182 Chapter 3. Look at the Ezponent 3.B.

An Introduction to p-adic Numbers 183

Lemma 3.B.39 shows that B, = 0 for all odd numbers n > 1 {this can also hence it is enough to show that we can find k such that SL’;}':) — Bop € Zp.
be deduced from the generating function, by checking that the function z But
oy + § is even). We saw during the proof of theorem 3.B.41 that Bernoulli 2n-1
San(p") __1 W+ 1\ B ken—g)
numbers satisfy the crucial identity
> B’"*zmlg i B
—qh 1 = /n+1
= 1"+ 2"no 4
Splk) L,
4 1=
(k—1) n+1§( ) )Blk =i is certainly in Z, for k large enough, as each of the terms in the sum is in Z,.
This proves the claim.
This identity will play a crucial role in the proof of the following beautiful Finally, a standard argument shows that S,(p) = 0 (mod p) unless p—1
result, which considerably refines theorem 3.B.41. divides 7, in which case S,(p) = —1 (mod p). Combining this with the result
of the previous paragraph, we deduce that By, € Z, unless p — 1{2n, in which
Theorem 3.B.46. (von Staudi-Clausen theorem) For alln > 1 we have case By, + % € Zp- All in all, this shows that the rational number By, +
b a-1f2n % belongs to Z, for all primes p and so it is an integer. The result
By + Z i& 7. follows. [}
oot P
Proof. Let p be any prime. We need the following elernentary result: The following classical result is considerably more difficult to prove. The
proof is actually rather mysterious, though elementary. It is highly related to
Lemma 3.B.47. Ifn is even, then for all k > 1 we have So(p**1) = p- S, (p*)
the existence of p-adic analogues of the Riemann zeta function, though this
(mod p*1y. may not be apparent at all...
Proof. This is a simple application of the binomial formula:
‘Theorem 3.B.48. (Kummer’s congruences) Let m,n be positive integers and
pP=1p-1 let p be a prime such that p — 1 does not divide n, but p — 1 divides m — 1.
Sa(p**h) = G+ip" Then Bn — Ba ¢ p7,.
J=0 =0
el pol
Proof. It suffices to prove that ‘—25 = z’;;’:l‘ {mod p) whenever p— 1 does not
= ("l - 7Y divide k& > 1. Note that it is not even clear that 5;} € Zy, but this will be the
=0 =0
case (of course, this is clear if & is odd, as then B = 1}. Let us fix an integer
=pSa(®) + 50— VP
50 (0
n
1< a<p~—1such that a is a primitive root moduloe p. The following lemma
i acial.
= pSa(p*) (mod pk“),
a primitive oot mod p. If p— 1 does not divide n, then ¢” # 1 and
Note that if p is odd, then the hypothesis that n is even is useless. (]
=2 (—1)m
Next, we claim that S—"”z‘f—p—) — Byy € Zy for all primes p and all n. Note s =30
=
=L
that the previous lemma implies that % — Sg';’# €Z,forall k> 1,
184 Chapter 8. Look at the Erponent 8.B. An Introduction to p-adic Numbers 185

Lemma 3.B.49. There ezist b, € Z; such that (as formal series) for all n > 0.
Itis now easy to conclude. Recall that p — 1 does not divide &, so p does
1= Sba(eX - not divide a* ~ 1. The previous relatwn and the fact that bj,s;. € Z, for
n20 all j,n show that Uy € Z, and so 4 € Z,. The same argument shows that

Proof. By changing the variable Y = e — 1, this is equivalent to IZ‘%;)K; € Zy. Also, Fermat’s little theorem yields sjn= 5jn4p-1 (modp) for
all j > 0and all n > 1. Hence Up = Unip-g (modp) for all n > 1. As

e
1+ Y)“ -1 Y T =MR a**P=1= of (mod p), this congruence is equivalent to

B, B, k+p—~1
(a L N
) % o (a 1)—hk
s (mod p).
The binomial formula and the fact that ged(a,p) = 1 yield the existence of
9 € YZy[Y] such that (1+7Y)* — 1 =a¥ (14 ¢(Y)). We then have
The deslrcd result follows from this congruence and the fact that p does not
a 11 1 1 divide of — 1. a
GFyF=Ti~¥ ¥ (1“(_(](1’) - 1) =2
nzl el Remark 3.B.50. Here is the real meaning of this proof. As we have already
The result follows, as g € YZ[Y]. 0 said, the crucial fact is that fo(X) = T — % € LollX], say folX) =
w0 bnX™. This allows us to define a measure yz, on Zy {i.e. continuous
Let use denote Uy = (a™ 1)+ %1 and observe that by definition of (Bp)n linear form on the space of continuous functions g : Zy — Qp) by
we have 1 1
a
i S PP /z,,g”“ = %})un(g)bn
To fully exploit lemma 3.B.49, let us denote
for all continuous functions g : Z, — Q. Here a,(g) is the nth Mahler coeffi-
i Z( 1) J( ) cient of g and the series converges by Mahler’s theorem (as limyo an(g) =
sn,k=k!-[x
and by € Zp). Since b, € Z, for all n, Mahler’s theorem shows that

so that Xk

AREVEDPENES % ( L gua) > mip vy(4(2))

k>0
for all continuous maps g.
Replacing these relations in lemma 3.B.49 and identifying coefficients yields®
Note that s, = an(z), so the equality Uy, = Ej>0 b;8;n established
during the proof can be written as -

“SNote that there are no convergence issues, as s,y = 0 for n. > k. Unt1 :/ ETo
z,
186 Chapter 3. Look at the Ezponent 3.B. An Introduction to p-adic Numbers 187

Since vp(z"™ — 2™P~1) > 1 for all x € Z, and n > 1, the previous paragraph Tt is thus enough to prove that (hfil)B"H,,k”kfi P’ forall 3<k<n+1.
shows that U, = Upip—1 (mod p) for all n > 1. As pBnyy—k € Ly (by the von Staudt-Clausen theorem) and (1) € Zp. it is
This new interpretation of the proof of Kummer’s congruence yields other enough to check that # € Z,. This is easy and left to the reader {here one
nontrivial congruences. Assume for instance that m = n (mod p¥ (p—1)) and crucially uses the hypothesis p > 3). ]
m,n > N. By Euler’s theorem we have vy(a™ — 2"} > N+ 1 forall z € Z,
and we deduce that Uy, = U, (mod p¥*?). This observation is the beginning Applying the previous lemma to n = ©(p*) — 2 and noting that B, -1=0
of the construction of the p-adic zeta function of Kubota and Leopoldt. (as n is even), we obtain
We end this addendum with an application of the previous results to =1
harmonic numbers, It is standard that for any prime p > 3 we have DRI
k=1
= pB ey 5 (mod ).
-1
L

It remains to use Kummer’s congruence to obtain Bypry2 = %Bp‘g

5

5
=0 (modp?) and Z % =0 (mod p). (mod p)
k=1 and finally
= 2p
The following result considerably refines these congruences. >
= 2 5=FBs
3 'p—3
(modg?).
P
Theorem 3.B.51. For oll primes p > 3 we have
One could apply a similar method for the second congruence, but we
prefer to deduce it from the first one. Note that
P*117P"1 1 1 -1 ;

Proof. Euler’s theorem yields

QZ%‘E(EJ’H):
k=1 k=1 L iew
Using this identity and the congruence we have just proved, it is enough to
S:—l p—1 ) prove that
Zk"’(’%) 2 (mod p?).
ko) k=1 s 1 1
kz; (hk(p—k) + P) =0 (mod p%)
We will use the following general result;

Lemma 3.B.52. Ifp> 3 andn > 1, then

P
1
@Lmzfl (mod p)

1" +2" 4+ .+ (p—- 1) =pBy + 70 Bay (mod p?).

=0 (mod p).
Proof. The left-hand side is equal to
The last congruence is also standard and it is proved using primitive roots
1 ol n+1 np? By a3 7 P* mod p.
( 5 )B'rw 1k’ = pB + __{7,?1}__1 + Z P Buti7 o
"+1k71
 Chapter 4

Primes and Squares

This chapter is concerned with arithmetic properties of primes of the form

4k +1. It is very elementary and most problems use the following two results.

Theorem 4.1. (Fermat) Any prime number of the form 4k +1 can be written
as the sum of squares of two integers.
There are many proofs of this classical result, see for instance the first
example in chapter 4 of (3]. For a different proof, using properties of Gauss
and Jacobi sums, see the addendum 9.A.
Proposition 4.2. Let p be a prime of the form 4k + 3 and let @ and b be
integers such that p divides a® + b2, Then p divides a and b
Proof. If p does not divide a, there exists c such that ca = 1 (mod p). Then
(b¢)? = ~1 (mod p}, thus

(- =" =1 (mod p),

the last congruence being Eerm?t’s little theorem. But this is clearly impossi-
ble, since by hypothesis (—1)"Z = —1. The result follows. a]
An easy consequence of proposition 4.2 is that u,,(a2 -+ b7) is even for any
prime p of the form 4k + 3 and for any integers a,b. This is a very useful tool
when studying some diophantine equations, as the following problems show.
190 Chapter 4. Primes and Squares 191

1. Prove that the number 4mn — m — n cannot be a perfect square if m Proof. We clearly have no solutions for y < —1, so let us suppose that y+2 > 0.
and n are positive integers. Taking the equation modulo 4, we easily obtain that y = 1 (mod 4). The key
Fermat point is to rewrite the equation as z? + 112 = 37 + 27 or, by factoring the
right-hand side, as
Proof. If 4mn — m —n = 2% for some integer z, then
2?4+ 117 = (y +2)(5° - 2° + 4y* — 8y° + 16y” — 32y + 64).
(4m - 1)(4n — 1) = (22)* + 1. Since y = 1 (mod 4), we have y + 2 = 3 (mod 4), thus there exists a prime ¢
such that v,{y + 2) is odd. Note that ¢ does not divide
But there exists a prime p = 3 (mod 4) such that pl4m — 1. Then p divides
(22)% + 1, contradicting proposition 4.2. [m} 1 - 20 + ay* — 8% + 1647 — 32y + 64,
2. Prove that the equation 3’ 2% — 4 has no integer solutions. as otherwise g would divide 7-64 and z°+112, a contradiction. Thus ve(y”+27)
Balkan Olympiad 1998 is 0dd, which is impossible, as it equals vy(z? + 11%) and ¢ = 3 (mod 4). The
result follows. O

Proof. Write the equation as &® +2° = % 4 6%. We claim that there is always
4. Find all pairs (m,n) of positive integers such that
a prime p = 3 (mod 4) such that v,(z® ++ 2°) is odd. Since for any such prime
p we have v,(y2+6?) = 0 (mod 2), we will thus get a contradiction. Note that m? - 1j3™ + (n! — 2)™,
z is odd, otherwise 7 = 2z1, y = 2y; and § divides 37 + 1, which is impossible.
If = 1 (mod 4), there is a prime p = 3 (mod 4) such that vy(z +2) =1
Gabriel Dospinescu
(mod 2). One cannot have p|z?* — 223 +42% - 8z + 16 (otherwise p would divide
5-2%), 50 v(w® + 2%) = vp{z + 2) = 1 (mod 2) and we are done. If x = —1
Proof. First, assume that n > 2. Then m cannot be odd, since otherwise
(mod 4), then 2% 2 3 +...+16 = 3 (mod 4) and we can repeat the argument
8 would divide m? — 1, but 3™ + (n! — 2)"™ is odd. So m is even. But then
by taking a prime p = 3 (mod 4) such that vp{(z*~22%+ - +16) = 1 (mod 2).
m2~1=—1 (mod 4), so there exists a prime p = —1 (mod 4) dividing m2—1.
The claim being proved, the result follows. ]
But then p divides 3™ + (n! — 2)™ and since m is even, this implies that p
divides 3 and n! — 2. Thus 3 divides n! — 2, a contradiction.
Proof. We can work modulo 11: since 2 = z (mod 11) for any =, we deduce
Thus, we must have n = 1 or n = 2. If n = 1 then either m? — 1|3 + 1
that #° = 0,1,~1 (mod 11) for any z. On the other hand, the guadrati
and m is even or m® — 1|3 — 1 and m is odd. In the first case we can use
residues mod 11 are trivial to find: those are 0,1,4,9,5,3. One readily checks
the same argument as before to get a contradiction (choose a prime p = —1
that the equation has no solution mod 11 using these observations. So, it does
(mod 4) dividing m? — 1), while the second case is impossible since 3™ — 1 is
not have integral solutions either. a
not a multiple of 8 when m is odd. Thus n = 2 and m? — 1|37, Thus there
is k < m such that (m — 1){m + 1) = 3*. But then m — 1,7+ 1 are powers
3. Solve in integers the equation =y 4 7.
of 3 which differ by 2. Thus clearly m — 1 = 1 and m = 2. We deduce that
Titu Andreescu, USA TST 2008 m =n = 2 is the only solution of the problem. a
192 Chapter 4. Primes and Squares 193

z°+ 2 Yy ones are equal to 2. The equation becomes 5™ — 16 = k%, As k is odd, a

5. Find all pairs (z,y) of positive integers such that the number is
consideration mod 8 shows that m is even. But then (5™/2 —k)(5™/2 4+ k) = 16,
a divisor of 1995. which easily implies that m = 2. Thus the sequence (as,az,...,a,) consists
Bulgaria 1995 of two numbers equal to 3 and the remaining equal to 2. Clearly, all such
sequences are solutions of the problem. )
Proof. Note that 1995 = 3-5-7-19. The key observation is that 3,7, 19 are aéll
primes of the form 4k + 3. If p is any of these primes and if p divides ”—;’%—, 7. Prove that there are infinitely many pairs of consecutive numbers, no
then p divides 2 + 32 and so it divides z,y. Writing z = pa1,y = py1, we two of which have any prime factor of the form 4k + 3.
obtain
21y steod Proof. Since no number of the form n? + 1 has a prime factor of the form
T~y m-—y; 4k +3, it is clear that the pairs ((n?+1)2, (n® + 1)? + 1) yield solutions of the
Doing this for every prime factor p of -——-’f-zif; and noting that 22 + 2 =z~ y problem. ]
2,2
has no solutions with z,y > 1, we just have to solve the equation I—T%’L = 5. The following problem is trickier and its solution uses the theory of Pell
This is easy, since we can write it as (22 — 5)° + (2y + 5)% = 50 and since equations.
50 can only be written as sum of two squares in two ways 1 + 49 = 25 4 25,
Thus 2y +5 € {~7,~5,~1,1,5,7} and since y is positive we must have y = 1. 8. Let p be an odd prime. Prove that 1 {mod 4) if and only if there
But then z = 2 or z == 3. Putting everything together, we deduce that the are integers z,y such that 2% — py? = —1.
solutions are all pairs (2k, k), {3k, k) with & € {1,3,7,19,21,57,133,399}. O
Proof. One direction follows directly from proposition 4.2, so let us assume
6. Find all n-tuples (a1,as,...,an) of positive integers such that
that 1 {mod 4). The key point is to consider the positive Peil equation
(ar!—1)(ag! —1).. (gl — 1) — 16 7% — py® = 1. By the general theory of Pell equations, this has a smallest
nontrivial solution {(zo, %) {so % > 0 is minimal among all solutions with
is a perfect square. y # 0). Note that 2o is odd, since otherwise 4 would divide 32 + 1. If
zp = 20 + 1, we deduce that yo is even and a® +a = pb? for b = %l I
Gabriel Dospinescu
p divides a, then a = pc® and @+ 1 = d° for some integers ¢,d {because a
and a + 1 are relatively prime}, so that & — pc? = 1. But obviously ¢ < yg,
Proof. Suppose that
contradicting the minimality of (zo, yg). Thus p divides a+1 and we can write
(a1} — )(az! — 1) {an! — 1) — 16 = K2 a=c? a+1=pd? for some integers c,d. But then ¢ — pd® = —1 and we are
done. =)
First, we claim that a; € {2,3} for all 4. It is clear that a; # 1 for all 4, so
assume that a; > 3 for some ¢. Then a;! — 1 = 3 (mod 4), thus there is a ‘We continue with another beauty, which became classical in mathematical
prime p = 3 (mod 4) such that p divides a;! — 1. Then p divides k2 + 4%, a contests. It has the nice feature of having a completely elementary solution
contradiction. Say m among the numbers ¢; are equal to 3 and the remaining that uses quite a lot of different ideas.
194 Chapier {. Primes and Squares 195

