 ≤ x. This is the observation that will allow us to ﬁnd a good lattice. but the idea is always the same. z.

Consider the numbers z1 = a + bi. t) ∈ L ∩ A and we will clearly have x2 + y 2 + z 2 + t2 = p. It follows form the miraculous identity: (a2 + b2 + c2 + d2 )(x2 + y 2 + z 2 + t2 ) = (ax + by + cz + dt)2 = (ay − bx + ct − dz)2 + (az − bt + dy − cx)2 + (at + bz − cy − dx)2 . but how could one think at such an identity? The eternal question. everything would be nice if the product of two sums of 4 squares is always a sum of 4 squares. thus A ∩ L is not empty. Of course. y. Well. It suﬃces then to choose a point (x. but the proof is not obvious at all. Thus the theorem is proved for prime numbers. N= z3 z4 −z 4 z 3 . c. z2 = c + di.also prove (by employing the non-elementary results stated before this problem) that the volume of the convex body A = {x = (a. But surprise! We have MN = z 1 z3 − z2 z 4 z1 z 4 + z2 z 3 z1 z2 −z 2 z 1 . b. z4 = z + ti. very nice. Hopefully.. it is the case. z3 = x + yi.. −z1 z4 + z2 z3 z1 z3 − z2 z 4 190 . Of course. z. let us reason only with four. Introduce the matrices M= We have det(M ) = |z1 |2 + |z2 |2 = a2 + b2 + c2 + d2 and similarly det(N ) = x2 + y 2 + z 2 + t2 . It is then normal to try to express (a2 + b2 + c2 + d2 )(x2 + y 2 + z 2 + t2 ) as det(M N ). this time there is a very nice reason: instead of thinking in eight variables. d) ∈ R4 | a2 + b2 + c2 + d2 < 2p} is equal to 2π 2 p2 > 16V ol(P ).

i = 1. thus {Y ∈ Rn | |A−1 Y |i < ci . . We need to prove that there exists a non-zero vector X that also belongs to the region {Y ∈ Rn | |A−1 Y |i < ci . . Suppose A = (aij )1≤i≤m is a matrix with m.and so det(M N ) is again a sum of 4 squares. . Then there n are integers x1 . . But observe that {Y ∈ Rn | |A−1 Y |i < ci . i = 1. . Then one can ﬁnd x1 . such that j=1 aij xj < ci for all i = 1. c2 . . n} is 1 2n c1 . . i = 1. . j=1 n Solution. n} (here A−1 Y = (|A−1 Y |1 . (Minkowski’s linear forms’ theorem) Let A = (aij ) be a n × n invertible matrix of real numbers and suppose that c1 . . . . n real 1≤j≤n numbers and a ≥ 1 is a real number.. . . cn . x2 . . n} which has volume 2n c1 . xn . By Minkowski’s theoa convex body of volume det A rem. . And here is a nice consequence of the previous theorem. . Solution. . . . . . which will verify the conditions of the problem. i = 1. cn are positive real numbers such that c1 c2 . whose determinant equals 1 or −1 In 0 191 . cn > | det A|. All we need to do is to apply the result in problem 6 for A Im the invertible matrix . Problem 6. such that n aij xj < a− m for all 1 ≤ i ≤ m. . . . . not all 0. . . . . . We have some beautiful results of Minkowski that deserve to be presented after this small introduction to geometry of numbers. . . n. . x2 . . . . |A−1 Y |n ). . . Let us concentrate a little bit more on approximations of real numbers. this body will contain a non-zero lattice point. The identity appears thus naturally. . .. xn integers between −a and a. n} is exactly the image trough A−1 of the parallelepiped {Y ∈ Rn | − ci < Yi < ci . not all 0. Problem 7. . . cn > 2n .

3 Miklos Schweitzer competition 5. Then prove that it is also a sum of squares of three natural numbers (the use of three squares theorem is forbidden!). Consider a disc of radius R. Then prove that √ 1 R2 +1 ≤r< 1 . such that 2 ax + by + cz = 0 and max{|x|. |z|} < √ a + 1. 1 ≤ i ≤ n. Incredibly.j≤n 192 n aij xi xj > 0. Davenport-Cassels lemma 3. he can see at least a point situated at the exterior of the disc. Imo Shortlist 2. Then there exist x. b. Then . Prove that we can ﬁnd three integers x. one plants a circular tree of radius r. cm+i = a. . b. j ≤ n) are rational numbers such that for any x = (x1 . z. xn ) ∈ Rn \ {0} we have 1≤i. Suppose that a natural number is the sum of three squares of rational numbers. 6. not all 0. c are positive integers such that ac = b2 + 1. |y|. . . Suppose that r is optimal with respect to the following property: if one regards from the origin. Suppose that a. t integers such that a = x2 + y 2 . Suppose that a. R AMM 4. except for the origin. Prove that the equation ax2 + 2bxy + cy 2 = 1 is solvable in integers. c are positive integers such that ac = b2 + 1. At each lattice point of this disc. b = z 2 + t2 . y. Suppose that aij (1 ≤ i. but the proof ends here!!! Proposed problems 1. b. y. c are positive integers such that a > b > c.and make the choice c1 = · · · = cm = a− m . z. c = xz + yt. . Suppose that a.

f (xn )) with f ∈ Z[X] is dense in Rn . r) is non-empty. . xn . .j≤n can ﬁnd an integer matrix B such that A = B · t B. . Let us denote A(C. . . Suppose that a. . . Deduce the result from problem 1. Then the set of n-tuples (f (x1 ). . xn are algebraic integers such that for any 1 ≤ i ≤ n there is at least a conjugate of xi which is not between x1 . x2 . . . x2 . xn such that √ n aij xi xj < n det A. Suppose that x1 . .j≤n where A = (aij ). 11. bn some integers verifying that all pairs (ai . Let n ≥ 5 and a1 . Mathlinks contest (after an ENS entrance exam problem) 10.there are integers (not all zero) x1 . . . r) the set of points w on the unit sphere in C Rn with the property that |wk| ≥ for any non-zero vector k ∈ Zn k r (here wk is the usual scalar product and k is the Euclidean norm of the vector k ∈ Zn ). . 8. an . 193 . f (x2 ). Kurschak contest 9. . bn+1 ) = (a1 . b1 )). . . Prove that one can ﬁnd 1 < |i − j| < n − 1 such that |ai bj − aj bi | = 1. . b1 . . prove that if A = (aij )1≤i. xn ) ∈ Rn \ {0}. x2 . Using the non-elementary results presented in the topic. but if r < n − 1 there is no such C > 0. . . . 1 ≤ i ≤ n (here (an+1 . . Prove that if r > n − 1 there exists C > 0 such that A(C. .j≤n is a symmetric integer matrix such that aij xi xj > 0 for all x = (x1 . 1≤i. then we 1≤i. b are rational numbers such that the equation ax2 + by 2 = 1 has at least one rational solution. . Then it has inﬁnitely many rational solutions. Minkowski 7. bi ) are diﬀerent and |ai bi+1 − ai+1 bi | = 1. . . .

y ∈ [0. 1] and ax + by. d). b. Let a. cx + dy ∈ Z. j ≤ n + 1. Bulgaria Olympiad 13. d be positive integers such that there are 2004 pairs (x. Nikolai Nikolov.Korea TST 12. ﬁnd (b. Prove that it contains n + 1 points Ai (xi . China TST 1988 194 . yi ) such that xi − xj . yi − yj ∈ Z for all 1 ≤ i. c. A polygon of area greater than n is given in a plane. y) such that x. c) = 6. If (a.

the smallest possible integer d for which n|ad −1 is called the order of a modulo n. We have seen or will see a few such examples in our journey through the world of numbers: clever congruences that readily solve Diophantine equations. which may seem really trivial after this introduction. but in fact they are much more. we ﬁnd that n|agcd(k. then d|k.d) − 1. properties of the primes of the form 4k + 3 or even complex numbers and analysis. But from the deﬁnition of d it follows that d ≤ gcd(k. This is a consequence of the above property and of Euler’s theorem. d). All these can be called ”tricks”. THE BETTER Often. the problem has an 195 . as you are going to see soon. before that. since from Euler’s theorem we have n|aϕ(n) − 1 so such numbers d indeed exist.THE SMALLER. because n|ak − 1 and n|ad − 1. It may seem contradictory for us to talk about simple ideas and then say ”a fundamental concept”. we note a ﬁrst application of this simple observation: on (a)|ϕ(n). The deﬁnition is correct. But. Nice and easy. Well. the order of an element. There is a simple property of on (a). Indeed. which has important consequences: if k is a positive integer such that n|ak − 1. an old and nice problem. which cannot hold unless d|k. The complexity of this concept will follow from the examples. n) = 1. The reason for which we say it is a simple idea can be easily guessed from the deﬁnition: given are the positive integer n > 1 and the integer a such that gcd(a. In this unit. what we are going to discuss about is the bridge between simplicity and complexity. But could such a simple idea be good at anything? The answer is positive and will follow from the solutions of the problems to come. In what follows we will denote by on (a) the order of a modulo n. Now. But do not get excited so easily. we will discuss a fundamental concept in number theory. a large amount of simple ideas can solve very diﬃcult problems.

Example 1. But this follows from the introduction. we obtain exactly n|ϕ(an − 1). but this does not mean that it is obvious. But this is impossible. then op (2)|2n and so p|2op (2) −1|22 −1. Indeed. Prove that any prime factor of the nth Fermat number n 22 + 1 is congruent to 1 modulo 2n+1 . Solution. Show that there are inﬁnitely Solution. n. We will prove that in fact k = n + 1. suppose that d|gcd(22 + 1. p|22 n+1 − 1 and consequently op (2)|2n+1 . This ensures the existence of a positive integer k ≤ n + 1 such that op (2) + 2k . but in fact any two diﬀerent Fermat’s numbers are relatively prime. The second part is a direct consequence of the ﬁrst. Then n many prime numbers of the form 2n k + 1 for any ﬁxed n. 22 n n+k + 1). Combining 196 . Example 2. Prove that n|ϕ(an − 1) for all positive integers a. Indeed.extremely short solution. if this is not the case. it is enough to prove that there exists an inﬁnite set of Fermat’s numbers (22 nk n n + 1)nk >a any two relatively prime. It appeared in Saint Petersburg Mathematical Olympiad and also in Gazeta Matematica. we have found that op (2) = 2n+1 and we have to prove that op (2)|p − 1 to ﬁnish the ﬁrst part of the question. What is oan −1 (a)? It may seem a silly question. The proof is easy. Then we could take a prime factor of each such Fermat’s number and apply the ﬁrst part to obtain that each such prime is of the form 2n k + 1. since of course oan −1 (a) = n. Using the observation in the introduction. But not only it is easy to ﬁnd such a sequence of Fermat’s coprime numbers. Here is another beautiful application of the order of an element. Let us consider a prime p such that p|22 + 1. since p|22 +1. Then d|22 n+1 − 1 and so d|22 n+k − 1. Indeed. Therefore. It is the ﬁrst case case of Dirichlet’s theorem that we intend to discuss and is also a classical property.

a) If p|m. Example 3. so the only choice is p|q−1. the following problem is not straightforward and this explains probably its presence on a Korean TST in 2003. hence q = p. let fp (x) = xp−1 + xp−2 + · · · + x + 1. But this implies that oq (m)|q and consequently oq (m) ∈ {1. Because q|1 + m + · · · + mp−1 . This is not as easy as in the ﬁrst example. Hence both parts of the problem are solved. b) Prove that there are inﬁnitely many numbers n such that pn + 1 is prime. In fact. Though classical. m) = 1. we deduce that q|p. q|m − 1 and because q|fp (m).this with d|22 n+k + 1. . m−1) = 1. which is clearly impossible. There are many other approaches. just like in the previous problem. . For a prime p. If oq (m) = p. So. then q ≡ 1 (mod p). Moreover. Then we have of course q|mp − 1. if gcd(q. a) is straightforward. but we like the following one: take m1 = p and mk = pf (m1 )fp (m2 ) . Take such a prime divisor q. Solution. More diﬃcult is b). then q|m−1 and because q|1+m+· · ·+mp−1 . we obtain a contradiction. just by trial and error. it is not diﬃcult to ﬁnd such a sequence. we will prove that any prime factor of fp (m) is relatively prime with m(m − 1). Now. But we are tempted to use a) and to explore the properties of fp (m). it is clear that gcd(q. But we have seen while solving a) that this is not possible. Let us prove that fp (mk ) 197 . But p|m and we ﬁnd that q|m. we need to ﬁnd a sequence (mk )k≥1 of multiples of p such that fp (mk ) are pairwise relatively prime. Otherwise. p}. it follows that q|p. prove that there exists a prime factor of fp (m) that is relatively prime with m(m − 1). fp (mk−1 ). We continue with another special case of the well-known diﬃcult theorem of Dirichlet on arithmetical sequences. Anyway. let us take a prime q|fp (m) for a certain positive integer m divisible by p.

. The following problem became classical and variants of it have been given in contests for years. v2 (172 − 1) = 4 + k and the order is found by solving the equation k + 4 = 2005. Solution. Using the factoring 172 − 1 = (17 − 1)(17 + 1)(172 + 1) . Find all primes p. but not straightforward application of the order of an element is the following divisibility problem. . 5|2003q + 1 and it is easy to deduce that q is even. we also need some properties of the prime numbers. but not a multiple of 4. 2. . Let us suppose that p ≤ q. . But this is not diﬃcult. . The problem actually asks for o22005 (17). . . 2004}. o22005 (17) = 22001 is the answer to the problem. The solution ends here. The order of an element has done its job. Gabriel Dospinescu Solution. Thus. (172 k k−1 k + 1). Find the smallest number n with the property that 22005 |17n − 1. Example 5. it is time to work with exponents. so o22005 (17) = 2k . Fortunately. Example 4. . It seems to be a favorite Olympiad problem. Thus. Here. we proceed by ﬁnding the exponent of 2 in each factor of this product. that we have already studied in a previous unit. We discuss ﬁrst the trivial case p = 2. We know that o22005 (17)|ϕ(22005 ) = 22004 . We have 22005 |172 − 1. this is easy. Another simple. since fp (m1 )fp (m2 ) . 198 k t . .is relatively prime to fp (m1 ). Now. fp (mk−1 )|fp (mk ) − fp (0)|fp (mk ) − 1. q such that p2 + 1|2003q + 1 and q 2 + 1|2003p + 1. fp (m2 ). . where k ∈ {1. fp (mk−1 ). . In this case. because for all i ≥ 0 the number 172 + 1 is a multiple of 2. since it uses elementary facts and the method is more than beautiful. .

if for example p = q. r|pq + 1. or (2003)) = 1. which combined with p|q r + 1 yields p = 2. Therefore the only pair is (2. Example 6. proposed by Reid Barton for the USA TST in 2003. If op (q) is odd. Indeed. Reid Barton. r) such that p|q r + 1. which 199 . The case r = 13 is also impossible. Find all ordered triples of primes (p. Because r|20032q − 1. that is q|(p − 1)(p + 1).2). So. Then or (2003)|2 and r|20032 − 1 = 23 · 3 · 7 · 11 · 13 · 167. suppose that p > 2 and let r be a prime factor of p2 + 1. using the order of an element. r > 2. This implies that p2 + 1 ≡ 2 (mod q). q|rp + 1. since there are too many possible values for r. The ﬁrst condition p|q r + 1 implies p|q 2r − 1 and so op (q)|2r. it follows that or (2003)|2q. we have either r = 2 or q|or (2003)..hence q = 2. It is quite clear that p. We will prove that we cannot have p. we have found that for any prime factor r of p2 + 1. It follows that p = 2.. must be of the form 4k + 1 or equal to 2 and now we do not have many possibilities: r ∈ {2. Now. q. It seems that this is a dead end. TST USA 2003 Solution. the problem is not very diﬃcult and the solution follows naturally. Suppose this is the case. contradicting the assumption p > 2. which is a solution of the problem. Anyway. q. Combining this with the assumption that p ≤ q yields q|p + 1 and in fact q = p + 1. which in turn implies q|r−1. Let us see. Another simple observation narrows the number of possible cases: because r|p2 + 1. Suppose that (q. but not divisible by 4 and because any odd prime factor of it is congruent to 1 modulo q. we must have p2 + 1 ≡ 2 (mod q). More diﬃcult is the following problem. q. it follows that p|q r − 1. 13}. because 2003q + 1 ≡ 2 (mod 13) and r|2003q + 1. r are distinct. Because p2 +1 is even. then the relation p|q r + 1 is impossible.

we ﬁnd that op (12)|216 . Moreover. Therefore p ≥ 1 + 216 . If q|or (2). we might try to see if this number divides 122 + 1. In the same way. or (2)|2 and r|3. But it is wellknown that 216 + 1 is a prime (and if you do not believe. q|r2 + 1 and r|2q + 1. Thus. Example 7. It is immediate to verify that this triple satisﬁes all conditions of the problem. or (2)|2q. so you will probably be bored before ﬁnding its smallest prime factor. which contradicts the already established result that q is odd. you can check. Exactly as in the solution of the ﬁrst example. that is r = 2. Find the smallest prime factor of the number 122 + 1. Yes. all solutions are given by cyclic permutations of the components of this triple. We cannot have p|q −1. Could we have op (q) = 2r? No. As a matter of fact. Thus. Can you ﬁnd the smallest prime factor of the number 22 + 1. which does not need a single division. Therefore. then q|r − 1 and so q|r2 +1−(r2 −1) = 2. Because p|122 − 1. q. It remains to see whether = −1. But here is a beautiful and short solution. Suppose without loss of generality that p = 2. we ﬁnd that q| and 2 2 p+1 r| . Then q is odd. just by looking at the largest among 2 p. Then 122 + 1 = 2q−1 · 3 2 + 1 ≡ 3 2 + 1 3 (mod q). Let q = 216 + 1. This is clearly impossible. because p|q r +1 and p = 2. Thus. q+1 r+1 p|q + 1 and in fact p| . we ﬁnd that op (12) = 216 and so 216 |p − 1. yielding the triple (2. op (q) is either 2 or 2r.is impossible. So. so indeed 200 15 15 q−1 q−1 16 15 15 5 . you will probably ﬁnd it. So. with a large amount of work. false.5. r. But what about the number 122 +1? It has more than 30000 digits. Solution.3). Similarly. Let p be this prime number. this implies that r = 3 and q = 5. it is not that diﬃcult). But this is done in q the unit Quadratic reciprocity and the answer is positive. the only possibility is op (q) = 2 and so p|q 2 −1. since this would imply that 2r|p − 1 and so 0 ≡ pq + 1 (mod r) ≡ 2 (mod r). our assumption was wrong and indeed one of the three primes must equal 2.

