Problems in Discrete Geometry

and Extremal Combinatorics

Zilin Jiang
Department of Mathematical Sciences
Carnegie Mellon University

A thesis submitted in partial fulfillment for the degree of

Doctor of Philosophy

April 2016
.Ÿ⌘Ñ6Õ
 Acknowledgements

I am especially grateful to my advisor and coauthor Boris Bukh, without whom

I would not have been able to finish my PhD. I would like to thank Po-Shen
Loh, who first introduced extremal combinatorics to me. I would also like to
thank other faculty members of math, physics and computer science departments,
Tom Bohman, James Cummings, Alan Frieze, Robert Griffiths, Rami Grossberg,
Bill Hrusa, Gautam Iyer, Giovanni Leoni, Gary Miller, Ryan O’donnell, Ernest
Schimmerling and Dejan Slepčev, from whom I have learnt a lot.
It is a pleasure to acknowledge the positive influence to various aspects of my life
from my fellow graduate students, Devi Borg, Joe Briggs, Sam Cohn, Chris Cox,
Nicolás Garcı́a, Nate Ince, Slav Kirov, Jing Liu, Pan Liu, Clive Newstead, Yue
Pu, Matteo Rinaldi, Daniel Rodrı́guez, Thames SaeSue, Xiaofei Shi, Sebastien
Vasey, Yan Xu, Jing Zhang, Linan Zhang and Andy Zucker. In addition, I would
like to thank my friends Ti↵any Bao, Tianyuan Deng, Duo Ding, Wenying Gan,
Staren Gong, Henry Hu, Yi Hu, Tao Huang, Yangyang Huang, Zhijun Huang,
Ruizhang Jin, Aoxi Li, Jason Li, Yun Ling, Bin Liu, Lexi Liu, Mingqian Liu,
Victoria Lv, Xuele Ma, Yutian Nie, Neal Pei, Sharon Romero, Lily Shen, Roger
Shen, Rujun Song, Vincent Sun, Xuening Sun, Maigo Wang, Hong Wang, Zhiyu
Wang, Vena Wei, Susan Weng, Liang Wu, Ben Yang, Emma Yang, Yaoqing Yang,
Jenkin Xia, Tianying Xia, Zhukai Xu, Gaochi Zhang, Ruixiang Zhang, Ruixun
Zhang, Yu Zhao, Yufei Zhao, Hai Zhou, Ye Zhou, Catherine Zhu, Charlene Zhu,
Ted Zhu and Ye Zhu, for their advices and encouragements through my ups and
downs.
Finally, I would like to thank my parents for their unconditional love.
Chapter 2 is a version of the paper A slight improvement to the colored Bárány’s
theorem, which has been published in Electron. J. Combin.. Chapter 3 is a
version of the paper A bound on the number of edges in graphs without an even
cycle co-authored with Boris Bukh, which has been accepted to Comb. Probab.
Comp.. Chapter 4 is a version of the paper Bipartite algebraic graphs without
quadrilaterals co-authored with Boris Bukh, which has been submitted. All these
works were carried out under the supervision of Boris Bukh.

4
 Abstract

We study several problems in discrete geometry and extremal combinatorics.

Discrete geometry studies the combinatorial properties of finite sets of simple
geometric objects. One theme of the field is geometric Ramsey theory. Given m
geometric objects, we want to select a not too small subset forming a configura-
tion that is “regular” in some sense.
The first problem that we study is a colored variant of the point selection problem
in discrete geometry. Given d+1 disjoint n-point sets, P0 , . . . , Pd , in Rd , a colorful
simplex is the convex hull of d + 1 points each of which comes from a distinct Pi .
We establish a “positive-fraction theorem” that asserts the existence of a point
2d
common to at least (d+1)(d+1)!
nd+1 colorful simplices.
Extremal combinatorics studies the maximum or minimum size of discrete struc-
tures under given constraints. For a graph H, define the Turán number ex(n, H)
as the maximum number of edges that an H-free graph on n vertices can have.
When H is bipartite, the problem of pinning down the order of magnitude of
ex(n, H) remains in general as one of the central open problems in combina-
torics.
The second and the third problems respectively consider the Tuán number for
two classes of graphs: cycles of even length and complete bipartite graphs. For
p
cycles of even length C2k , we show that for ex(n, C2k )  80 k log k·n1+1/k +O(n).
For complete bipartite graphs Ks,t , motivated by the algebraic constructions of
the extremal Ks,t -free graphs, we restrict our attention algebraically constructed
graphs. We conjecture that every algebraic hypersurface that gives rise to a Ks,t -
free graph is equivalent, in a suitable sense, to a hypersurface of low degree. We
establish a version of this conjecture for K2,2 -free graphs.
Contents

1 Introduction 1

2 A slight improvement to the colored Bárány’s theorem 3

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Proof of the Theorem 2.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 A bound on the number of edges in graphs without an even cycle 11

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2 Graph exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 Trilayered graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3 ⇥-graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.3.1 Well-placed ⇥-graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.2 Finding a trilayered subgraph of large minimum degree . . . . . . . . 15
3.3.3 Locating well-placed ⇥-graphs in trilayered graphs . . . . . . . . . . . 19
3.4 Proof of Theorem 3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

4 Bipartite algebraic graphs without quadrilaterals 25

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.2 Conjectures on the (s, t)-grid-free case . . . . . . . . . . . . . . . . . . . . . 28
4.3 Results on the (1, t)-grid-free case . . . . . . . . . . . . . . . . . . . . . . . . 30
4.4 Results on the (2, 2)-grid-free case . . . . . . . . . . . . . . . . . . . . . . . . 31
4.5 Fields of finite characteristic . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

Bibliography 36

i
List of Figures

2.1 3 red points, 3 green points and 3 blue points are placed in the plane. The
2
point marked by a square is contained in 6 (= 9
· 33 ) colorful triangles. . . . 5
2
2.2 The bird’s-eye view of a triangulation of S with a 2-simplex containing 1
and the cone over part of the triangulation. . . . . . . . . . . . . . . . . . . . 6
2.3 An illustration of an 1-simplex , @ , cone(@ ) and a homeomorphism from
to cone(@ ). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4 The illustration shows a cone over part of @cone( ) with apex c and a point
v on the boundary, and how a point w on the line segment [v, c) are mapped
under h. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 An illustration of the partition, the result of filling in against c, and f (cone( ))mod2. 9

3.1 A schematic drawing of a tyilayered graph. . . . . . . . . . . . . . . . . . . . 14

3.2 An arbitrary long path P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

ii
Chapter 1

Introduction

Discrete geometry studies the combinatorial properties of finite sets of points, lines, circles,
planes, or other simple geometric objects. For example, one can ask, “what is the maximum
number of incidences between m points and n lines?”, or “what is the minimum possible
number of distinct distances occurring among n points in the plane?”. We refer the interested
readers to the book of Matoušek [Mat02].
One theme of the field is geometric Ramsey theory. Given m geometric objects, we
want to select a not too small subset forming a configuration that is “regular” in some
sense. In many cases, we obtain “positive-fraction theorems”: the regular configuration has
n
size at least ⌦(m). Consider n points in Rd in general position, and draw all the m = d+1
simplices with vertices at the given points. The point selection problem asks for a point of Rd
common to as many simplices as possible. In the early eighties, Bárány [Bár82] established
a “positive-fraction theorem” that asserts the existence of a point common to at least cd · m
of these simplices, where cd > 0 is a constant depending only on d. Gromov [Gro10] made a
major breakthrough in improving the lower bound of cd by introducing a topological proof
method. Later Krasev [Kar12] found a very short and elegant proof of Gromov’s bound.
The problem that we study in Chapter 2 is a colored variant of the point selection
problem. Given d + 1 disjoint n-point sets, P0 , . . . , Pd , in Rd , a colorful simplex is the convex
hull of d+1 points each of which comes from a distinct Pi . The colored point selection problem
asks for a point of Rd covered by at least c0d · nd+1 of the colorful simplices. As it generalizes
the original point selection problem, one can show that cd  c0d . Karasev [Kar12] indeed
1
proved that c0d (d+1)!
. Based on his method that combines probabilistic and topological
2d
arguments, we improve his result to c0d (d+1)(d+1)! matching Gromov’s bound for cd .
Extremal combinatorics studies the maximum or minimum size of discrete structures
under given constraints. The basic statement of extremal combinatorics is Mentel’s theo-
rem [Man07], proved in 1907, which states that any graph on n vertices with no triangle

1
contains at most n2 /4 edges. The natural generalization of this theorem to clique is proved
by Turán [Tur41] in 1941. Turán’s theorem states that, on a given vertex set, the graph with
the most edges with no clique of size t is the complete and balanced (t 1)-partite graph,
in that the part sizes are as equal as possible.
For a graph F , define the Turán number ex(n, F ) as the maximum number of edges that
a graph on n vertices can have without containing a copy of F . For general graphs F , we
still do not known how to compute the Turán number exactly, but if we are satisfied with an
approximate answer, the theory becomes quite simple: Erdős and Stone [ES46] showed in
1 n
1946 that if the chromatic number (H) = t, then ex(n, Kr )  1 t 1 2
+ o(n2 ). When
F is not bipartite, this gives asymptotic result for the Turán number. When F is bipartite,
the problem of pinning down the order of magnitude of ex(n, F ) remains in general as one
of the central open problems in combinatorics. Most of the study of ex(n, F ) for bipartite F
has been concentrated on trees, complete bipartite graphs, and cycles of even length.
We address the cycles of even length C2k in Chapter 3. A general bound of ex(n, C2k ) 
k· n1+1/k + O(n), for some unspecified constant k , was asserted by Erdős [Erd64, p.33].
The first proof was by Bondy and Simonovits [BS74, Lemma 2], who showed the bound for
k = 20k. This was improved by Verstraëte [Ver00] to k = 8(k 1) and by Pikhurko [Pik12]
to k = k 1. Pikhurko asked whether limk!1 k /k
can be 0. Inspired by the proof of
p
Pikhurko, we answer this question in the positive by improving k to 80 k log k.
In the final chapter, Chapter 4, we further study the Turán number for complete bipartite
graphs Ks,t . So which graphs are Ks,t -free with a maximum number of edges? The question
was considered by Füredi in his unpublished manuscript [Für88] asserting that every K2,2 -
free graph with q vertices (for q q0 ) and 12 q(q + 1)2 edges is obtained from a projective
plane via a polarity with q + 1 absolute elements. Motivated by the algebraic constructions
that match the upper bounds in the cases that ex(n, Ks,t ) has been solved asymptotically, we
restrict our attention to algebraic bipartite graphs defined over algebraically closed fields. We
conjecture that every algebraic hypersurface that gives rise to a Ks,t -free graph is equivalent,
in a suitable sense, to a hypersurface of low degree. We establish a version of this conjecture
for K2,2 -free graphs.
In the following chapters, backgrounds and motivations of the problems will be discussed
in more details in their own introductions.

2
Chapter 2

A slight improvement to the colored

Bárány’s theorem

2.1 Introduction
n
Let P ⇢ Rd be a set of n points. Every d + 1 of them span a simplex, for a total of d+1
simplices. The point selection problem asks for a point contained in as many simplices as
possible. Boros and Füredi [BF84] showed for d = 2 that there always exists a point in R2
contained in at least 29 n3 O(n2 ) simplices. A short and clever proof of this result was
given by Bukh [Buk06]. Bárány [Bár82] generalized this result to higher dimensions:

Theorem 2.1 (Bárány [Bár82]). There exists a point in Rd that is contained in at least
n
cd d+1
O(nd ) simplices, where cd > 0 is a constant depending only on the dimension d.

This general result, the Bárány’s theorem, is also known as the first selection lemma. We
will henceforth denote by cd the largest possible constant for which the Bárány’s theorem
holds true. Bukh, Matoušek and Nivasch [BMN10] used a specific construction called the
stretched grid to prove that the constant c2 = 29 in the planar case found by Boros and
d!
Füredi [BF84] is the best possible. In fact, they proved that cd  (d+1)d
. On the other hand,
d
Bárány’s proof in [Bár82] implies that cd (d + 1) , and Wagner [Wag03] improved it to
d2 +1
cd (d+1) d+1 .