9. Find all positive integers n such that the number 2™ — 1 has a multiple
of the form m?2 +9.
IMO 1999 Shortlist
Proof. Assume that 27 — 1 divides m? + 9 for some integer m and that n > 2.
Then the only prime factor of 2 — 1 which is 3 modulo 4 is 3. Indeed, if p
is such a prime, then p divides m? + 3% and so p divides 3. Now, if n has
a nontrivial odd divisor d, then 2¢ —1 = ~1 (mod 4) and 3 does not divide
24 _ 1, Thus 2¢ — 1 has a prime factor p = 3 (mod 4) different from 3. Since
2¢ — 1 divides 2" — 1, Choosing m = 3a, it is enough to find a such that
(224 1)(2% +1)...(27°" + 1) divides a2+ 1. this is impossible, by the previous
remark. Thus, n must be a power of 2. Conversely, assume that n = 2% and
observe that
2 —1=3- (1)@ +1) - (287 4 1)
Choosing m = 3a, it is enough to find a such that (224+1)(2*+1) --- (22
divides a?+1. The crucial point is that the Fermat numbers 22" 41 are pairwise
relatively prime, so by the Chinese Remainder Theorem it is sufficient to prove
that for any i there is @ such that ¢® + 1 = 0 (mod 2% + 1). But this is clear,
since @ = 22" works. Thus, the answer to the problem is: all powers of 2.
The next problems are concerned with properties of sums of two squares.
A crucial fact is that the set of numbers which can be written as a sum of two
squares of integers is closed under multiplication, by Lagrange’s formula
Proof. One direction being clear, let us prove that if n is the sum of the squares
of two rational numbers, then it is also the sum of the squares of two integers.
Thus, we know that na? = b+ ¢ for some integers a, b, ¢ with a # 0. For any
prime p, we deduce that
wp(n) + 2up(a) = vp(8 + ),
so that vp(n) has the same parity as v,(b% + ¢?). Since for any prime p = 3
(mod 4) we have vy(b* + ¢*) = 0 (mod 2), it follows that for any such prime
p we have vp(n} = 0 (mod 2). The result now follows from the preliminary
discussion. o
Remark 4.3. Actually, the following result holds: if n > 2 is an integer and if
an integer can be written as a sum of n squares of rational numbers, then it
can also be written as a sum of n squares of integers. For n = 3, this follows
41) from Davenport-Cassels’” lemma ([3], chapter 13, example 12), while for n > 4,
this follows from the famous theorem of Lagrange, according to which any
nonnegative integer is the sum of four squares ({3], chapter 13, example 5). )
O 11. Prove that each prime p of the form 4k+ 1 can be represented in exactly
one way as the sum of the squares of two integers, up to the order and
signs of the terms.
Fermat

(6 + B9 + &%) = (ad + be)? + (ac ~ bd)®. Proof. The fact that any such prime p is a sum of two squares is the content
of che;m»em 4.1. Let us focus on the uniqueness part. Assume that we have
Actually, combining theorem 4.1 and proposition 4.2, it is easy to completely p=a?+y? =22 +% Then (z —2)(z +2) = (t— y)(t +v). We will need the
characterize the elements of this set: they are precisely the nonnegative inte- following very useful
gers 1 such that vp(n) is even for all primes p = 3 (mod 4).
Lemma 4.4. If a,b,c,d
are nonzero integers such that ab = cd, then there
10. Prove that a positive integer can be written as the sum of two perfect are integers m,n, p,q such that a = mn,b = pq,c = mp,d = ng.
squares if and only if it can be written as the sum of the squares of two
Proof. Let & = % = % be the representation of the fraction 2 in lowest terms.
rational numbers.
Since ap = nc, we have pic, so we can write ¢ = mp for some m. Then also
Fermat @ = mn. Doing the same with dp = nb yields the conclusion. a
196 Chapter 4. Primes and Squares 197

Coming back to the problem, assume that || # |2} and |¢| # |y|. Then Proof. We will show that 3% = m? + n2 4 1 has infinitely many solutions.
by the lemma we can find nonzero integers mi,n1, p1,g1 such that Indeed, start with the observation that 3> = 22+ 22+ 1. On the other hand,
if {k,m,n) is a solution with m > n, then (2k,3*m — n,3%n + m) is also a
r—z=many, T+z=pq@, t—y=mp, t+y=ng. solution. Indeed, this follows from

But then (FFm—n) +@FFn+mP+1= 3% 1 mi+ad)+1=9% 1+1=3% O

_ mm +piq _ gL - mpy
= 5 Y= ) s In the following two problems we will use the fact that the density of the
so that using Lagrange’s identity we obtain
set of positive integers all of whose prime factors are of the form 4k + 1 (or
4k +3) is zero. This is a nontrivial result, for a proof of which we refer to [3},
dp = 4fa® +47) = (md +ad)(n + 5. chapter 4, example 10.

13. It is a long standing conjecture of Erdés that the equation

We may assume that p divides m? -+ g3, so that nf + p} is equal to 1,2 or 4.
As 11,01, p1. ) are nonzero, we obtain a contradiction unless ny = py = 1.
But in this case we get © — 2z = £(t —y) and @ + 2 = (¢ + ), thus

{l=t, 121} = {I¢l. 11} has solutions in positive integers for all positive integers n. Prove that
the set of those n for which this statement is true has density 1.
and we are done again. )
Proof. We look for solutions with y = z and z = na for some positive integer
We continue with an easy exercise in Lagrange’s formula. a. The equation becomes dzy = n{y + 2z) or, equivalently, y{da — 1) = 2na.
Thus, if we can find a prime factor p of n of the form p = 4a — 1, then we can
12. Prove that the equation 3% = m? + n? + 1 has infinitely many solutions take y = 222 and we have a solution in positive integers. Thus, it is enough
in positive integers. to prove that the set of integers having at least one prime factor of the form
Saint-Petersburg Olympiad 4k — 1 has density 1. which has already been discussed. a

14. Let T be the set of positive integers n for which the equation n? = g% +5%
Proof. Guided by the formula
has solutions in positive integers. Prove that T has density 1.

Pl = @ F 1) T 1), Moshe Laub, AMM 6583

we will choose k = 2% Since all factors in the product are sums of two Proof. We will prove that n € T if and only if n has at least one prime factor
squares and since the set of numbers which are sums of two squares is stable of the form 4k + 1. Suppose first that n € T and choose positive integers a, b
by multiplication, it follows that 3% _ 1 is always a sum of two squares. Since such that n? = a? + 5%, If all prime factors p of n are of the form 4k — 1, then
it is trivially not a perfect square (it is of the form 3k + 2), the conclusion for any such p we have pja® + 5%, so that p divides a and p divides b. Dividing
follows. [m} the previous relation by p* and repeating the argument, we deduce that poeln)
198 Chapter 4. Primes and Squares 199

divides @ and b. Since this happens for all pin, it follows that n divides a Note that
and b, which is clearly impossible. Conversely, if n has a prime factor p = 2 i2
i (modp)=¢*-p [‘j{
(mod 4), by Fermat’s theorem we can find integers ¢, d such that p = 2 + d?%. P
We may assume that ¢, d are positive (they are nonzero since primes are not and since
perfect squares). But then %i’ _pECE+1)
=5
it remains to prove that the sum of the quadratic residues mod p is pk. But if
L1, T2, Tpst BTE the nonzero quadratic residues mod p (any residue mod p
and so n € 7. Now, the density of those numbers which are not divisible by
any prime of the form 4k + 1 is 0 and we are done. O is implicitly ta.ken between 0 and p— 1), then p—z1,p— x3,...,p— Top are a
permutation of the z;’s (since—1 is a quadratic residue mod p, which follows
‘We continue with two very nice problems concerning primes of the form from p=1 (meod 4)). Thus
4k + 1. The method used in the solution of the following problem is standard.
2% 2%
15. Let p be a prime number of the form 4k + 1. Prove that Yr=Y
=t =1
-z
> |V and the result follows.

The following functional equation is rather nonstandard.

O

Proof: Write p = 4k + 1 and note that

16. Find all functions f : Z* — Z with the properties:
i & 2%
SWi=3
ge=1
3 1=3i=1 kZJZ‘Z
F=li¥<gp
3 1 fla) = f(b) whenever a divides b.
2. for all positive integers @ and b,

since the inequality ¢ < /jp with 1 < j < k& implies that ¢ < 2k. On the other 1(ab) + f(&® +8) = fla) + f(B).
hand, the condition j 2 2 g equivalent fo § > 1+ [ }, since % is not an
Gabriel Dospinescu, Mathlinks Contest
integer. Thus we can also wntc
2% 2 2% o2 Proof. Since 1 divides any integer, it follows that f(1) > f(z) for all z. Let

s
gl i=1 (- [5]) - £ [7]
L4 P
k= f(1).
"The first step is to prove that f(n) only depends on the prime factors of
7 and not their multiplicities, i.e.
and it remains to prove that

HOENARIFS
pin
|

200 Chapter 4. Primes and Squares

201

Indeed, consider two positive integers a,m and choose b = am. Thus The following challenging problem requires some estimates about prime
numbers
e that
thatfollow from Dirichlet’s theorem, , forfor which
whick we refer the reader to
F(@®m) + f(a*(m?® + 1)) = f(a) + flam)
and by the first condition f(a) 2 f(a®(m? + 1)) and f(am) > f(a®m). Thus
17. Prove that the equation 2® = n! 4+ 1 has only finitely many solutions
both of these inequalities must be equalities and so f(am) = f (a®m) for any in
nonnegative integers. ’
positive integers a, m. This immediately proves the claim.
In the second step, we prove that f(n) does not depend on the prime Proaf. Suppose that (z,n) is & solution of the equation z® = n! + 1. Then
factors of n that are congruent to 1 or 2 modulo 4. Indeed, the proof of the
previous claim shows that for all n and = we have f(n) = Ffn(z?+1)). Tn
nl=(z* —1)(z® + D' + 1).
particular, f(r) = f(2n) and so we don’t have to care about possible powers Let A, be the set of prime numbers 7 < nof the form 4k + 3. The
key point is
of 2 in the prime factorization of n. Also, if p = 1 (mod 4) and « is chosen that no p € A, can divide 22 + 1 or z* + 1, so that p*(*) must divide 22 — 1
such that p|z? + 1, then Thus we obtain - )
fin) = fnp) = f(n(a® +1) = f(n) Vnlza?—1> I .
pEAn
and so f{np) = f(n). In conclusion, we have Then, using that u,(n!} > n/p— 1 (by Legendre’s formulaj, we obtain
1
fin) =1 ( - annn>1n\‘/r§> Z (;)Al) Inp.
pin.pePs PEAn
A classical inequality of Erdés (theorem 3.A.3, chapter 3} yields
where P is the set of prime numbers of the form 4k + 3. Let py, p, . .. be the
elements of P; and define g(A) = f([1,c 4 Pe). We obtain a function g defined
on the set of all subsets of N with integral values. Moreover, we claim that Z Imp<In Hp <mnind.
9@ =k, 51 C S = g(51) = ¢(8e) and finally PEAL p<n

9(S1) + g{Se) = 9(S1 U S2) + 9(5: 1 82) We deduce that

The first two relations are obvious. For the third one, note that if the sets
of prime factors congruent to 3 mod 4 of a and of b are 4, B, then the set of
prime factors of the form 47 + 3 of ab is exactly AU B and the set of prime for any solution (z,n).
factors of the form 45 + 3 of @® +5% is AN B. If g({n}) = k,, for some integers N.ovvu it remains to prove that there are only finitely many such integers
kn < k, then an easy induction on [A| shows that 7. This is a consequence of the proof of Dirichlet’s theorem, which establish
es,
among many other things, that
9(A) =3 ko= (14] — V. 1 np 1
'
acA
Inn E T 3

]
Conversely, any choice of such &, yields a.corresponding g and the previous PEAR
construction yields a solution f of the equation. This ends the solution. I for n — oo, See addendum 7.A for a proof. O
 4.1. Vi
Notes 203
202 Chapter 4. Primes and Squares

It is really amazing that the following result has a purely elementary 4.1 Notes
proof. We present here the beautiful idea of John H.E.Cohn and we refer the
Ma.ny of the solutions to the problems in this chapter were provided
reader to [18] for other similar results (including the fact that 1 and 144 are by the
following people: Alexandru Chirvisitu {problem 14}, Daniel Harrer
the only squares in the Fibonacci sequence). (problems
2’,3)’ Benjamin Gunby (problems 1, 6, 9), Fedja Nazarov (problems
12, 15)
VIS. Let Lo =2, L1 = 1 and L0 = Lnqy + Ly be Lucas’s famous sequence. Gergji Zaimi (problems 5, 7, 13, 16). :
Then the only n > 1 for which L, is a perfect square is n = 3.
Cohn’s theorem

Proof. 1f z; and z3 are the roots of the polynomial X? — X ~ 1, then L, =

27 + 2, which combined with zyzy = —1 yields Lo, = L2 ~ 2(~1)". This
already shows that if L, is a perfect square, then n is odd, for 42+ 2 is never
a perfect square.
The case when n is odd is much more subtle. There are two key properties
of the Lucas numbers that make everything work. The first is that Ly = 3
(mod 4) whenever k is an even number not a multiple of 6 (use the previous
formula for Lg, and the fact that L, is odd whenever n is not a multiple of
3). Call such a number & good. The second key ingredient is the fact that Ly
divides Lok + Ln for all good numbers k and all nonnegative integers n. This
follows from k = 0 (mod 2), the equality #3122 = —1 and the computation

Lnpab+ Lo = a7(@f + 1) + a5 (afF +1) =

nik,
ARt k) 4 R+ o) = DL
Assume now that 7 = 1 (mod 4) and n > 1. Then we can write n =
1+ 28k for a nonnegative number r and a good number k. Applying the
second key point 37 times, we deduce that L, = —L; = ~1 (mod Lg). As
Ly = 3 (mod 4), it has a prime divisor of the form 45 + 3 and so L, cannot
be a perfect square. .
Assume finally that n = 3 (mod 4) and n > 3. Then we can write
n = 3+ 2. 3% for a nonnegative number r and a good number k. The
same argument shows that L, = —L3z = —4 (mod L) and we reach the same
conclusion, as numbers of the form 2% + 4 have no prime factors of the form
45 +3. a
 Chapter 5

75’s Lemma

All of the following problems fall to a rather handy inequality, known

as
T>'s lemma even though it is a special case of the Cauchy-Schwarz
inequality.
This result says that for all real numbers a1,82,...,a, and all positive
real
numbers 21,23, ..., 2, the following inequality holds

2 2 .2
B
a
B gy +ag+--- ta) 2
Gy ,
@t
1 X2 Ty Tyt Tz + o F T,

To get the reader familiar with this trick, we start with a series of more
direct applications.

1. Let 21,22, ..., Zn, 1, Y2, - ,¥n be positive real numbers such that

Tyt Bt H T 2 Ty ol + oo+ T

Prove that

Tt Tyt
N .
Ty < 2 T
[ Yn

Romeo Ilie, Romania 1999

206 Chapter 5. T3’s Lemma 207

Proof. Using Th’s lemma, we can write Proof. Ty's lemma gives a lower bound

kA (D) a (Ze?)?
2 : waz Lljy, z) Tt Tal
ve s D
the last inequality bemg‘ exax:tly the hypothesis. ] Since
2
Za(b-+c+d) = (Za) 7202 §3Za2,
2. Let a,b,c be nonzero real numbers such that @b+ be 4+ ca > 0. Prove
that it remains to prove that 3" a? > 1. But again by Cauchy-Schwarz we have
ab
FIETET 1= (ab+bc+cd+da)? < (a® + b2 + 2 + d%)?,
Tita Andreescu which proves that 3~ a? > 1 and finishes the solution. o

Proof. The trick is $o add % to each fraction, in order to exploit the identity 4. Prove that if the positive real numbers a, b, ¢ satisfy abc = 1, then
a 4 b . c >1
ab +1_ (a+b)? b+e+1l cta+l a+b+17 7
a?+b2 " 2 2a?+8%
Vasile Cartoaje, Gazeta Matematici
With this observation, the inequality becomes
Proof. The solution using 75’s lemma is straightforward:
Z(a+8)?
2 2
a+ b2 -
and it is then a trivial consequence of 7’s lemina, combined with the hypoth-
T et~
bte+1 S e
ab+c+1) 2T 23 Inayw
ab+> a

esis ab+be+ca > 0, O so that we only need to prove the inequality 3~ a? > 3" a. This follows from
Sa?> &2—“)—2 and 3" a > 3 (the first being a consequence of Cauchy-Schwarz,
3. Prove that for any positive real numbers a,b, ¢, d satisfying the second being the AM-GM inequality). a

ab+be+ced+da=1, 5. Let a,b, ¢ be real numbers such that

1 . 1 N 1 S
the following inequality holds
@+l B+l 21T
@ . B . & . & J1
Prove that
btct+d c+d+a dta+b a+btec”3 ab+bc+ea <

o
IMO 1990 Shorthist Titu Andreescu
208 Chapter 5. Ty’s Lemma
209

Proof. Observe that Proof. Using Ty’s lemma, it is sufficient to prove the inequality
2
Y=o
1
(a? +b° +c%)? .3 a?+
b 4+
@b+c)?+b2(ct+a)+c2a+b)2 " 4 ab+be+ca
On the other hand, we have
Note however that the obvious application of T3’s lemma fails. The previous
z - 52_. > M inequality is equivalent to
@417 a2+ +E+3
Combining the two inequalities immediately yields the result. [m] 4(a® + 6% + ){ab + be + ca) > 3@ b+ e + e+ a)? + Fla+ )?).

