Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

1 Introduction
A simple and well-known fact in graph theory is that every n-vertex tree has n − 1 edges.
This immediately implies that if an n-vertex graph G has no cycles, then G has at most n−1
edges. Another well-known result in graph theory, following quickly from Euler’s formula, is
that an n-vertex planar graph G with n ≥ 3 has at most 3n − 6 edges.
This class is about extremal graph theory, the study of results of this type. How many
edges can an n-vertex graph have, given that it satisfies some natural constraint? Our
ultimate goal is to prove the Erdős–Stone–Simonovits theorem, sometimes called the Fun-
damental Theorem of Extremal Graph Theory, which answers this question more or less
completely for a very wide range of constraints.
The question that will occupy us for the majority of the class is what happens when the
constraint is excluding a single “forbidden subgraph”.
Definition 1.1. Let H and G be graphs. We say that G is H-free if H is not a subgraph
of G (or, more formally, if G has no subgraph isomorphic to H). We will often also say that
G has no copy of H.
The basic question we will be attempting to answer is “how many edges can an n-vertex
H-free graph have?”. Because we will be using this notion over and over again, it’s best to
just give it a name. We use e(G) to denote the number of edges of a graph G.
Definition 1.2. The extremal number of H is defined as
ex(n, H) = max{e(G) | G is an n-vertex H-free graph}.
In other words, ex(n, H) is simply the most number of edges that an H-free graph on n
vertices can have. Note that this quantity is well-defined, since there are only finitely many
n-vertex graphs.
In this class, we will attempt to understand how the function ex(n, H) behaves when H is
some fixed graph, and when n tends to infinity. Additionally, we will often try to understand
which graphs G are the maximizers in the definition of ex(n, H); that is, which graphs G
have the most edges among all n-vertex H-free graphs.
Before getting into specific examples, let’s briefly think about what it means to prove
upper and lower bounds on ex(n, H). Since ex(n, H) is defined as the maximum of something,
to prove a lower bound on ex(n, H), it suffices to exhibit an n-vertex graph G with no copy
of H; such a G gives us the lower bound ex(n, H) ≥ e(G). On the other hand, to prove an
upper bound on ex(n, H), we need to prove that every n-vertex graph G with m edges has
a copy of H; this yields the upper bound ex(n, H) < m.

2 Forbidden cliques: Mantel’s and Turán’s theorems

The earliest result in extremal graph theory is due to Mantel, from more than 100 years ago.
Mantel studied (though not in this language) the extremal number of the triangle graph,
K3 . Let’s begin by coming up with a lower bound on ex(n, K3 ).

1
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

After playing around with it a bit, it’s pretty natural to come up with the following
construction. Let G = Ka,b be a complete bipartite graph, where a + b = n. Then G is
certainly triangle-free, since K3 is not bipartite. Moreover, the number of edges in G is
simply ab. So we find that
ex(n, K3 ) ≥ ab for all integers a, b with a + b = n.
Since we want as good a lower bound as possible, we want to pick a, b so that ab is maximized,
subject to the constraint that a + b = n. Using the AM-GM inequality, we see that
2
n2