Gromov [Gro10] further improved the lower bound on cd by topological means. His
2d
method gives cd (d+1)(d+1)!
. Matoušek and Wagner [MW14] provided an exposition of the
combinatorial component of Gromov’s approach in a combinatorial language, while Karasev
[Kar12] found a very elegant proof of Gromov’s bound, which he described as a “decoded
and refined” version of Gromov’s proof.

3
 The exact value of cd has been the subject of ongoing research and is unknown, except for
the planar case. Basit, Mustafa, Ray and Raza [BMRR10] and successively Matoušek and
Wagner [MW14] improved the Bárány’s theorem in R3 . Král’, Mach and Sereni [KMS12]
used flag algebras from extremal combinatorics and managed to further improve the lower
bound on c3 to more than 0.07480, whereas the best upper bound known is 0.09375.
However, in this chapter, we are concerned with a colored variant of the point selection
problem. Let P0 , . . . , Pd be d + 1 disjoint finite sets in Rd . A colorful simplex is the convex
hull of d + 1 points each of which comes from a distinct Pi . For the colored point selection
problem, we are concerned with the point(s) contained in many colorful simplices. Karasev
proved:

Theorem 2.2 (Karasev [Kar12]). Given a family of d + 1 absolutely continuous probability

measures m = (m0 , . . . , md ) on Rd , an m-simplex1 is the convex hull of d+1 points v0 , . . . , vd
with each point vi sampled independently according to probability measure mi . There exists
a point of Rd that is contained in an m-simplex with probability pd 1
(d+1)!
. In addition, if
2d
two probability measures coincide, then the probability can be improved to pd (d+1)(d+1)!
.

By a standard argument which we will provide immediately, a result on the colored point
selection problem follows:

Corollary 2.3. If P0 , . . . , Pd each contains n points, then there exists a point that is con-
1
tained in at least (d+1)!
· nd+1 colorful simplices.

Our result drops the additional assumption in theorem 2.2, hence improves corollary 2.3:

Theorem 2.4. There is a point in Rd that belongs to an m-simplex with probability pd

2d
(d+1)(d+1)!
.
2d
Corollary 2.5. There exists a point that is contained in at least (d+1)(d+1)!
· nd+1 colorful
simplices.

Proof of corollary 2.5 from the Theorem 2.4. Given d + 1 sets P0 , . . . , Pd in Rd each of which
contains n points. Let : Rd ! R be the bump function defined by (x1 , . . . , xd ) =
Qd 2
i=1 (xi ), where (x) = e 1/(1 x ) 1|x|<1 , and set n (x1 , . . . , xd ) = nd (nx1 , . . . , nxd ) for
n 2 N. It is a standard fact that and n are absolutely continuous probability measures
supported on [ 1, 1]d and [ 1/n, 1/n]d respectively.
(n) P
For each n 2 N and 0  k  d, define mk (x) := n1 p2Pk n (x p) for x 2 Rd . Note that
(n)
mk is an absolutely continuous probability measure supported on the Minkowski sum of Pk
1 An m-simplex is actually a simplex-valued random variable.

4
Figure 2.1: 3 red points, 3 green points and 3 blue points are placed in the plane. The point
marked by a square is contained in 6 (= 29 · 33 ) colorful triangles.

(n) (n)
and [ 1/n, 1/n]d . Let m(n) be the family of d + 1 probability measures m0 , . . . , md . By
the Theorem 2.4, there is a point p(n) of Rd that belongs to an m(n) -simplex with probability
2d
at least (d+1)(d+1)!
.
Because no point in a certain neighborhood of infinity is contained in any m(n) -simplex,
the set {p(n) : n 2 N} is bounded, and consequently the set has a limit point p. Suppose
p is contained in N colorful simplices. Let ✏ > 0 be the distance from p to all the colorful
simplices that do not contain p. Choose n large enough such that 1/n ⌧ ✏ and p(n) p ⌧ ✏.
By the choice of n, if p is not contained in a colorful simplex spanned by v0 , . . . , vd , then
p(n) is not contained the convex hull of v00 , . . . , vd0 for all vi0 2 vi + [ 1/n, 1/n]d . This implies
N
that the probability that p(n) is contained in an m(n) -simplex is at most nd+1 . Hence p is the
2d
desired point contained in N (d+1)(d+1)!
· nd+1 colorful simplices.

Readers who are familiar with Karasev’s work [Kar12] would notice that our proof of the
Theorem 2.4 heavily relies on his arguments. The author is deeply in debt to him.

2.2 Proof of the Theorem 2.4

In this section, we provide the proof of the Theorem 2.4. The topological terms in the proof
are standard, and can be found in [Mat03]. In addition to the notion of an m-simplex, in the
proof, we will often refer to an (mk , . . . , md )-face which means the convex hull of d k + 1
points vk , . . . , vd with each point vi sampled independently according to probability measure
mi . An m-simplex and an (mk , . . . , md )-face are both set-valued random variables.

Proof of the Theorem 2.4. To obtain a contradiction, we suppose that for any point v in
Rd , the probability that v belongs to an m-simplex is less than pd := 2d
(d+1)(d+1)!
. Since this

5
Figure 2.2: The bird’s-eye view of a triangulation of S 2 with a 2-simplex containing 1 and
the cone over part of the triangulation.

probability, as a function of point v, is continuous and uniformly tends to 0 as v goes to

infinity, there is an ✏ > 0 such that v is contained in an m-simples with probability at most
pd ✏ for all v in Rd .
Let S d := Rd [ {1} be the one-point compactification of the Euclidean space Rd . Take
= ✏/d. Choose a finite triangulation2 T of S d with one of the d-simplices containing 1
such that for 0 < k  d, any k-face of T intersects an (mk , . . . , md )-face with probability less
than and that the measure of any d-face of T under (md 1 + md ) /2 is less than . This
2
can be done by taking a sufficiently fine triangulation of S with one d-simplex having 1 in
its relative interior.
We use cone(·) as the cone functor3 with apex O. A triangulation T of S d naturally
extends to a triangulation cone(T ) of cone(S d ). We denote the k-skeleton4 of T and cone(T )
by T k and cone(T )k respectively.
We are going to define a continuous map f : cone(T )d ! S d . Put f (x) = x for all
x 2 S d = kT k ⇢ cone(T )d , and set f (O) = 1. We proceed to define f on cone( ) for all
2A triangulation T of a topological space X is a simplicial complex K, homeomorphic to X,
together with a homeomorphism h : kKk ! X. Since the finite triangulation of interest is an
extension of the triangulation of a d-simplex X in Rd and h is an identity map, we will freely use
topological notions such as “a k-face (as a subset of S d )” instead of “the image of a k-face in K
under h”. With such abuse of language, we can avoid going back and forth between the simplicial
complex and the topological space.
3 The cone over a space X is the quotient space cone(X) := (X ⇥ [0, 1]) / (X ⇥ {1}). The apex

is the equivalence class {(x, 1) : x 2 X}.

4 The k-skeleton of a simplicial complex consists of all simplices of of dimension at most k.

6
the k-faces of T inductively on dimension k of while we maintain the property that the
image of the boundary of cone( ) under f , that is f (@cone( )), intersects an (mk , . . . , md )-
face with probability at most (k + 1)!(pd ✏ + k ). We say f is economical over a k-face
of T d 1 if f and satisfy the above property. Unlike Karasev [Kar12], our inductive
construction of f follows the same pattern until k = d 2 instead of d 1. The main
innovation of this proof is a di↵erent construction for k = d 1, which enables us to remove
the additional assumption in theorem 2.2.
Note that for any 0-face in T , f (@cone( )) = f ({ , O}) = { , 1}. According to the
assumption at the beginning of the proof, f (@cone( )) intersects an (m0 , . . . , md )-face, that
is, an m-simplex, with probability at most pd ✏. Therefore f is economical over 0-faces of
T . This finishes the first step.
Suppose f is already defined on cone(T )k and it is economical over k-faces of T . We
are going to extend the domain of f to cone(T )k+1 . Indeed, we only need to define f on
cone( ) for every k-face of T .
Take any k-face of T . Suppose convex hull of vk , . . . , vd , denoted by conv(vk , . . . , vd ),
is an (mk , . . . , md )-face. Notice that the following statements are equivalent:
1. f (@cone( )) intersects conv(vk , . . . , vd );
2. for some v 2 f (@cone( )), the ray with initial point v in the direction vvk v intersects
conv(vk+1 , . . . , vd ).
We call the union of such rays the shadow of f (@cone( )) centered at vk . Since f is economical
over , the probability for an (mk , . . . , md )-face to meet f (@cone( )) is at most (k + 1)!(pd
✏ + k ), and so there exists vk 2 Rd such that the shadow of f (@cone( )) centered at vk
intersects conv(vk+1 , . . . , vd ) with probability at most (k + 1)!(pd ✏ + k ).
Now, we define f on cone( ). First, let g be the homeomorphism from cone( ) onto
the cone over @cone( ) with apex c such that g is an identity on @cone( ). This can be
done because cone( ) is homeomorphic to a (k + 1)-simplex and it is easy to find a
homeomorphism from to cone(@ ) that keeps @ fixed.

e1 e1
e1 e1

e0 e0
e0 e0 c c

Figure 2.3: An illustration of an 1-simplex , @ , cone(@ ) and a homeomorphism from

to cone(@ ).

Next, note that every point w in cone( ) except c is on a line segment [v, c) for a unique
point v on @cone( ). If t = vw/wc 2 [0, 1), then put h(w) = vf (v)+t·vvk f (v). In addition,

7
set h(c) = 1. The function h maps [v, c) onto [f (v), vk ) linearly and then takes the inversion
centered at vk with radius vk f (v) so that [f (v), vk ) gets mapped onto the ray with the initial
point f (v) in the direction vvk f (v). Evidently, h is a continuous map from cone(@cone( ))
onto the shadow of f (@cone) centered at vk that coincides with f on @cone( ).

c v v1
w w0 f (v) h(w)

@cone( ) f (@cone( ))

Figure 2.4: The illustration shows a cone over part of @cone( ) with apex c and a point v
on the boundary, and how a point w on the line segment [v, c) are mapped under h.

Define f on cone( ) to be the composition of g and h:

= f
@cone(
 ) / @cone(
 ) / f (@cone(
 ))
_ _ _

✏ g
✏ ✏
h
cone( ) / cone (@cone( )) / the shadow of f (@cone( )) centered at vk .

According to the commutative diagram above, f is well-defined on cone( ) in the sense that
it is compatible with its definition on cone(T )k . We use the phrase “fill in the boundary of
cone( ) against the center vk ” to represent the above process that extends the domain of f
from @cone( ) to cone( ).
To complete the inductive step, we must demonstrate that f is economical over (k + 1)-
faces of T . Pick any (k + 1)-face ⌧ of T . Let 0, . . . , k+1 be the k-faces of ⌧ . Observing that
f (@cone(⌧ )) = f (⌧ [ cone(@⌧ )) = ⌧ [ f (cone( 0 )) [ . . . [ f (cone( k+1 )) and that f (cone( i ))
is the shadow of f (@cone( i )) centered at vk i which intersects an (mk+1 , . . . , md )-face with
probability at most (k +1)!(pd ✏+k ), we obtain that the probability for an (mk+1 , . . . , md )-
face to intersect f (@cone(⌧ )) is dominated by + (k + 2)(k + 1)!(pd ✏ + k )  (k + 2)!(pd
✏ + (k + 1) ).
We have so far defined a continuous map f on cone(T )d 1
such that for any (d 1)-
face of T the probability for an (md 1 md )-face to intersect D := f (@cone( )) is at most
1
d!(pd ✏ + (d 1) ). We write f (X)mod2 := {y 2 f (X) : |f (y) \ X| = 1 (mod 2)}
for the set of points in f (X) whose fibers in X have an odd number of points. Set m̄ :=
(md 1 + md )/2. We are going to define f on cone( ) such that m̄ (f (cone( ))mod2) is less
1
than d+1
.

8
 A 3

B B 2 2

A 1

c c

Figure 2.5: An illustration of the partition, the result of filling in against c, and
f (cone( ))mod2.