The following problems are a bit less straightforward. However, they do Unfortunately, the only reasonable way to prove this is to expand it in the
not require any heavy machinery. form
Do’ +8%) 2> O3530 2,2
252y + 1sabe Y
6. Prove that for any positive real numbers «,b, ¢,
Fortunately, this is trivial, since
1 ™
1 1 1 (a+b+c+ Vabe)?
ab T bre
T ora T 2 A DG et o) Z abla® + %) > 22112172 and Ea’b2 > acha. a
Tita Andreescu, MOSP 1999 It is possibie to solve the following problem using T3’s lemma, but the
proof is not really elegant. In the addendum we discuss some applications of
Proof. The solution using 7y's lemma is a bit tricky: the point is to look at
Holder’s inequality, which makes this problem really easy.
the denominator of the right-hand side, because

(a+ b+ cMe+a) = (a+ B+ (b+ c)a® + (c+ a)b? + 2abe. 8. Let a,b, ¢ be positive reals such that abc = 1. Show that

"This suggests writing the left-hand side in the following way 1 1 1 1

c* 4 a? N [ . (Vabe)?
FET20r B FEi RS
Ala+b) a?b+e) Blate) 2abe Titu Andreescu, USA TST 2010
A direct application of ’s lemma finishes the proof. =]
Proof. Start by making the substitution

1o
7. Prove that for any positive real numbers a, b, ¢ the following inequality
holds
Ty *TL
a \? b \? c V.3 d+pP+d
+ + 2 . The inequality becomes
bte cta a+b 4 ab+be+tca
23 s 23 1
Gabriel Dospinescu P
Grap v T o -3
210 Chapter 5. Ty’s Lemma 211

Applying Ty’s lemma, it is enough to prove that It is really not easy to prove the following inequality using T5’s lemma,
but the trick is worth remembering.
3(z§ + y% + z%)2 > }:fi(z +2)°.
9. Prove that for any positive real numbers a,b, ¢ the following inequality
The right-hand side is equal to 53" 2*y?-+4 3" x. Thus, we need to prove that holds
1 1 Sl 1 1 1
3215 + ()Z(wy)% > 5Z(a¢y)2 +4Zx. > .
3a+b+3b+c+3c+a - 2a+b+c+2b+c+a+2c+a+b
Note that M.O. Drambe

PRCILED DCIRIEES SN2

B LD W Proof. Choose three positive real numbers o, 3, and use Ty’s lemma in the
form
the first inequality being Chebyshev’s, the second one by the AM-GM inequal-
ity and the fact that zyz = 1. Thus, it suffices to prove that @ B v (a+B+1)°
3a+b 3b+c 3ct+aT alBa+7)+b(B88+0a)+c(3y+F)
3E:t5+2$2y2 Z4Zx.
Now, we impose the conditions

But 32° + 3%2% > 42% and so it is enough to prove that E.T,‘Ta > 3" . This da+y=2 38+a=1 3y+8=1
follows from the power-mean inequality and the fact that z +y+223. O
Solving this linear system yields the solution o = %,6 = %,7 = % Therefore,
Proof. As in the previous solution, we reduce the problem to proving the we obtain the inequality
following inequality
4 1 11 2 1 1
z 3 22 1 7345 7 3%b+c 7 3cra-Zatbie
GraE wrzr T praer -23 Proceeding in the same way with the two other terms of the left-hand side and
Using the AM-GM inequality, we can write adding up the resulting inequalities yields the desired result. =]

o y+z2 21'/_4:’3>r Proof. We can also use the trick of integrating polynomial inequalities to de-
Qyrar o T TE F3 duce fractional inequalities. Namely, the inequality

and two similar inequalities. Adding them yields the following estimate y+itz+ e 2 eyt y+2)

e3 N v3 v 2z 3 r+yt+z can be easily proved using 7»’s lemma, since it can be written in the form
(z+2y)? (+22)? (y+22)0° " 9
¥ 7
—Zzt+y+z
and we end up using once more the AM-GM inequality. a x ¥
212 Chapter 5. Ty’s Lemma
213

Using this inequality, we can write The proof is a bit tricky. If n is even, the inequality follows trivially from
the chain of inequalities
(Ratb=l g gSbte=l | yBeta—l o gatbbeol y gdbtota=l y glotath-l

Az1zp + -+ @a21)
< 43y tx3d )@t oyt g
for all 0 < ¢ < 1. Integrating this between 0 and 1 yields the desired inequality.
a Sz +zo+ o+ 3,03
the first one being obvious and the second one being the AM-GM inequality.
Remark 5.1. The technique used in the second proof looks unusual. Tt is
actually quite powerful and we refer the reader to [3], chapter 19 for many For n odd, things are subtler, but we can reduce the problem to the
case when n is even by the following mixing argument: we may assume that
more applications,
&1 2 T2, so that we trivially have
The following two problems are closely related and use a rather useful
inequality. 12y + TaT3 + T2y S T2 + 3123 + w324 < T1(T2 + T3) + (T2 + T5)74.
10. Prove that for all n > 4 and all 21, 22,..., @, > 0, Thus, replacing 1,22, ...,%» by T1,Z2 + 23,24, . .., Tp, We preserve the sum
of the ;’s while not decreasing the quantity 322+ -+ + @,11. Since n—1 is
zy x2 Tn
+ EERREE ~ 2 even, everything follows from the previous step. o
In+ Tz T+ T3 Tp1 B
11. Let 7 > 4 be an integer and let ay, a3, ...,a, be positive real numbers
Tournament of the Towns 1982
such that af + @3 + -+ + a2 = 1. Prove that
Proof, We start in the usual way by using 73's lemma:
Pe@ T@2 On 4
1 £ a2+1 aZ+1 * af+1°
2 plavar +ava + + n/an).
+ En
Tp+ T2 Ty Loy + T
Mircea Becheanu and Bogdan Enescu, Romanian TST 2002
(@1 4 mp bk 3p)?
T1(@n + x2) + w22y + 73) + A ZalBaoa +21) Proof. Applying To’s lemma we obtain

Since

1T+ T2) + z2(F1 F T3) A F T (Bt 31) = ez Fapzz + oo+ xpw), >CRE Zaz(aa?Za+1D) 2T Yaal,
(L) +1)
it remains to prove the following very useful Since I a? = 1, we have " a?a?,, < 1/4 by lemma 5.2. The result follows. [J
Lemma 5.2. Ifn >4 and xy,29,. .., @y, are nonnegative real numbers, then We give two proofs for the following beautiful problem. The first one
is a standard application of T5’s lemma, the second one uses a very useful
(@1 + 32+ + 30 2 Harzs + 2223 + - + Tu1). technique.
214 Chapter 5. Ty’s Lemma 215

12. Prove that for any positive real numbers a, b, ¢, the following inequality The problem asks to prove that under this assumption we have z +y+2 > 1.
holds " b R Assume that this is not the case,so z +y + 2z < 1. Let
“+ + 2L
Va2 18 Vb +B8ca Ve +8Bab T ¥ zZ
= >z, Y= >
z+y+z PR z
Hojoo Lee, IMO 2001
sothat
X +Y + Z =1 and
Proof. The most natural application of T's lemma turns out to work very
smoothly. Indeed, 1 1 1

Z a ‘_Z o* > (a+b+ef? )

e (X"’ Hl> (Y” _l) (z?'l)'
a? F 8he ava? +8be ~ Y av/a?
+ 8be ‘We deduce that

The only problem is to show that 512X2y?2?

SX+YNY +Z2HZ+ X)X +Y + 2)(2¥ + X+ Z)(2Z + X +Y),
Savatshe < (a+b+ o)
which is impossible since X +Y > 2/ XY, 2X + Y + Z > 4¥/X2YZ and the
This suggests using Cauchy-Schwarz and, indeed, we have
similar inequalities obtained by cyclic permutations. )

(Z a\/a7fl—§‘_8fi)2 < \/5 \/fzfi + 24ab(;. There are two traps in the following problem, making the problem harder
than it appears at first sight.
Thus, the problem is solved if we can prove the inequality
13. Let a,b,c be positive real numbers such that ab 4+ bec+ ca = 3. Prove
er +2abe < (a+b+ ).
that

Fortunately, after expansion, this becomes Y a(b— )2 > 0, which is obvious. S

2040 2+
A 2c+aZT
[m}
T.Q. Anh
Proof. Make the substitution

\/ a? / b2 ZA/ 2 Proof. The inequality we have to establish seems to be opposite to the usual
x applications of 73's lemma. This is actually not a very serious problem, since
i YTVE
s TTVETR
we can always change the terms of the sum a bit and change the sign of the
i—g—"‘ and the two similar ones yields the inequality. Indeed, note that
Multiplying the relation 517 -1 =
relation i L .

()G = - = -
GE) e —1) =512, e
2a+5
_1(
2
¥
2a+b2 )"
216 Chapter 5. Ty’s Lemma 217

Thus, the inequality is equivalent to is close to n — 2. This can be done easily, by taking n — 1 of the variables
equal to some z very close to 0, and the last variable equal to z'~", The
>1 difficult part is proving that F(a,.a: @) < n — 2 holds for any sequence
2c+at ™ a1,42,...,0, as in the statement. Using the basic inequality, this reduces to
proving that
And now, another trap: one would be tempted to use Tp's lemma in the
obvious form, but it is not difficult to check that this does not work. Instead,
w 1
——— < -2
we take advantage of the hypothesis ab + bc + ca = 3 to write ; Tra)tas)
ovie | (E00)
for any positive numbers a; with ajas - = 1. To prove this, write a; = —‘—,
with z,4; = ;. Subtracting 1 from every fraction, we obtain the eqluvalent
Zmufl ZZabMJ* 6+3.6% inequality
TiTipt + Tiiv +h
Thus, it remains to prove that E(ab)g/ ? > 3, which is immediate by the
power-mean inequality and the hypothesis. 1 — (zet m)(min + i)
This is too complicated to try a Cauchy-Schwarz approach, but it simplifies a
The following problem is also quite tricky. lot if we observe that
14. Determine the best constant k, such that for all positive real numbers Tl + TTigg + Ty o z2,;
41,0z, - .., @y, satisfying ayay - -+ an = 1, the following inequality holds @itz @ Fowe) | mAoe | @ +Zier (T + Tip2)
ajag . az03 N i k
Since
@ +a)a+a) (@ +a3)(ad Tay Ut @ e T T E
Ei
> E
z;
=
T+ Tiv1 T3+ xoF - o, L
Gabriel Dospinescu, Mircea Lascu it remains to prove the inequality

Proof. The basic inequality that will be used is

E T
(@i + s {@ap1 +ig2) ~
(@ +y)(e? + ) 2 ay(L+ (1 +y)-
T»’s lemina reduces this to the easier inequality
This reduces immediately to {z? ~y Yz —y) 20, w}uch is clear. This already
shows that for n = 2 the maximum value is k, ince (L+z){(1+y)>4
(Z -'”*)2 > Z 2?42 Z-’l'izi+1 + ZIi1§+z.
if 2y = 1.
Let us assume now that n > 2. We will prove that ky = n — 2. To prove Expanding the left-hand side makes the previous inequality obvious and fin-
that k, = n — 2, it is enough to exhibit sequences aj, dz, ..., Gn for which the ishes the proof of the fact that k, =n— 2 for n > 3. [m]
value of the expression
The following difficult problem seems to be exactly the opposite of what
n
Flai, (a1, a9, az,..., W)= aiain is usually called a standard application of T3's lemma. It turns out that we
T e
) = ; (a2 + aig1)(0fy,
(@2 +£ @)) can actually apply 72’s lemma, but in a very nontrivial way.
218 Chapter 5. Tp’s Lemma 219

15. Prove that for any positive real numbers a,b, ¢, is the obvious equality case in the original inequality}. Thus, we should have

<8 F{})=%+vandalso f (3) = u, since } would be a minimum of uz+v~ f(z)

(Qa+b+¢)? | (2b+cta) (2c+a+b)?
on [0,1]. A straightforward, but tedious computation shows that « = 4 and
S0+ (b4 WP +{cta)} 20+ (atb) T
v = % ‘We need to prove now that this pair (u,v) really works, ie. that
2003 f(z) < ux + v holds for any z € [0,1]. Clearing denominators and expanding
Titu Andreescu and Zuming Feng, USAMO
everything yields (after a tedious computation) the equivalent inequality
Proof. Writing
362° — 1522 — 2z + 1> 0.

But the conditions we imposed were made so that the left-hand side is divisible
we can write the inequality as
by (3z — 1)2. Doing the euclidean division shows that the left-hand side is
@+a)? | (2+y)? . (24 2) 2
<8 1+ 2z 12.£ 2y + 1;22 S? (32— 1)?(4z + 1), which is obviously positive. Finally, we just have to add the
2+ 22 2+ y? 24 72 2?42 yt+2 24272 inequalities f(z) < wux+ v for z € {a,b.¢} to end the proof. m}

-1
2+2
2
o1
yi+2
2
-1~1)2
2242 based
The following problems are harder than the previous ones. They are still
on T3's lemma, but applied in more subtle ways and often combined
with other tricks.
In order to prove the last inequality, we use T2’s lemma in the obvious form
and so it suffices thus to show that
16. Let a,b,¢,d be positive real numbers such that abed = 1. Prove that
(z+y+2z
z2 4y+ 22 SR S —— PR S
(T+a)2 (1+82 (+0? {1+d7 =
Expanding (z +y + ¢ — 3)%, we reduce this to proving that
Vasile Cartoaje
Nat-12) e 4y ay+1220.
Proof. The proof using T5’s lemma is rather mysterious. By performing the
This is not obvious, but the observation that 72 +4 > 4z reduces it to proving substitution
that S @y = 23, which is problem 7 in chapter 1. 0 £
a=2, b:f, c=— d:f,
ions, x ¥ z 13
Proof. 'This solution uses the linearization method. To simplify computat
we may assume (since the inequality is homogeneous) that a+b+c = 1. Define we obtain the equivalent inequality
the map (417 2
z? 2
1) = 5T2y [ O A ey
The inequality can also be written as f(a)+ f(b)+f(c) < 8. We will try to find Of course, an immediate application of T3’s lemma fails rather badly, so we
w,v such that f(x) < uz -+ v for any z & [0, 1], with equality for 2 = 3 (which need something more clever. Let us apply Ty’s lemma for the first two and
 220 Chapter 5. Ty’s Lemma 221

then for the last two terms of the inequality. If we try to prove the stronger The following problem is a bit tricky, especially because of the strange
inequality hypotheses.
(z+y)? (z+1)?
[N I sl e T 17. Let n > 16 be a positive integer and suppose that the positive numbers
1,82, .., 0, $2LSTY @1+ g+ +an = 1 and a; + 2a3+ - + nay, = 2.
we easily realize! that it is equivalent to (y + 2)(t + z) < (¢ +y)(2 + 1) Prove that
Unfortunately there is no reason to have (y -+ z)(t + ) < (z +y)(z + ). The
miracle is that if this fails, then we can apply 7b’s lemma for the first and (a2 = a1)V2+ (a3 — a2)V3 + -+ (an — an_1)v/7 < 0.
fourth terms of the initial inequality and then for the middle terms. In this
case, we obtain the stronger inequality Gabriel Dospinescu
(z+1)* (y+2)? 1 Proof. The first step is to apply Abel’s summation formula to rewrite the
@+ry)l+ @+t (y+a22+z+t2 inequality as

A similar argument shows that this holds precisely when (z + y)(z + 1) < aVE+a(VB- V2 4+ e (Vi - VA=1) > e
(y -+ 2){( +t), in particular whenever (y + 2)(t + ) < (2 +y)(2 +1) fails. The
Using the obvious bound a1v/2 > a1{v/2 — 1) and the formula
result follows. a

R
- 1
Proof. This solution is not natural, either, but it is very elegant. We claim
that for all positive numbers a, b we have
1 - 1 - 1 an application of T3’s lemma shows that the left-hand side is greater than

(1+a? " (1+b62 " 1+ab (al tagt oot ap)?