a+b
ab ≤ = .
2 4
Moreover, equality holds if and only if a = b = n/2. If n is odd, then we can’t have a = b =
n/2 if a and b are both integers; the product ab is maximized when a = bn/2c , b = dn/2e.
But in any case, we find that  2
n
ex(n, K3 ) ≥ ,
4
with the example of a K3 -free n-vertex graph with bn2 /4c edges given by the complete
bipartite graph Kbn/2c,dn/2e .
Mantel’s theorem says that this is the best we can do.
Theorem 2.1 (Mantel 1907). ex(n, K3 ) = bn2 /4c. Moreover, the unique n-vertex triangle-
free graph with bn2 /4c edges is Kbn/2c,dn/2e .
We won’t prove this right now. Instead, we’ll first generalize Mantel’s theorem, and then
prove the generalization.
Almost 40 years after Mantel, Turán started thinking about similar questions, and it is
thanks to his work that the field of extremal graph theory exists at all. Turán was studying
what happens when, rather than excluding a triangle, we exclude some larger complete graph
(also known as a clique). Namely, he was studying ex(n, Kr ) for r ≥ 3.
Again, there is a natural type of example we can come up with to lower-bound ex(n, Kr ).
Namely, let G be a complete (r − 1)-partite graph on n vertices, namely a graph obtained
by splitting the n vertices into r − 1 parts, then putting all edges between pairs of vertices
in different parts and no edges within a part. Then G certainly will not have a copy of Kr :
by the pigeonhole principle, if we take any r vertices in G, two of them must lie in the same
part, and thus there cannot be an edge between them. Moreover, another simple application
of the AM-GM inequality (or Jensen’s inequality) shows that the way to do this in order to
maximize the number of edges of G is to make all the parts have as equal sizes as possible,
namely to make each part have size either bn/(r − 1)c or dn/(r − 1)e. This motivates the
following definition.
Definition 2.2. The Turán graph Tr−1 (n) is the n-vertex complete (r − 1)-partite graph
with all parts of size either bn/(r − 1)c or dn/(r − 1)e. We denote its number of edges by
tr−1 (n) := e(Tr−1 (n)).

2
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Remark. In case n is divisible by r − 1, then every part of the Turán graph Tr−1 (n) has
exactly n/(r − 1) vertices in each part, so
2 
r − 2 n2
      2
r−1 n 1 n
tr−1 (n) = · = = 1− .
2 r−1 r−1 2 r−1 2

In case n is not divisible by r − 1, the formula is a little messier, involving the remainder of
n when divided by r − 1. However, we still have that for all fixed r and n → ∞,
  2
1 n
tr−1 = 1 − + o(1) ,
r−1 2

where o(1) represents a quantity that tends to 0 as n tends to infinity. In other words, if we
fix ε > 0, then for any sufficiently large n, we have that
  2   2
1 n 1 n
1− −ε ≤ tr−1 (n) ≤ 1 − +ε .
r−1 2 r−1 2

One other useful observation is that for any n, if we delete one vertex from each of the r − 1
parts of Tr−1 (n), we obtain a copy of Tr−1 (n − r + 1). Moreover, each non-deleted
vertex is
adjacent to exactly r − 2 deleted vertices. So we delete (r − 2)(n − r + 1) + r−12
edges to
obtain Tr−1 (n − r + 1) from Tr−1 (n). This shows that
 
r−1
tr−1 (n) = tr−1 (n − r + 1) + (r − 2)(n − r + 1) + . (1)
2

Note that T2 (n) = Kbn/2c,dn/2e , so Mantel’s theorem can be rephrased as saying that
ex(n, K3 ) = t2 (n) and that T2 (n) is the unique n-vertex K3 -free graph with t2 (n) edges.
Turán’s theorem generalizes this to ex(n, Kr ) for all r ≥ 3.

Theorem 2.3 (Turán 1941). For every r ≥ 3, we have ex(n, Kr ) = tr−1 (n). Moreover, the
unique n-vertex Kr -free graph with tr−1 (n) edges is Tr−1 (n).

Proof. We proceed by induction, with steps of size r − 1. So we need r − 1 base cases,

corresponding to n = 1, 2, . . . , r − 1. But the theorem holds
for such n, because for such n,
any n-vertex graph has no Kr subgraph. So ex(n, Kr ) = n2 for 1 ≤ n ≤ r − 1. Moreover,
Tr−1 (n) is exactly Kn in these cases. This proves the base cases of the induction.
Now let n > r − 1, and assume the theorem is true for n − r + 1. Let G be an n-vertex
graph with no copy of Kr and as many edges as possible. G must contain a copy of Kr−1 ,
for otherwise we could add an edge and get a Kr -free graph with strictly more edges. Let K
be some such Kr−1 subgraph, and let F ⊆ G be the subgraph obtained by deleting K. We
r−1


know that e(K) = 2 . By induction, we know that

e(F ) ≤ tr−1 (n − r + 1).