Fix a point s in Rd \ D. For any point t in Rd \ D, if a generic piecewise linear path from
s to t intersects with D an odd number of times, then put t in B, otherwise put it in A. Here
the number of intersections of a piecewise linear path L and D might not be the cardinality
P
of L \ D. Instead, the number of intersections is precisely x2L\D |f 1 (x) \ @cone( )|, that
is, it takes the multiplicity into account. Thus we have partitioned Rd \ D into A and B
such that any generic piecewise linear path from a point in A to a point in B meets D an
odd number of times. Suppose a := md 1 (A), b := md (A) and x := m̄(A) = (a + b)/2. The
probability that an (md 1 md )-face intersects with D is at least a(1 b) + (1 a)b. Hence
1 1
a(1 b) + (1 a)b < d!(pd ✏ + (d 1) ) < 2 d+1
1 d+1
. Because a(1 b) + (1 a)b =
2 1
(a + b) 2ab (a + b) (a + b) /2 = 2x(1 x), either x or 1 x is less than d+1
. In other
1 1
words, one of m̄(A) and m̄(B) is less than d+1
.
We may assume that m̄(B) < d+1
.
Fix a point c 2 A. Again, we fill in the boundary of cone( ) against the center c. For
any generic point x 2 A, the line segment [c, x] intersects with D an even number of times.
For every v on @cone( ), the ray with the initial point f (v) in the direction vcf (v) covers x
once if and only if the line segment [c, x] intersects with D at f (v). Because f (cone( )) is
the union of such rays, the number of times that x is covered by f (cone( )) is exactly the
number of intersections between [c, x] and D. This implies that x is not in f (cone( ))mod2.
Therefore f (cone( ))mod2 is a subset of B [ D almost surely. Noticing that m̄(D) = 0, the
1
extension of f has the desired property m̄ (f (cone( ))mod2) < d+1 .
Pick any d-face ⌧ of T . Suppose the (d 1)-faces of ⌧ are 0 , . . . , d. By a parity
argument, we have

f (@cone(⌧ ))mod2 = [⌧ [ f (cone( 0 )) [ . . . [ f (cone( d ))] mod2

⇢ ⌧ [ f (cone( 0 ))mod2 [ . . . [ f (cone( d ))mod2.

9
 1
Therefore m̄ (f (@cone(⌧ ))mod2) is less than + (d + 1) d+1 = 1, and so the degree of f on
@cone(⌧ ), denoted by deg (f, @cone(⌧ )), is even. Because
X X
deg (f, @cone(⌧ )) = 2 deg (f, cone( )) + deg (f, T ) = deg (f, T ) (mod 2),
⌧

where the first sum and the second sum are over all d-faces and all (d 1)-faces of T
respectively, we know that deg (f, T ) is even, which contradicts with the fact that f is
identity on T .

10
Chapter 3

A bound on the number of edges in

graphs without an even cycle

3.1 Introduction
Let ex(n, F ) be the largest number of edges in an n-vertex graph that contains no copy of
a fixed graph F . The systematic study of ex(n, F ) was started by Turán [Tur41] over 70
years ago, and it has developed into a central problem in extremal graph theory (see surveys
[FS13, Kee11, Sid95]).
The function ex(n, F ) exhibits a dichotomy: if F is not bipartite, then ex(n, F ) grows
quadratically in n, and is fairly well understood; if F is bipartite, ex(n, F ) is subquadratic,
and for very few F the order of magnitude is known. The simplest classes of bipartite graphs
are trees, complete bipartite graphs, and cycles of even length. Most of the study of ex(n, F )
for bipartite F has been concentrated on these classes. In this chapter, we address the even
cycles. For an overview of the status of ex(n, F ) for complete bipartite graphs see [BBK13].
For a thorough survey on bipartite Turán problems see [FS13].
The first bound on the problem is due to Erdős [Erd38] who showed that ex(n, C4 ) =
O(n3/2 ). Thanks to the works of Erdős and Rényi [ER62], Brown [Bro66, Section 3], and
Kövari, Sós and Turán [KST54] it is now known that

ex(n, C4 ) = (1/2 + o(1))n3/2 .

The current best bounds for ex(n, C6 ) for large values of n are

0.5338n4/3 < ex(n, C6 )  0.6272n4/3

due to Füredi, Naor and Verstraëte [FNV06].

11
 1+1/k
A general bound of ex(n, C2k )  kn , for some unspecified constant k, was asserted
by Erdős [Erd64, p. 33]. The first proof was by Bondy and Simonovits [BS74, Lemma 2],
who showed that ex(n, C2k )  20kn1+1/k for all sufficiently large n. This was improved by
Verstraëte [Ver00] to 8(k 1)n1+1/k and by Pikhurko [Pik12] to (k 1)n1+1/k + O(n). The
principal result of the chapter is an improvement of these bounds:
2
Theorem 3.1. If G is an n-vertex graph that contains no C2k and n (2k)8k , then
p
ex(n, C2k )  80 k log k · n1+1/k + 10k 2 n.

It is our duty to point out that the improvement o↵ered by the Theorem 3.1 is of uncertain
value because we still do not know if n1+1/k is the correct order of magnitude for ex(n, C2k ).
Only for k = 2, 3, 5 are constructions of C2k -free graphs with ⌦(n1+1/k ) edges known [Ben66,
Wen91, LU95, MM05]. We stress again that the situation is completely di↵erent for odd
cycles, where the value of ex(n, C2k+1 ) is known exactly for all large n [Sim68].
Our proof is inspired by that of Pikhurko [Pik12]. Apart from a couple of lemmas that
we quote from [Pik12], the present chapter is self-contained. However, we advise the readers
to at least skim [Pik12] to see the main idea in a simpler setting.
Pikhurko’s proof builds a breadth-first search tree, and then argues that a pair of adjacent
levels of the tree cannot contain a ⇥-graph1 . It is then deduced that each level must be at
least /(k 1) times larger than the previous, where is the (minimum) degree. The bound
on ex(n, C2k ) then follows. The estimate of /(k 1) is sharp when one restricts one’s
attention to a pair of levels.
In our proof, we use three adjacent levels. We find a ⇥-graph satisfying an extra technical
condition that permits an extension of Pikhurko’s argument. Annoyingly, this extension
requires a bound on the maximum degree. To achieve such a bound we use a modification
of breadth-first search that avoids the high-degree vertices.
What we really prove in this chapter is the following:

Theorem 3.2. Suppose k 4. If G is a biparite n-vertex graph of minimum degree at least

2
2d + 5k , where
p
d max(20 k log k · n1/k , (2k)8k ), (3.1)

then G contains C2k .

Theorem 3.1 follows from Theorem 3.2 and two well-known facts: every graph contains
a bipartite subgraph with half of the edges, and every graph of average degree davg contains
a subgraph of minimum degree at least 12 davg .
1 We recall the definition of a ⇥-graph in Section 3.3

12
 The rest of the chapter is organized as follows. We present our modification of breadth-
first search in Section 3.2. In Section 3.3, which is the heart of the chapter, we explain
how to find ⇥-graphs in triples of consecutive levels. Finally, in Section 3.4 we assemble the
pieces of the proof.

3.2 Graph exploration

Our aim is to have vertices of degree at most d for some k ⌧ ⌧ d1/k . The particular
choice is fairly flexible; we choose to use := k 3 .
Let G be a graph, and let x be any vertex of G. We start our exploration with the set
V0 = {x}, and mark the vertex x as explored. Suppose V0 , V1 , . . . , Vi 1 are the sets explored
in the first i steps respectively. We then define Vi as follows:
1. Let Vi0 consist of those neighbors of Vi 1 that have not yet been explored. Let Bgi be
the set of those vertices in Vi0 that have more than d unexplored neighbors, and let
Smi = Vi0 \ Bgi .
2. Define 8
<V 0 if |Bgi | > 1
|Vi0 | ,
i 2k
Vi =
:Smi if |Bgi |  1
|Vi0 | .
2k

The vertices of Vi are then marked as explored.

1
We call sets V0 , V1 , . . . levels of G. A level Vi is big if |Bgi | > 2k
|Vi0 |, and is normal otherwise.

Lemma 3.3. If  d, and G is a bipartite graph of minimum degree at least , then each
0
v 2 Vi+1 has at least neighbors in Vi [ Vi+2 .

Proof. Fix a vertex v 2 V (G). We will show, by induction on i, that if v 62 V1 [ · · · [ Vi , then

v has at least neighbors in V (G) \ (V1 [ · · · [ Vi 1 ). The base case i = 1 is clear. Suppose
i > 1. If v 2 Bgi , then v has d neighbors in the required set. Otherwise, v is not in Vi0
and hence has no neighbors in Vi 1 . Hence, v has as many neighbors in V (G)\(V1 [· · ·[Vi 1 )
as in V (G) \ (V1 [ · · · [ Vi 2 ), and our claim follows from the induction hypothesis.
0
If v 2 Vi+1 , then the neighbors of v are a subset of V1 [ · · · [ Vi [ Vi+2 . Hence, at least
0
of these neighbors lie in Vi [ Vi+2 .

3.2.1 Trilayered graphs

We abstract out the properties of a triple of consecutive levels into the following definition.
A trilayered graph with layers V1 , V2 , V3 is a graph G on a vertex set V1 [ V2 [ V3 such
that the only edges in G are between V1 and V2 , and between V2 and V3 . If V10 ⇢ V1 ,

13
V20 ⇢ V2 and V30 ⇢ V3 , then we denote by G[V10 , V20 , V30 ] the trilayered subgraph induced by
three sets V10 , V20 , V30 . Because the graph G that has been explored is bipartite, there are no
0
edges inside each level. Therefore any three sets Vi 1 , Vi , Vi+1 from the exploration process
naturally form a trilayered graph; these graphs and their subgraphs are the only trilayered
graphs that appear in this chapter.
We say that a trilayered graph has minimum degree at least [A : B, C : D] if each vertex
in V1 has at least A neighbors in V2 , each vertex in V2 has at least B neighbors in V1 , each
vertex in V2 has at least C neighbors in V3 , and each vertex in V3 has at least D neighbors
in V2 .

A C
! !

V1 V2 V3

B D

Figure 3.1: A schematic drawing of a tyilayered graph.

3.3 ⇥-graphs
A ⇥-graph is a cycle of length at least 2k with a chord. We shall use several lemmas from
the previous works.

Lemma 3.4 (Lemma 2.1 in [Pik12], also Lemma 2 in [Ver00]). Let F be a ⇥-graph and
1  l  |V (F )| 1. Let V (F ) = W [ Z be an arbitrary partition of its vertex set into
two non-empty parts such that every path in F of length l that begins in W necessarily ends
in W . Then F is bipartite with parts W and Z.

Lemma 3.5 (Lemma 2.2 in [Pik12]). Let k 3. Any bipartite graph H of minimum degree
at least k contains a ⇥-graph.

Corollary 3.6. Let k 3. Any bipartite graph H of average degree at least 2k contains a
⇥-graph.

For a graph G and a set Y ⇢ V (G), let G[Y ] denote the graph induced on Y . For
disjoint Y, Z ⇢ V (G), let G[Y, Z] denote the bipartite subgraph of G that is induced by the
bipartition Y [ Z.

14
3.3.1 Well-placed ⇥-graphs
Suppose G is a trilayered graph with layers V1 , V2 , V3 . We say that a ⇥-graph F ⇢ G is well-
placed if each vertex of V (F ) \ V2 is adjacent to some vertex in V1 \ V (F ). The condition
ensures that, for each vertex v of F in V2 there exists a path from the root x to v that avoids
F.

Lemma 3.7. Suppose G is a trilayered graph with layers V1 , V2 , V3 such that the degree of
every vertex in V2 is at least 2d + 5k 2 , and no vertex in V2 has more than d neighbors in
V3 . Suppose t is a nonnegative integer, and let F = d · e(V1 , V2 )/8k |V3 |. Assume that

F 2, (3.2a)
e(V1 , V2 ) 2kF |V1 | , (3.2b)
e(V1 , V2 ) 8k(t + 1)2 (2 k)2k 1
|V1 | , (3.2c)
e(V1 , V2 ) 8(et/F )t k |V2 | , (3.2d)
e(V1 , V2 ) 20(t + 1)2 |V2 | . (3.2e)

Then at least one of the following holds:

1. There is a ⇥-graph in G[V1 , V2 ].
2. There is a well-placed ⇥-graph in G[V1 , V2 , V3 ].

The proof of Lemma 3.7 is in two parts: finding trilayered subgraph of large minimum
degree (Lemmas 3.8 and 3.9), and finding a well-placed ⇥-graph inside that trilayered graph
(Lemma 3.10).