2. . If the start is correct. Problems for training 1. number 3 201 n . . Let p be a nonzero polynomial with integral coeﬃcients. But then qr i=1 i=1 n 22 +1 ki ≥ 2n+1 . Prove that any positive divisor of the 2p + 1 is of the form 2kp + 1. Now everything becomes sim2n . pr prime numbers. even if it uses only this simple idea. Now. Prove that n is a prime number. 2005 Solution. . It follows that 1 ≡ 1 + 2n+1 r r i=1 ki qi (mod 22n+2 ) and so ki qi ≥ 2n+1 . Ok. since we have > (1+2n+1 )k1 +···+kr and so k1 +· · ·+kr ≤ This shows that qr ≤ 2(n + 1) and the proof ﬁnishes here. let us write 22 + 1 = pk1 . here is a problem that will keep you occupied for a certain period of time. . reduce the relation 22 + 1 = pk1 . Let a. pkr r 1 with p1 < . Yet. The reader will not imagine how simple this problem re- ally is. Prove that for any n > 1 the largest prime factor of n 22 n + 1 is at least equal to n · 2n+2 + 1. Prove that there are at most ﬁnitely many numbers n for which p(n) and 22 + 1 are not relatively prime. n+1 ple. pkr r 1 r n n modulo 22n+2 . where x < n − 1 and x|n − 1. . . China TST. you must be already tired of this old fashioned idea that any prime factor of 22 + 1 is congruent to 1 modulo 2n+1 . We know that we can ﬁnd qi ∈ N such that pi = 1 + 2n+1 qi .3 q−1 2 + 1 ≡ 0 (mod 2) and 216 + 1 is the smallest prime factor of the 15 number 122 + 1. Example 8. Indeed. n > 2 be positive integers such that n|an−1 − 1 and n does not divide any of the numbers ax − 1... . Let p > 3 be a prime. 3.

a n−1 p n n n − 1) = 1. where k is odd. Prove that any prime factor of n is congruent to 1 modulo F . Laurentiu Panaitopol. Gazeta Matematica 7. Suppose further that there is an integer a such that n|an−1 − 1 and for all primes p dividing n − 1 we have gcd(n. Find the least possible value of m + n. Let p be a prime and q > 5 a prime factor of the number 2p + 3p . d < n.Fermat 4. Let n be a positive integer such that n − 1 = F R. TST Romania 8. Find the smallest number n such that 21989 |mn − 1. Prove that any positive divisor of the number an − bn is either of the form nk + 1 or divides a number of the form ad − bd . Let a > b > 1 and n > 1 be positive integers. IMO 1978 10. 5. Bulgaria. 1997 6. Let 0 < m < n be integers such that 1978m and 1978n have the same last three digits. Let m > 1 be an odd number. Prove that q > p. Find oq (k) in terms of n and v2 (m) J. van de Lune 12. Prove that there is a positive integer n such that op (n) = d. Find all positive integers m. R) = 1. Let q = k · 2m + 1 be a divisor of the number 22 + 1. with d|n. Find the smallest repunit divisible by 19. n for which n|1 + m3 + m2·3 . IMO 1989 Shortlist 9. where all the prime factors of F are known and gcd(F. Let p be a prime number and let d a positive divisor of p − 1. 202 . 11.

with n ≥ 1. b) If k = 1 then let l = vp (npd −1). Pocklington. opj (n) = pd for 2 ≤ j ≤ l and opj (n) = pj−l+1 d. Let a > 1 be an integer and let us deﬁne op (a) = 0 if p|a. Prove that there are only ﬁnitely many positive integers n such that all prime factors of an − 1 are in A. Let a > 1 be a positive integer. AMM 14. . Prove that op (n) = d. Lehmer Test 13. Let d = op (n) and let k = vp (nd − 1). IMO 2003 17. Prove that for any prime p there is a prime number q that does not divide any of the numbers np − p. . 5. Prove that for inﬁnitely many n the largest prime factor of an − 1 is greater than n loga n. Let A be a ﬁnite set of prime numbers and let a ≥ 2 be a positive integer. Jerrold W Grossman. . op (a) Jon Froemke. 7. Prove p−1 that the function f : {2. f (p) = is unbounded. Iran Olympiad 16. 15. 11. for all j ≥ l.Proth. Gabriel Dospinescu 203 . } → N. 3. a) If k > 1 then opj (n) = d for j ≤ k and opj (n) = pj−k d for all j ≥ k.

one cannot start a topic about approximation theorems without talking ﬁrst about Kronecker’s theorem. This is the reason for which we will try to discuss some classical approximation theorems. Prove that the sequence ([n 2003])n≥1 contains arbitrarily long geometric progressions with arbitrarily large ratio. the consequences that will be discussed here are almost elementary. We will prove that there are arbitrarily long geometric sequences with ratio p. Instead. IMO TST Romania Solution. we will discuss two beautiful problems.DENSITY AND REGULAR DISTRIBUTION Everyone knows that ({na})n≥1 is dense in [0. let us prove that we can ﬁnd a positive integer m such that √ √ [pk m 2003] = pk [m 2003] for all 1 ≤ k ≤ n. Given n ≥ 3. but because we will prove a much stronger result about the sequence ({na})n≥1 . Of course.1] if a is an irrational number. The natural question: what fraction of numbers between 1 and n have this property (speaking here about large values of n) is much more diﬃcult and to answer it we need some stronger tools. Various applications of this nice result have appeared in diﬀerent contests and will probably make the object of Olympiad problems in the future. A simple one is as follows: using Kronecker’s theorem one can easily prove that for any positive integer a that is not a power of 10 there exists n such that an begins with 2006. The proofs are non-trivial and require some heavy-duty analysis. If the existence of such a number is proved. a classical theorem of Kronecker. Let us take p a very large number. there are some examples in which this result is ineﬀective. Yet. then the conclusion is immediate. consequences of this theorem. not only because it is very well-known. Radu Gologan. √ Example 1. We skip the proof. particularly the very eﬃcient Weil criterion and its consequences. Yet. But observe that 204 .

n a −1 205 . It follows m and. First of all. Deﬁne the sequence xn as the number formed by the ﬁrst k digits of the number [an ] with n ≥ 1.√ √ √ [pk m 2003] = pk [m 2003] is equivalent to [pk {m 2003}] = 0. The proof of this claim is not diﬃcult. xk = 10p−k Another remark is the following: there is a positive integer r such that xrT > 10k−1 . that x1 . xk · 10p−k + xk+1 . Example 2. xp . it follows that k − 1 + {log[arT ]} < log(1 + 10k−1 ) for all r. . xk · 10p−k ≤ m < (x1 . Then m = x1 . since p = 1 + [log m]. The existence of a positive integer m with the last p property is ensured by Kronecker’s theorem. . Mathlinks Contest Solution. arT − 1 It suﬃces now to consider a sequence of positive integers (rn ) such 1 that 1 − < {rn T log a} (the existence is a simple consequence of Kron necker’s lemma) and we will deduce that: = {rT log a} − log log 1 + 1 10k−1 + 1 arn T + log rn T > 1 for all n. . . xp . Prove that this sequence is not eventually periodical. xk + 1) · 10p−k . Using the ﬁrst observation. . . but useful remarks. we ﬁnd that for all r > 0 we have xrT = 10k−1 . but which is again a simple consequence of Kronecker’s theorem. let us write m = x1 x2 . Indeed. . . The solution is based on certain simple. . the number formed with the ﬁrst k digits of a number m is [10k−1+{log m} ]. . hence x1 . or to √ 1 {m 2003} < n . Indeed. Here is a problem that is apparently very diﬃcult. with p ≥ k. . Consider k ≥ 1 and a such that log a is irrational. Gabriel Dospinescu. . assuming the contrary. Thus log 1 + 1 10k−1 > log[arT ] − [log[arT ]] > log(arT − 1) − [log arT ] arT . the claim is proved.

we can multiply all its elements by a suitable integer such that the elements of the new set are as close to integers as we want. . Before passing to the quantitative results stated at the beginning of this chapter. 2. . . Example 3. There exists a positive integer n and integers p1 . . 1) = s=1 Js . it can even help us reducing the problem to integers. But this contradicts Kronecker’s theorem. which turns out to be very useful when dealing with real numbers and their properties.The last inequality is clearly impossible. . yet surprising result. N N . α2 . . It follows that for n > r we have xnT = xrT . xk be some real numbers and let ε > 0. thus {log[anT ]} ≥ log 1 + This shows that log 1 + 1 10k−1 ≤ log[anT ] − [log[anT ]] ≤ nT log a − [log anT ] = {nT log a} for all n > r. 1) in ε N intervals. as we will see in one of the examples. N 1 10k−1 . . where 206 . Now. (Dirichlet) Let x1 . we must speak about a simple. [0. n} a sequence of k positive integers α1 . . . Js = s−1 s . . But ﬁrst. . . choose n = Nk + 1 and associate to any positive integer q ∈ {1. let us state and prove this result. pk such that |nxi − pi | < ε for all i. x2 . . . Solution. Sometimes. . Finally. . Thus we need to prove that if we have a ﬁnite set of real numbers. 1 Let us choose an integer N > and partition the interval [0. assume the existence of such an r. αk . p2 .

We claim that if ε > 0 is small enough. Now. . and some integers p1 . pk have 207 . x2 . p2 . if they are all rational. Putnam Competition and later on in a Japanese Mathematical Olympiad.αi = s if and only if {qxi } ∈ Js . assume that they are real numbers. the corresponding p1 . . . j = i. . The one we are going to discuss has had a circuitous itinerary between world’s Olympiads: proposed for IMO long time ago. otherwise. . We obtain at most N k sequences corresponding to these numbers and so by Dirichlet’s criterion we can ﬁnd 1 ≤ u < v ≤ n such that the same sequence is associated to u and v. pk such that |nxi − pi | < ε for all i. Of course. This means that for all 1 ≤ i ≤ k we have |{uxi } − {vxi }| < 1 ≤ ε. Example 4. . . Thus assume that at least one of the numbers is irrational. the problem for integers is very well-known and easy: it suﬃces to observe that in this case all numbers xi have the same parity and the use of inﬁnite descent solves the problem (either they are all even and in this case we divide each one by two and obtain a new set with smaller sum and the same properties. such that the sum of the numbers in each group is the same. pi = [vxi ] − [uxi ]. there are not many such problems. . . . which is clearly much more subtle. p2 . we subtract 1 from each one and then divide by 2). . it appeared next at the W. Prove that all numbers are equal. And here is how we can use this result in problems where it is more comfortable to work with integers. it suﬃces to multiply by their common denominator and apply the ﬁrst case. N It suﬃces thus to pick n = v − u. x2n+1 be positive real numbers with the property: for any 1 ≤ i ≤ 2n + 1 one can make two groups of n numbers by using xj . Consider ε > 0. But don’t kid yourself. . First of all. Let x1 . Solution.L. a positive integer n.

. . Then aij pj = j=i j=i aij (nxj − pj ) ≤ 2nε. Thus if we choose ε < 1 . Now. . p2 . pk j=i have the same property. . But 2 > |nN | max |xi − xj |. x2 . x2 . it is not diﬃcult to prove that the sequence (nN )N >2n is unbounded. x2n+1 . . ai ∈ [a. . because of the ﬁrst case). . . hence maxi. is famous and deserves to be discussed because of its beauty and apparent simplicity. lim |{i| 1 ≤ i ≤ n.j N solved. N Because at least one of the numbers x1 . 1}). . pN such that |nN xi − pN | ≤ .1]. Since they are all integers. Thus we have proved that for any 1 N > 2n there are integers nN . 1 n→∞ n lim n 1 f (ak ) = k=1 208 0 f (x)dx. 1] → R. Then the following statements are equivalent: a) For any real numbers 0 ≤ a ≤ b ≤ 1. . they must be equal (again. .the same property as x1 . Let (an )n≥1 be a sequence of real numbers from the interval [0. take some i and write the partition condition in the form aij nxj = 0 or j=i j=i aij (nxj − pj ) = − j=i aij pj (where of course aij ∈ {−1. b]}| = b − a. The following criterion. . let us turn to more quantitative results about the set of fractional parts of natural multiples of diﬀerent real numbers. .j |xi − xj | = 0 and the problem is i. x2n+1 is irrational. Weil criterion. due to Weil. n n→∞ b) For any continuous function f : [0. Indeed. then 2n aij pj = 0 and so p1 .

containing yet all the necessary ingredients. 1]. But then if we write 1 n n 1 f (ak ) − k=1 0 f (x)dx ≤ 1 n n 1 |f (ak ) − g(ak )| + k=1 1 0 |f (x) − g(x)|dx 1 + n n g(ak ) − k=1 0 g(x)dx and apply the result in b) for the function g. there is a well-known and easy to verify property of continuous functions: they can be uniformly approximated with piecewise functions.1]. consider two sequences of continuous functions fk . b] with 0 < a < b < 1. More subtle is that b) implies a). The fact that b) implies c) is immediate. Let us consider the subinterval I = [a. k=1 In this case we will say that the sequence is equidistributed. By linearity. j=1 . Therefore 1 n n j=1 |{i| 1 ≤ i ≤ n. We will present just a sketch of the solution. b − (being aﬃne otherwise). Next. we easily deduce that b) is true for any continuous function. 1 n→∞ n lim n e2iπpak = 0. b]}| 1 fk (aj ) ≤ ≤ n n 209 n gk (aj ). given ε > 0. That is. Now. we can ﬁnd a piecewise function g such that |g(x) − f (x)| < ε for all x ∈ [0. ai ∈ [a. First of all. this remains true for any piecewise function. 1] and 1 on a + . we observe that a) says precisely that b) is true for the characteristic function of any sub-interval of [0. k k while gk has ”the same” properties but is greater than or equal to λI (the characteristic function of I = [a. [b. b]). a].c) For any positive integer p ≥ 1. gk such that fk 1 1 is zero on [0.

let us prove that c) implies b). Finally. Of course. both having equal values at 0 and 1 and such that 1 |f (x) − g(x)| ≤ h(x) and 0 h(x)dx ≤ ε. 1] → R continuous. The reader has already seen how to adapt this proof for the case a = 0 or b = 1.But from the hypothesis. one can easily see that b) is true for any continuous function. we deduce that b) is true for continuous functions f for which f (0) = f (1). We invite the reader to ﬁnd an elementary proof in order to appreciate the power of Weil’s criterion. h. ai ∈ [a. k Now. The above inequalities show that actually for all suﬃciently large positive integer n |{i| 1 ≤ i ≤ n. a linearity argument allows us to assume that b) is true for any trigonometric polynomials of any degree. Before presenting the second problem. we need another deﬁnition: we say that the sequence 210 . The ﬁrst problem that we discuss is in fact the most common result about equidistribution. let us take ε > 0 and k suﬃciently large. Now. 1] → R satisfying f (0) = f (1) can be uniformly approximated by trigonometric polynomials (this is a really non-trivial result due to Weierstrass). b]}| − b + a ≤ 2ε n and the conclusion follows. given f : [0. it is immediate that for any ε > 0 we can ﬁnd two continuous functions g. Using the same arguments as those used to prove that b) implies a). 1 n and 1 n n 1 n 1 fk (aj ) → j=1 0 fk (x)dx = b − a − 1 k gk (aj ) → j=1 0 1 gk (x)dx = b − a + . Because any continuous function f : [0.