3
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Finally, each vertex of F cannot be adjacent to every vertex of K, for otherwise we would
get a Kr . So the number of edges between F and K is at most (r − 2)(n − r + 1). So
 
r−1
e(G) ≤ + tr−1 (n − r + 1) + (r − 2)(n − r + 1) = tr−1 (n),
2

by (1).
If e(G) = tr−1 (n), then every inequality above must be an equality. In particular, the
induction hypothesis implies that F ∼ = Tr−1 (n − r + 1). Moreover, each vertex in F must be
adjacent to exactly r − 2 vertices in K, since we assume we have equality in the number of
edges. Moreover, given two adjacnet vertices in F , they cannot be non-adjacent to the same
vertex of K, for otherwise we could take the remaining r − 2 vertices and these two to get
a Kr . So this implies that each part of F is associated to exactly one missed vertex. So by
adding this missed vertex to its part, we see that G ∼= Tr−1 (n).
On the homework, you will see many different proofs of Turán’s theorem. It is one of
those amazing mathematical theorems with dozens of different, and differently informative,
proofs. It is also extremely useful, as you’ll see on the homework!
Before moving on, let me just mention one convenient way to think about Turán’s theorem
is as follows. Note that
n2 n n2
 
n
= − = (1 + o(1)) .
2 2 2 2
This shows that
  2   
1 n 1 n
tr−1 (n) = 1 − + o(1) = 1− + o(1) .
r−1 2 r−1 2

Note that an n-vertex graph can have anywhere between 0 and n2 edges. So Turán’s theorem

implies that a Kr -free n-vertex graph can have at most, asymptotically, a 1−1/(r−1) fraction
of all possible edges.

3 Beyond Turán’s theorem

Turán’s theorem is great, and tells us exactly what ex(n, Kr ) is for any r. But we started this
class by asking about ex(n, H) for general H; what can we say about that? In general, we’d
probably expect this problem to be really hard, and the answer should depend in complicated
ways on the fixed graph H.
But it turns out that’s not the case! Kind of amazingly, the answer depends, essentially,
on a single parameter of the graph H—its chromatic number.

Theorem 3.1 (Erdős–Stone(–Simonovits) 1946 (1966)). For any graph H,

  
1 n
ex(n, H) = 1 − + o(1) .
χ(H) − 1 2

4
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Remark.

• This is sometimes called the Fundamental Theorem of Extremal Graph Theory, for
hopefully obvious reasons: it more or less completely resolves the main question that
we started with.

• The history (and naming) of this theorem is a bit confusing. Erdős and Stone proved
a special case of it in 1946. In 1966, Erdős and Simonovits realized that the special
case actually implies (with a one-line implication) the general case, which had not been
really studied before. We will soon see the special case, and how it implies the general
case.

• Notice that if H is bipartite (i.e. if χ(H) = 2), then 1 − 1/(χ(H) − 1) = 0. So the

theorem simply says that if H is bipartite, then
 
n
ex(n, H) = o(1) · ,
2

which we usually write as ex(n, H) = o(n2 ). In other words, if G is an n-vertex graph

containing no copy of some fixed bipartite graph H, then G must have very few edges—
its number of edges grows sub-quadratically in n. Said differently, the fraction of all
possible edges that we can put in such a graph is a vanishingly small fraction; the
fraction tends to 0 as n → ∞.
Already this statement is far from obvious, and we’ll soon prove it. In fact, as we’ll
see, proving the statement for bipartite H implies, in a certain sense, the full Erdős–
Stone–Simonovits theorem.