3.3.2 Finding a trilayered subgraph of large minimum degree

The disjoint union of two bipartite graphs shows that a trilayered graph with many edges
need not contain a trilayered subgraph of large minimum degree. We show that, in contrast,
if a trilayered graph contains no ⇥-graph between two of its levels, then it must contain a
subgraph of large minimum degree. The next lemma demonstrates a weaker version of this
claim: it leaves open a possibility that the graph contains a denser trilayered subgraph. In
that case, we can iterate inside that subgraph, which is done in Lemma 3.9.

Lemma 3.8. Let a, A, B, C, D be positive real numbers. Suppose G is a trilayered graph with
layers V1 , V2 , V3 and the degree of every vertex in V2 is at least d + 4k 2 + C. Assume also
that
a · e(V1 , V2 ) (A + k + 1) |V1 | + B |V2 | . (3.3)
Then one of the following holds:

15
 1. There is a ⇥-graph in G[V1 , V2 ].
2. There exist non-empty subsets V10 ⇢ V1 , V20 ⇢ V2 , V30 ⇢ V3 such that the induced
trilayered subgraph G[V10 , V20 , V30 ] has minimum degree at least [A : B, C : D].
3. There is a subset Ve2 ⇢ V2 such that e(V1 , Ve2 ) (1 a)e(V1 , V2 ), and Ve2  D |V3 | /d.

Proof. We suppose that alternative 1 does not hold. Then, by Corollary 3.6, the average
degree of every subgraph of G[V1 , V2 ] is at most 2k.
Consider the process that aims to construct a subgraph satisfying 2. The process starts
with V10 = V1 , V20 = V2 and V30 = V3 , and at each step removes one of the vertices that violate
the minimum degree condition on G[V10 , V20 , V30 ]. The process stops when either no vertices
are left, or the minimum degree of G[V10 , V20 , V30 ] is at least [A : B, C : D]. Since in the latter
case we are done, we assume that this process eventually removes every vertex of G.
Let R be the vertices of V2 that were removed because at the time of removal they had
fewer than C neighbors in V30 . Put

E 0 := {uv 2 E(G) : u 2 V2 , v 2 V3 , and v was removed before u},

S := {v 2 V2 : v has at least 4k 2 neighbors in V1 }.

Note that |E 0 |  D |V3 |. We cannot have |S| |V1 | /k, for otherwise the average degree of
4k
the bipartite graph G[V1 , S] would be at least 1+1/k 2k. So |S|  |V1 | /k.
The average degree condition on G[V1 , S] implies that

e(V1 , S)  k(|V1 | + |S|)  (k + 1) |V1 | .

Let u be any vertex in R \ S. Since it is connected to at least (d + 4k 2 + C) 4k 2 = d + C

vertices of V3 , it must be adjacent to at least d edges of E 0 . Thus,

|R \ S|  |E 0 | /d  D |V3 | /d.

Assume that the conclusion 3 does not hold with Ve2 = R \ S. Then e(V1 , R \ S) <
(1 a)e(V1 , V2 ). Since the total number of edges between V1 and V2 that were removed due
to the minimal degree conditions on V1 and V2 is at most A |V1 | and B |V2 | respectively, we
conclude that

e(V1 , V2 )  e(V1 , S) + e(V1 , R \ S) + A |V1 | + B |V2 |

< (k + 1) |V1 | + (1 a)e(V1 , V2 ) + A |V1 | + B |V2 | ,

implying that
a · e(V1 , V2 ) < (A + k + 1) |V1 | + B |V2 | .
The contradiction with (3.3) completes the proof.

16
Remark 3.1. The next lemma can be eliminated at the cost of obtaining the bound ex(n, C2k ) =
p
O(k 2/3 n1+1/k ) in place of ex(n, C2k ) = O( k log k · n1+1/k ). To do that, we can set B ⇡ k 2/3 ,
D ⇡ k 1/3 and a = 1/2. One can then show that when applied to trilayered graphs arising
from the exploration process the alternative 3 leads to a subgraph of average degree 2k.
The two remaining alternatives are dealt by Corollary 3.6 and Lemma 3.10. However, it is
possible to obtain a better bound by iterating the preceding lemma.

Lemma 3.9. Let C be a positive real number. Suppose G is a trilayered graph with layers V1 ,
V2 , V3 , and the degree of every vertex in V2 is at least d+4k 2 +C. Let F = d · e(V1 , V2 )/8k |V3 |,
and assume that F and e(V1 , V2 ) satisfy (3.2) for some integer t 1. Then one of the
following holds:
1. There is a ⇥-graph in G[V1 , V2 ].
2. There exist numbers A, B, D and non-empty subsets V10 ⇢ V1 , V20 ⇢ V2 , V30 ⇢ V3
such that the induced trilayered subgraph G[V10 , V20 , V30 ] has minimum degree at least
[A : B, C : D], with the following inequalities that bind A, B, and D:

B 5, (B 4)D 2k, A 2k( D)D 1 . (3.4)

Proof. Assume, for the sake of contradiction, that neither 1 nor 2 hold. With hindsight, set
1 (0) (1)
aj = t j+1 for j = 0, . . . , t 1. We shall define a sequence of sets V2 = V2 ◆ V2 ◆ ··· ◆
(t)
V2 inductively. We denote by
(i) (i)
di := e(V1 , V2 )/ V2

(i) (0) (1) (t)

the average degree from V2 into V1 . The sequence V2 , V2 , . . . , V2 will be constructed
so as to satisfy
(i+1) (i)
e(V1 , V2 ) (1 ai )e(V1 , V2 ), (3.5)
i
Y
di+1 d i · F ai (1 aj ). (3.6)
j=0

Note that (3.5) and the choice of a0 , . . . , ai imply that

(i) e(V1 , V2 )
e(V1 , V2 ) . (3.7)
t+1
(0) (i)
The sequence starts with V2 = V2 . Assume V2 has been defined. We proceed to define
(i+1)
V2 . Put
(i) ✓ ◆
ai e(V1 , V2 ) 1 8k
A= k 1, B= ad
4 i i
+ 5, D = min 2k, .
2 |V1 | ai d i

17
With help of (3.7) and (3.2c) it is easy to check that the inequalities (3.4) hold for this choice
of constants.
In addition,
(i) (i) (i)
(A + k + 1) |V1 | + B V2 = 34 ai e(V1 , V2 ) + 5 V2
(3.2e)
3 (i) 1
 a e(V1 , V2 )
4 i
+ 4(t+1)2
e(V1 , V2 )
(3.7)
(i)
 ai e(V1 , V2 ).
h i
(i)
So, the condition (3.3) of Lemma 3.8 is satisfied for the graph G V1 , V2 , V3 . By Lemma 3.8
(i+1) (i) (i+1)
there is a subset V2 ⇢ V2 satisfying (3.5) and V2  D |V3 | /d.
(i+1)
Next we show that the set V2 satisfies inequality (3.6). Indeed, we have
(i+1) (i)
e(V1 , V2 ) (1 ai )e(V1 , V2 ) d (i)
di+1 = (1 ai )ai di e(V1 , V2 )
V2
(i+1) D |V3 | /d 8k |V3 |
i 1 i
(3.5) d · e(V1 , V2 ) Y Y
(1 ai )ai di (1 aj ) = d i · F a i (1 aj ).
8k |V3 | j=0 j=0

Iterative application of (3.6) implies

t 1
Y t 1
Y
t t j t e 1 (F/e)t
dt d0 F aj (1 aj ) d0 F = d0 . (3.8)
j=0 j=0
t j+1 (t + 1)!
h i
(t) (t)
If we have V2 < |V1 |, then the average degree of induced subgraph G V 1 , V2 is
greater than
(3.7) (3.2c)
(t)
e(V1 , V2 )/ |V1 | e(V1 , V2 )/(t + 1) |V1 | 2k,

which by Corollary 3.6 leads to outcome 1. h i

(t) (t)
If V2 |V1 | and dt 4k, then the average degree of G V1 , V2 is at least dt /2 2k
(t)
because dt is the average degree of V2 into V1 , again leading to the outcome 1. So, we may
assume that dt < 4k. Since (t + 1)!  2tt we deduce from (3.8) that

d0 < 4k(t + 1)!(e/F )t  8k(et/F )t .

This contradicts (3.2d), and so the proof is complete.

18
3.3.3 Locating well-placed ⇥-graphs in trilayered graphs
We come to the central argument of the chapter. It shows how to embed well-placed ⇥-
graphs into trilayered graphs of large minimum degree. Or rather, it shows how to embed
well-placed ⇥-graphs into regular trilayered graphs; the contortions of the previous two
p
lemmas, and the factor of log k in the final bound, come from authors’ inability to deal
with irregular graphs.

Lemma 3.10. Let A, B, D be positive real numbers. Let G be a trilayered graph with layers
V1 , V2 , V3 of minimum degree at least [A : B, d + k : D]. Suppose that no vertex in V2 has
more than d neighbors in V3 . Assume also that

B 5 (3.9)
(B 4)D 2k 2 (3.10)
A 2k( D)D 1 . (3.11)

Then G contains a well-placed ⇥-graph.

Proof. Assume, for the sake of contradiction, that G contains no well-placed ⇥-graphs.
Leaning on this assumption, we shall build an arbitrary long path P of the form

• • • • • • V3

• • • • • • • • V2

v0 • v1 • v2 • V1

Figure 3.2: An arbitrary long path P .

where, for each i, vertices vi and vi+1 are joined by a path of length 2D that alternates
between V2 and V3 . Since the graph is finite, this would be a contradiction.
While building the path, we maintain the following property:

Every v 2 P \ V2 has at least one neighbor in V1 \ P . (3.12)

We call a path satisfying (3.12) good.

We construct the path inductively. We begin by picking v0 arbitrarily from V1 . Suppose
a good path P = v0 ! v1 ! · · · ! vl 1 has been constructed, and we wish to find a path
extension v0 ! v1 ! · · · ! vl 1 ! vl .

19
 There are at least about A ways to extend the path by a single vertex. The idea of the
following argument shows that many of these extensions can be extended to another vertex,
and then another, and so on.
For each i = 1, 2, . . . , 2D 1 we shall define a family Qi of good paths that satisfy
1. Each path in Qi is of the form v0 ! v1 ! · · · ! vl 1 ! u, where vl 1 ! u is a
path of length i that alternates between V2 and V3 . The vertex u is called a terminal
of the path. The set of terminals of the paths in Qi is denoted by T (Qi ). Note that
T (Qi ) ⇢ V2 for odd i and T (Qi ) ⇢ V3 for even i.
2. For each i, the paths in Qi have distinct terminals.
3. For odd-numbered indices, we have the inequality
✓ ◆i Y ✓ ◆
1 j
|Q2i+1 | 3k + A 1 . (3.13)
ji
D

4. For even-numbered indices, we have the inequality

e(T (Q2i ), V2 ) d |Q2i 1 | . (3.14)

Let t := dB/2e. We will repeatedly use the following straightforward fact, which we call
the small-degree argument: whenever Q is a good path and u 2 V2 \ Q is adjacent to the
terminal of Q, then u is adjacent to fewer than t vertices in V1 \ Q. Indeed, if vertex u were
adjacent to vj1 , vj2 , . . . , vjt 2 V1 \ Q with j1 < j2 < . . . < jk , then vj2 ! u (along path Q)
and the edge uvj2 would form a cycle of total length at least
(3.10)
2D(t 2) + 2 2D(B/2 2) + 2 2k.