. we deserve a reward: this is a simple consequence of Weil’s criterion..).(an )n≥1 is uniformly distributed mod 1 if the sequence of fractional parts of an is equidistributed. Indeed. . . Let a be an irrational number. 1. . . which is clearly equivalent to the existence of p ≥ 1 such that 2007 · 10p > 2n ≥ 2006 · 10p . again a reward: this is going to be equally easy (of course. . .. So. it suﬃces to prove that c) is veriﬁed. a2 . 9} such that 2n − 2006a1 a2 . Well. Example 6. . . ap ∈ {0. the reader needs some rest before looking at some deeper results. after so much work. Well. This can be rewritten in the form log 2007 + p > n log 2 ≥ log 2006 + p 2006 2007 > {n log 2} > log . But this is just a geometric series!!! A one-line computation shows that (∗) is trivially satisﬁed and thus we have the desired result. Indeed. Then the sequence (na)n≥1 is uniformly distributed mod 1. here is the classical example. which reduces to proving that 1 lim n→∞ n n e2iπpka = 0 k=1 (∗) for all integers p ≥ 1. What is the density of the set of positive integers n for which 2n begins with 2006? Solution. It’s probably time to solve the problem presented in the very beginning of this note: how to compute the density of those numbers n for which 2n begins with 2006 (for example). . 1000 1000 Thus the density of the desired set is exactly the density of the set of This implies [n log 2] = p + 3 hence log 211 . 2n begins with 2006 if and only if there is a p ≥ 1 and some digits a1 . Example 5. Solution. ap .

Then the sequence (xn ) is also equidistributed. . from a certain time on) one should try to discover what is behind such great results. Unbelievable. We will limit to give the main idea of the 212 . a diﬃcult proof that uses the ”well-known” but not easy to remember inequality of Van der Corput.positive integers n satisfying log 2007 2006 > {n log 2} > log . zn and any h ∈ {1. (Van der Corput) For any complex numbers z1 . . much more is true. . Example 7. but true! Not to mention that the proof of this inequality is anything but easy. n}): n 2 p−1 n−r n h 2 i=1 zi ≤ (n + h − 1) 2 r=1 (h − r)Re i=1 zi zi+r +h i=1 |zi |2 . We have seen a beautiful proof of the fact that if a is irrational. (Van der Corput) Let (xn ) be a sequence of real numbers such that the sequences (xn+p − xn )n≥1 are equidistributed for all p ≥ 1. the last set has density log answer to our problem. This is not an Olympiad problem!!! But mathematics is not just Olympiad and from time to time (in fact. . . . but this much more is also much more diﬃcult to prove. 1000 1000 2007 and so this is the 2006 From example 5. n}. . z2 . This is the reason for which we present a proof of this theorem. . The ﬁrst is due to Van der Corput and shows how a brilliant combination of algebraic manipulations and Weil’s criterion can yield diﬃcult and nice results. then (na)n≥1 is uniformly distributed mod 1. . 2. 2. Actually. Lemma. the following inequality is true (with the convention that zi = 0 for any integer i not in {1. . . . The next two examples are two important theorems.

by using this lemma and Weil’s criterion... if they have the patience to expand i=1 j=0 zi−j and see that it is nothing else than p−1 n−r n 2 r=1 (p − r)Re i=1 zi zi+r +p i=1 |zi |2 . You would have never guessed: the Cauchy Schwarz inequality!!! The simple observation that n n+h−1 h−1 h i=1 zi = i=1 j=0 zi−j allows us to write (via Cauchy Schwarz’s inequality): n 2 n+h−1 h−1 2 h 2 i=1 zi ≤ (n + h − 1) i=1 j=0 zi−j . . this is where the readers will get some satisfacn+p−1 p−1 2 tion. Also.proof. k=1 Fix such a p ≥ 1 and take for the moment a positive real number h and ε ∈ (0. the idea is to prove that 1 n→∞ n lim n e2iπpxk = 0 for all p ≥ 1.. we have: 1 n n 2 zj j=1 ≤ 1 n+h−1 · hn + 2 n2 h2 213  h−1 i=1  (h − i)Re  n−i j=1  zj · zi+j  . Of course. denote zj = e2iπpxj .. the computations being very technical. 1) (h may depend on ε). Wishing them good luck with the computations. And next ? Well. Using the lemma. The idea behind this fundamental inequality is another fundamental one. we will now prove Van der Corput’s theorem.

j=1 Therefore 1 n n 2 zj j=1 1 n+h−1 hn + 2εn ≤ 2· n h2 h−1 (h − i) i=1 < 2(1 + ε) n+h−1 (1 + ε) < nh h 2(1 + ε) . . but focus on the following result. 2. 214 . we deduce ε2 for all suﬃciently large n. h − 1 we deduce that for all suﬃciently large n we have n−i e2iπp(xj −xi+j ) ≤ εn. by choosing h > that for all suﬃciently large n we have 1 n n zj ≤ ε. Now.Now. let us regard   n−i  n−i j=1  e2iπp(xj −xi+j )  ≤ n−i Re  j=1 zj · zi+j  = Re  e2iπp(xj −xi+j ) . j=1 Using Weil’s criterion for the sequences (xn+i − xn ) for i = 1. . but why not taking one more step once we are already here? Let us prove the following weaker (but as the reader will probably agree. This was surely the most diﬃcult result of this unit. . absolutely nontrivial) version of famous theorem of Weil. It is related to the equidistribution of the sequence ({f (n)})n≥1 where is a real polynomial having at least one irrational coeﬃcient except for the free term. . j=1 This shows that Weil’s criterion is veriﬁed and thus (xn ) is equidistributed. We will not prove this here.

But this is exactly the induction hypothesis applied to the polynomial (whose leading coeﬃcient is clearly irrational) f (X + p) − f (X). The proof by induction ﬁnishes here. (Weil) Let f be a polynomial with real coeﬃcients and irrational leading coeﬃcient. Tuymaada Olympiad 215 . Indeed.Example 8. Now.  n≥1 p. Prove that the sequence consisting of the ﬁrst digit of 2n + 3n is not periodical. if the result holds for polynomials of degree at most k. Then the sequence ({f (n)})n≥1 is equidistributed. if f has degree 1. the proof by induction is immediate. Indeed. Prove that by using diﬀerent terms of the sequence [n2 2006] one can construct geometric sequences of any length. Let x be an irrational number and let f (t) = min({t}. Iran 2004 4. Compute sup  min |p − q 3|. Problems for training  √ 1.q∈N p+q=n Putnam Competition √ 2. {1 − t}). 3. Prove that given any ε > 0 one can ﬁnd a positive integer n such that f (n2 x) < ε. then the conclusion is immediate (see example 5). the sequence (f (n + p) − (f (n)) is equidistributed. The reader has probably noticed that this is an immediate consequence of Van der Corput’s theorem (but just imagine the amount of work done to arrive at this conclusion!!!). it suﬃces (by Van der Corput’s theorem) to prove that for all positive integers p.

b be positive real numbers such that {na} + {nb} < 1 for all n. [n2 2]. Suppose that A = {n1 . [nk 3] are both geometrical sequences. Give a necessary and suﬃcient condition ensuring n |f (k)| the existence of a constant c > 0 such that > c ln n for all n. Romanian Contest. . continuous. Then at least one of them is an integer. 2005 216 . Does the sequence sin(n2 ) + sin(n3 ) converge? Gabriel Dospinescu 8. n2 . Then prove that A has zero density. Let f be a polynomial with integral coeﬃcients and let a be an irrational number. . . and periodical function such n |f (k)| that the sequence is bounded. n + 1] for all n ≥ n0 . . 2. starting from √ √ the origin. k = 1. Prove that for every k one can ﬁnd distinct positive inte√ √ √ √ gers n1 . . . It can only jump over distances equal to 2 and 2005. Let a. } is a set of positive integers such that the sequence (cos nk )k≥1 is convergent. . . be in the set A = {[na]| n ≥ 1}? Is it true that any set of positive integers with positive density contains an inﬁnite arithmetical sequence? 9. . . Marian Tetiva 6. A ﬂea moves in the positive direction of an axis. . . Suppose that f is a real. . Prove that f (k) = 0 for k k=1 n≥1 all positive integers k. 10. [nk 2] and [n1 3]. . . Prove that there exists n0 such that the ﬂea will be able to arrive in any interval [n. . k k=1 Gabriel Dospinescu 7.5. . Can all numbers f (k). . After a romanian IMO TST problem 11. n2 . nk such that [n1 2]. √ √ [n2 3].

. Let z1 . Putnam 13. . Prove that the sets A = {[na]| n ≥ 1}.12. . Prove that for any ε > 0 there are inﬁnitely many positive integers n such that ε+ k k k k |z1 + z2 + · · · + zn | < max{|z1 |. B = {[nb]| n ≥ 1}. c be positive real numbers. . z2 . Let a. . C = {[nc]| n ≥ 1} cannot form a partition of the set of positive integers. . |z2 |. . . b. zn be arbitrary complex numbers. 217 . |zn |}.

This can work surprisingly well. just the careful analysis of the structure of the numbers and their digits.... 1. Also. Note that among the ﬁrst 40 numbers. Baltic.. probably. Let x be this number. The easiest of these techniques is.. as we will see in the following examples.s(x) + 12 occur as sum of digits 218 . Obviously. x. s(x) + 1. On the other hand.THE SUM OF DIGITS OF A POSITIVE INTEGER Problems about the sum of digits of a positive integer often occur in mathematical contests because of their diﬃculty and the lack of standard ways to tackle the problem. there are several more or less known results about sum of digits. Prove that among any 79 consecutive numbers.. The following ”formula” can be checked easily: s(n) = n − 9 k≥1 n 10k (1) From (1) we can easily deduce some well-known results about s(n) such as s(n) ≡ n( mod 9) and s(m + n) ≤ s(m) + s(n). We will work only in base 10 and we will denote the decimal sum of digits of the positive integer x by s(x). results which may oﬀer simple ways to tackle hard problems. which can hardly be used in applications. Unfortunately. x+39 are among our numbers. (1) is a clumsy formula. there are exactly 4 multiples of 10. This is why a synthesis of the most frequent techniques that occur in such cases would be useful. one can choose at least one whose sum of digits is a multiple of 13. This is what we will discuss about in the following. x+1. so s(x). 1997 Solution.. We have selected several representative problems to show how the main results and techniques work and why they are so important.. it is clear that the last but one digit of one of them is at least 6.

2. N . s(N + i) = 2 + 9 + 0 + 9(11! − 1) + i − 2 = i + 11! while for 12 ≤ i ≤ 21. For 2 ≤ i ≤ 11. the following is true: a positive integer cannot be written as the sum of two numbers with the same sum of digits iﬀ all of its digits (eventually) excepting the ﬁrst are 9 and the sum of its digits is odd. So. In truth.in some of our numbers. but this situation can occur at most once in our case.. The main idea is that among s(n + 2). N +2. In fact. 2. How many positive integers n ≤ 102005 can be written as the sum of 2 positive integers with the same sum of digits? Adrian Zahariuc Solution. At ﬁrst glance. s(N + i) = 2 + 9 + 1 + (i − 12) = i. The answer here is not trivial at all. which still do not require any special result or technique. so our numbers have the desired property. 3. For N = 21. we make transports at a + 10 only when the last but one digit of a is 9. we can choose N +1. 2000 Solution. For i = 1 we have nothing to prove. 219 . for N > 21. it might seem almost impossible to ﬁnd the exact number of positive integers with this property. Answer: 102005 − 9023. . one of these numbers is a multiple of 13 and we are done. Find the greatest N such that one can ﬁnd N consecutive numbers with the property that the sum of digits of the k-th number is divisible by k. N + 21... Tournament of Towns. Obviously. We will continue with two harder problems. we have no solution... for k = 1. which is impossible since they should all be even. s(n + 12) and s(n + 22) there are two consecutive numbers. where N = 291 · 1011! − 12.. namely 21..

Then let us chose a = 10p−1 (c+1). n − a = b1 b2 . B = B1 B2 . . Suppose there are a.. Back to the original problem. say I1 and I2 . It follows that s(n) = s(a) + s(b) = 2s(a). It is clear that the numbers A = A1 A2 . which is impossible since s(n) is odd. then n must have a digit which is not the ﬁrst and is not equal to 9.. If s(n) is odd...Let n be such a number. (ai + bi + 1)/2 if i ∈ I1 or (ai + bi − 1)/2 if i ∈ I2 and Bi = ai + bi − Ai . can be written as the sum of 2 positive integers with the same sum of digits. So. Proof. If the s(n) is even.. k} : 2 does not divide ai + bi } is even. All we have to do now is take one-byone a ”unity” from a number and give it to the other until the 2 numbers have the same sum of digits.... so it can be divided into 2 sets with the same number of elements.. .. Now we will prove that any number n which is not one of the numbers stated above.Ak . This is clear enough. We will start with the following: Lemma. so s(a) + s(n − a) = 9 + s(n) ≡ 0(mod2) ⇒ s(a) ≡ s(n − a)(mod2) which proves our claim.bk The lemma shows that the number of elements of the set I = {i ∈ {1. let us do this rigorously. 2.. take a = 0. Let c be the value of this digit and p its position (from right to left). Let a = a1 a2 . k deﬁne Ai = (ai + bi )/2 if i ∈ I. At the adding a + (n − a) = n there is exactly one transport. There is a ≤ n such that s(a) ≡ s(n − a)(mod2).Bk 220 . This will happen since they have the same parity.. For i = 1. otherwise it would have one of the forbidden forms.ak .. 2. there are no transports. The main fact is that when we add a + b = n. b ∈ Z+ such that n = a + b and s(a) = s(b).

2002 Solution.99 2 = 99. 4 and 7. Find all the possible values of the sum of digits of a perfect square.. We have a ≡ b ≡ a + b ≡ n(mod9).. 4 or 7 mod 9..99 82 00.99 12 = 99. 1. Vasile Zidaru and Mircea Lascu...have the property that s(A) = s(B) and A + B = n. Sometimes it is combined with some simple inequalities such as s(n) ≤ 9( lg n + 1). but perfect squares do have something to do with remainders mod 9! In fact. What does sum of digits has to do with perfect squares? Apparently.99 2 ) = 9n n n−1 n−1 n 99. JBMO TST..00 1 ⇒ s(99.. To prove that all such numbers work. 1. The proof is complete. we deduce that the sum of digits of a perfect square must be congruent to 0. we will use a small and very common (but worth to remember!) trick: use numbers that consist almost only of 9-s.... it is very easy to prove that the only possible values of a perfect square mod 9 are 0... 1995 Solution. Find all n for which one can ﬁnd a and b such that s(a) = s(b) = s(a + b) = n.. we can take a = b = 10k − 1 and we are done since s(10k − 1) = s(2 · 10k − 2) = 9k. If n = 9k. So. We have the following identities: 99.00 81 ⇒ s(99.. nothing.99 12 ) = 9n + 1 n−1 n−2 n−2 221 n−1 . This is probably the most famous property of the function s and it has a series of remarkable applications. Iberoamerican. 5.99 8 00.. so 9 divides n. We have previously seen that s(n) ≡ n( mod 9). Some immediate applications are the following: 4.

. so s(3n+1 ) − s(3n ) ≥ 9.99 42 ) = 9n + 7 n−1 n−2 n−2 n−1 and since s(0) = 0.00 36 ⇒ s(99. s(4) = 4 and s(25) = 7 the proof is complete. so 9n − 9N ≤ 9 + 9(n + 1) lg 3. 3k+1 − 3k ≥ 9(n − N ) ⇒ s(3n+1 ) ≥ 9(n − N ).00 64 ⇒ s(99. Finally. 000.99. in Kvant... ∀n ≥ N . 000 = 180. n ≥ N + 1. Solution. s(3n+1 ) − s(3n ) ≡ 0(mod9). for n ≥ 2.99 22 ) = 9n + 4 n−1 2 n−2 n−2 n−1 99. 7... IMO 1975 Solution.. Compute s(s(s(44444444 ))). If so far we have studied some remarkable properties of the function s. It follows that n 1481 ≡ 7(mod9).99 88 00. Prove that for any N there is n ≥ N such that s(3n ) ≥ s(3n+1 ). we present a beautiful problem which appeared in the Russian Olympiad and. But... 6. s(s(s(n))) ≡ s(s(n)) ≡ s(n) ≡ n(mod9) and since 44444444 ≡ 74444 = 7 · 73 the only possible answer is 7.99 4 = 99. On the other hand. This is obviously a contradiction.99 22 = 99. for all n ≥ N + 1..99 84 00... which were quite well-known. later. so s(s(s(44444444 ))) ≤ 12. k=N +1 But s(3n+1 ) ≤ 9( lg 3n+1 + 1). Using the inequality s(n) ≤ 9( lg n + 1) several times we have s(44444444 ) ≤ 9( lg 44444444 + 1) < 9 · 20. s(1) = 1. 000 + 1) ≤ 36. Suppose by way of contradiction that there is one N such that s(3n+1 ) − s(3n ) > 0. s(s(44444444 )) ≤ 9( lg s(44444444 ) + 1) ≤ 9(lg 180.. ∀n ≥ N . it is time to present some problems and 222 ...