Most of the rest of the class will be spent on proving the Erdős–Stone–Simonovits theo-
rem. To do
so, we’ll prove upper and lower bounds on ex(n, H) of the form (1 − 1/(χ(H) −
1) + o(1)) n2 . In fact, we can easily dispense with the lower bound.

Proposition 3.2. For any fixed graph H and integer n,

  
1 n
ex(n, H) ≥ tχ(H)−1 (n) = 1 − + o(1) .
χ(H) − 1 2

Proof. We claim that the Turán graph Tχ(H)−1 (n) has no copy of H. Indeed, suppose we had
some vertices in Tχ(H)−1 (n) that defined a copy of H. Give the parts of Tχ(H)−1 (n) names,
say V1 , . . . , Vχ(H)−1 . Then note that any two vertices of H that lie in the same part Vi cannot
be adjacent in H, since Tχ(H)−1 (n) has no edges inside any part Vi . Said differently, if we
assign to any vertex v of H the number i so that v ∈ Vi , then two adjacent vertices are
assigned different numbers. In other words, this yields a proper coloring of H with χ(H) − 1
colors. But this contradicts the definition of the chromatic number.

5
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

4 Extremal numbers of bipartite graphs

4.1 Upper bounds
Let H be a bipartite graph. Recall that the Erdős–Stone–Simonovits theorem implies that
in this case, ex(n, H) = o(n2 ), or equivalently that
ex(n, H)
lim = 0.
n→∞ n2
This is pretty surprising! For example, the complete bipartite graph Kbn/2c,dn/2e has bn2 /4c
edges and no copy of any odd cycle C2`+1 . Thus, ex(n, C2`+1 ) ≥ bn2 /4c for all `. But the
four-cycle (or any other even cycle) is bipartite, so the Erdős–Stone–Simonovits theorem
implies that ex(n, C4 ) = o(n2 ). What’s up with that?
We will shortly prove that in fact, ex(n, C4 ) ≤ O(n3/2 ). In case you haven’t seen it before,
the big-O notation means that ex(n, C4 ) ≤ Cn3/2 for some absolute constant C, which we
won’t specify. In other words, we will shortly prove that if G is an n-vertex graph with at
least Cn3/2 edges, then G has a copy of C4 , assuming C is an appropriately chosen constant.
As a warm-up, we will begin with an easier special case of this result, which is the case when
G is d-regular (i.e. every vertex in G has degree d). Recall that in any graph, the sum of the
degrees of all the vertices equals twice the number of edges, so if G is√d-regular then it has
dn/2 edges. Thus, if G has Cn3/2 edges and is d-regular, then d = 2C n.
√
Proposition 4.1. Let G be a d-regular n-vertex graph. If d ≥ 2 n, then G contains a copy
of C4 .
Proof. Suppose for contradiction that G is C4 -free. We count the number of copies of K1,2
in G, where K1,2 = consists of one central vertex adjacent to two outer vertices.
On the one hand, if we sum over all possibilities for the central vertex, we see that
X X deg(v)  
d
#(K1,2 in G) = #(K1,2 with central vertex v) = =n .
2 2
v∈V (G) v∈V (G)

On the other hand, suppose we fix some u, w ∈ V (G). We claim that they can be the outer
vertices of at most one copy of K1,2 . Indeed, if not, then we would have two K1,2 s agreeing
on the outer vertices, which yields a copy of C4 , a contradiction. So we conclude that
 
X X n
#(K1,2 in G) = #(K1,2 with outer vertices u, w) ≤ 1= .
2
u,w∈V (G) u,w∈V (G)
distinct distinct

Rearranging, we see that

     
d n d n−1
n ≤ ⇐⇒ ≤ ⇐⇒ d(d − 1) ≤ n − 1.
2 2 2 2
√
But if d ≥ 2 n and n ≥ 0, then this is a contradiction.