As uvj3 is a chord of the cycle, and u is adjacent to vj1 that is not on the cycle, that would
contradict the assumption that G contains no well-placed ⇥-graph.
The set Q1 consists of all paths of the form P u for u 2 V2 \ P . Let us check that the
preceding conditions hold for Q1 . Vertex vl 1 cannot be adjacent to k or more vertices in
P \ V2 , for otherwise G would contain a well-placed ⇥-graph with a chord through vl 1 . So,
|Q1 | A k. Next, consider any u 2 V2 \ P that is a neighbor of vl 1 . By the small-degree
argument vertex u cannot be adjacent to t or more vertices of P \ V1 , and P u is good.
Suppose Q2i 1 has been defined, and we wish to define Q2i . Consider an arbitrary path
Q = v0 ! v1 ! · · · ! vl 1 ! u 2 Q2i 1 . Vertex u cannot have k or more neighbors
in Q \ V3 , for otherwise G would contain a well-placed ⇥-graph with a chord through u.
Hence, there are at least d edges of the form uw, where w 2 V3 \ Q. As we vary u we obtain
a family of at least d |Q2i 1 | paths. We let Q2i consist of any maximal subfamily of such

20
paths with distinct terminals. The condition (3.14) follows automatically as each vertex of
T (Q2i 1 ) has at least d neighbors in T (Q2i ).
Suppose Q2i has been defined, and we wish to define Q2i+1 . Consider an arbitrary path
Q = v0 ! v1 ! · · · ! vl 1 ! u 2 Q2i . An edge uw is called long if w 2 P , and w
is at a distance exceeding 2k from u along path Q. If uw is a long edge, then from u to Q
there is only one edge, namely the edge to the predecessor of u on Q, for otherwise there
is a well-placed ⇥-graph. Also, at most i neighbors of u lie on the path vl 1 ! u. Since
deg u D, it follows that there are at least (1 i/D) deg u short edges from u that miss
vl 1 ! u. Thus there is a set W of at least (1 i/D)e(T (Q2i ), V2 ) walks (not necessarily
paths!) of the form v0 ! v1 ! · · · ! vl 1 ! uw such that vl 1 ! uw is a path and w
occurs only among the last 2k vertices of the walk.
From the maximum degree condition on V2 it follows that walks in W have at least
(1 i/D)e(T (Q2i ), V2 )/ d distinct terminals. A walk fails to be a path only if the terminal
vertex lies on P . However, since the edge uw is short, this can happen for at most 2k possible
terminals. Hence, there is a Q2i+1 ⇢ W of size

|Q2i+1 | (1 i/D)e(T (Q2i ), V2 )/ d 2k (3.15)

that consists of paths with distinct terminals. It remains to check that every path in Q2i+1
is good. The only way that Q = v0 ! · · · ! vl 1 ! uw 2 Q2i+1 may fail to be good is if
w has no neighbors in V1 \ Q. By the small-degree argument w has fewer than t neighbors
in V1 . Since w has at least B neighbors in V1 and B t + 2, we conclude that w has at
least two neighbors in V1 outside the path. Of course, the same is true for every terminal of
a path in Q2i+1 . The condition (3.13) for Q2i+1 follows from (3.15), (3.14) and from validity
of (3.13) for Q2i 1 .
Note that Q2D 1 is non-empty. Let Q = v0 ! · · · ! vl 1 ! u 2 Q2D 1 be an arbitrary
path. Note that since 2D 1 is odd, u 2 V2 . By the property of terminals of Vi (odd i)
that we noted in the previous paragraph, there are two vertices in V1 \ Q that are neighbors
of u. Let vl be any of them, and let the new path be Qvl = v0 ! · · · ! vl 1 ! uvl . This
path can fail to be good if there is a vertex w on the path Q that is good in Q, but is bad
in Qvl . By the small-degree argument, w is adjacent to fewer than t vertices in Q \ V1 that
precede w in Q. The same argument applied to the reversal of the path Qvl shows that w is
adjacent to fewer than t vertices in Q \ V1 that succeeds w in Q. Since 2t 2 < B, the path
Qvl is good.
Hence, it is possible to build an arbitrarily long path in G. This contradicts the finiteness
of G.

21
 Lemma 3.7 follows from Lemmas 3.9 and 3.10 by setting C = d + k, in view of inequality
4k 2 + k  5k 2 . We lose k 2 k here for cosmetic reason: 5k 2 is tidier than 4k 2 + k.

3.4 Proof of Theorem 3.2

Suppose that G is a bipartite graph of minimum degree at least 2d + 5k 2 and contains no
C2k . Pick a root vertex x arbitrarily, and let V0 , V1 , . . . , Vk 1 be the levels obtained from the
exploration process in Section 3.2.

Lemma 3.11. For 1  i  k 1, the graph G[Vi 1 , Vi , Vi+1 ] contains no well-placed ⇥-graph.

Proof. The following proof is almost an exact repetition of the proof of Claim 3.1 from
[Pik12] (which is also reproduced as Lemma 3.12 below).
Suppose, for the sake of contradiction, that a well-placed ⇥-graph F ⇢ G[Vi 1 , Vi , Vi+1 ]
exists. Let Y = Vi \ V (F ). Since F is well-placed, for every vertex of Y there is a path
avoiding V (F ) of length i to the vertex x. The union of these paths forms a tree T with x
as a root. Let y be the vertex farthest from x such that every vertex of Y is a T -descendant
of y. Paths that connect x to Y branch at y. Pick one such branch, and let W ⇢ Y be the
set of all the T -descendants of that branch. Let Z = V (F ) \ W . From W 6= Vi \ V (F ) it
follows that Z is not an independent set of F , and so W [ Z is not a bipartition of F .
Let ` be the distance between x and y. We have ` < i and 2k 2i + 2` < 2k  |V (F )|.
By Lemma 3.4 in F there is a path P of length 2k 2i + 2` that starts at some w 2 W and
ends in z 2 Z. Since the length of P is even, z 2 Y . Let Pw and Pz be unique paths in T
that connect y to respectively w and z. They intersect only at y. Each of Pw and Pz has
length i `. The union of paths P, Pw , Pz forms a 2k-cycle in G.

The same argument (with a di↵erent Y ) also proves the next lemma.

Lemma 3.12 (Claim 3.1 in [Pik12]). For 1  i  k 1, neither of G[Vi ] and G[Vi , Vi+1 ]
contains a bipartite ⇥-graph.

The next step is to show that the levels V0 , V1 , V2 , . . . increase in size. We shall show by
induction on i that

e(Vi , Vi+1 ) d |Vi | , (3.16)

e(Vi , Vi+1 )  2k |Vi+1 | , (3.17)
0 0
e(Vi , Vi+1 )  2k Vi+1 , (3.18)
|Vi+1 | (2k) 1 d |Vi | , (3.19)
d2
|Vi+1 | 400k log k
|Vi 1 | . (3.20)

22
To prove Theorem 3.2, we only need (3.20); the remaining inequalities play auxiliary roles in
derivation of (3.20). Clearly, these inequalities hold for i = 0 since each vertex of V1 sends
only one edge to V0 .

Proof of (3.16). By Lemma 3.3 the degree of every vertex in Vi is at least 2d + 4k, and so
induc.
0
e(Vi , Vi+1 ) (2d + 4k) |Vi | e(Vi 1 , Vi ) (2d + 2k) |Vi | .

We next distinguish two cases depending on whether Vi+1 is big (in the sense of the definition
0
from Section 3.2). If Vi+1 is big, then e(Vi , Vi+1 ) = e(Vi , Vi+1 ), and (3.16) follows. If Vi+1 is
normal, then Corollary 3.6 and Lemma 3.12 imply that

1 0 0
e(Vi , Bgi+1 )  k(|Vi | + Bgi+1 )  k |Vi | + 2k
Vi+1  k |Vi | + 12 e(Vi , Vi+1 )

and so

0 1 0
e(Vi , Vi+1 ) = e(Vi , Vi+1 ) e(Vi , Bgi+1 ) 2
e(Vi , Vi+1 ) k |Vi | d |Vi |

implying (3.16).

Proof of (3.17). Consider the graph G[Vi , Vi+1 ]. Inequality (3.16) asserts that the average
degree of Vi is at least d 2k. If (3.17) does not hold, then the average degree of Vi+1 is at
least 2k as well, contradicting Corollary 3.6 and Lemma 3.12.
0
Proof of (3.18). The argument is the same as for (3.17) with G[Vi , Vi+1 ] in place of G[Vi , Vi+1 ].

Proof of (3.19). This follows from (3.17) and (3.16).

Proof of (3.20) in the case Vi is a normal level. We assume that (3.20) does not hold and
0
will derive a contradiction. Consider the trilayered graph G[Vi 1 , Vi , Vi+1 ]. Let t = 2 log k.
Suppose momentarily that the inequalities (3.2) in Lemma 3.7 hold. Then since Vi is normal,
0
each vertex in Vi has at most d neighbors in Vi+1 , and so Lemma 3.7 applies. However,
the lemma’s conclusion contradicts Lemmas 3.11 and 3.12. Hence, to prove (3.20) it suffices
0
to verify inequalities (3.2a–3.2d) with F = d · e(Vi 1 , Vi )/8k Vi+1 .
We may assume that
F 2e2 log k, (3.21)

and in particular that (3.2a) holds. Indeed, if (3.21) were not true, then inequality (3.16)
0
would imply Vi+1 (d2 /16e2 k log k) |Vi 1 |, and thus

0
|Vi+1 | (1 1
k
) Vi+1 (d2 /32e2 k log k) |Vi 1 | ,

23
and so (3.20) would follow in view of 32e2  400.
Inequality (3.2b) is implied by (3.19). Indeed,
(3.19) (3.19)
0
e(Vi 1 , Vi ) = 8k Vi+1 F/d 8k |Vi+1 | F/d 4F |Vi | 2k 1 dF |Vi 1 | ,

and d k 2 by the definition of d from (3.1).

Inequality (3.2c) is implied by (3.1) and (3.16).
Next, suppose (3.2d) were not true. Since F/t e2 by (3.21), we would then conclude
(3.19)
|Vi+1 | (2k) 1 d |Vi | d(16k 2 ) 1 (F/et)t e(Vi 1 , Vi )
(3.16)
d(16k 2 ) 1 e2 log k e(Vi 1 , Vi ) 1 2
16
d |Vi 1 | ,

and so (3.20) would follow.

Finally, (3.2e) is a consequence of (3.16).

Proof of (3.20) in the case Vi is a big level. We have

(3.18)
0 0 0
|Vi+1 | 1
2
Vi+1 (4k) 1 e(Vi , Vi+1 ) (4k) 1 e(Bgi , Vi+1 ) (4k) 1
d |Bgi |
(3.19)
(8k 2 ) 1
d |Vi | (16k 3 ) 1
d2 |Vi 1 | = 1 2
16
d |Vi 1 | ,

and so (3.20) holds.

Proof of Theorem 3.2. If k is even, then k/2 applications of (3.20) yield

dk
|Vk | .
(400k log k)k/2

If k is odd, then (k 1)/2 applications of (3.20) yield

dk 1 dk
|Vk | |V1 | .
(400k log k)(k
1)/2 (400k log k)(k 1)/2

p
Either way, since |Vk | < n we conclude that d < 20 k log k · n1/k .

24
Chapter 4

Bipartite algebraic graphs without

quadrilaterals

4.1 Introduction
The Turán number ex(n, F ) is the maximum number of edges in an F -free graph1 on n
vertices. The first systematic study of ex(n, F ) was initiated by Turán [Tur41], who solved
the case when F = Kt is a complete graph on t vertices. Turán’s theorem states that, on
a given vertex set, the Kt -free graph with the most edges is the complete and balanced
(t 1)-partite graph, in that the part sizes are as equal as possible.
For general graphs F , we still do not know how to compute the Turán number exactly, but
if we are satisfied with an approximate answer, the theory becomes quite simple: Erdős and
Stone [ES46] showed that if the chromatic number (F ) = t, then ex(n, F ) = ex(n, Kt ) +
o(n2 ) = 1 t 1 1 n2 + o(n2 ). When F is not bipartite, this gives an asymptotic result
for the Turán number. On the other hand, for all but few bipartite graphs F , the order
of ex(n, F ) is not known. Most of the research on this problem focused on two classes of
graphs: complete bipartite graphs and cycles of even length. A comprehensive survey is by
Füredi and Simonovits [FS13].
Suppose G is a Ks,t -free graph with s  t. The Kövari–Sós–Turán theorem [KST54]
p
implies an upper bound ex(n, Ks,t )  12 s t 1 · n2 1/s + o(n2 1/s ), which was improved by
Füredi [Für96b] to
1p
ex(n, Ks,t )  s
t s + 1 · n2 1/s
+ o(n2 1/s
).
2
Despite the lack of progress on the Turán problem for complete bipartite graphs, there are
certain complete bipartite graphs for which the problem has been solved asymptotically, or
1 We say a graph is F -free if it does not have a subgraph isomorphic to F .