.aj−1 (aj − 1) 99.99(9 − a1 ).. The problem in trivial if we know the previous result. Proof. such as the following problem. All we have to do is write x = a1 a2 ....99).... Compute s(9 · 99 · 999 · . Only to assure our readers that this problem did not appear on the ISL out of nowhere. The previous result is by no means hard.99 N ) = 9 · 2n . then s(x(10n − 1)) = 9n. let’s solve it.aj with aj = 0 (we can ignore the ﬁnal 0-s of x) and note that x(10n − 1) = a1 a2 ..+2 2n−1 n−1 < 102 − 1 n so s(99. The ﬁrst result is the following: Statement. We will see that constructing such a number is hard. First. 9... such numbers are called Niven numbers and they are an important research source in number theory.. · 99.. If 1 ≤ x ≤ 10n . IMO Shortlist.. 2n USAMO. Now.(9 − aj )(10 − aj )... We have N = 9 · 99 · 999 · ... · 99.. there are very hard applications of this apparently unimportant result.. but we will see that it can be the key in many situations. Prove that for any n there is a positive integer n which is divisible by its sum of digits. The idea is very simple. n−j which obviously has the sum of digits equal to 9n. but interesting and hard.99 < 101+2+. A ﬁrst application is: 8. 1998 Solution.results which are less familiar. 2n However. 1992 Solution. we 223 .

The following result turned out to have several consequences. Take the number N = M − 10p−k a. we decide that the required number p should be aa. k Proof.. Statement. Any multiple of 11... we have s(N ) < s(M ) < k. M = a1 a2 .aa b ≤ 10t − 1..11. Suppose by way of contradiction that the statement is false and take M to be the smallest multiple of a such that s(M ) < k..aa b · (10t − 1).. 224 .aa b ≤ 10t − 1. we will require s + t = n − 1 and 9t| aa. we have s(N ) = s(M ) < k and if ak+1 = 9.11 (it n can be easily proved by induction that 3k+2 |103 − 1). 2·3k We continue our investigations in ﬁnding suitable techniques for problems involving sum of digits with a very beautiful result... Obviously. If we also take in account the condition aa.ap .. 9. We will try to prove that s(N ) < k.. N is a multiple of a... 9t|10t − 1.11 has sum of digits at least k.. Note that s(ia) = ik k for i = 1. We will use the extremal principle.aa b · (10t − 1).. Therefore... So M ≥ 10a > 10k .. We now use the fact that s for t a power of 3. Therefore.... when we can take the number 11. with p ≥ k + 1 and ap = 0. the choice p = 11. let us take t = 3k where k is chosen such that 3k < n < 3k+1 .22(102·3 k and p = − 1) otherwise becomes natural. most of them being very hard.11 2(103 − 1) when n ≤ 2 · 3k s n−3k −1 k 22. Due to the trick that we should search numbers with many equal digits and the last result. .. 2. where a = 11.. But this is not hard at all since if ak+1 < 9.will get rid of the case n = 3k . with aa. we would contradict the minimality of M and the proof would be complete. So.. This number will have s s k s + t + 1 digits and its sum of digits will be 9t. In this way.

the solution follows directly from our result. Denote by f (n) the smallest k for which one can ﬁnd a set X ⊂ Z+ of cardinality n with the property 225 lg n n→∞ − 1 ≤ n ⇒ 10 lg n − 1|n!. But. we will present a hard. Again. which might seem simple. due to our ”theorem” we can chose the sequence an = n(10m − 1). 11. Tournament of Towns. Let us remind that this is the last problem of ISL 1999. 13. 12. 1999 Solution. such that all its members have the sum of digits greater than k. we have lim s(n!) =∞ lnk ln n Gabriel Dospinescu Solution. We can choose the number 101989 − 1. from which our conclusion follows . Due to the simple fact that 10 easily. Let n be a ﬁxed positive integer. Let S be the set of positive integers whose decimal representation contains only of at most 1988 1-s and the rest 0-s. but which might be unsolvable without it. whose multiples have sum of digits greater than 1988. IMO Shortlist. The oﬃcial solution is indeed one for such a problem. Prove that there is a positive integer which does not divide any member of S. but beautiful. 10. we have that s(n!) ≥ lg n .We will show 3 applications of this fact. problem from the USAMO. 1988 Solution. Prove that for every k. as a ﬁnal proof of the utility of these two results. Prove that for any k > 0. there is an inﬁnite arithmetical sequence having the ratio relatively prime to 10. so the hardest. Now. where m > k and we are done.

. 2 Consider the set X = {j(10l −1) : 1 ≤ j ≤ n}. Let m be the largest integer such that n ≥ 10m − 1.. Then. m|ai+1 +ai+2 +. so f (n) ≤ 9l and the RHS is proved.+aj so we are done. If one of then is a multiple of m..... We will use the following well-known Lemma. USAMO 2005 Solution. + am . Any set M = {a1 . Gabriel Dospinescu and Titu Andreescu. am } has a nonempty subset whose sum of elements is divisible by m. there are 2 of them congruent mod m. We will prove that lg(n + 1) ≤ f (n) ≤ 9 lg n(n + 1) +1 . From the lemma. Let l be the smallest integer such that 10l − 1 ≥ n(n + 1) .. them we are done. say Y . . it follows that s x∈Y x = 9l for all nonempty subsets Y of X. 2 which is enough to establish our claim.. Consider the sums a1 . a1 + a2 .. whose sum of elements is divisible by 10m − 1. Prove that C1 lg n < f (n) < C2 lg n for some constants C1 and C2 .that s x∈Y x =k for all nonempty subsets Y of X. a2 . a1 + a2 + .. say the i-th and the j-th. Proof. it follows that any set X with n elements has a subset. Otherwise.. By our 226 . By the previous inequality and our ﬁrst statement.

and the proof is complete. It follows that m|m and m |m. for suﬃciently large p. ∀c ≥ max{x. b}. Let us solve now the harder problem.second statement. Now. Let a = 2x 5y m and b = 2z 5t m . then the problem becomes trivial. where (m. This means the the sum of the ﬁrst p digits of b/a is bounded. For any k ≥ 1. a a so. 14. Suppose a = max{a. it follows that an = bn. Let a. 10) = 1. Adrian Zahariuc and Gabriel Dospinescu Solution. It follows that s(ank ) is bounded. On the other hand. b ∈ Z+ such that s(an) = s(bn) for all n ∈ Z+ . 10k b b ≤ bnk < 10k + b. there is a nk such that 10k ≤ ank ≤ 10k + a − 1. y} 227 . The last solved problem is one we consider to be very hard. so m = m . and which uses diﬀerent techniques than the ones we have mentioned so far. Then. If (max{a. 10) = (m . by Euler’s theorem. which could only happen when this fraction has ﬁnitely many decimals. so a = b. we can write the hypothesis as s(2z 5u mn2c−x 5c−y ) = s(2x 5y mn2c−x 5c−y ) = s(mn). We start with an observation. we can prove the same result about a/b. 10)=1. b}. so s(bnk ) is bounded as well. so there is an n such that an = 10ϕ(a) − 1 and since numbers consisting only of 9-s have the sum of digits greater than all previous numbers. the ﬁrst (nonzero) digits of b/a are exactly the same as the ﬁrst p digits of bnk for large enough k. Analogously. a|10ϕ(a) − 1. it follows that s x∈Y x ≥ m ⇒ f (n) ≥ m. Prove that lg a − lg b ∈ Z.

So p = 0 and hence z + c − x = u + c − y. so a = 10x−z b and the proof is complete.. It follows that s(m) = s(k p m) = s(k 2p m) = s(k 3p m) = . This problem can be nicely extended to any base. then for any sequence of digits. Finally. if p = max{z + c − x. 1). although there are some very important diﬀerences. Prove that among any 18 consecutive 2-digit numbers there is at least one Niven number.. u + c − y}. 1961 2. u + c − y} − min{z + c − x. which is a contradiction. such that tn m starts with the selected sequence of digits. we get that there is a k ∈ {2. we will use the following: Lemma. If lg t ∈ R − Q. Let t = ap . Prove that among any 39 there is one whose sum of digits is divisible by 11. Proof. so lg t ∈ R − Q unless p = 0. But lg tn m = n lg t + m and by Kronecker’s theorem {{n lg t} : n ∈ Z+ } is dense in (0. so the proof is complete. The proof of the general case is quite similar. 1). we invite our readers to solve this proposed problems: Proposed Problems 1. USSR.Now. The upmetioned methods are just a point to start from in solving such problems since the variety of problems involving sum of digits is very large. there is a n ∈ Z+ . Now. The techniques are useful only when they are applied creatively. The lemma implies the very important result that s(tn m) is unbounded for p = 0. If we will prove that {{lg tn m} : n ∈ Z+ } is dense in (0. 228 . then we are done. 5} such that s(mn) = s(mk p n) for all n ∈ Z+ .

Training. except itself). *** 10. Are there arbitrarily long arithmetical sequences whose terms have the same sum of digits? What about inﬁnite aritmetical sequences? *** 7. Then.. 1997 3. 1970 5. diﬀerent form 6 (which is equal to the sum of its divisors. 000? USSR. Are there positive integers n such that s(n) = 1. we calculate the sum of digits of the new number and so on. 000. We start with a perfect number. Kurschak. Prove that there are arbitrarily long sequences of consecutive numbers which do not contain any Niven number. Find all x such that s(x) = s(2x) = s(3x) = . Prove that we will eventually get 1. = s(x2 ). 1989 6. and calculate its sum of digits. Are there p ∈ Z[X] such that n→∞ lim s(p(n)) = ∞? *** 9. *** 229 . Prove that for any positive integer n there are inﬁnitely many numbers m which do not contain any zero.Tournament of Towns. *** 8. such that s(n) = s(mn).. 1985 4. Prove that n→∞ lim s(2n ) = ∞. USSR. 000 and s(n2 ) = 1.

Prove that there are inﬁnitely many special numbers 10n + b iﬀ b − 1 is special.. + < 3. Let (an )n≥1 be a sequence such that s(an ) ≥ n. Find a Niven number with 100 digits.2 a1 a2 an Can we replace 3. Christopher D. Prove that there are inﬁnitely many x ∈ Z+ such that s(x) + s(x2 ) = s(x3 ). Gabriel Dospinescu 13. A number m is called special if there is a k such that f (k) = m. Deﬁne f (n) = n + s(n).2 by 3? Laurentiu Panaitopol 14. 1990 230 .. Sankt Petersburg. < n50 such that n1 + s(n1 ) = n2 + s(n2 ) = . = n50 + s(n50 ) Poland. b. = n50 − n49 . Prove that for any n.. c and d are primes such that 2 < a ≤ c and a = b.. Prove that a = c and b = d. Prove that one can ﬁnd n1 < n2 < . Study whether we can choose the numbers in the previous problem such that n2 − n1 = n3 − n2 = . Gabriel Dospinescu 12. We now that there is one M ∈ Z such that the numbers an + b and cn + d have the same sum of digits for any n > M and base between 2 and a − 1. a. Gabriel Dospinescu 16. we have 1 1 1 + + .... 1999 15.11. Long 17..

Prove that the sequence s( an + b ) contains a constant subsequence. Let k ∈ Z+ . b > 0. 2003 21. Laurentiu Panaitopol. Prove that there is a positive integer m such that the equation n + s(n) = m has exactly k solutions. Are there 19 positive integers with the same sum of digits. which add up to 1999? Rusia. there is a number n ∈ Z+ such that αn ∈ S. Romanian IMO TST. Romanian IMO TST. Prove that n≥1 s(n) 10 = ln 10. Rusia. Prove that S contains numbers with arbitrarily large sum of digits. 2002 23. Find the smallest positive integer which can be expressed at the same time as the sum of 2002 numbers with the same sum of digits and as the sum of 2003 numbers with the same sum of digits. ﬁnd s(3n). Shallit 231 . Let S be a set such that for any α ∈ R−Q. 2002 25.18. 1999 24. n(n + 1) 9 O. Gabriel Dospinescu 20. where M is given. Let a be a positive integer such that s(an + n) = 1 + s(n) for any n > M . Mihai Manea. Prove that a is a power of 10. Let a. 1999 22. Gabriel Dospinescu 19. Rusia. If s(n) = 100 and s(44n) = 800.

Is it true that s(2n ) = ∞? n→∞ ln n lim 232 . b ∈ Z such that lg a − lg b ∈ Z / with the property that s(ak ) = s(bk ) for any k ∈ {1. Prove that for any n. Prove that n→∞ lim lg a − lg b ∈ Z.. 4. For any a > 1. 2. 5. Let a. we have n→∞ lim s(an ) = ∞ (proved only for a couple of values.Open Questions 1.. . namely 2.. b ∈ Z+ such that s(an ) = s(bn ) for all n ∈ Z+ . 4. n}. 6. 8). 2. Is it true that s(n!) = ∞? n ln ln n 3. there are a.

Indeed. multiplying by the common denominator of all coeﬃcients of polynomials h. of course. Let f. As usual. r are polynomials from Q[X] and deg r < deg f . Solution. while using analysis they can be solved quickly. We will develop our skills in ”hunting” these sequences by solving ﬁrst some easy problems (anyway. we begin with a classic beautiful problem. Sometimes. Now. ”easy” is a relative concept: try to solve them elementary and you will see if they really are easy) and after that we will attack the chestnuts.ANALYSIS AGAINST NUMBER THEORY? ”Olympiad problems can be solved without the aid of analysis or linear algebra” is a sentence always heard when speaking about the elementary problems given in contests. which has lots of applications and extensions. Practically. were h. The diﬃculty is in ﬁnding those sequences. Example 1. ”quickly” only if you see the sequence that hides after each problem. Their elementary solutions are very tricky and sometimes extremely difﬁcult. (bn )n≥1 of integer numbers r(an ) and a positive integer N such that bn = N (we could have some f (an ) 233 . Prove that f divides g in Q[X]. These sequences must become constant and from here the problem is much easier. the hypothesis becomes: there exists two inﬁnite sequences (an )n≥1 . this is easy. Well. g ∈ Z[X] be two non-constant polynomials such that f (n)|g(n) for an inﬁnite natural numbers n. but the true nature and essence of some of these problems is in analysis and this is the reason for which such type of problems are always the highlight of a contest. but most of the time this is a very diﬃcult task. we need to look at the remainder of g when divided with f in Q[X]! Let us write g = f h + r. This is true. our aim is to exhibit convergent sequences formed by integer numbers. r.

The next problem we are going to discuss is a particular case of a much more general and classical result: if f is a polynomial with integer coeﬃcients. Well. let us say a = x2 . so the convergence is useless. but we need another sequence. the sequence is formed by integer numbers. which is practically the same thing with r = 0 (don’t forget that any non-zero polynomial has only a ﬁnite number of roots!).problems with the roots of f . this is the same as r(an ) = 0 from a certain point n0 . it follows that → 0. but it’s not a sequence of integers. since √ this sequence certainly converges to a (the reader has already noticed √ why it wasn’t useless to ﬁnd that xn −n a is convergent. This implies that from a certain point. Let a = 0. an is not a r(an ) root of f ). Thus. thus (bn )n≥1 f (an ) is a sequence of integer numbers that tends to 0. And yes. c = y 2 . We won’t discuss here this general result (the reader will ﬁnd a proof in the chapter about arithmetic properties of polynomials). A simple induction shows that xn = xM + (n − M )x and so (xM − M x + nx)2 = x2 n2 + bn + c for 234 . Example 2. a must be a perfect square. so from a certain point. but they are in ﬁnite number and the sequence (an )n≥1 tends to inﬁnity. y ∈ Z such that a = x2 . b. Solution. we used this to establish the convergence of (xn+1 − xn )n≥1 ). it’s not that useless. b = 2xy. The easiest way is to work with (xn+1 − xn )n≥1 . Let us begin by writing an2 + bn + c = x2 for a certain n √ sequence of nonnegative integers (xn )n≥1 . We could expect that xn −n a converges. Since deg r < deg f . then there exists a polynomial g ∈ Q[X] such that f (x) = g k (x). it converges. all the terms of these sequence are 0. the number an2 + bn + c is a perfect square. so it is constant from a certain point. √ we can ﬁnd a number M such that if n ≥ M then xn+1 = xn + a. Thus. c be integers such that for any natural number n. k > 1 is a natural number and k f (n) ∈ Q for all natural numbers n. In fact. This time. But in this moment the problem is solved. Prove that there exist x.

A simple identiﬁcation of coeﬃcients ﬁnishes the solution. but it really doesn’t seem to be related with mathematical analysis. Let a. Prove that a = b. So. let us supaxn + bxn xn+1 + bxn+1 a pose that a > b and let us write = c in the form axn + bxn b a b 1+ a xn+1 1+ axn+1 −xn xn = c. What does the problem say in fact? That we can ﬁnd a sequence of positive integers (xn )n≥1 such that axn +bxn = 2cn . Let us see if this is true or not. Being a sequence of integer numbers. it must become constant. as we will see. What could be the convergent sequence here? We see that (xn )n≥1 is appreciatively kn for a certain constant k. From where could we ﬁnd xn+1 −xn ? Certainly. Example 3. The following problem is based on the same idea. this is immediate. Thus. Laurentiu Panaitopol Solution. Why is it so easy? It would be easy if we could show that xn → ∞. Fortunately. it is closely related to the concept of convergence. since 2axn > 2cn ⇒ xn > n loga c. Now. c > 1 be positive integers such that for any positive integer n there exists a positive integer k such that ak +bk = 2cn . by writing that axn+1 +bxn+1 = 2cn+1 axn+1 + bxn+1 and after that considering the value = c. In fact.all n ≥ M . b. from where it is easy to see that axn+1 −xn converges to c. since we can take y = xM − M x. so there exist M such that for all n ≥ M we have 235 . we found that axn+1 −xn converges. we could expect that the sequence (xn+1 − xn )n≥1 converges.