6
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

To prove the real result, we will need one extraordinarily useful analytic tool, called
Jensen’s inequality. We will actually only need the following special case. For a real number
x and a positive integer r, we extend the definition of the binomial coefficient as
 
x x(x − 1)(x − 2) · · · (x − r + 1)
= .
r r!
Lemma 4.2 ((Consequence of) Jensen’s inequality). PLet r ≥ 1 be a positive integer, and let
1 n
x1 , . . . , xn be non-negative integers. Suppose that n i=1 xi ≥ r. Then
n    1 Pn 
X xi i=1 x i
≥n n .
i=1
r r

The point of this is that if we add up terms of the form xri , we can only decrease
the

sum if we replace each xi by the average of all the xi . One says that the function x 7→ xr is
convex : the sum of its values is minimized when all the variables are equal (to their average).
We won’t prove Jensen’s inequality in class, but its proof is on the homework if you’re
interested. Once we have Jensen’s inequality, we can easily prove the full result that
ex(n, H) ≤ O(n3/2 ). In fact, we will prove the following much more general result.
Theorem 4.3 (Kővári–Sós–Turán 1954). For positive integers s ≤ t, we have

ex(n, Ks,t ) ≤ O(n2−1/s ).

Here, the implicit constant may depend on s and t (which we think of as fixed).
Proof. We proceed much as in the proof of Proposition 4.1. Let G be an n-vertex graph
with at least Cn2−1/s edges, where C is some large constant we will pick later. Let d be the
average degree in G, so that d = n2 e(G) ≥ 2Cn1−1/s . Suppose for contradiction that G is
Ks,t -free. We count the number of copies of K1,s in G in two ways. First, by summing over
the options for the central vertex, we have that
X deg(v)  
d
#(K1,s in G) = ≥n ,
s s
v∈V (G)

using Lemma 4.2, as well as the fact that d ≥ s by picking C sufficiently large. On the other
hand, by counting over the s outer vertices of K1,s , we have that every u1 , . . . , us ∈ V (G)
can be the outer vertices of at most t − 1 copies of K1,s . So
 
X n
#(K1,s in G) ≤ (t − 1) = (t − 1) .
s
u1 ,...,us ∈V (G)
distinct

Combining these, we see that

   
n d
(t−1) ≥n ⇐⇒ (t−1)(n−1)(n−2) · · · (n−s+1) ≥ d(d−1) · · · (d−s+1).
s s

7
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Now, if n is very large (which is the regime we care about anyway), all this subtracting stuff
doesn’t matter. So this is roughly equivalent to

(t − 1)ns−1 ≥ ds ⇐⇒ d ≤ (t − 1)1/s n1−1/s .

If C is sufficiently large, then this is a contradiction. Moreover, if C is sufficiently large, then

the slightly sketchy step above where we dropped the subtractions is also OK, and we get
the desired contradiction.
Note that C4 = K2,2 , so in the case s = t = 2, we indeed get the claimed bound of
ex(n, C4 ) ≤ O(n3/2 ).
There are a number of important consequences of the Kővári–Sós–Turán theorem. The
first is that it immediately gives us a bound on ex(n, H) for all bipartite H. Indeed, note
that if H1 is a subgraph of H2 , then

ex(n, H1 ) ≤ ex(n, H2 )

for all n, as any H1 -free graph is also H2 -free. Now, if H is a bipartite graph, then H is a
subgraph of Ks,t for some s ≤ t. So

ex(n, H) ≤ ex(n, Ks,t ) ≤ O(n2−1/s ).

In particular, this proves that ex(n, H) = o(n2 ) = o(1) · n2 for bipartite H. Indeed,

ex(n, H) O(n2−1/s )
lim ≤ lim = lim O(n−1/s ) = 0.
n→∞ n2 n→∞ n2 n→∞

Recall that this was a consequence of the Erdős–Stone–Simonovits theorem.