25
even exactly. The constructions that match the upper bounds in these cases are all similar
to one another. Each of the constructions is a bipartite graph G based on an algebraic
hypersurface2 H. Both partite sets of G are Fsp and the edge set is defined by: u ⇠ v if and
only if (u, v) 2 H. In short, G = Fsp , Fsp , H(Fp ) , where H(Fp ) denotes the Fp -points of H.
Note that G has n := 2ps vertices.
In the previous works of Erdős, Rényi and Sós [ERS66], Brown [Bro66], Füredi [Für96a],
Kollár, Rónyai and Szabó [KRS96] and Alon, Rónyai and Szabó [ARS99], various hypersur-
faces were used to define Ks,t -free graphs. Their equations were

x1 y1 + x2 y2 = 1, for K2,2 ; (4.1a)

(x1 y1 )2 + (x2 y2 )2 + (x3 y3 )2 = 1, for K3,3 ; (4.1b)
(Ns ⇡s )(x1 + y1 , x2 + y2 , . . . , xs + ys ) = 1, for Ks,t with t s! + 1; (4.1c)
(Ns 1 ⇡s 1 )(x2 + y2 , x3 + y3 , . . . , xs + ys ) = x1 y1 , for Ks,t with t (s 1)! + 1, (4.1d)

where ⇡s : Fsp ! Fps is an Fp -linear isomorphism and Ns (↵) is the field norm, Ns (↵) :=
s
↵(p 1)/(p 1)
.
Clearly, the coefficients in (4.1a) and (4.1b) are integers and even independent of p.
With some work, one can show that both (4.1c) and (4.1d) are polynomial equations of
degree  s with coefficients in Fp . Therefore each equation in (4.1) can be written as
F (x, y) := F (x1 , . . . , xs , y1 , . . . , ys ) = 0 for some F (x, y) 2 Fp [x, y] of bounded degree. The
previous works directly count the number of Fp solutions to F (x, y) = 0 and yield |H(Fp )| =
⇥(p2s 1 ) = ⇥(n2 1/s ), for each prime3 p.

Definition 1. Given two sets P1 and P2 , a set V ⇢ P1 ⇥ P2 is said to contain an (s, t)-grid
if there exist S ⇢ P1 , T ⇢ P2 such that s = |S|, t = |T | and S ⇥ T ⇢ V . Otherwise, we say
that V is (s, t)-grid-free.

Observe that every F (x, y) derived from (4.1) is symmetric in xi and yi for all i. We
know that (u, v) 2 H if and only if (v, u) 2 H for all u, v 2 Fsp . The resulting bipartite graph
G = Fsp , Fsp , H(Fp ) would be an extremal Ks,t -free graph if H(Fp ) had been (s, t)-grid-free.
So which graphs are Ks,t -free with a maximum number of edges? The question was
considered by Zoltán Füredi in his unpublished manuscript [Für88] asserting that every
K2,2 -free graph with q vertices (for q q0 ) and 12 q(q + 1)2 edges is obtained from a projective
2 An algebraic hypersurface in a space of dimension n is an algebraic subvariety of dimension
n 1. The terminology from algebraic geometry used throughout the article is standard, and can
be found in [Sha13].
3 We need p ⌘ 3 (mod 4) for (4.1b) to get the correct number of F points on H. If p ⌘ 1
p
(mod 4), then the right hand side of (4.1b) should be replaced by a quadratic non-residue in Fp .

26
plane via a polarity with q + 1 absolute elements. This loosely amounts to saying that all
extremal K2,2 -free graphs are defined by generalization of (4.1a).
However, classification of all extremal Ks,t -free graphs seems out of reach. We restrict
our attention to algebraically constructed graphs. Given a field F and a hypersurface H
defined over F, it is natural to ask when H(F) is (s, t)-grid-free. Because the general case
is difficult, we work with algebraically closed fields K in this chapter. Denote by Ps (K) the
s-dimensional projective space over K. We are interested in hypersurface H in Ps (K)⇥Ps (K).
Since standard machinery from model theory, to be discussed in Section 4.5, allows us to
transfer certain results over C (the field of complex numbers) to algebraically closed fields
of large characteristic, our focus will be on the K = C case. We use Ps for the s-dimensional
complex projective space and As := Ps \ {x0 = 0} for the s-dimensional complex affine space.
Note that even if H contains (s, t)-grids, one may remove a few points from the projective
space to destroy all (s, t)-grids in H. For example, the homogenization of (4.1b) is

(x1 y0 x0 y1 )2 + (x2 y0 x0 y2 )2 + (x3 y0 x0 y3 )2 = x20 y02 .

The equation defines hypersurface H in P3 ⇥ P3 . Let V := {x0 = x21 + x22 + x23 = 0} be a

variety in P3 . Since V ⇥ P3 ⇢ H, H contains a lot of (3, 3)-grids. However, H \ (A2 ⇥ A2 )
is (3, 3)-grid-free.

Definition 2. A set V ⇢ Ps ⇥ Ps is almost-(s, t)-grid-free if there are two nonempty Zariski-

open sets X, Y ⇢ Ps such that V \ (X ⇥ Y ) is (s, t)-grid-free.

Suppose the defining equation of H, say F (x, y), is of low degree in y. Heuristically, for
generic distinct u1 , . . . , us 2 Ps , by Bézout’s theorem, one would expect {F (u1 , y) = · · · =
F (us , y) = 0} to have few points. So we conjecture the following.

Informal conjecture. Every almost-(s, t)-grid-free hypersurface is equivalent, in a suitable

sense, to a hypersurface whose degree in y is bounded by some constant d := d(s, t).

The right equivalence notion depends on X and Y in Definition 2. We shall discuss

possible notions of equivalence in Section 4.2, and make three specific conjectures. Results
in support of these conjectures can be found in Section 4.3 and Section 4.4.
Before we make our conjectures precise, we note that an analogous situation occurs
for C2t -free graphs. The upper bound ex(n, C2t ) = O(n1+1/t ) first established by Bondy–
Simonovits [BS74] has been matched only for t = 2, 3, 5. The t = 2 case was already
mentioned above because C4 = K2,2 . The constructions for t = 3, 5 are also algebraic (see
[Ben66, FNV06] for t = 3 and [Ben66, Wen91] for t = 5). Also, a conjecture in a similar

27
spirit about algebraic graphs of girth eight was made by Dmytrenko, Lazebnik and Williford
[DLW07]. It was recently resolved by Hou, Lappano and Lazebnik [HLL15].
The chapter is organized as follows. In Section 4.2 we flesh out the informal conjecture
above, in Section 4.3 we briefly discuss the s = 1 case, in Section 4.4 we partially resolve
the s = t = 2 case, and finally in Section 4.5, we consider algebraically closed fields of large
characteristic.

4.2 Conjectures on the (s, t)-grid-free case

Given a field F, we denote by F[x] the set of homogeneous polynomials in F[x] and by
Fhom [x, y] the set of polynomials in F[x, y] that are separately homogeneous in x and y.
We might be tempted to guess the following instance of the informal conjecture.

False conjecture A. If H is almost-(s, t)-grid-free, then there exists F (x, y) 2 Chom [x, y]
of degree  d in y for some d = d(s, t) such that H = {F = 0}.

Unfortunately, Conjecture A is false because of the following example.

Example 4.1. Consider H0 := {x0 y0 + x1 y1 + x2 y2 = 0} and H1 defined by

x0 y0d + x1 y0d 1 y1 + x2 y0d 1 y2 + y0d f (y1 /y0 ) = 0, (4.2)

where f is a polynomial of degree d. One can check that both H0 and H1 \ {y0 = 0} are
(2, 2)-grid-free, whereas equation (4.2) can be of arbitrary large degree in y.
Behind Example 4.1 is the birational automorphism : P2 99K P2 defined by

(y0 : y1 : y2 ) := y0d : y0d 1 y1 : y0d 1 y2 + y0d f (y1 /y0 ) .

Note that id⇥ is a biregular map4 from H1 \{y0 = 0} to H0 \{y0 = 0}. Composition with the
automorphism increased the degree of H0 in y while preserving almost-(2, 2)-grid-freeness.
Here is another example illustrating the relationship between birational automorphisms and
(s, t)-grid-free hypersurfaces.
Example 4.2. Define H2 := {x0 y1 y2 + x1 y0 y2 + x2 y0 y1 = 0}. One can also check that H2 \
{y0 y1 y2 = 0} is (2, 2)-grid-free. Behind this example is the standard quadratic transformation
from P2 to itself given by (y0 : y1 : y2 ) = (y1 y2 : y0 y2 : y0 y1 ). Note that id⇥ is a biregular
map from H2 \ {y0 y1 y2 = 0} to H0 \ {y0 y1 y2 = 0}.
Let Cr (Ps ) be the group of birational automorphisms on Ps , also known as the Cremona
group. Evidently, the almost-(s, t)-grid-freeness is invariant under Cr (Ps ) ⇥ Cr (Ps ).
4A biregular map is a regular map whose inverse is also regular.

28
Proposition 4.1. If V1 ⇢ Ps ⇥ Ps is an almost-(s, t)-grid-free set, then so is V2 := ( X ⇥
Y )V1 for all X, Y 2 Cr (Ps ).

Remark 4.1. Though little is known about the structure of the Cremona group in 3 dimen-
sions and higher, the classical Noether–Castelnuovo theorem says that the Cremona group
Cr (P2 ) is generated by the group of projective linear transformations and the standard
quadratic transformation. The proof of this theorem, which is very delicate, can be found
in [AC02, Chapter 8].
We say that sets V1 , V2 ⇢ Ps ⇥ Ps are almost equal if there exist nonempty Zariski-open
sets X, Y ⇢ Ps such that V1 \ (X ⇥ Y ) = V2 \ (X ⇥ Y ). We believe that the only obstruction
to Conjecture A is the Cremona group.

Conjecture B. Suppose H is an irreducible hypersurface in Ps ⇥ Ps . If H is almost-(s, t)-

grid-free, then there exist 2 Cr (Ps ) and F (x, y) 2 Chom [x, y] of degree  d in y for some
d = d(s, t) such that H is almost equal to {F (id ⇥ ) = 0}.

Remark 4.2. The conjecture is false if the irreducibility of H is dropped. Take H0 and H1
from Example 4.1 and set f (y) = y d in (4.2), where d can be arbitrarily large. Because
both H0 and H1 are almost-(2, 2)-grid-free, we know that H0 [ H1 is almost-(2, 3)-grid-free.
However, one can show5 that for any 2 Cr (Ps ), the degree of (id ⇥ )(H0 [ H1 ) in y is d.
In fact, we believe in an even stronger conjecture.

Conjecture C. Suppose H is an irreducible hypersurface in Ps ⇥ Ps . Let X, Y be nonempty

Zariski-open subsets of Ps . If H \ (X ⇥ Y ) is (s, t)-grid-free, then there exist Y 0 ⇢ Ps , a
biregular map : Y ! Y 0 and F (x, y) 2 Chom [x, y] of degree  d in y for some d = d(s, t)
such that H \ (X ⇥ Y ) = {F (id ⇥ ) = 0} \ (X ⇥ Y ).

We prove Conjecture C if s = 1 and if s = t = 2, Y = P2 (see Section 4.3 and 4.4

respectively).
One special case is when H \ (As ⇥ As ) is (s, t)-grid-free. In this case, H can be seen as
an affine algebraic hypersurface in 2s-dimensional affine space. The group of automorphisms
of As , denoted by Aut (As ), is a subgroup of the Cremona group. In this special case, we
make a stronger conjecture.
5 Let R1 = P1 /Q1 , R2 = P2 /Q2 be rational functions such that 1 = (1 : R : R ) and set
1 2
0
d1 = deg P1 = deg Q1 , d2 = deg P2 = deg Q2 . On the one hand, H0 := (id ⇥ )H0 is defined by
x0 +x1 R1 +x2 R2 = 0, and so degy H00 d2 . On the other hand, H10 is defined by x0 +x1 R1 +x2 (R2 +
⇣ ⌘
R1d ) = 0 and so degy H10 = dd1 + d2 deg G, where G = gcd Qd1 Q2 , P1 Qd1 1 Q2 , Qd1 P2 + P1d Q2 .
⇣ ⌘
Since G = gcd Qd1 1 Q2 , Qd1 P2 + P1d Q2 , it follows that G divides Q22 . Hence we estimate that
deg G  2d2 and degy H10 dd1 d2 d d2 . So, degy (H00 [ H10 ) d.