This is rather easy to prove, though it requires some nasty computations. ey a,(\f +ViTD)
Namely, by clearing denominators and performing the obvious simplifications,
On the other hand, Cauchy-Schwarz together with the hypothesis give the
we reduce the problem to proving that estimate
ab(a® — ab+b?) — 2b+12 0, n-1

which is equivalent to
Dani< \} \lzm; <2
=1 i=1
abla — b)% + (ab~1)* > 0. and also
nei
Once we have this inequality, we deduce that ’ Saviti< Vi
i=1
Z > 1 . 1 _ 1 4 ab _
Therefore, taking into account that a; +as+--- +ap—1 = 1 —a,. we still need
(1+a)2
= 1+ab 1+4cd 1+ab 1+ab
to prove the inequality )
jul
and the result follows. (-an)?
't is convenient to denote o = ¥1% and b= g Vi
i
H
222 Chapter 5. Ty’s Lemma 223

But since The hard point is to figure out that such an inequality holds, since proving it
is a very easy matter. Note that the right-hand side is clearly positive, so by
1= ay + 202 + ... + nan ~ (a1 + ... + an) > (0~ Lag, squaring and canceling similar terms we obtain the equivalent inequality
we have an, < ,;{—1 and so it is enough to prove the previous inequality with a,
¥+ 6yz + 9%zt GuBy+2) | w2
replaced by ;1-1 But this is immediate for n > 16, which ends the solution, Qu+22 = 2+z Q2u+2)? =
O
Using this estimate we conclude:
‘We end this chapter with two challenging inequalities. The first one uses
a combination of 7%’s lemma and a rather subtle linearization technique. Zm\/8y2+zz §4E:cy—3zyzzfll+—z
18. Prove that for all positive numbers a, b, ¢ the following inequality holds
and since
’L+
Vacra
/--———7>1
VBardp ™"
Yy
+z T zty+z
it remains to prove that
Vo Quoc Ba Can
AEx;u—- x-?—% 52121»2239.
Proof. We use Ty's lemma in the form

[a_ (AP 1t is not difficult to check that the last inequality is actually equivalent to
V8 +e Ja®+o Schur’s inequality, which finishes the proof of this hard problem. jn}

g0 it remains to prove that The next problem requires some preliminaries. We will prove the following
beautiful discrete variant of a classical inequality of Wirtinger.
(X va)' = 3 vl Theorem 5.3. (Fan’s inequality) If z1.%2. ..., 2, are real numbers which add
up to zero, then
Define & = a,y = b,z = /¢, 50 that the inequality becomes
2+ 2y o5 27 >
Yevir A< (3a) i «72*'“‘*%)005?_Ilrz+12173+---+1u271-
Proof. We will actually mimic the proof of Wirtinger’s inequality by using
This is actually a very strong inequality and it is easy to see that it resists any
finite Fourier transforms (for more on this fascinating subject, the reader is
attempt to prove it using Cauchy-Schwarz (even if its form invites us to use
invited to read the addendum 7.A). Namely, if 21, 22,..., 2, IS a sequence of
such a technique). We will use a linearization technique, by approximating
complex numbers, define
following estimate: N 1 & _ 2imky
5 3 =7 e m oL
V82
+ 22 <3y+z— 5 i
Y+2 k=1
 224 Chapter 5. Tp’s Lemma 225

Using the fact that follows from the definitions. Thus, using Parseval’s identity we obtain

k=1 o=l
Kl E
if and ouly if is not a multiple of n (in which case it equals n), it is easy to
check that we can recover our sequence from the sequence of its finite Fourier =S - PR
ii transforms using the inversion formula 7

n n
STl =3 1w and another application of Parseval's identity yields the desired inequality,. O
=t ket
We are now ready to prove the following version of Shapiro’s inequality.
Indeed,
The proof is based on a very tricky application of 73’s lemma combined with
n Fan’s inequality.
N 1 N L T3 3
15]? == fizz%zkec G

LYy
=3 3 kiks
1 o~ _zintky kg 19. Let a, = . Prove that for all z3,22,....2, € [ai" . the
2c0s 221
K1k 3
Shapiro inequality holds:

= Z Zh1 Thy ry
+
T2 N
+eet
Fy=ky Z2+x3 23t Tg Ty +x2

=3 |ul
k Vasile Cértoaje, Gabriel Dospinescu

Now, the proof of Fan's inequality is very easy: write the inequality in the Proof. First, we write the inequality in the form
form 5
5 " 5 _ Eis1 @iy 1
S (o — zpa1)? > (2 - 2003;) Sat.
& &
R PNi (1 _ ,2) .
T Tiel T+ Tipz 2 aZ
Note that the hypothesis @1 +wo+ - -+, = ( can be writts n in the ves v simple
form &, = 0. Now, let? yj = &j — Tj+1 and observe that §; = (1— Note that
2T'his is the analogue of the derivative of a function when establishing discrete inequalities.

|
|
|
226 Chapter 5. Ty’s Lemma

Using Ty's lemma, it suffices to prove that Addendum 5.A Holder’s Inequality
in Action
(%)
1~l2 S (Z=)"23
fon in(z,u,.]+rl+2)—%z(z,+zw1)2 : .
The purpose of this addendum is to present some applications of Hélder’s
A crucial step is to note that we have the identity inequality, which are guite similar to the problems considered in this chap-
ter. Of course, Hélder’s inequality is important because of its applications
(@os + Tig2) - 3L Z(Iz + Tip1)%2
in
S wuwins + @irn) = (@ + @1 measure theory, probability theory and analysis and not really for the amus-
ing problems to be discussed here. Actually, we will not even deal with the
which allows us to perform the substitution z; + z;4y = 2b; and reduce the
classical version
problem to proving that

(1) (£0) 2 (o0 (145 )

1 1
(0] +af +-- +aB)F (b + B3+~ +1)7 > arby + asby + - + anb,
which holds for any positive real numbers @i, b;,p,g such that ‘l, + % =1, but
for nonnegative real numbers b;. rather with the following:
This simplifies drastically if we make another substitution
Theorem 5.A.1. Let a;; be positive real numbers. Then
&
(a1 +az+ - +aj) 2 (Yanag g+ + Hamazam)F.
A rather tedious, but straightforward computation shows that the previous
=1
inequality is equivalent to
Proof. This is a very easy application of the AM-GM
1 inequality. Indeed, just
(1+ 07) S d 22y e add up the following inequalities
i
Of course, we have ¢; + ¢z + - -+ + ¢, = 0. Finally, note that G31G21 - - Qpy
kf——— RSa1 axy
DR
1 2n V s8-8 =5 oot S’
1+ — =2cos —.
4 n
Thus, the result follows from Fan’s inequality. o k%0820 Gkn . Gin Gkn
5155 T 5 Sk
5.1 Notes where S; = ai1 +ain + -+ + .
o
We thank the following people for providing solutions: Alexandra We are now able to obtain a generalization of Ty’s lemma that turns out
Chirviisitu (problem 16), Xiangyi Huang (problem 14), Michael Rozenberg o be very handy in quite a lot of situations when 7's lemma fails. There are
(problem 15), Dusan Sobot (problems 1, 3), Gjergji Zaimi (problems 1, 10, of course much more general results than the following one, but our purpose
11). is not to delve into the greatest generality.
228 Chapter 5. Tp’s Lemma 5.A. Holder’s Inequality in Action
229

Theorem 5.A.2. Letay,ag,... 0y and 21,02, ... be positive real numbers A.3. For any positive real numbers a,b, ¢ the following
inequality holds *
and let g > p be two positive integers. Then
b+c + c+a " a+b >4
it¥
4 af ah o i@t ozt Fan)?9
Var+bhe VB2ieca VR tab o
y &y = (my o wg o+
Proaf. Note that by the previous theorem we have Pham Kim Hung

@ [
4 Proof. Let S be the left-hand side. Using Hélder’s inequality,
(z1 4zt -+ @) (w we can write
+ g
%
s2. (Z(b +eja® + bC)) >8a+b+c)f,
= (g bt
wn) e (X b )
g n3 50 it is enough to prove the stronger inequality
> (al* +al” -+«~+a${a"_')"“
(a+0+0)>23 (b +)(a® + bo).
and the result follows from the Power Mean inequality. ju}

The first theorem is particularly useful when working with sums of square An easy computation shows that this follows from Schur’s
inequality. o
(or cubic or ...) roots, which are a nightmare most of the times. Here are a
few examples that will probably convince you how useful this result is. Most A4, Let a.b,c be positive reals such that abe = 1. Show
that
of the problems are quite hard to solve by other means.
1 1 1
A.2. Prove that for all positive real numbers g, b, ¢ such that abe = 1 we have SEr 20 T el T S o >3
3

VETT+ B v+ V7 <2qa+b+0). Titu Andreescu, USA TST 2010

Vasile Cértoaje Proof. The substitution a = %,b = 5,0 = % reduces the proble
m to

Proof. This is quite a strong inequality, but the previous results are effective 23 . I
in such a context:
Cure? T Eerar TR 2%
(VP + T+ VB 37+ Y37 By theorem 5.4.2
< (14 1+ (e +b+e)((a® + The) + (b + Tea) + (¢ + Tab)) 3 3 3
z + ¥ z THy+z
and the result follows immediately from the inequality Cu+22 @raE @Rt 5
ab+be+ca
< a? b0+ a and the result follows from the AM-GM inequality. a
230 Chapter 5. Ty’s Lemma 5.A. Hélder’s Inequality in Action 231

A.5. Prove that if z3,29,...,2, are positive real numbers with product 1, Proof. This is a pretty tricky application of Hélder’s inequality:
then
41+ a1+ 1+ AP = (@ + 2+ B+ DB+ P+ 2P+ 1)
"t (1+xf )1+ 28) - (L +2f) SA+1+1+1)(@* + a2+ E+)
101 1\" > (1+ab+bc+ca)’.
la+at ottt —+—+F— 1] .
Ty E Tn
The result follows. O
Gabriel Dospinescu
A.8. Prove that for all positive real numbers a,b,¢ the following inequality
Proof. This follows easily by adding the inequalities holds
abe+ (@3 + DB + (@ + 1) 2 ab+ be + ca.
Proof. Using Hélder’s inequality, we can write

obtained from Hélder's inequality. 0 Y@+ DF+DE 1) 2 ab e

A.6. For any positive real numbers g, b, ¢ the following inequality holds We would like to have

%/a.a+b.a+b+c ., @+ Vab+ Vabe abe+ab+c > ab+bc+ca,

2 3 - 3
which is equivalent to (@ — 1)(b — 1) > 0. There is no reason for this to hold,
but at least one of the inequalities {a — 1)(b— 1) > 0, (b~ 1){c—~ 1} > 0 and
Kiran Kedlaya
(@ —1){c — 1} > 0 holds, as two of the numbers a — 1,5~ 1,¢ — 1 must have
the same sign. The result follows. a
Proof. Using theorem 5.A.1, we obtain

I vy » 3 A.9. Prove that for all positive real numbers a.b,c the following inequality
(a+a-+a) <a+ a—;-?—&vb> (a+b+c)> (aa» {,"3@(“+Q+ x“/dix,) . holds

(@ -+ P+ - +3) > (a+b+e)

One concludes by observing that \3/ ‘M‘:—H)) Z; Vab. o
Titu Andreescu, USAMO 2004
A.7. Prove that for all real numbers a, b, ¢ we have
Proof. Of course, we cannot apply theorem 5.A.1 directly in this case, but the
201+ a®)(1 + b3 (1 +¢?) 2 (1 +ab+
be + ca)®. form of the inequality is a temptation to find a way to apply theorem 5.A.1.
The key point is the inequality

Michael Rozenberg a®~a’+3>d+2,

232 Chapter 5. Ty’s Lemma 5.A. Holder's Inequality in Action 233

which is equivalent to (a* — 1)(a® - 1) > 0 and thus obvious. Using this, the so that after substituting z = 6%,y =42 and 2 = 2, it is enough to prove
the
result follows immediately from theorem 5.A.1, by writing inequality
(@+y+2)%>27(zy +yz + )@+t + 22).
S 42=ab+1%
1% 0
But this is immediate from the AM-GM inequality and the identity
A.10. Let a,b, ¢ be the sides of a triangle. Prove that
E2+y2+12+2(1’y+y2+2z):($+y+z)2. =]
abe abe abe .
E Zo+bte A.12. Prove that for all positive real numbers a,b,c.d
Vite—a i \/c+csz+\' a+b—(t‘a
41 P41 A1 Bl abed+1
Titu Andreescu, Gabriel Dospinescu
a+1 P+1 E+1 E+1° 2
Proof. This is a hard inequality. The point is to use Schur’s inequality Vasile Céartoaje. Gazeta Matematica

a-b)a—c) +bHb-a)b—e)+ Fle-a)c—b) 20, Proof. This is also a very hard problem. Of course, one
is again tempted to
use the theorem, but one needs a trick. The key point is that
for all positive
which can also be written as numbers a we have
a'3+l N “‘a‘1+1'
abela+b+c) 2 a®(b+e—a)+ P eta—b)+ Sla+b-c)
a2+1‘v 2
with the theorem 5.A.2. Indeed, observe that we can write Indeed, using the inequality

Za3(11+(7—a)—42——.—(> (@ + 1) < @+ )% + 1%,

which is just Cauchy-Schwarz, one reduces this to

The result follows. 2e®+ 17 > (a+ 1)2(a? + 1),

which, after division by (a +1)? and a small computation is equivale
A.11. Prove that for all a,b,¢ > 0 we have nt to

2 2 2y 4 (a—1">0.
(‘L+'L+f) > 27 (a + 54+ ).
b ¢ a Once we have this key inequality, the rest is an immediate application of
theorem 5.A.1. =]
Proof. If X is the square root of the left-hand side, then Holder’s inequality
yields ) s
X - (a®8? + BE 4 Pa?) 2 (a® + 02+ )Y,
 Chapter 6

Some Classical Problems in

Extremal Graph Theory

This elementary chapter is a variation on a classical topic in extre

mal
graph theory: Turan’s theorem on graphs withou
t cliques. Recall that if & is
a graph and k is a positive integer, then a k-cliq
ue is a set of k vertices, any
two of which are connected. A graph is called
k-free if it does not contain a
k-clique. We then have the following stand
ard result.
Theorem 6.1. (Turdn) The mazimal number of edges in a k-free
graph with
=2
n vertices is }% 4
n?—42
L2‘+ (;), where r is the remainder of n when
divided
byk—1.

In particular, the mzaximum number of edges in a k-free graph with n

vertices is at most H - %, which is really the estimate that we will constantly
use. We start with a series of rather direct applications of these two theorems.
The other problems are however different
in nature and more difficult.

1. Let z1.23,...,2, be real numbers.

2
Prove that there are at most nz
pairs (i,7) € {1,2,...,n}? such that i <jand
1 <|z;—z;} <2
MOSP 2001
236 Chapter 6. Some Classical Problems in Extremal Graph Theory 237

Proof. Consider the graph whose vertices are 1,2,...,n. Connect two vertices know 4;,,.... A;- Thus, if k-1959 — (k — 1) - 1999 > 1, we can always add
i, by an edge if (i,7) satisfies 1 < |m; — x| < 2." We claim that this graph one more person. If & < 48, then there are at least 79 people who know all
contains no triangle. If we manage to prove this, the result follows from of A ..., A;, and one of them is among Ay, As,.... Aiges. If k = 49, then
Turdn’s theorem. Suppose that the graph contains a triangle, thus we can there are at least 39 people who know all of Ay, Ay and (alth()llgfl that
find distinet a,b, ¢ such that person may not be among Ay, Az,..., Aigse) he completes a set of 50 all of
whom know each other. =}
1< |og—apf
<2, 1<|p—ad
<2, 1<|te—aaf <2
4. We ;re given 5n points in a plane and we connect some of them so that
By symmetry, we may assume that x, < 1z < z.. Then the previous in-
10n*+1 segments are drawn. We color these segments in 2 colors. Prove
equalitics become zp — T > 1,2 — Ty > 1, $0 ;. ~ zg > 2, contradicting the
jm} that we can find a monochromatic triangle.
inequality |z, — @e] < 2. The result follows.
Proof. Note that the corresponding graph (with vertices the 5n points and
2. Prove that if n points le on a unit circle, then at most segments
edges between points connected by a segment) contains a 6-clique. Indeed,
connecting them have length greater than V2. otherwise by Turén’s theorem it has at most —L 4 = 10n? edges, which
Poland 1997 contradicts the hypoth Now, consider a 6—chque G’ of cur graph G. The
edges of G' are colored in two colors and we claim that there is always a
Proof. Consider the E,raph whose vertices are the n points and connect two monochromatic triangle in G'. This is standard, but we recall the proof.
vertices if their distance is greater than V3. The point is that this graph Pick a vertex u; of G'. By the pigeonhole principle, we may assume that
contains no 4-clique. This is clear, since a chord of length +/2 subtends a V12,9193, U194 have the same color. If one of the edges vou3, vovy o1 ViU
central angle of §. Thus, by Turdn’s theorem the pumber of vertices is at has the same color as v;vs, we are done. Otherwise, the triangle vovgug is
{%} finishing the proof. o monochromatic and we win again. a
most

3. There are 1999 people participating in an exhibition. Out of any 50 Tlle following problem does not use Turén’s theorem, but it is still a very
people, at least two do not know each other. Prove that we can find at classical topic, namely Ramsey’s numbers.
least 41 people who each know at most 1958 other people.
Taiwan 1999 5. A group of people is called n-balanced if the following two conditions
are satisfied:
Proof. This is a special case of the proof of Zarankiewicz’s lemma. For the 2) among any three people, there are two who know each other:
reader’s convenience, let us recall the proof. Suppose that the conclusion does
b} among any n people, there are at least two not knowing each other.
not hold, so we can find at least 1959 people, say A1, ..., Aigso, each having
at least 1959 friends. Start with A;. There is a person among As, ..., Aigs0 Prove that there are always at most (—"M people in an n-balanced
that knows A;. Assuming that we found persons A, ..., Ay (i = 1) among group.
Ay,..., Argsg, every two knowing each other, let us try to add a new person
A;, i., to the group. But there ate at least k1959 — (k — 1) - 1999 persons that Dorel Mihet, Romanian TST 2008
238 Chapter 6. Some Classical Problems in Extremal Graph Theory 239

Proof. We will prove the result by induction. For n = 2, this is trivial, so Applying Cauchy-Schwarz and taking into account that
assume that it holds for n — 1. Consider an n-balanced group and pick an
n
arbitrary person P. Let A be the set of friends of P and let B be the set
of all the other persons in the group. By hypothesis, any two persons in B
S d(wi) =2k
i=1
are friends and B has at most n — 1 elements. By induction, 4 has at most
{n=B)ntt)
2 persons (by b) and the fact that P knows all clements of A, it vields the desired result. =]
follows that A is an n — 1-balanced group). Thus there are at most
The following problem is very similar to the previous one.
1){(n+2)
2 7. A graph with 7 vertices and k edges has no triangles. Prove that we can
persons in the group, which is enough to prove the inductive step. [m} choose a vertex such that the subgraph obtained by deleting this vertex

The following problem and the method of proof are absolute classics. and 2l its neighbors has at most k (l - —’; edges.