4.2 Lower bounds

How good are the upper bounds we proved? Let’s begin with the one we started with,
ex(n, C4 ) ≤ O(n3/2 ). Can we construct an n-vertex C4 -free graph with roughly that many
edges?
As it turns out, we can! The following construction was originally due to Eszter Klein in
1938 (as reported in a paper of Erdős). Note that this is before Turán’s theorem, so before
the birth of extremal graph theory! As such, no one really appreciated what this construction
was or meant, and it was later rediscovered by Erdős, Rényi, and Sós (and independently
Brown). These days, it is often called the “Erdős–Rényi” construction, which I find a little
odd, both because they weren’t the first to discover it, and because there are many other
things named after Erdős and Rényi.

Theorem 4.4 (Klein 1938). For every n ≥ 1, there is an n-vertex C4 -free graph with at
least n3/2 /64 edges.

8
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Proof. First, suppose that n = 2p2 for some prime p; we will later get rid of this assumption.
Consider the integers mod p, which form a field that we denote Fp . (If you don’t know what
the word “field” means, just believe me that among the integers mod p, we can use addition,
multiplication, and division and have them work basically the same way they do in R.)
Let F2p denote the two-dimensional plane over Fp , i.e. the set of points (x, y) with x, y ∈ Fp .
For m, b ∈ Fp , let `m,b denote the line y = mx + b in F2p . In other words, `m,b is the set of
points (x, y) ∈ F2p satisfying y = mx + b.
We define a bipartite graph G with parts P, L, where P = F2p and L = {`m,b : m, b ∈ Fp }.
The edges of G are given by incidence: we connect (x, y) ∈ P to `m,b ∈ L if and only if (x, y)
lies on the line `m,b , i.e. if and only if y = mx + b.
Note that |P | = |L| = p2 , so G has n = 2p2 vertices. Moreover, every line `m,b ∈ L has
exactly p points on it, so every vertex in L has degree p in G. Therefore, e(G) = p|L| =
p3 = (n/2)3/2 .
Finally, we claim that G is C4 -free. To see this, note that G is bipartite, so the only way
we could have a copy of C4 in G is to have distinct p1 , p2 ∈ P and distinct `1 `2 ∈ L so that
p1 `1 p2 `2 forms a 4-cycle. But this means that p1 lies on the lines `1 , `2 , and that p2 also lies
on both these lines. So we have two lines which intersect at two distinct points!
Using our intuition from R, we expect this to be impossible, and it’s impossible over Fp
as well. Formally, let p1 = (x1 , y1 ) and p2 = (x2 , y2 ), and `1 = `m1 ,b1 , `2 = `m2 ,b2 . Then we
have the equations
y1 = m1 x1 + b1 y2 = m1 x2 + b1
y1 = m2 x1 + b2 y2 = m2 x2 + b2
Rearranging the first column, we see that m1 x1 + b1 = m2 x1 + b2 , or equivalently that
(m2 − m1 )x1 = b1 − b2 . If m1 = m2 then this implies that b1 = b2 , contradicting that `1 , `2
are distinct. So we have that m1 6= m2 , so x1 = (b1 − b2 )/(m2 − m1 ). But from the second
column of equations, we conclude that x2 = (b1 − b2 )/(m2 − m1 ) as well, so x1 = x2 . But if
we plug this into any of the equations, we conclude that y1 = y2 , and thus that p1 = p2 , a
contradiction. So G is C4 -free.
The only remaining thing is to deal with the fact that n need not equal twice the square
of a prime. So let n be arbitrary. There is an important result in number theory, called
Bertrand’s postulate, which says√that there is always a prime between m and 2m for all
positive integers m. Let m = b n/4c, and let p be a prime between m and 2m, so that
n/8 ≤ 2p2 ≤ n. Using the construction above, we obtain a C4 -free graph G on 2p2 vertices
with p3 edges. We add to this graph n − 2p2 isolated vertices, and we obtain a new C4 -free
n-vertex graph with p3 ≥ (n/16)3/2 = n3/2 /64 edges.
Using these finite fields and finite geometries might seem like a neat trick, but it turns
out that it’s essentially the only thing one can do. Indeed, all constructions we know of
for C4 -free graphs with many edges use such techniques. Moreover, there is a powerful
result of Füredi, which says that for those n for which such a construction (appropriately
defined) exists, the unique C4 -free n-vertex graph with the most edges comes from such a
construction.