29
Conjecture D. Suppose H is an irreducible affine hypersurface in As ⇥ As . If H is (s, t)-
grid-free, then there exist 2 Aut (As ) and F (x, y) 2 C[x, y] of degree  d in y for some
d = d(s, t) such that H = {F (id ⇥ ) = 0}.
Remark 4.3. An automorphism 2 Aut (As ) is elementary if it has a form

: (x1 , . . . , xi 1 , xi , xi+1 , . . . , xs ) 7! (x1 , . . . , xi 1 , cxi + f, xi+1 , . . . , xs ),

where 0 6= c 2 C, f 2 C[x1 , . . . , xi 1 , xi+1 , . . . , xs ]. The tame subgroup is the subgroup

of Aut (As ) generated by all the elementary automorphisms, and the elements from this
subgroup are called tame automorphisms, while non-tame automorphisms are called wild. In
Example 4.1, we used a tame automorphism to make a counterexample to Conjecture A. It
is known [Jun42, vdK53] that all the elements of Aut (A2 ) are tame. However, in the case
of 3 dimensions, the following automorphism constructed by Nagata (see [Nag72]):

(x, y, z) = x + (x2 yz)z, y + 2(x2 yz)x + (x2 yz)z, z

was shown [SU03, SU04] to be wild. See also [Kur10]. We note that the question on the
existence of wild automorphisms remains open for higher dimensions.

4.3 Results on the (1, t)-grid-free case

As for the s = 1 case, one is able to fully characterize (1, t)-grid-free hypersurfaces. We
always assume that H is a hypersurface in P1 ⇥ P1 and X, Y are nonempty Zariski-open
subsets of P1 throughout this section.
Theorem 4.2. Suppose H = {F = 0}, where F (x, y) 2 Chom [x, y]. Let

F (x, y) = f (x)g(y)h1 (x, y)r1 h2 (x, y)r2 . . . hn (x, y)rn (4.3)

be the factorization of F such that h1 , h2 , . . . , hn are distinct irreducible polynomials depend-

ing on both x and y. Let di be the degree of hi in y. Then H \ (X ⇥ Y ) is (1, t)-grid-free if
and only if {f = 0} \ X = ; and |{g = 0} \ Y | + d1 + d2 + · · · + dn < t.
Proof. Clearly, if H \ (X ⇥ Y ) is (1, t)-grid-free, then {f = 0} \ X is empty. For fixed u 2 P1 ,
consider the following n + n + n + n + n2 systems of equations in y:

deg hi (u, y) < di , i = 1, 2, . . . , n;

y 2 P1 \ Y and hi (u, y) = 0, i = 1, 2, . . . , n;
hi (u, y) = @y hi (u, y) = 0, i = 1, 2, . . . , n;
hi (u, y) = g(y) = 0, i = 1, 2, . . . , n;
hi (u, y) = hj (u, y) = 0, i 6= j.

30
Since hi ’s are irreducible and distinct, Bézout’s theorem tells us that each of these systems
has no solution in P1 for a generic u. So for a generic u 2 P1 , F (u, y) = 0 has exactly
M := |{g = 0} \ Y | + d1 + d2 + · · · + dn distinct solutions in Y . The conclusion follows as
M is the maximal number of distinct solutions.

The informal conjecture thus holds when s = 1 as Theorem 4.2 implies:

Corollary 4.3. If H \ (X ⇥ Y ) is (1, t)-grid-free, then there exists F (x, y) 2 Chom [x, y] of
degree < t in y such that H \ (X ⇥ Y ) = {F = 0} \ (X ⇥ Y ).

Proof. Let H = {F = 0}, and let f, g and h1 , h2 , . . . , hn be the factors of F as in (4.3).

Suppose m := |{g = 0} \ Y | and {g = 0} \ Y = {v 1 , v 2 , . . . , v m }. Let gi (y) 2 Chom [y] be
linear such that {gi = 0} = {v i }. By Theorem 4.2,

F̃ (x, y) := g1 (y)g2 (y) . . . gm (y)h1 (x, y)h2 (x, y) . . . hn (x, y)

is of degree m + d1 + d2 + . . . dn < t in y. Clearly, F̃ = F on X ⇥ Y .

Conjectures B, C and D follow from the corollary in the s = 1 case. The birational
map becomes trivial in those conjectures since Cr (P1 ) consists only of projective linear
transformations.

4.4 Results on the (2, 2)-grid-free case

Throughout the section we assume that H is a hypersurface in P2 ⇥ P2 and X is a nonempty
Zariski-open subset of P2 .

Theorem 4.4. If H \ (X ⇥ P2 ) is (2, 2)-grid-free, then there exists F (x, y) 2 Chom [x, y] of
degree  2 in y such that H \ (X ⇥ P2 ) = {F = 0} \ (X ⇥ P2 ).

The theorem resolves Conjecture C for s = t = 2, Y = P2 . Note that the birational map
: Y ! Y 0 in the conjecture becomes trivial since biregular automorphisms of P2 are linear.
Our argument uses a reduction to an intersection problem of plane algebraic curves.
The key ingredient is a theorem by Moura [Mou04] on the intersection multiplicity of plane
algebraic curves.

Theorem 4.5 (Moura [Mou04]). Denote by Iv (C1 , C2 ) the intersection multiplicity of alge-
braic curves C1 and C2 at v. For a generic point v on an irreducible algebraic curve C1 of
degree d1 in P2 ,
8
< 1 (d2 + 3d ) if d1 > d2 ;
2 2
6 C2 , deg C2  d2 } = 2
max {Iv (C1 , C2 ) : C1 ⇢
C2 :d1 d2 1 (d2 3d1 + 2) if d1  d2 .
2 1

31
Corollary 4.6. For a generic point v on an algebraic curve C in P2 , any algebraic curve C 0
with v 2 C 0 intersects with C at another point unless C is irreducible of degree  2.

Proof. Suppose C has more than one irreducible components. Let C1 and C2 be any two of
them. Since C1 \ C2 is finite, we can pick a generic point v on C1 \ C2 . Now any algebraic
curve C 0 containing v intersects C at another point on C2 . So, C is irreducible.
Let d and d0 be the degrees of C and C 0 respectively. By Theorem 4.5, one can check
that Iv (C, C 0 ) < dd0 for a generic point v 2 C for all d > 2. From Bézout’s theorem, we
deduce that C intersects C 0 at another point unless d  2.

In our proof of Theorem 4.4, we think of H as a family of algebraic curves in P2 , each

of which is indexed by u 2 X and is defined by C(u) := {v 2 P2 : (u, v) 2 H}. We call
algebraic curve C(u) the section of H at u. A hypersurface H is (2, 2)-grid-free if and only
if C(u) and C(u0 ) intersect at  1 point for all distinct u, u0 2 X. The last piece that we
need for our proof is a technical lemma on generic sections of irreducible hypersurfaces.

Lemma 4.7. Suppose H1 and H2 are two di↵erent irreducible hypersurfaces in P2 ⇥ P2

defined by h1 (x, y), h2 (x, y) 2 Chom [x, y] \ (Chom [x] [ Chom [y]) respectively. Denote the section
of Hi at u by Ci (u) for i = 1, 2. For generic u 2 P2 , C1 (u) and C2 (u) share no common
irreducible components, and moreover, each Ci (u) is a reduced6 algebraic curve.

Proof of Theorem 4.4 assuming Lemma 4.7. Suppose H \ (X ⇥ P2 ) is (2, 2)-grid-free. Take
an arbitrary u 2 X and consider algebraic curve C(u) in P2 . We claim that every v 2 C(u) is
an intersection of C(u) and C(u0 ) for some u0 2 X \ {u}. Define D(v) := {F (x, v) = 0} \ X.
Since P2 \ X is Zariski-closed, the set D(v) is either empty or infinite. However, u 2 D(v)
and the claim is equivalent to |D(v)| 2.
Now pick a generic v 2 C(u). We know that point v is an intersection of C(u) and C(u0 )
for some u0 2 X \ {u} and it is the only intersection because H \ (X ⇥ P2 ) is (2, 2)-grid-free.
We apply Corollary 4.6 to C(u) and C(u0 ) and get that C(u) is irreducible of degree  2.
Suppose H is defined by F (x, y) 2 Chom [x, y] and

F (x, y) = f (x)g(y)h1 (x, y)r1 h2 (x, y)r2 . . . hn (x, y)rn

is the factorization of F such that h1 , h2 , . . . , hn are distinct irreducible polynomials in

Chom [x, y] \ (Chom [x] [ Chom [y]). The set {f = 0} \ X is either empty or infinite. So,
for H \ (X ⇥ P2 ) to be (2, 2)-grid-free we must have {f = 0} \ X = ;. Similarly, we know
that {g = 0} = ;, that is, g(y) is a nonzero constant.
6 The algebraic curve Ci (u) is reduced in the sense that its defining equation hi (u, y) is square-
free.

32
 Without loss of generality, we may assume that f (x) = g(y) = 1 and that F (x, y) is
square-free, that is, r1 = r2 = · · · = rn = 1. Let Ci (u) be the section of Hi := {hi = 0}
at u for i = 1, 2, . . . , n. From Lemma 4.7, we know that, for a generic u 2 X, Ci (u) and
Cj (u) have no common irreducible components for all i 6= j. Therefore C(u) = [ni=1 Ci (u)
has at least n irreducible components, and so n = 1. Now C(u) = C1 (u) = {h1 (u, y) = 0}
for all u 2 X. By Lemma 4.7, h1 (u, y) is square-free for generic u. This and the fact that
C(u) is irreducible of degree  2 imply that deg h1 (u, y)  2 for a generic u 2 X, and so
degy h1 (x, y)  2.

Proof of Lemma 4.7. Let d1 , d2 be the degrees of h1 , h2 in y respectively. Suppose on the

contrary that C1 (u) and C2 (u) share common irreducible components for a generic u 2 P2 .
So, h1 (u, y) and h2 (u, y) have a common divisor in C[y]. Therefore there exist two nonzero
polynomials g1u (y) 2 Chom [y] of degree < d2 and g2u (y) 2 Chom [y] of degree < d1 such that

h1 (u, y)g1u (y) + h2 (u, y)g2u (y) = 0. (4.4)

By treating the coefficients of g1u (y) and g2u (y) as variables, we can view equation (4.4)
d1 +d2 +1 d1 +1 d2 +1
as a homogeneous system of M := 2
linear equations involving N := 2
+ 2
variables. Note that the coefficient in the ith equation of the jth variable, say cij , is a
polynomial of u, that is, cij = cij (u) for some cij (x) 2 C[x] that depends on h1 , h2 only.
Because the system of linear equations has a nontrivial solution and clearly M > N , the
rank of its coefficient matrix (cij (u)) is < N . Using the determinants of all N ⇥ N minors of
M
matrix (cij (u)), we can rewrite the statement that matrix (cij (u)) is of rank < N as L := N
polynomial equations of entries in the matrix, say

Pk (cij (u)) = 0, for all k 2 [L], (4.5)

where Pk (cij (x)) is a polynomial of x independent of u. Since (4.5) holds for a generic u 2 P2 ,
we have
Pk (cij (x)) = 0 in C[x], for all k 2 [L], (4.6)
Reversing the argument above, we can deduce that the rank of matrix (cij (x)), over the
quotient field C(x), is < N , and so there exist two nonzero polynomials g1x (y) 2 C(x)hom [y]
of degree < d2 and g2x (y) 2 C(x)hom [y] of degree < d1 such that

h1 (x, y)g1x (y) + hx2 (x, y)g2 (y) = 0. (4.7)

Multiplying (4.7) by the common denominator of g1x (y) and g2x (y), we get two nonzero
polynomials g1 (x, y) 2 C[x, y] of degree < d2 in y and g2 (x, y) 2 C[x, y] of degree < d1
in y such that
h1 (x, y)g1 (x, y) + h2 (x, y)g2 (x, y) = 0, (4.8)

33
which is impossible as gcd (h1 , h2 ) = 1 and degy h1 (x, y) = d1 > degy g2 (x, y).
/ Chom [x], the
It remains to prove that C1 (u) is reduced for generic u. Because h1 (x, y) 2
polynomial h01 (x, y) := @h1 (x, y)/@y0 might be assumed to be nonzero. Again, we assume, on
the contrary, that h1 (u, y) is not square-free for a generic u 2 P2 . This implies that h1 (u, y)
and h01 (u, y) have a common divisor. The same linear-algebraic argument, applied to h1 and
h01 instead of h1 and h2 , then yields a contradiction.

We can adapt the proof of Theorem 4.4 to the case when P2 \ Y is finite. In this case,
we obtain a weaker result though.