USAMO 1995
6. Prove that a graph with n vertices and k edges has at least %(M —n?)
triangles. Proof. Each edge zy gets killed when we remove z and its neighbors, when
APMO 1989 we remove y and its neighbors, or when we remove any common neighbor of
z and y and its neighbors. Thus, this edge is killed a total of d(z) + d(y)
Proof. Let ©1,2g,...,T, be the vertices of the graph G and let d(z) be the times. Summing over all edges and taking into account that d{z) appears as
degree of z. Observe that for a given edge ¢ = z;z; with endpoints z;,7;, a summand exactly d{z) times, we obtain that the average number of edges
there are at least d(z;) + d(z;) — n triangles containing . Indeed, there are killed by removing a vertex and its neighbors is 2 37, d(x)?. Since
d(a;) + d{z;) — 2 edges having as endpoints one of z;,z; and another vertex
among the n—2 remaining vertices, so there are at least d(z;)+d{(z;)—2—{(n—2) S d(z) =2,
triangles containing 2, z; as vertices. Summing over all edges and taking into z
account that we count three times every triangle in this way, shows that the
we have
number of triangles is at least (we denote by E{G) the set of edges of G)
1 2 (2%
PO (;) .
% S (dle)
+ dizg) - n). =
e=z,3;€B(G) Thus, on average we kill at least 47{‘; edges. The resuit follows. m]
The previous sum is also equal to
8. A graph G has n vertices and contains no complete subgraph with four
vertices. Prove that G contains at most % triangles.
% (id(z*)Q nk) Ivan Borsenco, Mathematical Reflections
240 Chapter 6. Some Classical Problems in Extremal Graph Theory 241

Proof. The result is easy for n = 2,3, 4, so let us assume that » > 4 and that find X € Fy, such that v4 = Mve. Then A = 1, as {v1,vg) = (v1,v9) = 1. We
the result holds for all k < n. We may assume that G contains a triangle ABC. conclude that v; = w4, a contradiction.
Let G be the subgraph formed by the vertices of G different from A, B,C. For the second part, take any prime p and set n = p?. The graph G,, in
Since (4 has no 4-clique, it has at most Q—;-Bfi edges by Turén’s theorem. By the first part of the problem has n vertices. Let us evaluate the number of
the inductive hypothesis, G has at most (l;-.,g)z triangles. Any other triangle edges. If v = (z,y) is a nonzero vector in V, then the equation (w,v) = 1 has
in G consists either of a vertex of G and an edge of ABC or of an edge of p solutions. Thus the number of edges is @ and it is immediate to check
(71 and a vertex of ABC. Moreover, any edge of (1 forms a triangle with at that this is greater than "T‘M —-n= %3 — p?. The result follows. ]
most one vertex of ABC and any vertex of G forms a triangle with at most
one edge of ABC, because (7 contains no 4-clique. Thus G contains at most ‘We continue with two very nice problems concerning graphs with n2 + 1
edges and 2n vertices. Note that this is the first case when Turdn’s theorem
-3 (=37 tn-3+l=1 ensures the existence of a triangle. The following results show that we can do
7 g tnTirlEg better.
establishing the inductive step. Note that the result is optimal if 7 is a multiple
of 3, since we can consider the tripartite graph Ky 3.5/30/3- [} 10. Prove that a graph with 2n vertices and n? + 1 edges contains two tri-
angles sharing a common edge.
1t is a standard result that a graph with n > 4 vertices and more than
Chinese TST 1987
mfif"& edges has a four-cycle (see, for instance [3], example 3, chapter 22).
The following nice problem shows that this result is almost optimal. Proof. We will prove the result by induction. For n = 2, this is trivial. Assume
now that the result holds for n and consider a graph G with 2n + 2 vertices
9. a) Let pbe aprime. Consider the graph whose vertices are the ordered
and n? 4+ 2n + 2 edges. By Turdn’s theorem and the pigeonhole principle,
pairs (z,y) with z,y € {0,1,...,p— 1} and whose edges join vertices
we can find a triangle zyz such that d(z) = d(y) (mod 2). Consider the 2n
(z,y) and (&',¢/) if and only if zz' + yy’ = 1 (mod p). Prove that
points different from z,y. If there are at least n® + 1 edges among them, we
this graph does not contain a 4-cycle .
are done by the inductive hypothesis. If not, then there are at least 2n 4 2
b} Prove that for infinitely many n there is a graph G, with n vertices edges whose endpoints are or y. But since d(z) = d(y) (med 2}, it follows
and at least "%’Z — n edges that does not contain a 4-cycle. that there are actually at least 2n + 2 edges whose endpoints are z or y and
Hungary-Israel Competition 2001 which are different from the edge xy. Let A; be the set of vertices v # 2,y
that are connected to « and let A, be the set of vertices v # x,y that are
Proof. One can give a rather down-to-earth proof of a), based on explicit com- connected to y. Then the previous result implies that |41} + Az} > 2n + 2.
putations, but we prefer the following approach. Consider the Fy, = Z/pZ- Since |A1 U 42| < 2n, it follows that |A; N 45] > 2 and so we can find two
vector space V = ]F;Z, and the standard inmer product {(z,y),(<,%)) = distinct vertices z # ¢ that are connected to both z,y. Then the triangies zyz
xz’ + yy'. We are asked to prove that we cannot find four distinet vectors and xyt share a common edge and we are done. O
v1,v9,v3,04 € V such that (vi,via) = 1 for all 1 < i < 4 (here v5 = vy).
If such vectors existed, vo and vy would be both orthogenal to the nonzero { 11. A graph has 2n vertices and n® + 1 edges. Prove that it contains at least
vector vy — v3 and so they would be contained in a line of V. Thus, we can n triangles.
242 Chapter 6. Some Classical Problems in Extremal Graph Theory 243

Proof. Again, the proof is by induction. For n = 2 this is trivial, so assume Proof. Let G be the graph whose vertices are the aborigines, two of them being
that the result holds for n and consider a graph with 2n -+ 2 vertices and connected if they are friends. Let B; be the set of aborigixles who receive a
n? + 2n + 2 edges. By Turan’s theorem we can find a triangle 1@ow3. Let bead of the i-th color. Then by condition ii) each B; forms a clique in G ax;ld
A; be the set of vertices v # 1,72,%3 that are connected to ;. There are by i) each edge is in one of these cliques. Conversely, if we have a collection
obviously at least of k cliques that cover every edge of G, then we can give a bead of color i to
every aborigine in the i-th clique and satisfy the conditions of the order.
§ = | AL N Ag] + 142 N As] + 143 N Ay The previous paragraph shows that we need to find the smallest number
F(n) of cliques that cover the edges of any graph on n vertices G. Clearly
triangles different from zywsxs, since any element, of 4; N A; forms a triangle
that fi ‘(2) = 1 and f(3) = 2. To find an optimal graph, consider a complete
with the vertices @;x;. Thus, if we have S > n, we are done. Assume
bipartite graph on n vertices. The key point is that such a graph
S < n — 1. Then, using the inclusion-exclusion principle, it follows that has no
triangles, so if we want to cover its edges by cliques, we need at least E(G)
2 [A U Ag U dgf > |Ai] + [ Ao + 1 45) —n 4 1,
- 12 cliques, where E(G) is the number of edges. Therefore, we have a clear bound

thus one of the numbers [A;] + {Az],|As] + |4a]. [As] + |41} is smaller than on f(n}, namely f(n) > [’7;3}
2n 1, say |A1] + |As|. But this means that there are at least - The hard part is to prove the opposite inequality. We will prove this by
induction, going from n to n + 2 (which, combined with the first two values
4242 (2n+ 1) mal+l of f(n), will be enough to conclude). Consider a graph G with 7 + 2 vertices.
Thus, by induction we find at We may assume that at least two vertices are connected, say z,y. Conside;
edges among the vertices Ag, Aq,. ., Aznsz.
the graph G obtained
least » triangles among these vertices and adding the triangle A;A2A3 yields by deleting z,y and all edges adjacent to these two
The inductive step is thus proved and the problem
velzrticcs, By induction, we know that we can cover its edges by f{n) cliques.
at léast n + 1 triangles.
o If v is a vertex of Gy, consider the subgraph spanned by z,5,v. By addmg
solved.
either a triangle or an edge, we can cover all edges of this subgraph incident
The following problems are more difficult. The next one concerns cover- to v. Finally, by adding one more edge to cover the edge zy, we covered all
ings of the edges of a graph by cliques. edges of G by using at most f(n) +n + 1 cliques (by the way, all the cliques
used were only edges or triangles). Therefore
T 12. There are n aborigines on an island. Any two of them are either friends
or enemies. One day they receive an order saying that all citizens should
fr+2)< fn)+n+1.
make and wear a necklace with zero or more stones so that
Since
i) for any pair of friends there exists a color such that each of the two (n+202]_ [a?
persons has a stone of that color; | = |7t
ii) for any pair of enemies there does not exist such a color.
it follows that f(n) < ['—,}z] Therefore the answer is {"Tz} 0
What is the least number of colors of stones required (considering all
possible relationships between the inhabitants of the island)? It 1: f)ot hard to guess the answer of the following problem, but proving
Belarus 2001 that this is the correct answer is quite a pain in the neck.
244 Chapter 6. Some Classical Problems in Extremal Graph Theory
13. What is the least number of edges in a connected n-vertex graph such
that any edge belongs to a triangle?
Paul Erdés, AMM E 3255
Proof. Let f(n) be the desired number and let g(n) = @} ‘We will prove
that f(n) = g(n) for all n. To save words, call good a connected graph in
which every edge belongs to a triangle.
We can easily establish that f(n) < g(n) by explicit eonstructions: if
n = 2k + 1, consider & triangles sharing a common vertex, this being the only
common point of any two triangles. This is & good graph with 3k edges, so
F(2k+1) < 3k = g(2k + 1). If n. = 2k, start from the above good graph with
edges AX, BX. We
order 2k — 1, choose an edge AB and add a vertex X and
obtain a good graph with n vertices and g(n) edges.
The hard point is proving that any good graph on n vertices has at least
g(n) edges. We will prove this by strong induction. For n = 3,4, this is easily
checked. The crucial observation that makes the induction work is that a good
graph with n vertices and less than g(n) edges must have a vertex of degree
2. This is trivial, since by connectedness all vertices have degree at least 1
and since every edge is in a triangle, every vertex must have degree at least 2.
However, the sum of the degrees of all vertices is twice the number of edges,
thus smaller than 3n.
Suppose now that f(j) = ¢(§) for all j < n and let us prove that f(n) >
g(n). Suppose G is a connected graph with f(n) edges such that each edge
belongs to a triangle and suppose f(n) < g(n). There is a vertex x of G of
degree 2, which by assumption must be in some triangle zyz. Suppose the edge
yz is in a second triangle. Then the graph G obtained from G by removing =
and the edges zy and xz is connected, has n — 1 vertices, every edge belongs to
a triangle, and f(n) — 2 < g(n) — 2 < g(n — 1) edges. This is a contradiction.
Thus yz is not contained in another triangle. Form a new graph G” by deleting
the vertex @ and collapsing the vertices y and z into a single new vertex w,
where a vertex v of G is joined to w if in G it was joined to either of y or z.
Clearly G" is a connected graph with 7 - 2 vertices and every edge of G” is
contained in a triangle. In forming G, we lost three edges y, yz, and zz. We
would also lose an edge if some vertex v were adjacent to both y and z, but by
245
assumption this does not occur. Thus G” has f(n) — 3 < g(n) ~3 = g(n —2)
edges. Again this contradicts the induction hypothesis. Thus we must have
£() 2 g(n). g
The following result is classical and very nice. We give two proofs, the
first one being an explicit inductive construction, the second one using the
powerful probabilistic method of Erd8s. We refer the reader to the addendum
6.B for more details on this versatile tool.
14. Prove that for every n there is a graph with no triangles and whose
chromatic number is at least n.
Mycielski’s theorem
Before starting the proof, let us recall some definitions. A k-proper coloring
of a graph G is a coloring of its vertices with at most & colors such that each
vertex receives one color and no two adjacent vertices have the same color.
The chromatic number of a graph is the smallest number & for which G has a
proper k-coloring.
Proof. We will prove this by induction on n. Forn = 1 everything is clear,
so assume that we have a graph G with no triangles and chromatic numbe;
at least n and let us construct another graph with no triangles and chromatic
number at least n+1. Let x(G) be the chromatic number of G. We may assume
that x{G) = n {otherwise keep G for the inductive step). Let vy,...,v; be
the vertices of G and consider the following new graph G': the set of vertices
consists of vy, ..., vk, together with vertices z1,..., 24,y {such that the 2%k + 1
vertices thus obtained are distinet). Connect z; with all neighbors of v; and
connect y with z,...,2z. Also, keep the edges in G. We claim that &' is
triangle free and x(G') > n + 1. It is clear by construction that G’ has no
triangles. Also, we easily have x{G'} < x(G) + L, since any proper n~coloring
of G extends to a proper n + l-coloring of G' (assign the same color to T; as
to v; and assign the color n + 1 to y).
The difficult part is proving that this is actually an equality. Suppose
thus that G' has chromatic number at most n and take any pmpe; n-coloring
of . We will construct a proper 7 — 1-coloring of G, which will contradict
246 Chapter 6. Some Classical Problems in Extremal Graph Theory 247

the fact that (7 has chromatic number n. Assume, without loss of generality, independent vertices in ('Z) ways and the probability that a given set of a
that has color n and let A be the set of vertices of & whose color is n and vertices is independent is certainly (1 — p)(;)) An easy computation shows
for each v; € A change its color with that of z;. We claim that this gives a that the last quantity tends to 0 as N — oo. Thus for N sufficiently large
proper n — l-coloring of G. Note that no two vertices of A4 are adjacent, as we have P(a(G) > @) < } and so there is a graph G, n = G with X < 12‘5
all these vertices have the same color. Now, if v; € A is adjacent to v; £ A, and (@) < a. Now, delete one vertex from each short cycle arbitrarily. The
then v; is adjacent to «;, 5o that they have different colors. Consequently, we remaining graph Gy has at least N/2 vertices, no cycles of length at most n
found a n — 1-proper coloring of G, which is impossible. We deduce that the and independence number smaller than ¢. Thus
chromatic number of G’ is at least n + 1. ]
N/2
Proof. We will actually prove a stronger result, which is a classical theorem of X(G1) = NI N~ k
Erd6s. We will use the probabilistic method, for which the reader is referred
for sufficiently large N. The result follows. a
to the addendum 6.B.
‘We end this chapter with three gems in extremal graph theory, all taken
Theorem 6.2. For any positive integers k,n there exists a graph with chro-
matic number greater than k and such that each cycle has length greater than from mathematical contests.
. W15, For a pair A = (z1,51) and B = (22,12} of points on the coordinate
Take N sufficiently large and consider random graphs with N vertices plane, let
G p, each edge appearing independently, with probability p = Ne1 for some (A, B) =z — o+ ly1 — el
0 < c< % If X is the number of cycles of length at most n in Gy (which We call a pair (A, B) of (unordered) points harmonic if 1 < d(4,B) < 2.
cycles we will call short in the sequel), then clearly
Determine the maximum number of harmonic pairs among 100 points
n ren in the plane.
BX] <3 N pf< f?fi;; USA TST 2006
i3
Proof. Consider the graph with vertices on the 100 points and whose edges
since there are at most N cycles of length 4, each appearing with probability
conuect points of harmonic pairs. The key point is that this graph contains no
. Since the last quantity is smaller than ’:— for sufficiently large N, it follows
5-clique. Indeed, if P(z;,y;) are five vertices, any two of which are connected,
by Markov’s inequality that
then for all i # j we have 1 < d(F;, P;) < 2. Order these points such that
N 1 @ < ¢ < -+ < 5. It is easy to see that among the numbers ¥1,v2,....45
P
(r=3)<3
> =)<z one can find three forming a monotonic sequence. Call these three numbers
Yiys Yizs Yig- Then
Let a(G) be the size of the largest independent set in G (the independence
number of the graph) and let x(G) be the chromatic number of G. Then it APy, Pg) +d(Py, Py} = d(Py, Pig)-
is easy to see that x(G) > ;{%) if G has N vertices. Taking a = [% InN], This is however impossible, since d{(F;,, P;,), d(P;,, Py,),d{(Ps,, Pi) € (1,2
the probability that a(G) > a is at most (%) - (1~ )& (one can choose a This finishes the proof of the fact that our graph contains no Ks. By Turdn’s
248 Chapter 6. Some Classical Problems in Extremal Graph Theory 249