[a1 n] + [a2 n] + · · · + [ak n]. According to our assumption. This is obvious. they are the prime numbers. what are the most eﬃcient numbers n to be used? Yes. Of course. the solution of such a problem is better to be indirect. This means that for all n ≥ M we also have 1+ b a b 1+ a xn+1 xn = 1.axn+1 −xn = c. this implies that for all n ≥ N there exist a natural number xn such that [a1 pn ] + [a2 pn ] + · · · + [ak pn ] = xn pn . there is a number N such that if n ≥ N then pn ≥ M . attention must be paid to details! Example 4. . . then it must divide [a1 n] + [a2 n] + · · · + [ak n]. our assumption was wrong and we must have a ≤ b. Thus. Prove that there exits inﬁnitely many natural numbers n such that n and [a1 n] + [a2 n] + · · · + [ak n] are relatively prime. . Now. you have already guessed what is the convergent sequence! Yes. pn 236 . let us assume that there exists a number M such that for all n ≥ M the numbers n and [a1 n] + [a2 n] + · · · + [ak n] are not relatively prime. Arhimede Magazine Solution. Gabriel Dospinescu. [a1 pn ] + [a2 pn ] + · · · + [ak pn ] since tends to n ≥ N a1 + a2 + · · · + ak . So. since if n is prime and it is not relatively prime with . a2 . But. This suggests us to consider the sequence of prime numbers (pn )n≥1 . Since this sequence is inﬁnite. But this is impossible. we conclude that a = b. in which ﬁnding the right convergent sequence of integers in not diﬃcult at all. it is (xn )n≥N . Another easy example is the following problem. . And now. since a > b. Let a1 . Due to symmetry in a and b. ak be positive real numbers such that at least one of them is not an integer.

k and n ≥ P . Then. k. Let a. xn = √ a · 2n + b. Of course. . Then. Suppose that a = 0. Well. This implies the existence of a natural number N such that for all n ≥ P we have 2xn = xn+2 . we argue by contradiction. . Thus. A classical result of Schur states that for any non-constant polynomial f with integer coeﬃcients. Consequently. This shows. but very hard to ﬁnd. According to the hypothesis.Thus. b ∈ Z such that for all natural numbers n the number a · 2n + b is a perfect square. we start to have some experience in ”guessing” the sequences. which is impossible for a = 0. Suppose that f is a polynomial with integer coeﬃcients and (an ) is a strictly increasing sequence of natural numbers such that 237 . f (2). there exists a sequence of positive integers (xn )n≥1 such that for all natural numbers n. Example 5. Again. But this is the same as {a1 pn } + {a2 pn } + · · · + {ak pn } = 0. a direct computation shows that lim (2xn − xn+2 ) = n→∞ 0. this says that ai pn ∈ Z for all i = 1. The next problem we are going to discuss may seem obvious after reading the solution. . otherwise for large values of n the number a·2n +b is negative. which contradicts the hypothesis. f (3). we can ﬁnd a number P such that for xn = a1 + a2 + · · · + ak for all n ≥ P . a > 0. it’s just that type of problem whose solution is very short. we were wrong again and the problem statement is right! Step by step. Example 6. it’s time to solve some more diﬃcult problems. Poland TST Solution. In fact. a and 2a are both perfect squares. that our assumption was wrong and indeed a = 0. But 2xn = xn+2 is equivalent with b = 0. is inﬁnite. as usually. the set of prime numbers dividing at least one of the numbers f (1). the conclusion is immediate: ai ∈ Z for all i = 1. Then. Prove that a = 0. of course. The following problem is a generalization of this result.

by taking k = 1 we clearly have 2 deg(f ) n≥1 1 = ∞.an ≤ f (n) for all n. The same idea is employed in the following problem. pkαN and thus 1 N 1 ≤ ak n 1 α1 . if the conclusion of the problem is not true. .αN ∈Z+ pkα1 1 . .... pkαN 1 N pk 1 j = . . . . . .α2 . 1. . Then the set of prime numbers dividing at least one term of the sequence is inﬁnite... Let a. . .. Example 7. . we also have 1 ≥ ak n 1 = ∞. it suﬃces to remark that we have actually 1 α1 . Prove that there is a multiple of a which contains all digits 0.. . . pr such that any term of the sequence is of the form pkα1 . . p2 . we have 1 α1 . b − 1 when written in base b. we can ﬁnd p1 . On the other hand. pr and any k > 0.. (f (n))k Thus. pkαN N < ∞. b ≥ 2 be natural numbers. Adapted after a Putnam problem 238 . .α2 . . ki k−1 pj p j=1 j n On the other hand. pkαN N < ∞. .αN ∈Z+ N j=1 i≥0 pkα1 . Solution.. Indeed. .αN ∈Z+ n≥1 pkα1 1 ... p2 . (f (n))k n≥1 n≥1 which is clearly impossible.α2 .. . . The idea is very nice: for any ﬁnite set of prime numbers p1 .

So. It’s beauty explains probably the preference for this problem. in a Russian Team Selection Test in 1997 and also in the Bulgarian Olympiad in 2003. Example 8. the problem states that there exists a sequence of positive integers (xn )n≥1 such that f (xn ) = 2n . since the relation xn 239 . trying to ﬁnd the right convergent sequence. xn behaves like 2 k . Then. Then any multiple of a misses at least a digit when written in base b. f (x) behaves like xk . there are at most (b − 1)n numbers which have n digits in base b. it suﬃces to show 2 3 that the sum of inverses of all natural numbers missing at least one digit in base b is convergent and we will reach a contradiction. that is for large n. The conclusion b n follows. But of course. Since the sum of inverses of all 1 1 multiples of a diverges (because 1 + + + · · · = ∞). Now. Let’s suppose the contrary. all diﬀerent from j. We return to classical problems to discuss a beautiful problem. Then. So. = n a good possibly convergent sequence could be xn+k − 2xn . This is easy. Let us suppose that deg f = k > 1. Thus. For any n ≥ 1. for large values of x. Prove that deg f = 1. we could try ﬁrst to ”think n big”: we have xk ∼ 2n . Solution. Let f ∈ Z[X] be a polynomial with leading coeﬃcient 1 such that for any natural number n the equation f (x) = 2n has at least one natural solution. the sum of inverses of numbers that miss the digit j when written in base b−1 n b is at most equal to b . First. since each one of them is at least equal to bn−1 .Solution. it suﬃces to prove it for a ﬁxed (but arbitrary) digit j. that appeared in the Tournament of the Towns in 1982. the hard part: proving that this sequence is indeed convergent. which converges. we xn+k will show that converges to 2.

f (xn+k ) = 2k f (xn ) implies f (xn+k ) xk n+k and since f (x) = 1 and lim xn = ∞ x→∞ xk n→∞ (do you see why?), we ﬁnd that indeed lim xn+k = 2. n→∞ xn lim We will see that this will help us a lot. Indeed, let us write
k−1

xn+k xn

k

= 2k ·

f (xn ) xk n

f (x) = xk +
i=0

ai xi .

Then f (xn+k ) = 2k f (xn ) can be also written
k−1

ai (2k xi − xi ) n n+k xn+k − 2xn =
i=0 k−1 k−i−1 (2xn )i xn+k i=0

xn+k = 2. it follows that the right-hand But from the fact that lim n→∞ xn side in the above relation is also convergent. So, (xn+k − 2xn )n≥1 is convergent and it follows that there exist M, N such that for all n ≥ M we have xn+k = 2xn +N . But now the problem is almost done, since the last result combined with f (xn+k ) = 2k f (xn ) yields f (2xn + N ) = 2k f (xn ) for n ≥ M , that is f (2x + N ) = 2k f (x). So, an arithmetical property of the polynomial turned into an algebraic one using analysis. This algebraic property helps us to immediately solve the problem. Indeed, we see that if z is a complex root of f , then 2z + N, 4z + 3N, 8z + 7N, . . . are all roots of f . Since f is non-zero, this sequence must be ﬁnite and this can happen only for z = −N . Since −N is the only root of f , we deduce that f (x) = (x + N )k . But since the equation f (x) = 22k+1 has
240

natural roots, we ﬁnd that 2 k ∈ N , which implies, contradiction. Thus, our assumption was wrong and deg f = 1. The idea of the following problem is so beautiful, that after reading the solution the reader will have the impression that the problem is trivial. Wrong! The problem is really diﬃcult and to make again an experiment, we will ask the reader to struggle a lot before reading the solution. He will see the diﬃculty. Example 9. Let π(n) be the number of prime numbers smaller than or equal to n. Prove that there exist inﬁnitely many numbers n such that π(n)|n. AMM Solution. First, let us prove the following result, which is the key of the problem. Lemma. For any increasing sequence of positive integers (an )n≥1 an n such that lim = 0, the sequence contains all natural numn→∞ n an n≥1 bers. In particular, for inﬁnitely many n we have that n divides an . Proof. Even if it seems unbelievable, this is true and moreover the proof is extremely short. Let m ≥ 1 be a natural number. Consider the 1 amn ≥ . This set contains and it is bounded, since set A = n ≥ 1| mn m amk 1 amn = 0. Thus it has a maximal element k. If = , then lim n→∞ mn mk m n m is in the sequence . Otherwise, we have am(k+1) ≥ amk ≥ an n≥1 k + 1, which shows that k + 1 is also in the set, contradiction with the maximality of k. The lemma is proved. π(n) = 0. Fortunately, this is n→∞ n well-known and not diﬃcult to prove. There are easier proofs than the Thus, all we need to show is that lim following one, but we prefer to deduce it from a famous and beautiful result of Erdos.
241

1

Erdos’s theorem. We have
p≤n p prime

p ≤ 4n .

The proof of this result is magniﬁcient. The proof is by induction. For small values of n it is clear. Now, assume the inequality true for all values smaller than n and let us prove that
p≤n p prime

p ≤ 4n . If nis even,

we have nothing to prove, since =
p≤n p prime p≤n−1 p prime

p ≤ 4n−1 < 4n .

Now, assume that n = 2k + 1 and consider the binomial coeﬃcient 2k + 1 k = (k + 2) . . . (2k + 1) . k!

A simple application of the fact that 22k+1 =
i≥0

2k + 1 i

shows that 2k + 1 k ≤ 4k .

Thus, using the inductive hypothesis, we ﬁnd that p≤
p≤n p prime p≤k+1 p prime

p
k+2≤p≤2k+1 p prime

p ≤ 4k+1 · 4k = 4n .

π(n) = 0 is trivial. Indeed, ﬁx k ≥ 1. We n→∞ n have for all large n the inequality Now, the fact that lim n lg 4 ≥
k≤p≤n p prime

lg p ≥ lg k(π(n) − π(k)),

which shows that π(n) ≤ π(k) lg 4 + . n lg k
242

π(n) = 0. The problem is solved. n It is time now for the last problem, which is, as usual, very hard. We This shows of course that lim
n→∞

don’t exaggerate if we say that the following problem is exceptionally diﬃcult. Example 10. Let a, b > 1 be natural numbers such that for any natural number n, an − 1|bn − 1. Prove that b is a natural power of a. Marius Cavachi, AMM Solution. This time we will be able to ﬁnd the right convergent sequence only after some double recurrences. Let us see. So, initially we are given that there exists a sequence of positive integers (x1 )n≥1 n n b n 1 = b − 1 Then, x1 ∼ for large values of n. So, we such that xn n = an − 1 a could expect that the sequence (x2 )n≥1 , x2 = bx1 − ax2 is convergent. n n n n+1 Unfortunately, x2 = n bn+1 (a − 1) − an+1 (b − 1) + a − b , (an − 1)(an+1 − 1)

which is not necessarily convergent. But... if we look again at this b n sequence, we see that for large values of n it grows like , so a2 much slower. And this is the good idea: repeat this procedure until n b the ﬁnal sequence behaves like , where k is chosen such that ak+1 ak ≤ b < ak+1 . Thus, the ﬁnal sequence will converge to 0. Again, the hard part has just begun, since we have to prove that if we deﬁne
k+1 = 0. This isn’t easy at all. The idea xi+1 = bxi − ai xi n n n+1 then lim xn n→∞

is to compute x3 and after that to prove the following statement: for n any i ≥ 1 the sequence (xi )n≥1 has the form n ci bn + ci−1 a(i−1)n + · · · + c1 an + c0 (an+i−1 − 1)(an+i−2 − 1) . . . (an − 1) for some constants c0 , c1 , . . . , ci . Proving this is not so hard, the hard part was to think about it. How can we prove the statement otherwise than by
243

induction? And induction turns out to be quite easy. Supposing that the statement is true for i, then the corresponding statement for i+1 follows from xi+1 = bxi − ai xi n n n+1 directly (note that to make the diﬀerence, we just have to multiply the numerator ci bn +ci−1 a(i−1)n +· · ·+c1 an +c0 with b and an+i −1. Then, we proceed in the same way with the second fraction and the term bn+1 an+i will vanish). So, we have found a formula which shows that as soon as ai > b we have lim xi = 0. So, we have deduced n
n→∞

that lim xk+1 = 0. Another step of the solution is to take the minimal n
n→∞

index j such that lim xj = 0. Obviously, j > 1 and the recurrence n
n→∞ i relation xi+1 = bxi − ai xi n n n+1 shows that xn ∈ Z for all n, i. Thus, there

exists M such that whenever n ≥ M we have xj = 0. This is the same n b n−M j−1 as bxj−1 = aj xj−1 for all n ≥ M , which implies xj−1 = xM n n n+1 aj for all n ≥ M . Let us suppose that b is not a multiple of a. Since b n−M j−1 xM ∈ Z for all n ≥ M , we must have xj−1 = 0 and so M aj j j−1 xn = 0 for n ≥ M , which means lim xn = 0. But this contradicts
n→∞

the minimality of j. Since we have reached a contradiction, we must have a|b. Let us write b = ca. Then, the relation an − 1|bn − 1 implies an − 1|cn − 1. And now are ﬁnally done. Why? We have just seen that an − 1|cn − 1 for all n ≥ 1. But our previous argument applied for c instead of b shows that a|c. Thus, c = ad and we deduce again that a|d. Since this process cannot be inﬁnite, b must be a power of a. It worth saying that there exist an even stronger result: it is enough to suppose that an − 1|bn − 1 for an inﬁnite number n, but this is a much more diﬃcult problem. It follows from a result found by Bugeaud, Corvaja and Zannier in 2003: If a, b > 1 are multiplicatively independent in Q∗ (that is loga b ∈ Q), then for any ε > 0 there exists n0 = n0 (a, b, ε) such that gcd(an − 1, bn −

244

the proof is too advanced to be presented here. b = 35. Let p be a polynomial with integer coeﬃcients such that there exists a sequence of pair wise distinct positive integers (an )n≥1 such that p(a1 ) = 0. Prove that if f or g is surjective. the decimal representation of a + bn contains a sequence of consecutive digits which form the decimal representation of n (for example. 2002 6. where ϕ is the Euler totient function. Prove that for any given k > 0 there are inﬁnitely many numbers n such that ϕ(an + b) < kn. Unfortunately. Prove that there exists integer numbers a. 245 . Let f. b be natural numbers such that for any natural number n. 2. Romanian Olympiad 1991 3. if a = 600. Let a. Find the degree of this polynomial. Gabriel Dospinescu. Tournament of the Towns. b such that f (x) = (ax + b)k . Prove that b is a power of 10. Let f ∈ Z[X] be a polynomial of degree k such that for all n ∈ N we have k f (n) ∈ Z. .1) < 2εn for all n ≥ n0 . . Moldova TST 2004 5. Problems for training 1. Laurentiu Panaitopol. p(a2 ) = a1 . 2003 √ 4. . Let a. n = 16 we have 600 + 16 · 35 = 1160). Find all arithmetic progressions of positive integers (an )n≥1 such that for all n ≥ 1 the number a1 + a2 + · · · + an is a perfect square. p(a3 ) = a2 . then these functions have inﬁnitely many ﬁxed points. b > 1 be positive integers. . g : N∗ → N∗ two functions such that |f (n) − n| ≤ 2004 n and n2 + g 2 (n) = 2f 2 (n). Tournament of the Towns.

Gabriel Dospinescu 11. . 2a . Let f be a complex polynomial having the property that for all natural number n. 2 5 in base b. Mathlinks Contest 14. Let A be a set of natural numbers. . a · 2n + b is a divisor of cn + 1. . which contains at least one number among any 2006 consecutive natural numbers and let f a nonconstant polynomial with integer coeﬃcients. Putnam 12. . Let b an integer at least equal to 5 and deﬁne the number xn = 11 . the equation f (x) = an has at least a positive rational solution. the equation f (x) = n has at least a rational solution. IMO Shortlist 2004 8. Then f has degree at most 1. Gabriel Dospinescu. Gabriel Dospinescu 246 .Gabriel Dospinescu 7. Revue de Mathematiques Speciales 13. Laurentiu Panaitopol. c such that for any positive integer n. Find all triplets of integer numbers a. Find all complex polynomials f having the property: there exists a ≥ 2 a natural number such that for all suﬃciently large n. Prove that xn is a perfect square for n−1 n all suﬃciently large n if and only if b = 10. Then prove that a is also integer. . 3a . are integers. . Suppose that a is a real number such that all numbers 1a . 1 22 . Prove that there exists a √ number N such that for any n ≥ N there are at least ln ln n diﬀerent prime numbers dividing the number N ≤k≤n k∈A 2 f (k). . b.