Proposition 4.8. Suppose P2 \ Y = {v 1 , v 2 , . . . , v n }. If H \ (X ⇥ Y ) is (2, 2)-grid-free, then

either
1. there exists F (x, y) 2 Chom [x, y] of degree  2 in y such that H \ (X ⇥ Y ) = {F =
0} \ (X ⇥ Y ),
2. or there exists i 2 [n] such that P2 ⇥ {v i } ⇢ H.

Sketch of a proof. We follow the proof of Theorem 4.4 up to the point where we apply
Corollary 4.6. Note that Xi := {u 2 P2 : v i 2 C(u)} is Zariski-closed for all i 2 [n]. If none
of those Xi ’s equals P2 , then for a generic u 2 P2 , C(u) does not pass through any of the
points in P2 \ Y . The rest of the proof of Theorem 4.4 still holds and we end in the first
case. Otherwise Xi = P2 for some i 2 [n], which corresponds to the second case.

4.5 Fields of finite characteristic

A standard model-theoretic argument allows us to transfer statements over fields of charac-
teristic 0 to the fields of large characteristic.

Theorem 4.9. Let be a sentence in the language of rings. The following are equivalent.
1. is true in complex numbers.
2. is true in every algebraically closed field of characteristic zero.
3. is true in all algebraically closed fields of characteristic p for all sufficiently large
prime p.

The theorem is an application of the compactness theorem and the completeness of the
theory of algebraically closed field of fixed characteristic. We refer the readers to [Mar02,
Section 2.1] for further details of the theorem and related notions.
As quantifiers over all polynomials are not part of the language of rings, one has to limit
the degree of hypersurface H and the complexity of the open set X in Theorem 4.4. We now
formulate the analog over the fields of large characteristic.

34
Theorem 4.10. Let K be an algebraically closed field of large characteristic, let H be a
hypersurface in P2 (K) ⇥ P2 (K) of bounded degree, and let X be a Zariski-open subset of
P2 (K) of bounded complexity (i.e. X is a Zariski-open subset of P2 (K) that can be described
by some first order predicate in the language of rings of bounded length). If H is (2, 2)-
grid-free in X ⇥ P2 (K), then there exists F (x, y) 2 Khom [x, y] of degree  2 in y such that
H \ (X ⇥ P2 ) = {F = 0} \ (X ⇥ P2 ).

The proof essentially rewrites Theorem 4.4 as a sentence in the language of rings to which
Theorem 4.9 is applicable. We skip the tedious but routine proof.

35
Bibliography

[AC02] Maria Alberich-Carramiñana. Geometry of the plane Cremona maps, volume

1769 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 2002.

[ARS99] Noga Alon, Lajos Rónyai, and Tibor Szabó. Norm-graphs: variations and appli-
cations. J. Combin. Theory Ser. B, 76(2):280–290, 1999.

[Bár82] Imre Bárány. A generalization of Carathéodory’s theorem. Discrete Math., 40(2-

3):141–152, 1982.

[BBK13] Pavle V. M. Blagojević, Boris Bukh, and Roman Karasev. Turán numbers for
Ks,t -free graphs: topological obstructions and algebraic constructions. Israel J.
Math., 197(1):199–214, 2013. arXiv:1108.5254[math.CO].

[Ben66] Clark T. Benson. Minimal regular graphs of girths eight and twelve. Canad. J.
Math., 18:1091–1094, 1966.

[BF84] E. Boros and Z. Füredi. The number of triangles covering the center of an n-set.
Geom. Dedicata, 17(1):69–77, 1984.

[BMN10] Boris Bukh, Jiřı́ Matoušek, and Gabriel Nivasch. Stabbing simplices
by points and flats. Discrete Comput. Geom., 43(2):321–338, 2010.
arXiv:0804.4464[math.CO].

[BMRR10] Abdul Basit, Nabil H. Mustafa, Saurabh Ray, and Sarfraz Raza. Improving
the first selection lemma in R3 . In Computational geometry (SCG’10), pages
354–357. ACM, New York, 2010. http://paul.rutgers.edu/~abasit/papers/
firstsel.pdf.

[Bro66] W. G. Brown. On graphs that do not contain a Thomsen graph. Canad. Math.
Bull., 9:281–285, 1966.

[BS74] J. A. Bondy and M. Simonovits. Cycles of even length in graphs. J. Combina-

torial Theory Ser. B, 16:97–105, 1974.

36
[Buk06] Boris Bukh. A point in many triangles. Electron. J. Combin., 13(1):Note 10, 3
pp. (electronic), 2006.

[DLW07] Vasyl Dmytrenko, Felix Lazebnik, and Jason Williford. On monomial graphs of
girth eight. Finite Fields Appl., 13(4):828–842, 2007.

[ER62] P. Erdős and A. Rényi. On a problem in the theory of graphs. Magyar Tud.
Akad. Mat. Kutató Int. Közl., 7:623–641 (1963), 1962.

[Erd38] P. Erdős. On sequences of integers no one of which divides the product of two
others and on some related problems. Inst. Math. Mech. Univ. Tomsk, 2:74–82,
1938.

[Erd64] P. Erdős. Extremal problems in graph theory. In Theory of Graphs and its Appli-
cations (Proc. Sympos. Smolenice, 1963), pages 29–36. Publ. House Czechoslovak
Acad. Sci., Prague, 1964.

[ERS66] P. Erdős, A. Rényi, and V. T. Sós. On a problem of graph theory. Studia Sci.
Math. Hungar., 1:215–235, 1966.

[ES46] P. Erdös and A. H. Stone. On the structure of linear graphs. Bull. Amer. Math.
Soc., 52:1087–1091, 1946.

[FNV06] Zoltán Füredi, Assaf Naor, and Jacques Verstraëte. On the Turán number for
the hexagon. Adv. Math., 203(2):476–496, 2006.

[FS13] Zoltán Füredi and Miklós Simonovits. The history of degenerate (bipar-
tite) extremal graph problems. In Erdös centennial, volume 25 of Bolyai
Soc. Math. Stud., pages 169–264. János Bolyai Math. Soc., Budapest, 2013.
arXiv:1306.5167[math.CO].

[Für88] Zoltán Füredi. Quadrilateral-free graphs with maximum number of edges. 1988.
http://www.math.uiuc.edu/~z-furedi/PUBS/furedi_C4from1988.pdf.

[Für96a] Zoltán Füredi. New asymptotics for bipartite Turán numbers. J. Combin. Theory
Ser. A, 75(1):141–144, 1996.

[Für96b] Zoltán Füredi. An upper bound on Zarankiewicz’ problem. Combin. Probab.

Comput., 5(1):29–33, 1996.

37
[Gro10] Mikhail Gromov. Singularities, expanders and topology of maps. Part 2: From
combinatorics to topology via algebraic isoperimetry. Geom. Funct. Anal.,
20(2):416–526, 2010. http://www.ihes.fr/~gromov/PDF/morse2_gafa.pdf.

[HLL15] Xiang-dong Hou, Stephen D. Lappano, and Felix Lazebnik. Proof of a conjecture
on monomial graphs. July 2015. arXiv:1507.05306[math.CO].

[Jun42] Heinrich W. E. Jung. Über ganze birationale Transformationen der Ebene. J.

Reine Angew. Math., 184:161–174, 1942.

[Kar12] Roman Karasev. A simpler proof of the Boros–Füredi–Bárány–Pach–

Gromov theorem. Discrete Comput. Geom., 47(3):492–495, 2012.
arXiv:1012.5890[math.CO].

[Kee11] Peter Keevash. Hypergraph Turán problems. In Surveys in combinatorics 2011,

volume 392 of London Math. Soc. Lecture Note Ser., pages 83–139. Cambridge
Univ. Press, Cambridge, 2011. http://people.maths.ox.ac.uk/keevash/
papers/turan-survey.pdf.

[KMS12] Daniel Král’, Lukáš Mach, and Jean-Sébastien Sereni. A new lower bound
based on Gromov’s method of selecting heavily covered points. Discrete Comput.
Geom., 48(2):487–498, 2012. arXiv:1108.0297[math.CO].

[KRS96] János Kollár, Lajos Rónyai, and Tibor Szabó. Norm-graphs and bipartite Turán
numbers. Combinatorica, 16(3):399–406, 1996.

[KST54] T. Kövari, V. T. Sós, and P. Turán. On a problem of K. Zarankiewicz. Colloquium

Math., 3:50–57, 1954.

[Kur10] Shigeru Kuroda. Shestakov-Umirbaev reductions and Nagata’s conjecture

on a polynomial automorphism. Tohoku Math. J. (2), 62(1):75–115, 2010.
arXiv:0801.0117[math.AC].

[LU95] Felix Lazebnik and Vasiliy A. Ustimenko. Explicit construction of graphs with
an arbitrary large girth and of large size. Discrete Appl. Math., 60(1-3):275–284,
1995. ARIDAM VI and VII (New Brunswick, NJ, 1991/1992).

[Man07] Willem Mantel. Problem 28. Wiskundige Opgaven, 10(60-61):320, 1907.

[Mar02] David Marker. Model theory: an introduction, volume 217 of Graduate Texts in
Mathematics. Springer-Verlag, New York, 2002.

38
[Mat02] Jiřı́ Matoušek. Lectures on discrete geometry, volume 212 of Graduate Texts in
Mathematics. Springer-Verlag, New York, 2002.

[Mat03] Jiřı́ Matoušek. Using the Borsuk–Ulam theorem: lectures on topological methods
in combinatorics and geometry. Springer, 2003.

[MM05] Keith E. Mellinger and Dhruv Mubayi. Constructions of bipartite graphs from
finite geometries. J. Graph Theory, 49(1):1–10, 2005. http://homepages.math.
uic.edu/~mubayi/papers/ArcConstf.pdf.

[Mou04] Claire Moura. Local intersections of plane algebraic curves. Proc. Amer. Math.
Soc., 132(3):687–690, 2004.

[MW14] Jiřı́ Matoušek and Uli Wagner. On Gromov’s method of selecting

heavily covered points. Discrete Comput. Geom., 52(1):1–33, 2014.
arXiv:1102.3515[math.CO].

[Nag72] Masayoshi Nagata. On automorphism group of k[x, y], volume 5 of Lectures in

Mathematics, Department of Mathematics, Kyoto University. Kinokuniya Book-
Store Co. Ltd., Tokyo, 1972.

[Pik12] Oleg Pikhurko. A note on the Turán function of even cycles. Proc. Amer. Math.
Soc., 140(11):3687–3692, 2012. http://homepages.warwick.ac.uk/~maskat/
Papers/EvenCycle.pdf.

[Sha13] Igor R. Shafarevich. Basic algebraic geometry 1: varieties in projective space,

volume 1. Springer, Heidelberg, russian edition, 2013.

[Sid95] Alexander Sidorenko. What we know and what we do not know about Turán
numbers. Graphs Combin., 11(2):179–199, 1995.

[Sim68] M. Simonovits. A method for solving extremal problems in graph theory, stability
problems. In Theory of Graphs (Proc. Colloq., Tihany, 1966), pages 279–319.
Academic Press, New York, 1968.

[SU03] Ivan P. Shestakov and Ualbai U. Umirbaev. The Nagata automorphism is wild.
Proc. Natl. Acad. Sci. USA, 100(22):12561–12563 (electronic), 2003.

[SU04] Ivan P. Shestakov and Ualbai U. Umirbaev. The tame and the wild automor-
phisms of polynomial rings in three variables. J. Amer. Math. Soc., 17(1):197–227
(electronic), 2004.

39
[Tur41] Paul Turán. On an extremal problem in graph theory. Mat. Fiz. Lapok, 48(436-
452):137, 1941.

[vdK53] W. van der Kulk. On polynomial rings in two variables. Nieuw Arch. Wiskunde
(3), 1:33–41, 1953.

[Ver00] Jacques Verstraëte. On arithmetic progressions of cycle lengths in graphs. Com-

bin. Probab. Comput., 9(4):369–373, 2000.

[Wag03] Ulrich Wagner. On k-sets and applications. PhD thesis, Swiss Federal Institute
of Technology, ETH Zürich, 2003.

[Wen91] R. Wenger. Extremal graphs with no C 4 ’s, C 6 ’s, or C 10 ’s. J. Combin. Theory
Ser. B, 52(1):113–116, 1991.

40