9
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

So we conclude that ex(n, C4 ) = Θ(n3/2 ), where the big-Θ means that we have upper and
lower bounds that agree up to a constant factor. Since ex(n, K2,t ) ≥ ex(n, C4 ) for all t ≥ 2,
we conclude that ex(n, K2,t ) = Θ(n3/2 ) for all t ≥ 2.
What about ex(n, K3,3 )? We proved in Theorem 4.3 that ex(n, K3,3 ) ≤ O(n5/3 ). As it
turns out, this is also tight.
Theorem 4.5 (Brown 1966). For every n, there exists an n-vertex K3,3 -free graph G with
n5/3 /100 edges.
Proof sketch. I won’t present the proof in detail, but will explain the big idea. Suppose that
n = p3 . Construct a graph G with vertex set F3p , where we connect two vertices (x, y, z) and
(x0 , y 0 , z 0 ) by an edge if and only if
(x − x0 )2 + (y − y 0 )2 + (z − z 0 )2 = 1.
In other words, the neighborhood of any vertex looks like a “unit sphere” centered at that
vertex, except that “spheres” don’t really exist over Fp .
Nonetheless, if we were working in R3 , then we’d expect that any three unit spheres can
intersect in at most two points: two unit spheres can intersect in a circle, and that circle can
intersect a thid unit sphere in only two points. So we’d expect that G is K3,3 -free, since any
three vertices have at most two common neighbors.
Since we expect a sphere to be “two-dimensional”, we should expect every unit sphere
to have roughly p2 points on it, and this turns out to be true. So G has n = p3 vertices, and
every vertex has degree around p2 , so we expect e(G) ≈ p5 = n5/3 .
All of this intuition can be made precise, some of it with some annoyance. For example, it
turns out that this only really works if p ≡ 3 (mod 4). But the high-level idea is correct.
So we see that the Kővári–Sós–Turán theorem is best possible (up to the constant factor)
for s = 2 and s = 3. The case of s = 1 is much easier, but it’s also best possible there,
as you saw on the homework. So it is natural to conjecture, as many have done, that the
Kővári–Sós–Turán theorem is tight in general.
Conjecture 4.6 (Many people). For all s ≤ t,
ex(n, Ks,t ) = Θ(n2−1/s ).
Moreover, based on what I’ve told you so far, it is natural to expect that not only is this
conjecture proved, but that the constructions look kind of the same as above. You work with
the s-dimensional space Fsp over the field Fp , and use some kind of cleverly chosen polynomial
or set of polynomial equations to define the adjacency condition. However, despite many
people having this same idea, Conjecture 4.6 remains unproved. Moreover, many experts in
the field now even question whether it is true.
Nonetheless, some other things are known about ex(n, Ks,t ). Namely, it is known that
ex(n, Ks,t ) = Θ(n2−1/s ) if t is sufficiently large compared to s. The first result of this type is
due to Kollár, Rónyai, and Szabó in 1996, who proved that
ex(n, Ks,t ) = Θ(n2−1/s ) if t > s!.

10
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

To do this, they constructed a Ks,t -free graph, again using the space Fps , called the norm
graph. Their construction was later modified by Alon, Rónyai, and Szabó in 1999, who
defined a similar graph called the projective norm graph (again over Fsp ), which implies that

ex(n, Ks,t ) = Θ(n2−1/s ) if t > (s − 1)!.