Any inﬁnite arithmetical progression contains inﬁnitely many terms that are not powers of integers. Find all real polynomials such that the image of any repunit is also a repunit. Find all a. the number a · 4n + b · 6n + c · 9n is a perfect square. c such that for all suﬃciently large n. g two real polynomials of degree 2 such that for any real x. After a problem from Kvant 247 . Prove that there exists n0 such that for all n ≥ n0 we have an = a1 + a2 + · · · + an−1 . so is g(x).15. Let (an )n≥1 be an inﬁnite and strictly increasing sequence of positive integers such that for all n ≥ 2002. Let f. Tournament of the Towns. Kvant 16. 18. if f (x) is integer. After a Putnam problem 21. Find all pairs of natural numbers a. b such that for every positive integer n the number an + b is triangular if and only if n is triangular. 17. 20. Then there are integers m. 2002 22. Bulgarian Olympiad 19. Prove that in any strictly increasing sequence of positive integers (an )n≥1 which satisﬁes an < 100n for all n. Try to generalize the preceding problem (this is for the diehards!!!). one can ﬁnd inﬁnitely many terms containing at least 1986 consecutive 1. an |a1 + a2 + · · · + an−1 . b. n such that g(x) = mf (x) + n for all x.

Sa se arate ca exista k ∈ R astfel incat A = {xk | x ∈ B}.23. Gabriel Dospinescu 248 . Fie doua multimi ﬁnite de numere reale pozitive cu proprietatea ca xn | n ∈ R x∈A ⊂ x∈B xn | n ∈ R .

deﬁne the function a p wise. In the ﬁrst case. we will unfold some easier properties of Legendre’s symbol. let us assume that =1 p and consider x a solution of the equation x2 = a in Zp . Now. other- a = −1. q the following equality holds: p q q p = (−1) p−1 q−1 · 2 2 .QUADRATIC RECIPROCITY For an odd prime p. 1} by = 1 if the equation x2 = a has at least a solution in Zp and. otherwise we say that it is a non quadratic residue modulo p. let us present an useful theoretical (but not practical at all) way of a computing due to Euler. we p−1 will prove that there are exactly quadratic residues in Zp \ {0}. Until then. 2 This will enable us to conclude that quadratic residues are precisely the 249 p−1 . we say that a is a quadratic residue p modulo p. for any quadratic residue we have a 2 = 1 (mod p). we ﬁnd that a 2 = xp−1 = 1 (mod p). p Theorem. First. a p : Z → {−1. This states that for diﬀerent odd prime numbers p. Using Fermat’s theorem. We will prove this result and many other simple remarks concerning a quadratic residues in what follows. First. This is a nontrivial result whose proof will be sketched later. In p p−1 addition. Perhaps the most remarkable result involving this symbol is Gauss’s quadratic reciprocity law. This function is called Legendre’s symbol and plays a fundamental role in number theory. Thus the equality p−1 a = a 2 (mod p) holds for all quadratic residues a modulo p. The following identity is true: a p =a p−1 2 (mod p).

Lemma. a b the same Euler’s criterion implies that = whenever a ≡ b p p (mod p). Indeed. These two observations show that in fact 2 p−1 each of these polynomials has exactly zeros in Zp . p p If m is the number of positive integers x such that x < and < ax 2 2 a m. (mod p) < p. The relation It is now time to come back to Gauss’s celebrated quadratic reciprocity law. Thus each of the polynomials X 2 − 1 and X 2 + 1 p−1 has at most zeros in Zp . 1} is a group p morphism. the number of diﬀerent zeros of a polynomial cannot exceed its degree. it a allows a very quick proof of the fact that : Z → {−1. Also. ab a b = shows that while studying Lep p p gendre’s symbol. next that there are at least 2 p−1 all numbers i2 (mod p) with 1 ≤ i ≤ are quadratic residues and 2 p−1 they are all diﬀerent. p) = 1. This shows that there are exactly quadratic 2 residues in Zp \ {0} and also proves Euler’s criterion. it suﬃces to focus on the prime numbers only. But in a ﬁeld. we have ab p = (ab) p−1 2 (mod p) = a p−1 2 b p−1 2 (mod p) = a p b p . Note that Fermat’s little theorem implies that the polynomial − 1)(X p−1 2 + 1) has exactly p − 1 roots in the ﬁeld p−1 p−1 Zp . First of all. Indeed. Let p be an odd prime and let a ∈ Z such that gcd(a. We have said that Euler’s criterion is a very useful result.roots of the polynomial X p−1 2 − 1 and also that non quadratic residues p−1 2 are exactly the roots of the polynomial X X p−1 − 1 = (X p−1 2 + 1 (from Fermat’s little theorem). we will prove a lemma (due to Gauss). then = (−1) p 250 . Indeed. Let us observe 2 p−1 quadratic residues modulo p.

. On the other hand. x2 . yk all numbers 2 2 p smaller than . . 2 On the other hand. . 2 Observe that p−1 2 (ax) = a x=1 p−1 2 p−1 !≡ 2 a p p−1 ! (mod p). . . . . 2 that p − axi (mod p) = ayj (mod p). we ﬁnally deduce that 251 . We clearly have (ax) (mod p) ax (mod p)> p 2 m ax k (mod p)< p 2 (ax) (mod p) = i=1 axi (mod p) j=1 ayi (mod p). Indeed. 2 a p = (−1)m . Hence we can write 2 m k axi i=1 m (mod p) j=1 ayi (mod p) k = (−1)m i=1 p−1 2 (p − axi (mod p)) j=1 ayj (mod p) ≡ (−1) m i=1 i (mod p) = (−1)m p−1 ! (mod p). yj < . . the numbers p − axi (mod p) and ayi (mod p) p−1 (mod p). which is clearly true by the deﬁp nition of xi . . . . . p−1 2 (ax) = x=1 ax (mod p)> p 2 (ax) (mod p) ax (mod p)< p 2 (ax) (mod p). x2 . 2. . xm . xm be those numbers x for which x < and 2 p p−2 < ax (mod p) < p. it suﬃces to prove give a partition of 1. Combining these facts. . diﬀerent from x1 . .p Proof. Let x1 . Let k = − m and y1 . .

the number 23 +1 has at least n prime divisors of the form 8k + 3. the contradiction is immediate. This can be also written in p p−1 p2 −1 −1 2 the form = 1. Then p n+1 −2 −2 ≡ (2 2 )2 (mod p) and so = 1. or (−1) 2 + 8 = 1. but we think that they are still worthy not necessarily because they appeared in various contests. The identity number p. Example 1. Most of them are merely direct applications of the above results. 252 n n . we are now able to solve some interesting problems. Theorem. we arrive at the following result. the reader will enjoy proving the next two classical results. Example 2. Theorem. assume that p is a prime divisor of the form 8k −1 that divides 2n + 1. if n is even. Prove that for any positive integer n. the following identity holds: p q q p = (−1) p−1 q−1 · 2 2 2 p = (−1) p2 −1 8 holds for any odd prime . Using this powerful arsenal. Prove that the number 2n + 1 does not have prime divisors of the form 8k − 1.Using Gauss’s lemma. Of course. Vietnam TST 2004 Solution. (quadratic reciprocity law) For any diﬀerent odd primes p. since in this case we would have −1 ≡ (2 2 )2 (mod p) and p−1 −1 so −1 = (−1) 2 = = 1. Indeed. q. if p is of p p the form 8k − 1 the later cannot hold and this is the contradiction that solves the problem. Now. Based on the same idea and with a bit more work. assume that n is odd. Fortunately.

Using the same argument as the one in the p2 − 1 p − 1 previous problem. Thus. We will prove that if n is odd. then it has no prime divisors of the form 8k + 5 either. let p be a prime divisor of 2n + 1. we will prove that for all 1 ≤ i < j ≤ n − 1. with 1 ≤ i ≤ n − 1 has at least a prime divisor of the form 8k + 3 diﬀerent from 3. we deduce that 2n + 1 does not have prime divisors of the form 8k + 7. This cannot happen unless p = 3. It remains to show that each of the numbers 22·3 − 23 + 1. Now. assume that p is a prime number dividing gcd(22·3 − 23 + 1. Using the result of the previous problem.Gabriel Dospinescu Solution. . 22·3 − 23 + 1) = 3. we deduce that + is even. It would follow in this case that 23 + 1 has at least n − 1 distinct prime divisors of the form 8k+3 (from the previous remarks) and since it 253 n i i i i j j j j i i j j j i+1 i i j j i+1 j j i + 1. Indeed. . Then 2n ≡ −1 (mod p) and so −2 ≡ (2 )2 (mod p). let us solve the proposed problem. Indeed. But since v3 (gcd(22·3 − 23 + 1. it follows that gcd(22·3 − 23 + 1. 22·3 − 23 + 1)) = 1 (as one can immediately check). gcd(22·3 − 23 + 1. 22·3 − 23 + 1) = 3 and the claim is proved. The essential observation is the identity: 23 + 1 = (2 + 1)(22 − 2 + 1)(22·3 − 23 + 1) . . We will assume n > 2 (otherwise the veriﬁcation is trivial). )3 j−i−1 ≡ (−1)3 j−i−1 ≡ −1 (mod p). (22·3 n n−1 n+1 2 − 23 n−1 + 1) i Now. 22·3 − 23 + 1) We will then have p|23 23 ≡ (23 implying 0 ≡ 22·3 − 23 + 1 ≡ 1 − (−1) + 1 ≡ 3 (mod p). which cannot 8 2 happen if p is of the form 8k + 5.

.is also divisible by 3. In the framework of the above results. 5. which is clearly impossible. . this should not be so diﬃcult now. Unfortunately this cannot work by Fermat’s little theorem. If we search for odd numbers. This is immediate. so let us try q = 11. because brute force will take us nowhere. Fix i ∈ {1. q diﬀerent prime numbers. Thus at least a prime divisor of 22·3 − 23 + 1 is diﬀerent from 3 and is of the form 8k + 3 and so the claim is proved. then 22·3 − 23 + 1 ≡ 8 (mod 8). At ﬁrst glance. n−1} and observe that any prime factor of 22·3 − 23 + 1 is also a prime factor of 23 + 1 and thus. the conclusion would follow. Find a number n between 100 and 1997 such that n|2n + 2. and p 254 i i i i i i i i n i i . 1997 Solution. q|22p−1 + 1. 2. Example 3. it must be of the form 8k + 1 or 8k + 3. since the hard work has already been completed: we −2 = have shown that it suﬃces to have p|q 2q−1 . APMO. . The conclusion follows. for some prime number p. The ﬁrst step would be choosing n = 2p. the following problem seems trivial. with p. Because v3 (22·3 − 23 + 1) = 1. A small veriﬁcation shows that q = 3. from the ﬁrst remark. So let us try setting n = 2pq. So let us search for even numbers. 7 are not good choices. Far from being true! It is actually very tricky. it follows that if all prime divisors of 22·3 − 23 + 1 except for 3 are of the form 8k + 1. Also. In this case we ﬁnd p = 43 and so it suﬃces to show that pq|22pq−1 + 1 for q = 11 and p = 43. this result due to Sierpinski has been proved in the topic ”Look at the exponent!”). . then we will certainly fail (actually. must have p q p|22q−1 + 1 and q|22p−1 + 1. using Fermat’s little theorem. We need pq|22pq−1 + 1 and so we −2 −2 = = 1.

2005 Solution. |f (n)−n| ≤ 2004 n for all n. ii) 2f 2 (n) = n2 + g 2 (n) for all positive integers n. But as one can easily check. Because = −1. Thus there exist sequences of positive integers an . bn such that xn = an pn . ii) implies the relation √ 2b2 = a2 + 1. Using the condition i) we can ﬁnd xn such that g(xn ) = pn for all n.−2 = 1 in order to have pq|22pq−1 +1. Finally. Clearly. the last congruence shows that pn |xn pn and pn |f (xn ). It follows that 2f 2 (xn ) = x2 + p2 . Example 4. which can be rewritten as 2f 2 (xn ) ≡ x2 n n n 2 (mod pn ). Moldova TST. g : Z + → Z + functions with the properties: i) g is surjective. prove that f has inﬁnitely many ﬁxed points. It is obvious that for all n we have 2 pn = (−1) p2 −1 n 8 = −1. moreover. but we invite the reader to ﬁnd an elementary proof). Were we wrong when choosing to present the following example? It apparently has no connection with quadratic reciprocity. √ xn xn an that is n→∞ lim a2 + 1 √ n = 2. √ If. using the property |f (n) − n| ≤ 2004 n we infer n n that bn 2004 f (xn ) ≥ −1 = −1 . but let us take a closer look. Gabriel Dospinescu. Let f. f (xn ) = bn pn for all n. an 255 . all q these conditions are veriﬁed and the number 2 · 11 · 43 is a valid answer. Let pn be the sequence of prime numbers of the form 8k + 3 (the fact that there are inﬁnitely many such numbers is a trivial consequence of Dirichlet’s theorem.

. rk . . Using the Chinese remainder theorem. Let us assume ﬁrst that n ≥ 1 and consider some odd distinct primes r1 . Let us write a = 2e q1 q2 . it follows immediately that m i=1 2 pi m = i=1 (−1) p2 −1 i 8 = (−1) b2 −1 8 =1 and m j=1 qi pj m = j=1 (−1) pj −1 q −1 · i2 2 pj qi 256 = (−1) qi −1 b−1 · 2 2 b qi = b qi . . . but how to prove that they are the only ones? Example 5. diﬀerent from r1 . 1}. It characterizes numbers that are quadratic residues modulo all suﬃciently large prime numbers. not necessarily distinct. We continue with a diﬃcult classical result. . . . . rk each of them diﬀerent from q1 .The last relation immediately implies that lim an = 1. . . . pm with pi odd primes. perfect square are such numbers. The conclusion now follows. write b = p1 . One may assume that a is square-free. . . 1 ≤ i ≤ n − 1      b ≡ s (mod q ) n Now. n→∞ starting from a certain rank. Suppose that a ∈ N is not a perfect square. we can ﬁnd a positive integer b such that   b ≡ 1 (mod ri ). Therefore. Then a p = −1 for inﬁnitely many prime numbers p. Solution. qn . 1 ≤ i ≤ k      b ≡ 1 (mod 8). Of course. qn .  b ≡ 1qi . Using the quadratic reciprocity law. . that often proves very useful. such that = −1. we have an = 1 = bn that is f (pn ) = pn . . We will show that there exists a a prime p. where qi are diﬀerent odd primes and e ∈ {0. . Let s be a non p quadratic residue modulo qn .

. 2. . 2. ap−1 + ap−1 + · · · + ap−1 ≡ k 1 2 k p−1 (mod p). Example 6. a2 . Indeed. Suppose that a1 . . . a2004 are nonnegative integers such that an + an + · · · + an is a perfect square for all positive integers 1 2 2004 n. it follows 1 2 k = 1. and since ap−1 + ap−1 + · · · + ak is a perfect square. This time it is a problem almost trivial in the above framework and almost impossible to solve otherwise (we say almost because there is a beautiful. since it suﬃces to use Dirichlet’s theorem to ﬁnd inﬁnitely many primes p such p2 − 1 is odd. . 2. . . But this one is very simple. but very diﬃcult solution using analytical tools. . . } \ {r1 . . Using Fermat’s little theorem. . greater than any prime divisor of a1 a2 . And now the problem becomes that p 257 . Suppose that a1 . a2 . Because pi b ≡ 1 (mod ri ). . . . there exists i ∈ {1. Thus k is a perfect square.for all i ∈ {1. rk } and Thus. What is the minimal number of such integers that must equal 0? Gabriel Dospinescu. The only remaining case is a = 2. ak are positive integers such that an 1 + an + · · · + an is a perfect square for all n. Hence  m a = pi i=1 n m j=1 2 2  pj b qn = n m i=1 j=1 qi pj = i=1 b qi = s qn = −1. . In order to prove this. we will use the above result k and show that = 1 for all suﬃciently large prime p. that we will not present here). . . . that 8 As in other units. This is not a p diﬃcult task. . Mathlinks Contest Solution. consider a prime p. a = −1. . n}. . . we will now focus on a special case. We will show that k 2 k is a perfect square. 1 ≤ i ≤ k we also have pi ∈ {1. m} such that the claim is proved. . ak . . .