theorem, it has at moat —— = 3750 edges and so there are always at most ;. Summing over ¢ and taking into account that we count each triangle three
3750 harmonic pairs. times in this way, we obtain that
To finish the proof, it remains to exhibit a configuration having 3750
harmonic pairs. It is enough to distribute the 100 points near the points % /_ Zd(:c.)—fe\/r
{0, i%-) and (i%,()), having 25 points near each of these four points. O
Note that the constant appearing in this estimate in optimal, by taking for
16. For a finite graph G let f(G) be the number of triangles formed by the
G = K, and then letting n — co. This already shows that we are on the right
edges of G and let g(G) be the number of tetrahedra formed by the edges
path, since we solved the two dimensional version of the problem.
of G. Find the least constant ¢ such that g(G)® < c- f(G)* for any finite
To solve the original problem, all we have to do it to repeat some of the
graph G.
arguments in the previous paragraph. Namely, take a graph G with vertices
IMO Shortlist 2004 T1,..-,T, and fix a vertex z;. The number of tetrahedra containing z; as
a vertex is the number of triangles in the subgraph induced by the vertices
Proof. Let us begin by considering the case of a complete graph K. Then connected to ;. Let e; be the number of edges in this subgraph, then the
obviously previous paragraph shows that there are at most %e‘\/% tetrahedra containing
@; as vertex. Also, note that we have 3f(G) = }_e;, since for each vertex ;
there are ¢; triangles containing z; as vertex (and we count each triangle three
Thus s ( )3 times this way). However, one still needs an observation to end the proof: we
& (G gave an estimate for the number of tetrahedra having z; as vertex in terms
CZIER ) of e;, but there is also an obvious estimate: this number is at most (G
Therefore, there are at most {/ % f(G)e; tetrahedra containing z;. Summing
and this happens for all n. Taking the limit as n — o¢, we deduce that ¢ > 312
over ¢ and taking into account that we count each tetrahedron four times, we
The hard part is to prove that the value % actually holds for arbitrary
finally obtain
graphs G. To do this, we will first consider the two dimensional version of the
problem, which is comparing the number of triangles and edges in a graph. —
2 /2
The reason is simple: computing tetrahedra in a graph ¢ comes down to LGRS §/ 5/ (@ = {/37(6) - 3f(0).
computing triangles in all subgraphs induced by the vertices of G and taking
the sum of these numbers of triangles (and then dividing by 4, since each Taking the cube of the last inequality yields the desired estimate
tetrahedra is counted four times).
Therefore, consider first any graph G with n vertices z1,22,...,&, and
¢ edges. Let us bound the number of triangles in terms of e. If d(xz;) is
<y
g (G)<
the degree of the vertex ;, then there are at most (d(;‘>) < 1%‘)3 triangles and ends the proof. a
containing the vertex z;. But the number of triangles containing the vertex
z; is also obviously bounded by the number of edges of G, that is ¢. Thus for W17, Let k be a positive integer. A graph whose vertex set is the set of positive
each vertex ;, there are at most d(x;) - \/§ triangles containing the vertex integers does not contain any complete k x k bipartite subgraph. Prove
250 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.1. Notes 251

that there are arbitrarily long arithmetic progression of positive integers Fix a large integer N and consider all arithmetic progressions
with m terms,
such that no two elements of the progression are joined by an edge in all among 1,2,..., N. There are at least

S
this graph.

Proof. We will first prove the following result, of independent interest:

KoMaL
a=1 E G- -des e
a=1
such arithmetic progressions, since each is determined by a pair
(a, d) of posi-
Lemma 6.3. Let a > 2 and let Ko, be the complete bipartite graph with o tive integers such that a + (m — 1)d < N. This is greater than c(m)N? for
vertices in each half of the partition. There is a c(mstmlxt ¢ > 0 such that ¢ N
sufficiently large, where ¢{m) > 0 is a constant depending only
on m. Since
graph with n vertices and without K, has at most n?s edges. each of these progressions contributes an edge to the graph
Gy (the subgraph
of G induced by the vertices 1,2,...,N), and since each edge
Proof. Let G be such a graph and call 1,2,.. its vertices. Let d; be the is counted at
most m? times, we deduce that Gy has at least %’?1\:’2 edges for
degree of vertex 4. Let us count pairs (v, {v1,...,va}), where {v \Ug} i all suffi-
a set of a (distinct) vertices, all connected to v. For each v = ¢, there are ciently large N. Since E}n’l})N 25 eN?TF for all sufficiently large N (and
this
precisely ('f[) sets {v1,...,v,} sharing a pair with v. Thus we find 37, (i‘) for any constant ¢ > 0}, it follows by the previous lemma
that G contains a
pairs in fotal. On the other hand, for each set of a vertices {v1,...,u,} there K for all sufficiently large N. Since this contradicts the hypothes
is on G,
are at most a — 1 vertices v sharing a pair with {1,...,7,}. Thus the conclusion follows.
o

£(0)<e ()
=1
6.1
We
Notes
would like to thank the following people for providing
some of the problems presented in this chapter:
solutions to
Let ¢ be the number of edges in the graph and let f(z) = (i) fz>a—1 Alon Amit (problem 12),
and f(z) =0for 0 < < a— 1. Then f is convex and since »_d; = 2g, we Alexandru Chirvisitu (problem 10), Xiangyi Huang (problem 3), Szymon Ku-

()2 (2).
deduce that 5 bicius (problem 4), Joel Brewster Lewis (problems 2, 7, 15), Fédj; Nazarov
(problem 14), Richard Stong (problem 13). Gjergji Zaimi (problem 5).
a n

This yields the rough estimate

«
a»n“Zn<E7a+l) s
n

from which the result follows immediately. O

Coming back to the proof, fix an integer m > 1 and suppose that among
any m vertices forming an arithmetic progression, at least two are connected.
 6.A. Some Pearls of Extremal Graph Theory 253

Addendum 6.A Some Pearls of Extremal one common point for all i # j. Then the number of pairs (i,7) such that
Graph Theory P, € C; is at most m + 4max(n, (mn)%),
In applications, the following immediate consequence is very handy:
The purpose of this addendum is to present some besutiful theorems
from extremal graph theory, that nicely complement the more elementary Corollary 6.A.2. Let S be a set of n points in the plane, L a collectio
n of
results discussed in chapter 6. The main result is the famous Szemerédi-Trotter curves such that any two curves have at most one common point. Suppose
theorem, a deep result in incidence geometry which has a lot of wonderful that each curve in L contains at least k > 2 points of S. Then
the number
geometric consequences. For instance, it is the basic tool in dealing with of pairs of a point in S and a curve in L containing that point
2 is at most
natural (but very bard) questions such as; if we are given n points in the plane, 29 max (n, %{)
how many distinct distances do they always determine, what is the greatest
number of segments of length 1 (or triangles of area 1) they determine, etc. Proof. Let [ = ]L)J and l(;t ? be the number of such pairs. The theorem
Thanks to a brilliant observation of Elekes, it also plays an important role in gives p < [ + 4n% max(n3,15). By hypothesis, we have p > Uk, therefore
additive combinatorics, giving nontrivial bounds on the so-called sum-product
p=izp(1-}) > £. Hence our inequality becomes p < st mm:(ml,(p/k)%).
problem, a major research problem in modern combinatorics.
From this the result follows by considering two cases.
The results discussed in this addendum found wide ranges of applications O
and their extensions are a very hot topic of research. See the papers [1], {11], It is important to note that corollary 6.A.2 {and also Szemerédi-Trotte
r’s
[15], [22], [23], {36], [74}, [72], [73}, {75], [80], [81], and the excellent book theorem) is sharp up to a constant. Suppese that 2k2 < n and consider
the
[82] for more details, as we will only scratch the surface of the subject (but set of points (.y), where | <z < k,i<y< % and z,y are integers. Also,
hopefully that will be enough to convince the reader of the beauty of these consider the set of lines y = mz +b, where m, b are positive integers
such that
results). Also, some proofs in this addendum use the probabilistic method, so 1shb<Fandl<m< shz- It is easy to check that any such
line contains
the reader is invited to take a look at addendum 6.B for details on this method exactly & such points (namely all points (i, mi+b) with 1 < i < k). Moreover,
and for probabilistic vocabulary. there are about » points and about fizg lines. Also, if k > /n,
one can simply
pick about »/E lines and put % points on each of them.
6.A.1 The Szemerédi-Trotter theorem The original proof [81] of theorem 6.A.1 was very intricate. In a beautiful
paper [80], Székely gave an amazing proof using a graph-theoretic result
The famous Szemerédi-Trotter theorem deals with the number of inter- on
crossing numbers. We will follow $his path, but in order to do that we first
sections between geometric objects, giving a fairly nontrivial (and even sharp
need some terminology. Let G be a graph and consider an injective map
up to absolute constants) upper bound for the number of incidences between that
sends vertices of G to points of the plane. Draw an arc (in the plane) between
a family of points and a family of curves. Since we do not want to delve into
any two points that come from the endpoints of an edge of G, such that except
subtle topological considerations, let us agree that curves will always mean
for the endpoints this arc does not meet any vertex. Call such a map
arcs of circles or polygonal lines, as this appears in all applications we have in a drawing
of G. A crossing (or crossing point) is simply the intersection of two arcs
mind. not
at the image of a vertex. Such a point belongs therefore to at least two
arcs,
Theorem 6.A.1. (Szemerédi-Trotter) Consider n points P, Py, ...,P, and but it is not an image of a vertex of the graph (with a piece of paper and a
m curves C1,Ce,...,Cp 0 the plane. Suppose that C; and C; have af most pencil in front of you, all these definitions should be obvious!). The number
 254 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.A. Some Pearls of Extremal Graph Theory 255

of crossings is the total number of crossing points, counted with multiplicities Lemma 6.A.5. If G is a graph with e edges and n vertices, then ¢(G) > e~3n.
(i.e. if a crossing point belongs to k arcs, its multiplicity is (’7‘)) It is also the
Proof. First, we reduce the proof to the case when G is planar. To do this,
number of pairs of edges with no common endpoints and whose associated arcs
remove edges of G which cross other edges until this is no longer possible, so
intersect each other. After this dry list of obvious definitions, let us glorify
you end up with a graph G’ which is planar. Suppose that you removed k
one that will be very helpful in what follows:
edges of G. Each of them removes at least one crossing, so that if we accept
Definition 6.A.3. The crossing number of a graph G, denoted ¢(G), is the the truth of the lemma for G, we can write
minimal number of crossings over all possible drawings of G.
G2 c(G)+k>e(G)+k—3n=e—3n
Note that, by definition, a graph is planar if and only if its crossing mumber
is 0. The key ingredient in Székely’s proof of theorem 6.A.1 is the following Now, assume that G is planar, so ¢(G) = 0. We need to prove that ¢ < 3n.
deep theorem [1], that will be proved in the next section. Ifn < 2, this is clear, so suppose that n > 3. Let f be the number of (possibly
infinite) faces of G. By Euler’s formula we have n + f = e + 2. As every face
JTheorem 8.A.4. (Ajtai, Chvdtal, Newborn, Szemerédi, Leighton) Let G be
has at least three edges, we have 3f < 2e. We easily deduce that ¢ < 3n — 6,
a simple graph with ¢ edges and n vertices. If € > 4n, then «(G) > 6—;’:;,.
finishing the proof of the lemma.
Let us see why theorem 6.A.4 implies theorem 6.A.1. We may assume that Finally, let us recall the proof of Euler’s formula: if you start with a single
every curve C; contains at least one of the points Py, Py, ..., P,. Consider the vertex and try to reconstruct the graph, then every time you add an edge, you
graph G whose vertices are Py, P, ..., P, and whose edges connect adjacent either add a vertex or a face, so the quantity n + f — e — 2 does not change.
points on some C; (i.e. two points P;,P;, belonging to some Cj, such that Since initially it is obviously zero, it is zero all the time. in}
there is no other point P, between them on ). Let e be the number of edges
Now, we use the probabilistic method to improve the previous inequality.
of G and let ] be the number of incidences between points and curves (i
Take an arbitrary number p € {0, 1] and consider a random induced subgraph
the number of pairs (4, 7) such that P; € Cj). As every curve contains at least
H of G, by picking each vertex of G independently and with probability p. As
one point among Pi, Py...., Py, we have e = I —m (since if a curve contains
the probability of a given vertex to be in H is p, by linearity of expectation
5 points, it yields s 1 edges of G and edges coming from two different curves
we have E{v(H)] = np, where v{H) is the number of vertices of H. Also, since
are distinct). Now, since two curves intersect in at most one point, it follows
an edge appears in H with probability p?, we have Ele(H)] = p2, where e
that ¢(G) < m* If e < dn, we deduce that I < m + 4n and we are done.
(respectively e(H)) is the number of edges of G (respectively H).
Otherwise, the previous theorem yields m? > %;7 thus e < 4(nm)§ and so
The previous lemma and linearity of expectation yield
T < m+4(nm)}. The result follows.
Ele(H)] > Ele(H)] - 3E[(H)].
6.A.2 Proof of theorem 6.A.4 and a generalization
‘We cannot easily express E[c(H)] in terms of ¢(G), but we can at least say that
The proof of theorem 6.4.4 is simply beautiful: we will start with a rather Ele(H)] < p'c(G). Indeed, take a drawing of G with exactly (@) crossings
weak inequality obtained from Euler's formula and, using the probabilistic and observe that the probability that a crossing survives in H is p. The
method, we will improve it drastically by averaging. Let us start with the conclusion is that
weak inequality: pe(G) 2 Ele(H)] > p'e — 3np.
256 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.A. Some Pearls of Extremal Graph Theory 257

As p was arbitrary, it is tempting to choose a value that will optimize the crossing.s:. However, one has to pay attention to the fact that each
crossing is
previous inequality. This valueis p = gie‘, but unfortunately it is not necessarily counted in 20-1U(e~2) ypembers of the family, so in total we obtain at least
smaller than 1. However, this is far from being a subtle jssue: simply choose
something a bit smaller, namely p = 47:”. This finishes the proof of the theorem. oli=1)e; e? 4i-1g3

In geometric applications, it is useful to have a more flexible version of 26D -

theorem 6.A.4, which allows graphs with multiple edges. This is the objective
crossings in G;, proving the claim.
of the following theorem of Székely [80].
Finally, by definition of the crossing number, it is clear that
Theorem 6.A.6. Let G = (V, E) be a multigraph with n vertices and e edges.
If the maximal multiplicity of the edges of G is m, then e < 32nm or ¢(G) 2
& A2 3 el 2 Yoy = YD 4H§? - # 3ok
FTnTay * €8 €S 64 i€s

Proof. The idea is to reduce everything to the case of simple graphs, where Now, since
theorem 6.A.4 applies. The details are however a bit involved. For each ’ ez IBG) 2
1 <4< loggm +1, let Gi = (V, E;) be the subgraph of G using only those ' €S €S
edges with multiplicity between 2! and 2* (more precisely, two vertices a,b

(5o (59 () 5
it is natural to use Hélder’s inequality in the form
are joined by k edges in G; if they are joined by k edges in G and k € (2071, 2%)).
Let €; be the number of edges in G, without multiplicity. In order to apply
theorem 6.A.4, we will restrict to those i for which e; > 4n. Let S be the set
of these i. As G; has maximal multiplicity at most 2%, we have (under the €S €S €S
assumption that e > 32mn, which will be made from now on)
Combining this with the easy estimate
1{logy (m)] e
Y16 = > a2 ze—16mm> 2 ;
S
- Lflog,(m)]
Y 2 cuvim
€S ge]
€S i=1
yields
1 &3 3
(G) > g e = e
“(G) 28n28.32m 216y
finishing the proof.
Indeed, pick an arbitrary drawing of G; and for every pair of vertices of G; with o
a multi-edge between them, select 227" of these edges and then arbitrarily and
independently only one of them. We get a family of (2~!)% simple graphs, 6.A.3 An application to additive combinatorics
each having ¢; > 4n edges (as i € §) and so with crossing number at least . In [23] Elekes made a nice connection between theorem 6.A.1 and & famous
]
sz~ Thus, over this whole family of graphs we obtain af least 2("”“6%; problem of Erdés and Szemerédi, the sum-product problem. This asks for a
258 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.4. Some Pearls of Extremal Graph Theory 259

sharp lower bound of the expression max(]4 + A|,|A4 - Af) over all sets of real The variant of the sum-product problem in which instead of sets of real
numbers A with n elements. Here numbers one considers subsets of a finite field has generated
a huge amount
of work. The proofs are in general very technical, as
A+ A={a+blabecA}and A-A={a-bla,beA} Szemerédi-Trotter theorem over finite fields is wrong.
the analogue of the
One has nevertheless
the following deep theorem [11].
In particular, they made the very deep conjecture that for any ¢ > 0 there
exists ce > 0 such that for all sets A with sufficiently many elements we have Theorem 6.A.8. (Bourgain, Katz, Tao, Glibichuk, Konyagin)
Let £ > 0. There exist positive constants C, & such that for any
prime F3
max(jA + AL, A A]) 2 c. - |AP and any A C Fp with |Af < p'~ we have