Suppose that f ∈ Z[X] is a second degree polynomial such that for any prime p there exists at least an integer n for which p|f (n). Then b2 − 4ac ≡ (2an + b)2 and so b2 − 4ac p = 1. It suﬃces of course to prove that b2 − 4ac is a perfect square. Let p be an odd prime and let p−1 f (x) = i=1 i p X i−1 . This boils down to proving that it is a quadratic residue modulo any suﬃciently large prime. Some of the properties of Legendre’s symbol can be also found in the following problem. since we must ﬁnd the greatest perfect square smaller than 2004. Solution. 2004 258 . The following example is adapted after a problem given in Saint Petersburg Olympiad. Romanian TST. (mod p) This shows that our claim is true and ﬁnishes the solution. Example 8. Here is another nice application of this idea. Example 7. a) Prove that f is divisible by X − 1 but not by (X − 1)2 if and only if p ≡ 3 (mod 4). Pick a prime number p and an integer n such that p|f (n). Let f (x) = ax2 + bx + c be this polynomial.trivial. b) Prove that if p ≡ 5 (mod 8) then f is divisible by (X − 1)2 and not by (X − 1)3 . Prove that f has rational zeros. A quick computation shows that this is 442 = 1936 and so the desired minimal number is 68.

. . The same idea of summing up in reversed order allows us to write: p−1 i i=1 p−1 2 i p p−1 = i=1 (p − i) i p p−i p p−1 2 p−1 = (−1) 2(p − i) i=1 = −(−1) f (1) (we used again the fact that f (1) = 0). The second question is much more technical. Observe that p−1 p−1 p−1 f (1) = i=1 (i2 − 3i + 2) i p = i=1 i2 i p −3 i=1 i i p (once again we use the fact that f (1) = 0). Also. Observe that the condition p ≡ 5 (mod 8) implies. even though it uses the same main idea.Solution. by a). p−1 f (1) = i=1 (i − 1) i p = i=1 i i p . The ﬁrst question is not diﬃcult at all. that f is divisible by (X − 1)2 so actually p−1 f (1) = i=1 259 i2 i p . then p−1 f (1) = i=1 i i p p−1 ≡ i=1 i= p(p − 1) ≡ 1 (mod 2) 2 and so f is divisible by X − 1 but not by (X − 1)2 . Observe that p−1 f (1) = i=1 i p =0 p−1 quadratic and non qua2 p−1 by the simple fact that there are exactly dratic residues in {1. In this case f is divisible by (X − 1)2 . Hence for p ≡ 1 (mod 4) we must also have f (1) = 0. because f (1) = 0. p − 1}. . On the other hand. if p ≡ 3 (mod 4). 2. .

Note that p−1 2 i i=1 2 i p p−1 2 p−1 2 ≡ i=1 p−1 2 i ≡ i=1 2 i= p2 − 1 ≡ 1 (mod 2). On the other hand. using the fact that is −4). then the problem will be solved. If we prove that the last quantity is a multiple of 8. But note that f (1) = 0 implies p−1 2 p−1 2 0= i=1 2i p + i=1 2i − 1 p . 260 . we obtain that its value (2i − 1)2 i=1 2i − 1 p p−1 2 ≡ i=1 2i − 1 p (mod 8). p−1 2 i=1 2i p p−3 2 =1+ i=1 p−3 2 2i p p−3 2  =1+ i=1 p−1 2 2    p−1 −i  2   p =1+ i=1 p−1 2 2i + 1 p = i=1 2i − 1 p .Let us break this sum into two pieces and treat each of them independently. Therefore i=1 2i − 1 p = 0 and the problem is ﬁnally solved. Let us deal with p−1 2 p−1 2 (2i)2 i=1 2i p =4 2 p i2 i=1 i p . 8 so (2i)2 i=1 2i p 2 p ≡ ± (mod 8) = (−1) p2 −1 8 (actually. p−1 2 . Also.

Therefore the only solution is n = 1. call it p.Finally. Prove that 2003 does not n divide any term of the sequence. −(±1) = 3 3 p−1 (−1) 2 = ±1. ii) there is a positive integer m such that p|m2 − m + 25. 261 . hence n is odd. Solution. Obviously. Find all positive integers n such that 2n − 1|3n − 1. Prove that for any odd prime p. Mathlinks Contest 4. Problems for training 1. Let p be a prime of the form 4k + 1. Let x1 = 7 and let xn+1 = 2x2 − 1. Suppose that n > 1 is a solution. Because any odd prime diﬀerent from 3 is of one of the forms 12k±1. it follows that 2n − 1 has at least a prime divisor 3 of the form 12k ± 5. We will prove that n = 1 is the only solution to the problem. Compute p−1 k=1 k2 2k 2 −2 p p . Polish Olympiad 3. Valentin Vornicu. a diﬃcult problem. 3 p ±2 On the other hand = = −(±1). Therefore. Prove that the following statements are equivalent: i) there is a positive integer n such that p|n2 − n + 3. Let p be a prime number. the smallest positive quadratic √ non residue modulo p is smaller than 1 + p. 12k±5 and since 2n − 1 ≡ 7 (mod 12). Consequently. we ﬁnally obtain = (−1) 2 . which is the desired contradiction. Example 9. we must have = 1 and p p−1 p using the quadratic reciprocity law. 2. Then 2n − 1 cannot be a multiple of 3. 2n ≡ 8 (mod 12).

then ab (mod p) ∈ A. b ∈ B. Let a. . b are in the same set among A. Gazeta Matematica 6. B. none of which can be written in the form (ax2 + bxy + cy 2 )z for some integers x. What is the number of solutions to the equation a2 + b2 = 1 in Zp × Zp . . Prove that: p−1 i=1 p−1 ai2 + bi p =− a p . Find all such subsets A. Let be integers relatively prime with an odd prime p. . . India 8. What about the equation a2 − b2 = 1? 7. iii) If a ∈ A. . B. p − 1}. Prove that p|b2 − 4ac. ii) If a. then ab ∈ B. b. 2. 2. B are two diﬀerent non empty subsets of {1. Suppose that for a certain prime p. Gabriel Dospinescu 9. Prove that the number 3n + 2 does not have prime divisors of the form 24k + 13. c be positive integers such that b2 − 4ac is not a perfect square. IMO Shortlist 262 . . Prove that for any n > 1 there are n consecutive positive integers. Laurentiu Panaitopol. Suppose that p is an odd prime and that A. y and some positive integer z.Korea TST 2000 5. Compute 11. . f (k) . where f is a polynomial with integral p k=1 coeﬃcients and p is an odd prime. 10. . p − 1} for which i) A ∪ B = {1. the values the polynomial with integral coeﬃcients f (x) = ax2 + bx + c takes at 2p − 1 consecutive integers are perfect squares.

12. Here φ is Euler’s totient function. Prove that gcd(m. Prove that the greatest prime divisor q of 2p − 1 satisﬁes the inequality 2q > (6p)p . Suppose that φ(5m − 1) = 5n − 1 for a pair (m. Gabriel Dospinescu 263 . n) of positive integers. Taiwan TST 13. n) > 1. Let p be a prime of the form 4k+1 such that p2 |2p −2.

g : [a. 3) For any integrable functions f. one just have to apply the corresponding inequalk k ities for the numbers f a + (b − a) . H¨lder. a 2) For any integrable functions f. with n n 264 . let us state some properties of integrals that we will use here. b] → R and any real numbers α. First. but sometimes ﬁnding elementary solutions for them is a real challenge. That is why there are plenty of inequalities which can be solved with integrals. . . They seem elementary. β we have b b b (αf (x)+βg(x))dx = α a a f (x)dx+β a g(x) (linearity of integrals). the hard part is to ﬁnd the integral that hides after the elementary form of the inequality (and to be sincere. g a + (b − a) . Young have corresponding ino tegral inequalities. Instead. Jensen. Also. Minkowski. a comparable one. b] → R we have b f 2 (x)dx ≥ 0.SOLVING ELEMENTARY INEQUALITIES WITH INTEGRALS Why are integral pertinent for solving inequalities? Well. when we say integral. . which are derived immediately from the algebraic inequalities (indeed. 1) For any integrable function f : [a. And area is a measurable concept. g : [a. the idea of using integrals to solve elementary inequalities is practically inexistent in Olympiad books). the well-known elementary inequalities of Cauchy-Schwarz. b] → R such that f ≤ g we have b b f (x)dx ≤ a a g(x)dx (monotony for integrals). Chebyshev. some of them with a completely elementary statement. we say in fact area. there are beautiful and short solutions using integrals. Of course.

. To convince yourself of the strength of the integral. This magniﬁcent proof was found by H. a2 . 265 . It seems at ﬁrst glance that this is not a very intricate and diﬃcult theory. n The reader will take a look at the glossary if he doesn’t manage to state them. after all. an absolute classic whose solution by induction can be a real nightmare. . B = n a1 a2 . . Prove that for any a1 . since each integral is nonnegative. an . Truly wonderful. 2. . . we can ﬁnd an index k ∈ {1. . Then it is immediate to see that A 1 −1= G n k i=1 G ai 1 1 − t G dt + 1 n n ai i=k+1 G 1 1 − G t dt and the last quantity is clearly nonnegative. . . .k ∈ {1. Example 1. integrals are nice! This is also conﬁrmed by the following problem. n} and to use the fact that b a f (x)dx = lim b−a n→∞ n n f k=1 a+ k (b − a) . . isn’t it? So. . an ≥ 0 we have the inequality √ a1 + a2 + · · · + an ≥ n a1 a2 . . . 2. Let us suppose that a1 ≤ a2 ≤ · · · ≤ an and let A= √ a1 + a2 + · · · + an . an . Totally false! We will see how strong is this theory of integration and especially how hard it is to look beneath the elementary surface of a problem. n Solution. Alzer and published in the American Mathematical Monthly. n − 1} such that ak ≤ G ≤ ak+1 . . . . n Of course. take a look at the following beautiful proof of the AM-GM inequality using integrals.

j=1 0 1 ai aj ti+j−1 dt ≥ 0. because the formula n 2 n ai xi i=1 = i. we will see how easy is this problem if we manage to handle integrals and especially to see from where they come. . This form suggests us that we should use the ﬁrst property. a2 .j=1 ai aj ti+j−1  dt ≥ 0. .j=1 ai aj ti+j−1 dt. the hard part was translating the inequality. an the following inequality holds: n n i=1 j=1 ai aj ≥ 0. using the linearity of the integrals. Prove that for any real numbers a1 . After that. n i.j=1 266 ai aj xi xj . or. . i+j Poland Mathematical Olympiad Solution.j=1 ai aj ≥0 i+j is equivalent to n i. What follows are just formalities. The essential suggestion is the observation that ai aj = i+j 1 ai aj ti+j−1 dt. we will decide what is better to do.Example 2. to   1 n  0 i. So. 0 And now the problem is solved. . Now. that is we should ﬁnd an integrable function f such that n f 2 (t) = i. This isn’t hard.

. Example 3.solves the task. 1 + t + · · · + tn . And yes. this is a consequence of the Power Mean Inequality for integral functions. It is clear that for t > 1 we have xn = and for t < 1 we have xn = 1 1−t t 1 t−1 t un du 1 un du. using integrals it becomes trivial. The following problem has a long and quite complicated proof by induction. n+1 Walther Janous. b] → R. Crux Mathematicorum Solution. 1 This is how the inequality to be proved reduces to the more general inequality b k b f k (x)dx a k+1 f k+1 (x)dx a b−a ≤ b−a for all k ≥ 1 and any nonnegative integrable function f : [a. 267 . . We just have to take n f (x) = i=1 ai xi− 2 . Yet. 1 We continue the series of direct applications of classical integral inequalities with a problem proposed by Walther Janous and which may also put serious problems if not attacked appropriately. Let t ≥ 0 and the sequence (xn )n≥1 deﬁned by xn = Prove that x1 ≤ √ x2 ≤ √ 3 x3 ≤ √ 4 x4 ≤ .

1995 Solution. xk )n − k). Let x1 . The last relations can be immediately translated with integrals in the form x x x (y − x) y tm+n−2 dt ≥ y tm−1 dt y tn−1 dt. . x2 . Applying AM-GM inequality. n positive real numbers such that n ≤ km. Example 5. Prove that m(xn + xn + · · · + xn − k) ≥ n(xm xm . xk be positive real numbers and m. . . . xk . Austrian-Polish Competition . . Prove that for any positive real numbers x. . y and any positive integers m. . xm − 1). A nice blending of arithmetic and geometric inequality as well as integral calculus allows us to give a beautiful short proof of the following inequality. . proposed by Poland Solution. . 268 k (x1 x2 .Example 4. . And this follows from the integral form of Chebyshev inequality. n (n − 1)(m − 1)(xm+n + y m+n ) + (m + n − 1)(xm y n + xn y m ) ≥ mn(xm+n−1 y + y m+n−1 x). . We transform the inequality as follows: mn(x − y)(xm+n−1 − y m+n−1 ) ≥ (m + n − 1)(xm − y m )(xn − y n ) ⇔ xm+n−1 − y m+n−1 xm − y m xn − y n ≥ · (m + n − 1)(x − y) m(x − y) n(x − y) (we have assumed that x > y). 1 2 1 2 k k IMO Shortlist 1985. we ﬁnd that m(xn + · · · + xn − k) ≥ m(k 1 k Let P = √ k x1 x2 .

b. c such that a + b + c = 1 we have (ab + bc + ca) b c a + + b2 + b c2 + c a2 + a 3 ≥ . 4 Gabriel Dospinescu Solution. As in the previous problem. we infer that a b c + + ≥ 2 2 (x + b) (x + c) (x + a)2 a b c + + x+b x+c a+x 2 . Now. this isn’t diﬃcult. Well. 0 We have seen a rapid but diﬃcult proof for the following problem. Prove that for any positive real numbers a. using the Cauchy-Schwarz inequality. which is the same as Pn − 1 P mk − 1 ≥ . the problem originated by playing around with integral inequalities and the following solution will show how one can create diﬃcult problems starting from trivial ones. since it is just 1 0 a b c + + 2 2 (x + b) (x + c) (x + a)2 dx. we minor with 1 . Using again the same inequality. n mk This follows immediately from the fact that Px − 1 = x ln P 1 ext ln P dt. Consequently. the most important aspect a b c is to translate the expression 2 + 2 + 2 in the integral b +b c +c a +a language. x + ab + bc + ca a b c 1 + + ≥ 2 2 2 (x + b) (x + c) (x + a) (x + ab + bc + ca)2 269 b c a + + x+b x+c a+x .We have to prove that mkP n − mk ≥ nP mk − n. Example 6. Fortunately. using the Cauchy-Schwarz inequality.

Let us observe that n n i=1 j=1 ai aj = i+j 1 n i=1  ai  n j=1  aj  = i+j n 1 ai i=1 0 xi f (x)dx n 1 = 0 xf (x) i=1 ai xi−1 dx = 0 xf 2 (x)dx. . Consider the function f : R → R. yet here it is much more diﬃcult to see the relation with integral calculus. ∞) which verify n n i=1 j=1 1 − ai aj ≥ 0. .and we can integrate this to ﬁnd that a b c 1 + + ≥ . all we have to do is to notice that 4 ab + bc + ca + 1 ≤ . Example 7. a2 . Let n ≥ 2 and S the set of the sequences (a1 . an ) ⊂ [0. another question for the interested reader: can we prove the general case (solved in Cauchy Schwarz’s inequality topic) using integral calculus? It seems a diﬃcult problem. Gabriel Dospinescu Solution. . There is an important similarity between the following problem and example 2. b2 + b c2 + c a2 + a (ab + bc + ca)(ab + bc + ca + 1) Now. f (x) = a1 + a2 x + · · · + an xn−1 . 1+j n n Find the maximum value of the expression i=1 j=1 ai + aj . 3 Now. over all i+j sequences from S. 270 . .

since we must ﬁnd the maximal value of 1 2 0 (x + x2 + · · · + xn )f (x)dx 1 where xf 2 (x)dx ≤ M. we have 1 2 (x + x + · · · + x )f (x)dx 0 1 2 2 n = 0 1 xf 2 (x) 1 x(1 + x + · · · + xn−1 )2 dx = 0 xf 2 (x)dx 0 n n (1 + x + · · · + xn−1 )2 dx ≤ M 2 .j≤n 1 1 . the problem becomes easy. Indeed.So. This shows that 271 . if we denote M = 1≤i.j≤n an = 1. attained for a1 = a2 = · · · = i+j 1≤i. 0 This was the hard part: translating the properties of the sequences in S and also the conclusion. Now. we infer that i+j xf 2 (x)dx ≤ M. ai + aj ≤ 2M and now the conclusion easily i+j i=1 j=1 1 follows: the maximal value is 2 . 0 Well. we have the identity n n i=1 j=1 ai + aj =2 i+j =2 a1 an a1 an + ··· + + ··· + + ··· + 2 n+1 n+1 2n 1 (x + x2 + · · · + xn )f (x)dx. 0 On the other hand. Cauchy-Schwarz inequality for integrals is the way to proceed.