The following result treats the case when 2 — ¢ = % In (72}, Solymosi proves max({A + A}, |4 - Af) > ClA]H.
the case 14
2 — ¢ = 3.
The analogue of the Erdés-Szemerédi conjecture for finite
fields would be
Theorem 6.A.7. (Elekes) For any finite set of real numbers A we have
max(|4 + A, |A- A]) > c(e) min(JA]>~¢, g'~¢)
; LRV .
5A+A|-gA-A|>32§A} for subsets A of F,. This is far from being settled. The
following theorem can
be deduced from the sum-product theorerm.
Proof. Consider P = (A+ A) x {A- A) as a set of points in R?. Let L be the. Theorem 6.A.9. (Bourgain, Katz, Tao) Let 0 < o« < 2. There ezists ¢ > 0
set of lines 1, , with equations y = a(z - b} ranging over all choices ()f el‘emen?s
and C' > 0 such that there are at most Cn3~ incidences between n < p*
a,bof A. Note that all lines lop witha € A~ {0} and b€ A are distinct, ZO
points and n < p® lines in Fp.
1L} > {Al(lA] ~ 1). Also, I,y is incident with (b+c,a-¢) € P, for any c € . .
Thus we have at least |A] - L] incidences between L and P. Thus, by the Note that the previous theorem no longer holds if one
considers subsets
Szemerédi-Trotter theorem, we have A of Fy, where ¢ is not a prime number. It suffices to
consider a nontrivial
subfield of Fy, for which |4 + A] = |4} and |4 - Al = A]. Also,
it is necessary
|A]-{2] < L} + dmax(|P], (1P]- 1LDP). to have some upper bound on |4, since otherwise 4 —
F, would be again a
counterexarple for p large enough.
1 {P{ > |LP, then clearly
6.A.4 Some geometric applications
[Pl > AR (1Al - 1) 2 In this section, we present some nice geometric applic
ations of Szemerédi-
and we are done. Otherwise, the previous inequality yields
Trotter’s theorem. The first result deals with the number of triangles of unit
area spanned by n points in the plane. It is not difficul
t to construct examples
, 32 ST,
Pz (M) Vi VA
i which there are at least en? log n unit-area triangl
> e ¢ >0, but it is rather hard to give nontrivial upper
of such triangles. We will use an easy geometric argume
es, for an absolute constant
bounds for the number
nt combined with
and the result follows. corollary 6.A.2 to give such an upper bound.
 260 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.A. Some Pearls of Extremal Graph Theory 261

Theorem 6.A.10. There ezists an absolute constant ¢ > 0 with the following points of §. Thus the number of pairs of points that lie on average lines is
property: the number of tflan_qles of area 1 with vertices among n points in the bounded by
plane is smaller than cns.
: non 2 n2
" Proof. Let P be aset of n points in the plane. The essential observation is the 29 E 4" i+1 max (2_'§) <olt Z ?+ Z 2.5
following: if a,b are points of P, then all points p € P such that the triangle 14<i<log, ¥y 14<i<logy n logy n<i<log, 3fr
abp has area 1 are on the union of two fixed lines Lo p, L;_b. parallel to ab. This
is immediate. Now, consider those pairs (a,b) for which the lines L and Lf,, <2l ZinrZQ‘n <fi%+ffifif
contain at most 7§ points of P apiece. Each such pair determines at most 2n 13 L
214 2
7 i<log,2N 4 42
unit-area triangles and so these pairs contribute at most 20t . pl/3 = 2073
unit-ares triangles. On the other hand, by corollary 6.A.2, there are O(n%/%) proving the claim.
So we have at least n2/2 pairs of points of S that do not determine
incidences between lines | containing at least n'/? points of P and points of an
average line. Now, we have two possibilities: either some line contains more
P. However, given such a line I, there are O(n) pairs {a,b) such that I = Lgp
(as for any @ and [ there are at most two points b such that { = L,p). Thus than /2™ points of S, in which case the result is proved, or actually the n?/2
pairs (a,b) for which at least one of the lines Ly and L:L,h contains more than
pairs of points determine only lines that contain less than 2!4 points of S. But
then each such line contains at most 2% pairs of points of S, so there must
n1/3 points of P also determine O(n-n%/3) = O(n"/?) unit-area triangles. The be
O at least n?/2%% such lines. So again the result is proved. o
result follows.
We continue with another rather natural question: what is the maximal
Using a lot more work, one can improve the previous result to Onfite)
number of unit-distances spanued by » points in the plane? Erdés proved
for any & > 0 as shown in {22]. The exact order of growth of the number of unit-
the lower bound n!*Feien and the upper bound ¢n3/2 and conjectured
area triangles is unknown. If instead one considers the number of triangles (’,hé:\t
the true order of growth should be about r' " #=isn . This is far from being
with perimeter 1, the same kind of reasoning (but working with incidences
proved, but the following theorem gives a better upper bound. In dimension
between ellipses and points) yields & bound O(n'%/7).
3, Clarkson proved that the number of unit distances is O(n%*'f) foralle >
The second application is a famous result of Beck. It is again an easy
0, while Erdés, Pach and Hickerson gave examples with at least cn?/2
consequence of corollary 6.A.2. unit
distances. Also, note that already in dimension 4 one might have cn? unit
Theorem 6.A.11. (Beck) For any n poinis in the plane, either more than distarxces, as shown by placing n points on :c% + 22 = 1/2 and n points on
12714 of them are on a line or these points determine at least 2~2n? different 2} +23=1/2.
lines.
Theorem 6.A.12. (Spencer, Szemerédi, Trotter) Let § be a set of n points
Proof. Let S be a set of n points in the plane, Call a line average if it contains in the plane. Then there are e at most 16n*3 ordered pairs of poi ts (P, i
between 2!* and n/2' points of S. We claim that there are at most n?/2 pairs S such that PQ= 1. et A
of points of § that determine average lines. Indeed, for each i € [14,logz 5]
Proaf, Draw 2 unit circle around each point of S and consider the multigraph
there are at most 2° - max 308 )
(%, lines that contain between 2 and 2i*+1 G with vertex set § and in which P,Q € § are Jjoined if they are adjacent on
points of 8, by corollary 6.A.2. Any such line contains at most 4+ pairs of one of these circles. Note that there might be more than one edge between two
262 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.A. Some Pearls of Extremal Graph Theory 263

vertices and that G might have Joops (if there are circles containing exactly Theorem 6.A.13. (Székely) Any set of n points in the plane determines at
one point of §). If ¢ is the number of unit distances between points in S, least cn®/® distinct distances, for an absolute constant ¢ > 0.
then G has q edges, for if a circle centered at P € S contains z; points of
S, these points contribute z; edges to . Now, remove all circles containing
Proof. From now on c is an absolute constant that will change throughout the
at most two points of S, so we get a new multigraph with at least ¢ — 2n
proof without changing its name.
edges. We claim that the maximal multiplicity in this multigraph is at most
Fix n distinct points Pi, P3,..., P, in the plane and let ¢ be the number
2. Indeed, if there are at least 3 edges between P s Q € S, then there are at
of distinct distances among them. Let dy,dp,...,d, be these distances. We
least three circles containing P, @, so the circles of radius 1 centered at P,@
may assume that ¢ < n15, as otherwise we are done. Around each point P;
have at Jeast three common points, a contradiction. Next, for every pair of
draw ¢ circles, having radii dy,d; . d¢. Consider now the multigraph whose
vertices with exactly two edges between them, remove one edge (arbitrarily),
vertices are F,..., P, and whose edges are arcs connecting adjacent vertices
so that we end up with a simple graph having at least g ~ n edges. Now, if
on these circles. Note that there are at least n(n — 1) edges in this multigraph
g > 10n, then this simple graph has more than max($,4n) edges and so at
(for each P, there are n — 1 points on the union of the circles centered at I
least (g/4)%/(64n?) crossings by theorem 6.A.4. But since any two circles have
and if a circle contains s points, it contributes s edges). As usual, we remove
at most two common points, there cannot be more than 2{}) < n? crossings.
edges that come from circles containing at most two points. It is easy to see
The conclusion is that if ¢ > 10n, then ¢* < 4096n* and the result follows (if
that this removes at most 2n¢ edges, so we still have n? + o(n?) edges.
g < 10m, everything is clear). a
Unfortunately, the maximal multiplicity might be too big for the previous
theorems to yield the desired estimate. The crucial point is to show that there
‘We end this addendum with a rather technical result concerning a famous cannot be too many edges with very high multiplicity. This is the aim of the
following lemma. Before stating it, call a pair of points k-rich if there are at
question of Erdds: what is the least number of distinct distances determined
by n points in the plane? Though rather innocent-looking, this is an extremely
least k edges between them.
difficult problem. Erdds gave examples of configurations which determine at
distances, for some absolute con%am ¢. For more than thirty Lemma 6.A.14. There are af most ;k-fg_’ +ctnlogn edges joining k-rich pairs
of points.
years, the best result was Moser’s bound cn?/ 3 which is a rather tricky, but
very elementary argument. Note that this bound also follows from the previous
theorem: there are at least "22 ordered pairs of points and the previous theorem Proof. By definition, this number of edges is at most the number of pairs
{combined with a rescaling argument) shows that each distance appears at (d, e}, where d is the symmetry axis of an edge e of G such that d contains at
most 16743 times, thus the number of distinct distances is at lea cast k vertices. If 20 < \/n, there are at most "‘2 lines containing at least 2°
only in 1984 that Chung [15] improved Moser’s bound to en®/7 3 The next big by corollary 6.A.2. If such a line concams « points then, by definition
line bisects at most 2tu edges. So, as long as we are counting pairs
step was done in the paper {16], where bounds of the form [
(d, e) such that d contains between k and 4y/n vertices, the total number of
by rather difficult arguments. After a huge effort, a major breakthrough was
such pairs is at most
made in{36] where Guth and Katz prove that » points always determine at
2
least 2% distinct distances. Here, we prove a much weaker estimate, but i1 T
which is alruvdy highly nontrivial {following [80}). >
k/2<2<dym
wtg
264 Chapter 6. Some Classical Problems in Extremal Graph Theory

and an easy computation shows that this is bounded by Qk%f On the other Addendum 6.B Probabilities in
hand, for each a > 4y/n, it is easy to see that there are at most £ lines
Combinatorics
containing between a and 2a vertices. These lines yield at. most
This addendum presents some applications of probabilities in combina-
Z ;—7&‘“ < ctnlogn torics, or what is commonly called the probabilistic method. This is one of
2y/m<ign the most powerful tools in modern combinatorics, even though, just as with
the pigeonhole principle, the underlying idea is extremely easy. The proba-
pairs. Combining these two observations finishes the proof. [
bili method was used at first by Erdds, who proved a great deal of fairly
Finally, the previous lemma shows that there is an absolute constant ¢; nontrivial results using rather simple probabilistic arguments. For a much
such that if we delete all ¢v/E-rich edges, the resulting multigraph Gy has at deeper study of the method, a canonical reference is the excellent book {2] by
least en? edges. As clearly G has crossing number at most 2n%t2, it follows Alon and Spencer. Many of the examples we are going to discuss are taken
from theorem 6.A.6 that from this book (however, some of them are left as exercises in the book and
cn® are not really easy... ).
32 >
n2e1vE ‘We begin by recalling some useful notions from probability theory. In
and the result follows. [m] discrete combinatorics, one works with finite probability spaces, which makes
the discussion much more elementary, avoiding subtle issues from measure
theory. A finite probability space is the data of a finite set Q (called the
sample space) and of a map {called the probability distribution} P: Q — [0,1]
such that

¥ e
weR
1.

One obvious such map is the constant map fi, which is called the uniform
distribution. One defines for any subset A of SLZ; which we call an event, its
probability as
P(A) =3 Pla).
e
It is very easy to check that P(AUB)+ P(ANB) = P{A)+ P(B) and that the
probability of the complement of A is 1~ P{A). Given a probability space, a
random variable X is simply a map X : @ — R. The expectation (or mean
value) of X is

EX]= 3 X(@Pw)=Y - P(X ==).

wen w€R
 i2

266 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.B. Probabilities in Combinatorics 267

Note that the second sum is finite, as the image of X is finite. Another ex- it :';)Hows that P(Bj) < P(4;) (as B; C A; and P takes nonnegative
values)
tremely important notion is that of independence. Some events Ay, Aa, ..., A an

are called independent if for all I C {1,2,..., k} we have

P(U;A;) = P(U;B;) = 3 P(B;) < 3 P(4;).
Plrierd) = [T P(4). The rest is immediate.
i€l 2) is an easy consequence of the definition of expectation,
[m]
Two randem variables X,V are called independent if the events X = a and
Y = b are independent for all a,b. It is an easy exercise to check that in this ) When using the probabilistic method, one is naturally confronted with
case BE[XY] = E[X|E[Y}. estimating probabilities, which sometimes can be quite painful. A simple
The following theorem consists of two elementary, but incredibly powerful example is that if (2, P) is a probability space and if X is a random
variable
principles. The first one is basically the principle of the probabilistic method: on (2, then the definition gives that there exists w € Q such that
X{w) > E[X].
if you want to show the existence of an object with properties Pi,..., Py, it The following two inequalities are also very basic, but useful tools in estimati
ng
is enough to prove that there is a probability space (2, P) such that the sum probabilities:
of the probabilities that an object does not have property P is smaller than
Theorem 6.B.2. 1) (Markov’s inequality) If X is a random variable tak-
1. Of course, the difficult point is constructing the good probability space
ing nonnegative values and if a > 0, then
{though in practice the construction is naturally imposed by the statement of
the problem we are trying to solve) and estimating these probabilitiess The
second part of the theorem is called linearity of expectation and it Iso a P(X > a) < 1E[X].
@
very powerful result, as we will soon see.

Theorem 6.B.1. Let (2,P) be o finite probubility space. 2) (Chebyshev’s inequality) Let X be a random variable and let ¢ > 0. Then
1} For any subsets Ay, As,. .., Ay of Q we have
P(X ~ E[X]| 2 @) < (EIX?
1
- BIXP).
Proof. For the first part, simply note that

with equality if the events are pairwise disjoint. So, if Zle P(4;) < 1, EX]=Y P(X=2)z> S P(X =2)a = aP(X > a).
then UA; # Q. = z>a
2) If X1, Xa,...,Xk are random variables, t)ten For the second part, using Markov's inequality, we can write

E[X) + Xg + -+ Xy| = B[X))

+ E{Xo] + - + E[Xg].
POX - E[X]| 2 a) = P(X - E[X])* > o®) < %EI(X ~ E[X))?]
Proof. 1) Let By = Ay and let B; be the complement of USZ] 4; in 4; for all
4> 2. Then clearly U;B; = U;A; and the By's are pairwise disjoint. But then and an easy computation using linearity of expectation yields
the result. O
268 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.B. Probabilities in Combinatorics 269

As usual, knowledge comes with practice, so we will spend the remainder Proof. 1t is enough to check the equality for = € (0,1), since a polynomi
al
of this chapter by giving a lot of applications of the previous theorems. We vanishing on (0, 1) vanishes everywhere. Consider a random two-colo
ring of
start with some applications of the sub-additivity of probabilities. One can S such that the probability that a point is black is z and all these
events
prove the following inequality using standard techniques, but there is also a are independent. For a given P, the quantity 2% )(1 — z)P) ) is simply the
very elegant probabilistic proof: probability that the vertices of P are black and the points outside P are white.
As these events are disjoint, the sum over all P is the prebability that there
B.2. Let zy € [0,1] for 1 < 4,7 < n. Prove that is a polygon with black vertices and such that all points outside it are white.
But this probability is 1, since the convex hull of the black points is such
a
2 n polygon (note that there might be 0,1 or 2 black points and this is why
the
11 (1 -fix,]) +fi i-Jla-=zp ) 21 sumn is also taken over degenerate polygons).
) )
F=1 g} im] F=1
We continue with two problems which use the principle of the probabilis-
tic method, namely the fact that if the sum of probabilities
Proof. Consider a random binary matrix (ay) such that Play = 1) = zy of some events
is (strictly) less than 1, then there is a point in the probability space not
and all these events are independent. Then ]_[;LI (1[I, =) is simply the
belonging to any of these events.
probability that each column countains at least a zero, while the expression
e (1— H;':] (1— zlj)) is the probability that each row contains at least a B.4. Prove that there is a four-coloring of the set M = {1,2,...,1987}
1. Since all binary matrices have either at least one zero in every column or such that M contains no monochromatic arithmetic progression with
at least one one in every row, the result follows. O 10 terms.

Another application of theorem 6.B.1 is the following nice problem: IMO Shortlist 1987
Proof. Pick a random coloring and observe that the probability that M con-
B.3. Let S be a finite set of points in a plane, no three of which are collinear. tains a monochromatic progression of length at least 10 is bounded by
N/4°,
For each convex polygon P with vertices in S, let a(P) be the number where N is the number of progressions of length 10 contained in M. So the
of vertices of P! and let 5(P) be the number of points of S lying outside problem is solved if we prove that N < 4°. But there are [ 4287=§ progressions
P (i.e. outside its interior or border). Prove that for all real numbers =z, of length 10 contained in M and whose first term is i. So
1987
S a1 -2 =1, 1987 — ¢ 1
P
N<)y ——— < = .2000%

where here the sum is taken over possibly degenerate convex poly- So, it remains to check that 2855 < g. 21 which follows from
gons(polygons with 2, 1, or 0 vertices), too
IMO Shortlist 2006 5= (5%)2 < (272 =21 g < 9. 011, ]
"With the convention that a(P) = 0,1,2 if P is @, a point of S, respectively & segment Recall that the mm-th Ramsey number R(m) is the smallest positive integer
connesting two points of S. n such that in any coloring of the edges of K, (complete graph with
n vertices)
 :

270 Chapter 6. Some Classical Problems in Exztremal Graph Theory 6.8. Probabilities in Combinatorics 271

with two colors, one can find a monochromatic K, subgraph. Note that it B.6. Let Ay, Az,..., A, and By, B, ..., B, be distinct subsets of N such that
is not obvious that this is well defined, but one can prove without too much Ai (1 B; = @ for all i and (A; 0 B;) U(4; N B;) # 0 for all i # j. Prove
difficulty that any coloring of the edges of K(szz) with two colors contains a that for all p € [0,1]
—
monochromatic Ky, so that R(m) is well-defined and at most (2,:‘"’12) (which n
grows roughly as 4™). The following result of Erdés from 1947 is an absolute
classic. Amazingly, even after more than 60 years, it remains close to the best
Sl - p)E <1,
known lower bound (and 4™ remains close to the best known upper bound).
Tusza
B5. 1 2(3)71 > (), then R(m) > n. In particular, R(m) > 2% for all
m > 4.
Proof. Let X be the union of all A; and B; and consider a random subset S
Erdés of X such that the events & € § for z € X are independent and of probability
p (more formally, let © be the set of all subsets of X and define
Proof. By definition, saying that 2(mn) > n is the same as saying that we can
find a coloring of the edges of K, with no monochromatic K,,.
random coloring of the edges of K, with two colors, each having
Consider a
probability
P(8) =1 - )1,
1/2. 1f S is an m-element subset of the vertices of K, let Ag be the event that which is a probability measure on Q, because g x P(S) = 1 by the binomial
the corresponding subgraph of K, is monochromatic. We want tc prove that theorem). Consider the event Ej: A; C § C X — B;. By hypothesis, no two
there is an element in the probability space (i.e. a 2-coloring) which belongs events oceur at the same time, thus
to none of the events Ag. It is enough to check that

S P(As) <1
S PE)=PUE) <1
&
On the other hand, we have
n
However, it is clear that we have | %) choices of sets S and for each of them

P(Ag) = 2-(3), PE) = 30 PPl = gl gy lB,

A;CSCX-B;

Thus, as long as 21 5 (;‘L) we can find the desired coloring. The second the last equality being an easy computation left to the reader. Inserting this
part of the theorem is now easy: taking n = [27/2], we have (for m > 4) in the previous inequality yields the desired result. a

n) _nn=1)- ~m+1)<n"'
m) n! ml
<@m <ot m <2l O Remark 6.B.3. If all A; have a elements and all B; have b elements, we deduce
from Tusza’s inequality (by taking p = ;%) that n < ("T"',f_);;—b.
Here is a more delicate problem using similar ideas, but for which it is ‘We continue with a series of applications of the linearity of expectation
more difficult to find a good probabilistic interpretation. and Chebyshev’s inequality. First, a very simple problem:
272 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.B. Probabilities in Combinatorics 273

B.7. Let p,(k) be the number of permutations of {1,2,...,n} having exactly For what follows, in a tournament there is exactly one match between each
k ftixed points. Prove that pair of players and there is no draw. A Hamiltonian path in a tournament is
a permutation o of the players such that player o{i) beats player o{i + 1) for
Xn:kpn(k) =l all 7.
k=0
B.9. There exists a tournament with n players which hias at least .‘,—,I‘_lr Hamil-
MO 1987 tonian paths.

Proof. Let o have the uniform distribution on the set of permutations of Szele
{1,2,...,n}. If X(0) is the number of fixed points of o, then
Proof. Pick a random tournament (in which the results of each match occur
L3 kpalk) = BIX]. with equal probability), and let X be the number of Hamiltonian paths. If o is
a permutation of the players, let X,(T') be 1 if ¢ induces a Hamiltonian path
L1
in the tournament 7" and 0 otherwise. It is clear that, as random variables, we
On the other hand, N
have X = 3= X,. It is equally clear that E[X,] = 21~ (since the result of
X=3 X, the matches between o(i) and o(i +1) is imposed for all ). Thus EX]= 5,?—1;
il and the result follows. m}
where Xi(0) = 1,¢)=i- Then
A good exercise for the reader is to prove that the following result im-
1
E{X;| = P(o(i) =) = - plies (a weak form of) Turdn’s famous theorem on graphs without n-complete
subgraphs.
and the conclusion follows by linearity of expectation. a
B.10. Let di,dy,...,d, be the degrees of the vertices of a graph. Prove that
The next two applications are absolute classics.
one can find a subset S of vertices such that
B.8. Any graph with g edges contains a bipartite subgraph with at least g/2
edges.
1) S has af least 3°7, Ay elements.
Erdés 2) There are no edges between vertices in S.

Proof. Pick a random subset S of the set V of vertices, by including a vertex Proof. Consider a random permutation o of the vertices of our graph G, all
in § independently with probability 1/2 . Let X(S) be the number of edges zy permutations having equal probability # Let A; be the event: o{i) < o(j)
for which exactly one of z,¥ is in S. For a given edge zy, the probability that for any neighbor j of i. We claim that P(4;) = fi Indeed, we need to find
exactly one of z,y is in § is 1/2, so by linearity of expectation E{X] = ¢/2. the number of permutations ¢ of the vertices such that o{i) < o(j} for any
Thus one can find S such that X(S) > ¢/2. By construction, the subgraph neighbor j of &. If y1,. . ., yg, are the neighbors of ¢, there are (d"jrl) possibilities
comprising all edges with exactly one vertex in S is a solution. [ for the set {o(i),0(y1),...,0(ya,)}, di! ways to permute the elements of this
274 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.B. Probabilities in Combinatorics 275

set (and not (d; + 1)}, since (i) is the smallest element of the set) and finally from which the result follows, as ([ ]) is the largest among the binomial coef-
{n — d; — 1)! ways to permute the remaining vertices. So ficients. o
n (p—di—-Dldl 1
Actually, the proof shows the following more general result, known as
Plag) = <41+1> nl TG r the Lubell-Yamamoto-Meshalkin-inequality: let A;, 4o...., Ayj be subsets of
as claimed. If X is the random variable X (o) = Y7| loe4;, then {1,2,...,n} such that no A; is a subset of any A;. Then
n . RS SR R
X =3 Pl =3 =y (i) (2 Ga)
41 =l

Hence one can find o such that The following famous theorem of Bollobds generalizes this result further:

B.12. Let Ay, As,..., Ap and By, B, .. ., B, be distinct sets of positive integers
such that 4; fl B; = ¢ for all i, but A;nBy#0foralli+#j. Then

It is clear that the set of vertices ¢ such that ¢ € A; satisfies both properties. n

&
1
O — - < 1.

We continue with a beautiful proof, due to Lubell, of a famous theorem

of Sperner on maximal anti-chains. Bollobés
B.11. Let F be a family of subsets of {1,2,...,n} which does not contain two Proof. We may assume that all As and By’s are subsets of {1,2,...,N}.
elements A, B such that A C B. Then |F| < ({%]) Consider the uniform distribution on the set Sy of all permutamons of
Sperner’s theorem {1.2,...,N}. Let E; be the event consisting of all permutations & with the
following property: all elements of A; come before all elements of B; in the
Proof. Consider a random permutation o of {1,2,...,n} (with uniform dis- list 9(1),5(2),...,0(n). It is easy to see that the hypothesis implies that the
tribution) and let A; be the event that {o(1),0(2 )..... ,o(i)} € F. The hy- events E; are pdlrwme disjoint. Thus Y, P(E;) < 1. It remains to notice
pothesis on F implies that the random variable X (o) = 37| ls¢a, satisfies that P(E;)= (M
H Hy as any subset with [4;] elements of A; Ui B; is equally
X(o) < 1 for all o, so that its expectation E{X] § 1 On the other hand,
likely to form the first {4;] elements in the list 0(1),0(2),...,0(n). The result
E[X] = 37, P(A;) and the probability that {c(1),...,0({)} is a given set
follows. o
with 1 elements is flgfifl%)h thus P(4;) = 7%, where n; is the number of subsets
with 4 elements in F. Putting this togeth‘er yields the beautiful inequality We take a break here to discuss a very nice consequence of Bollobés
inequality. Recall that an r-uniform hypergraph is a pair (V, E), where Visa
ica
(n <1,
set and E is a collection of subsets of V, each subset having r elements. Let
Mg

g i K7 be the complete r-uniform hypergraph on ¢ vertices.

276 Chapter 6. Some Classical Problems in Extremal Graph Theory 6.8B. Probabilities in Combinatorics 277

B.13. Let G = (V, E) be a r-uniform hypergraph on n vertices. Suppose that On the other hand, the probability that i ¢ A+ B is clearly the n-th power
G contains no K7, but that if we add any r-element set to E, at least of the probability that a given integer is not in A, that is
(l_‘l)n‘
one K] appears. Then
P(i€A+B):17(]_iniz')flzl_
0 (1) (1) We deduce that
n

Bollobés
and the result follows from the inequality (1—1)" < 1. m]
Proof. Let [n] = {1,2,...,n} and let [n]) be the collection of r-element
subsets of [n]. If A € [r)®) — B, consider a copy C(A) of K7 that appears Often, the choices of the probability space and distribution are imposed
by adding 4 to E. Then 4 € C(A), yet A" isnot in C(A) if A’ # Aisin by common sense. However, in order to get better quantitative results, one
[n]™) ~E. So, if {A1,...,Am} = [#]®) ~ E and B is the complement of C(A4;), uses sometimes less natural probability distributions. The following problems
then (A4;); and (B;); satisfy the hypotheses of Bollobds® inequality, yielding illustrate this:
m< ("’:*’), The result follows, (]
B.15. Let V be a set with n elements and let F be a family of m subsets of
We continue with a very nice result from additive number theory. V', each having three elements. Prove that if 3m > n, then there exists
S C V having at least 2, /7 elements and such that no element of F
is contained in S.
B.14. Let A C Z/n?Z be a subset with n elements. Prove that there exists a
2
subset B C Z/n*Z with n elements such that |4+ B > &' Proof. Choose p € [0,1] and pick a random subset S of V such that P(v €
IMO Shortlist 1999 S} =pfor all v € V, all the events being independent. If n(S) is the number
of elements of F contained in S, then linearity of expectation yields
Proaf. Pick a random collection of n elements of Z/n?Z, each of the n elements
being taken with probability 1/n2 and all choices being independent. Consol- E[IS} = n(8)] = E[S] - E[n(S)] = np —mp®.
idate the distinct elements among the n chosen ones in a set B, which may This is maximal for p = /2= € [0,1] and is equal to 2. /Z. Thus, we can
have less than n elements. Consider then the random variable X = [4 + Bl find § such that |S] - n(S) > %/Z. Choose such S. For each ¢ € F with
As all of its three elements in S, delete arbitrarily one of these three elements.
A= Z lica+n, We thus end up with at least |S| — n(S) elements of V which obviously have
i€L/n2T the desired property. ]
we have by linearity of expectation B.16. The minimal degree of the vertices of a graph G with n vertices is d > 1.
Prove that there exists a subset S of the vertices such that
E[X]= 3 PleA+B).
i€B/nT 1) S| < n - 10D
1+in(d+1
 278 Chapter 6. Some Classical Problems in Exztremal Graph Theory 6.8. Probabilities in Combinatorics 279

2) Any vertex of G is either in § or neighbor of a vertex in S. z,y,z € §. Pick z € F} randomly with uniform probability and consider the
set
Noga Alon
A; ={a € Alaz (mod p) € 5}.
Proof. Consider a random subset S of the set of vertices V such that each This is easily seen to be a sum-free subset of A. To prove that one can choose
vertex is in S independently with probability p = %—’;‘12 if §' is the set of z such that {4;] > JT" it is enough to compute the expected value of |A,].
vertices which do not belong to § and which have no neighbor in S, then SUS’ Linearity of expectation shows that
satisfies the second condition. It is thus enough to show that
E[lA:f] =) Plaz (mod p) € 5)
EISus| <n. LTBEFD agA

and it is clear that for any @ € 4 we have P(az (mod p) € §) = i as

However,
{az) Jzeky is a permutation of Fj. Thus E|A.[] > 37714l
&1 and we are done a
ElIS 0 &1 < EISI + BiS) We end this chapter with a very beautiful application of Chebyshev’s
=Y PweS)+) Pues) inequality. Before attacking this problem, let us recall a few basic properties
of the variance. Note that for any constant ¢ and any random variable X we
<pn+n(l— )d+1 have Var(cX)= @Var(X). Also, it is not difficult to check that i XY are
independent variables, then Var(X +Y) = Var(X) + Var(Y). Finally, if X is
< Pl1+ind+1)
W
STTTER 0 with probability p and 1 with probability 1 — p, then Var(X) = p(1 ~ j23

the last inequality being equivalent (after dividing by n, canceling p and taking B.18. Let v1,v2,...,v, be n vectors in the plane, whose coordinates are integers
logarithms) to log{1 — p) < ~p, which is well-known. [ of absolute value less that fi‘ / ZT" Prove that there are disjoint subsets

A very beautiful but rather subtle application of the probabilistic method I,J of {1,2,....n} such that
is the following gem due to Erdés:
Su=Yu
B.17. Let A be a finite set of nonzero integers. Then A contains a subset B €l i€

of cardinality greater than %1 such that the equation z + y = 2 has no

Proof. Note first of all that it is enough to find two distinet such subsets I, A
solutions in B x B x B, .
for then it is enough to take out of each the common elements of . ,J. Now,
Erdés assume that we cannot find two such distinct subsets and write v; = {zs,31).
Consider a random binary sequence (a1, aa, ..., a,), each ¢; being 0 or 1 with
Proof. "The proof is quite tricky. Choose a prime p = 3k + 2 large enough, say probability 1/2 (all these events being independent). Consider the random
such that p > 2laf for all a € A, The subset S = {k+1,k+2,...,2k + 1} variables X = 3% a;z; and Y = 3", a;:. The point is that the values
of Fp is a sum-free subset, i.e. the equation & +y = z has no solutions with taken by (X,Y) are all different and that = Iot of values are concentrated near
280 Chapter 6. Some Classical Problems in Extremal Graph Theory

the expectation of (X,Y). To be more precise, let m, = E[X], m, = E[Y]

and observe that the discussion preceding the problem yields
2 n
Var(X) -= Var (Z a,zl)AT= Z A
1 < 100506

by hypothesis. Let 0% = % Using Chebyshev’s inequality, we obtain Chapter 7

P(X ~mq| 2 20) <p PY-mlz2)<q 1
1
<
hence Complex Combinatorics
P(|X —mg| < 20,]Y —my| < 20) >

SIE
This means that at least half of the (pairwise distinct) points (X, Y') are located
in the square |X — m;| < 20, [Y —my| < 20. But this square contains at There is a class of combinatorial problems, most of the times with number-
most (1 + 40)? laitice points, so theoretic flavor, which have very elegant solutions using finite Fourier trans-
forms or versions of it. The main observation is the identity
(1+40)? 2 277,
=
which is easily seen to be impossible (as say 2502 > (1 + 40)? for n > 16} liza (modky =7 ZZ’('H'),
unless 2 = 1, in which case we are easily done. o =0
which holds for any primitive root of unity 2 of order k. This gives & rather
powerful approach to problems concerning the distribution mod & of the sums
of subsets of a given set or to tiling problems. Actually, this relation is a
very special case of a much broader theory, that of representations of finite
groups. The case of finite abelian groups is particularly easy and useful in
combinatorial problems and we refer the reader to addendum 7.A for a more
detailed discussion. For the next problems the previous relation is actually
sufficient.

7.1 Tiling and coloring problems

We start with two tiling problems which have rather elegant solutions
using complex numbers. Of course, the idea is similar to the usual coloring
method, but we think it is neater.