to present a new method of treating inequalities involving fractions. here is another application of this idea. z = tc− 3 and integrate the inequality as t ranges between 0 and 1. c the following inequality holds: 1 1 1 3 1 1 1 + + + ≥ + + 3a 3b 3c a + b + c 2a + b 2b + a 2b + c + 1 1 1 + + . the reader has noticed that this is stronger than Popoviciu’s inequality. Example 9. b. b. Some relevant examples will be treated revealing that bunching could be a great pain for the reader wanting to use it. c the following inequality holds: 1 1 1 + + +2 3a 3b 3c ≥3 1 1 1 + + 2a + b 2b + c 2c + a . Of course. And surprise. let us write Schur’s inequality in the form: x3 + y 3 + z 3 + 3xyz ≥ x2 y + y 2 x + y 2 z + z 2 y + z 2 x + x2 z where x = ta− 3 . And what if we say that this is Schur’s inequality revisited? Indeed. Gabriel Dospinescu 272 1 1 1 1 1 1 + + a + 2b b + 2c c + 2a . In the same category. Example 8.Two more words about fractions.. This is why we designed this topic. It is time to say it again. Prove that for any positive real numbers a. 2c + b 2c + a 2a + c Gabriel Dospinescu Solution. since what we get is exactly the desired inequality. Prove that for any positive real numbers a. y = tb− 3 . so it seems that classical methods will have no chances. We have already said that bunching is a mathematical crime..

we see that the problem asks in fact to prove that x3 + y 3 + z 3 + 2(x2 y + y 2 z + z 2 x) ≥ 3(xy 2 + yz 2 + zx2 ) for any positive real numbers x. n. we haven’t tried). Therefore. z = x + n for some nonnegative real numbers m. . If the previous problem could be solved using bunching (or not? Anyway. Let a1 ≥ a2 ≥ · · · ≥ an+1 = 0 and let b1 .. it suﬃces to prove that (n − m)3 + m3 ≥ (n − m)m2 . If only this was easy to see. In fact. bn ∈ [0. but in the sequel we seemed to neglect the last concept. this one is surely impossible to solve in this manner. . Example 10. we said that there is a deep relation between integrals and areas. We ask the reader to accept our apologizes and bring to their attention two mathematical gems. . 273 . these problems are discrete forms of Young and Steﬀensen inequalities for integrals. Simple computations show that the inequality is equivalent to 2x(m2 − mn + n2 ) + (n − m)3 + m3 ≥ (n − m)m2 . which is the same as (via the substitution t = t ≥ −1. which is immediate. z. .. b2 . With the experience from the previous problem. Prove that if n n−m 3 ) t + 1 ≥ t for all m k= i=1 bi + 1. y.Solution. in which they will surely have the occasion to play around with areas. z) and write y = x + m. y. Let us assume that x = min(x. Starting this topic. 1].

a2 . So. ai  (where bj = 0). . ai ) (we need n rectangles of k heights a1 . this is evident. ai .  Sn . let us consider a coordinate system XOY and let us draw the rectangles R1 .  j=1 b j . 0. 0. Thus. where the vertices Si are of     . . areas!!!) could lead to a neat solution. S2 . . looking in a picture. where b k= a f (x)dx.   the points  j=1 0 bj . . b] → R such that f is decreasing and 0 ≤ g ≤ 1 we have a+k b f (x)dx ≥ a a f (x)g(x)dx.then n k ai bi ≤ i=1 i=1 ai . Saint Petersburg Olympiad. . . a2 b2 . (i. . (i − 1. . . so that to view i=1 i−1 i i−1 i ai as a sum of areas) and the rectangles S1 . 0). . Sn can be covered with the rectangles R1 . by looking again at the picture. probably an argument using areas (this is how we avoid integrals and argue with their discrete forms. Rn such that the vertices of Ri are the points (i − 1. an and a1 b1 . 1996 Solution. . The very experienced reader has already seen a resemblance with Steﬀensen’s inequality: for any continuous functions f. Intuitively. . Let us make it rigorous. . .  j=1 bj . We have made this choice because we need two sets of pair j=1 wise disjoint rectangles with the same heights and areas a1 . a2 . an bn so that we can compare the areas of the unions of the rectangles in the two sets. . . 0). R2 . (i. g : [a. we ﬁnd immediately what we have to show: that the set of rectangles S1 . . . . . Rk+1 . . Since 274 . ai ). S2 . an and weights 1. . .. . . . So.  j=1 bj . R2 .

. Sn is j=1 bj < k + 1 (and the weight of R1 . it suﬃces to take the given sequence (xn )n≥0 as the one to one increasing function in Young’s inequality and the sequence (yn )n≥0 m m as the inverse of f . . consider the vertical rectangles with weight 1 and heights x0 . . the weight for S1 . Let (xn )n≥0 be an increasing sequence of nonnegative integers such that for all k ∈ N the number of indices i ∈ N for which xi ≤ k is yk < ∞. . R2 . But if we consider a horizontal line y = p and an index r such that ar ≥ p > ar+1 . which is at least b1 + b2 + · · · + bp . S2 . B] and any a ∈ (0. . . Example 11. And now the second problem. y1 . S2 . probably again a geometrical solution is hiding behind some rectangles. experienced reader will see immediately a similarity with Young’s inequality: for any strictly increasing one to one map f : [0. Again. Balkan Mathematical Olympiad. it is enough to prove this for any horizontal line. Prove that for any m. x1 . Sn . 1999 Solution. . A). . Indeed. . Rk+1 is p. . . yn . Then in a similar way one can prove that the set of these 275 . .n the weight of the union of S1 . n ∈ N we have the inequality m n xi + i=0 j=0 yj ≥ (m + 1)(n + 1). . R2 . b ∈ (0. then the corresponding weight for the set R1 . . . . . Indeed. And the problem is solved. Rk+1 is k + 1). B) we have the inequality a b f (x)dx + 0 0 f −1 (x)dx ≥ ab. . . A] → [0. . . xm and the rectangles with weight 1 and heights y0 . given this time in a Balkan Mathematical Olympiad. . . . Just view i=0 xi and j=0 yj as the corresponding integrals and the similarity will be obvious. Thus.

Let us deﬁne the functions fi . Now. 0. Thus. It will be diﬃcult to solve the following beautiful problems using integrals. b2 .j≤n 0 fi (x)gj (x)dx = 1≤i. Poland.j≤n ∞ min(ai . ∞) → R. ∞ g(x) = i=1 gi (x). since the idea is very well hidden. x ∈ [0. t ∈ [0. aj ) and 0 g 2 (x)dx = 1≤i. a2 . gi : [0. . there is such a solution and it is more than beautiful. . Prove that for any a1 .j≤n |ai − bj |. bj ). bn ≥ 0 the following inequality holds (|ai − aj | + |bi − bj |) ≤ 1≤i<j≤n 1≤i. . t > ai and gi (x) = 1. fi (x) = 1. Yet. let us deﬁne n n f (x) = i=1 fi (x). the sum of their areas is at least the are of this rectangle. . .j≤n min(ai .rectangles covers the rectangle of sides m + 1 and n + 1. x > bi . A similar computation shows that ∞ f 2 (x)dx = 0 1≤i. bi ].j≤n 276 min(bi . . an . Example 12. 0. Also. ai ]. .j≤n  fi (x)gj (x) dx f (x)g(x)dx = 0 ∞ 0 = 1≤i. . . bj ). let us compute 0 ∞ f (x)g(x)dx. b1 . We see that ∞   1≤i. 1999 Solution.

. .j≤n |bi − bj | ≤ 2 1≤i.j≤n 1≤i. x2 . Mathlinks Contest 277 .j≤n min(ai . this is easy to ﬁnd for the author of the problem. . bj ) ≥ 2 1≤i. . . the problem is solved. xn be real numbers such that n ai xi = 0. but in a contest things change! Example 13. . a2 . b) Prove that we have equality in the above inequality if and only if there exist a partition A1 . k} we have j∈Ai xj = 0 and aj1 = aj2 if j1 . Let a1 . Of course.. remember that 2 min(x. A2 . n} such that for all i ∈ {1. . here is a diﬃcult problem.. . an > 0 and let x1 . . . aj ) + 1≤i. . Now. i=1 a) Prove that the inequality 1≤i<j≤n xi xj |ai − aj | ≤ 0 holds. Gabriel Dospinescu. y) = x + y − |x − y| and the last inequality becomes |ai − aj | + 1≤i.j≤n |ai − bj | and since |ai − aj | = 2 1≤i. Using this idea. whose elementary solution is awful and which has a 3-lines solution using the above idea. . .Since ∞ ∞ ∞ ∞ f 2 (x)dx+ 0 0 g 2 (x)dx = 0 (f 2 (x)+g 2 (x))dx ≥ 2 0 f (x)g(x)dx.j≤n 1≤i<j≤n |ai − aj |. . 2. bj ). 2. . we ﬁnd that min(ai . . .j≤n 1≤i. . . Ak of the set {1. .j≤n min(bi . j2 ∈ Ai .

.j≤n Since min(ai . . 1≤i. let us compute ∞ f= i=1 xi λ[0. . . Ak is a partition of the set {1.ai ] (x)λ[0. b2 .ai ] . . we conclude that xi xj |ai − aj | ≤ 0. let b1 . Let λA be the characteristic function of the set A. bk the distinct numbers that appear among a1 . Let us consider the function n f : [0. . . aj ). . . . an > 0 and let Ai = {j ∈ {1. A2 .j≤n xi xj min(ai .bi ] = 0  i=1 j∈Ai 278 . a2 . n}| aj = bi }. . ∞ f 2 (x)dx = 0 1≤i. n} and we also have  k  xj  λ[0. Then A1 . 1≤i<j≤n Let us suppose that we have equality. . Now. We ﬁnd that ∞ f 2 (x)dx = 0 0 and so f (x) = 0 almost anywhere. Now.j≤n xi xj 0 λ[0. . .aa ] (x)dx = 1≤i. 2. aj ) = and n n ai + aj − |ai − aj | 2 xi xj (ai + aj ) = 2 1≤i.Solution.j≤n i=1 xi i=1 ai xi = 0. Hence xi xj min(ai . . . aj ) ≥ 0. . 2. . ∞) → R. . .

. . aj bi ). And since we have proved the nice inequality xi xj min(ai . . bi ) − 1 and xi = Solution. we can simply put ri = 0). 279 . aj bi ) − bi bj = min(ai bj . a1 . rj ) = bi bj min aj ai − 1. Prove the following inequality min(ai aj . bj = 0. xn . rj ). . a2 . aj bi ) − min(ai aj . . an > 0 let’s make a step further and give the magniﬁcent proof found by Ravi B. The ﬁrst one is when at least one of the two inequalities ai ≥ bi and aj ≥ bj becomes an equality. by any chance. which leaves us with only two cases.j≤n for any numbers x1 .almost anywhere. This case is trivial. k}. x2 . . 1999 max(ai .j≤n 1≤i. . aj bi ) − min(ai aj . . −1 bi bj = bi bj min ai aj . bi bj ) ≤ 1≤i. . Zbaganu. The crucial observation is the following identity: min(ai bj . bi bj −1 = min(ai bj . solution based on this result: Example 14. Then xi xj min(ri .j≤n min(ai bj . aj ) ≥ 0 1≤i. bi bj ) = xi xj min(ri . 2. . USAMO. bi bj ). . bi ) sgn(ai − bi ) (if. (see mathlinks site) for one of the most diﬃcult inequalities ever given in a contest. from where we easily conclude that xj = 0 for all i ∈ {1. . one of ai . Let us deﬁne the numbers ri = min(ai . i∈Ai The conclusion follows. but let us remark that actually we may assume that ai ≥ bi and aj ≥ bj . so let us assume the contrary. Proving this relation can be achieved by distinguishing 4 cases. G.

which is a consequence of this last hard inequality.j≤n |xi − xj |. We obtain: min(xi .j≤n n 1≤i. xn some positive real numbers such that |1 − xi xj | = 1≤i. Taking into account that |1 − xi xj | = 1≤i. the last inequality being nothing else than the main ingredient of the preceding problem. Consider bi = 1 in the inequality from example 14. use the formula min(u. v) = above inequality in the form n 2n i=1 xi − 1≤i. since otherwise the problem is really extremely hard.j xi xj min(ri .j≤n min(1. xj ) ≥ 1≤i. . Prove that i=1 xi = n.j≤n 1≤i.Now.j≤n 1≤i. Consider this a hint and try to solve it. Finally.j≤n |1 − xi xj |. . . rj ) ≥ 0. Let x1 . xi xj ). Example 15. u + v − |u − v| and rewrite the 2 n 2 Now. .j≤n |xi − xj |. Gabriel Dospinescu Solution. bi bj ) = i.j≤n |xi − xj | ≥ n + i=1 2 xi − 1≤i. we ﬁnally obtain n n 2 2n i=1 xi ≥ n + i=1 280 2 xi .j≤n min(ai aj . here is a very funny problem.j≤n 1≤i. we can write min(ai bj . aj bi ) − 1≤i. . x2 .

Journal 4. Let a continuous and monotonically increasing function f : [0. 1] → R such that f (0) = 0 and f (1) = 1. Prove the following inequality an + bn + 2 a+b 2 n ≥2· an + an−1 b + · · · + abn−1 + bn n+1 for any positive integer n and any nonnegative real numbers a. 10 Sankt Petersburg. for a ∈ (−1. Harris Kwong. 281 .which can be rewritten as n 2 xi − n i=1 ≤0 Therefore n xi = n. Folklore 3. b. 1991 5. Show that for all a. Prove that 9 10 f k=1 k 10 + k=1 f −1 k 10 ≤ 99 . 0) we have the reversed inequality. b ∈ N∗ ln bn + 1 an + 1 < 1 1 1 b + + ··· + < ln . an + 1 an + 2 bn a (n + 1)a+1 − 1 a+1 2. College Math. Prove that for any real number x n n n k=0 x2k ≥ (n + 1) k=1 x2k−1 . i=1 Problems for practice 1. Also. Prove that for any a > 0 and any positive integer n the inequality 1a + 2a + · · · + na < holds.

Prove that k−j j−1 αi αi+1 1≤i≤n 1≤j≤k ≥ k nk−2 n k−1 αi i=1 . aj ) ≥ i=1 ai + 2n − i=1 ai . x+1 x+2 x + 2n Komal 8. . α2 . Prove that the function f : [0. ab + bc + ca Marius and Sorin Radulescu 282 . . Prove that for any real numbers a1 . Komal 9. a2 . Crux Mathematicorum 11. . . . . . .j≤n min(ai . For all positive real number x and all positive integer n we have: 2n 0 x 2n 2n 2n 1 2 2n − + − ··· + > 0. is bounded.j=1 ij ai aj ≥ i+j−1 n 2 ai i=1 . c such that a + b = c = 1 we have: 1+ 1 a b 1+ 1 b c 1+ 1 c a ≥1+ 1 . α1 . Let k ∈ N. .Mihai Onucu Drambe 6. b. . Prove that if a1 ≤ a2 ≤ · · · ≤ an ≤ 2a1 the the following inequality holds n 2 n 2 an 1≤i. Gabriel Dospinescu 7. an n i. αn+1 = α1 . Hassan A. Shah Ali. Prove that for any positive real numbers a. 1) → R deﬁned by f (x) = log2 (1 − x) + x + x2 + x4 + x8 + . 10.

. . G the corresponding means 2 A A for 1 − xi . a) = 0 tdt . . bn ≥ 0 the inequality holds   1≤i. a2 . . Prove that for any positive integer n we have n 1> k=1 2k 1 ϕ(k) ln k > 1 − n. . p2 . i a) Let us denote ∞ I(x. pn some positive numbers which add up to 1 and x1 . Prove that for all a1 . b1 . i=1 Adapted after an IMC 2000 problem 14. Prove that ≥ . xn some positive real numbers. k 2 2 −1 Gabriel Dospinescu 15. . aj )  1≤i. . . x2 . G G Oral examination ENS 283 . . bj ) ≥  13. . . Prove that for any x1 ≥ x2 ≥ · · · ≥ xn > 0 we have n i=1 x2 + x2 + · · · + x2 n i i+1 ≤π i n xi . Let also n n A= i=1 ai xi and G = i=1 xpi . A). (1 + t)(x + at)2 Prove that A ln = G n pi (xi − A)2 I(xi . an . . . bj ) . Let p1 .j≤n  min(ai . .j≤n   1≤i. . Let ϕ be Euler’s totient function. .j≤n  min(ai . Don Zagier min(bi . where ϕ(1) = 1. i=1 Deduce the arithmetic-geometric inequality. 1 b) Suppose that xi ≤ and deﬁne A .12. b2 .

. Prove that for any a1 . i i=1 Hilbert’s inequality 284 . xn such that n i=1 n 1 = . Prove that if xi+1 − xi ≤ h for all 1 ≤ i ≤ 2n then 1−h < 2 2n x2i (x2i+1 − x2i−1 ) < i=1 1+h . xi + xj 2 Gabriel Dospinescu 17. . 1996 19. a2 . Prove that we can ﬁnd a constant c such that for any x ≥ 1 and any positive integer n we have n k=1 1≤i. an ≥ 0 we have the following inequality 1≤i.j≤n (k 2 1 c kx − ≤ . Prove that for any positive real numbers x1 . . 2 + x) 2 x IMC. . Let 0 = x1 < · · · < x2n+1 = 1 some real numbers. 2 Turkey TST. 1996 18. .16. 1 + xi 2 we have the inequality n2 1 ≥ . . . x2 . .j≤n ai aj ≤π i+j n a2 .