So, for example, we know that ex(n, K4,7 ) = Θ(n7/4 ), but have no such lower bound for
ex(n, K4,4 ).
For about 20 years, the Alon–Rónayi–Szabó result was the best known. But very recently,
Bukh proved the following theorem.
2/3
Theorem 4.7 (Bukh 2021). Suppose s ≥ 2 and t ≥ 9s · s4s are integers. Then

ex(n, Ks,t ) = Θ(n2−1/s ).

The key point is that for large s, the previous bound on t, namely (s − 1)!, grew super-
exponentially in s. But Bukh’s bound, for large s, grows merely exponentially in s. The key
to Bukh’s construction is again to work over Fsp , but not to pick a clever polynomial. Instead,
he picks a random polynomial, and then uses arguments from probability, combinatorics, and
algebraic geometry to prove that the resulting graph is Ks,t -free with positive probability.

11
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

4.3 Non-complete bipartite graphs

So far, we have focused on complete bipartite graphs Ks,t . But it is natural to ask about
ex(n, H) for general bipartite graphs H. What can we say about this question?
The answer is, not very much. We already saw a general-purpose upper bound, namely
ex(n, H) ≤ ex(n, Ks,t ) ≤ O(n2−1/s ) if H is a subgraph of Ks,t . But for many specific bipartite
graphs, much better upper bounds are known, using arguments specific to the graph at hand.
For example, the following is known about the extremal numbers of even cycles.
Theorem 4.8 (Erdős (unpublished), Bondy–Simonovits 1974). For every ` ≥ 2, we have

ex(n, C2` ) ≤ O n1+1/` .

Note that in case ` = 2, this matches the bound ex(n, C4 ) ≤ O(n3/2 ) that we saw earlier.
It is again widely conjectured that this bound is tight, but it is only known to be tight in
case ` ∈ {2, 3, 5}. This is pretty remarkable: we know that

ex(n, C4 ) = Θ(n3/2 ) ex(n, C6 ) = Θ(n4/3 ) ex(n, C10 ) = Θ(n6/5 )

but we have no idea what the value of ex(n, C8 ) is!

There is a general-purpose lower bound that is known. In general, algebraic techniques
like the ones described above are the best techniques we have for constructing lower bounds,
but they often rely on specific structures that we can exploit. The following bound holds for
any bipartite graph. We use the notation big-Ω to denote the opposite of big-O—it means
that the left-hand side is at least as large as the right-hand side, up to a constant factor.
Given a graph H, we define its 2-density to be
e(F ) − 1
m2 (H) := max .
F ⊆H v(F ) − 2
Theorem 4.9. For any bipartite H, we have

ex(n, H) ≥ Ω(n2−1/m2 (H) ).

The proof of Theorem 4.9 uses the probabilistic method, and I won’t cover it in class. But
at a high level, the idea is to pick a random graph G with n vertices and roughly n2−1/m2 (H)
edges. One can then show that with positive probability, the number of copies of H in G is
less than half the number of edges of G. By deleting a single edge from each copy of H, we
obtain a graph with half as many edges—so still Ω(n2−1/m2 (H) )—and no copy of H.
To end this section, let me just mention two remarkable conjectures of Erdős and Si-
monovits, which roughly say that the behavior of ex(n, H) for general bipartite H is very
complicated.
Conjecture 4.10 (Erdős–Simonovits rational exponents conjecture). For every bipartite H,
there exists some rational number α ∈ [1, 2) so that

ex(n, H) = Θ(nα ).

12
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Moreover, the converse holds: for every rational α ∈ [1, 2), there exists some bipartite graph
H so that
ex(n, H) = Θ(nα )

The first part of this conjecture is doubted by some experts, though no one has any idea
how to prove or disprove it. However, the second part of the conjecture—that there exists
a graph for any rational α—is widely believed to be true, and we are in fact fairly close to
proving it. Every few months, a new paper appears finding a new infinite set of rational
numbers that are now known to be “achievable”, i.e. to be the exponent of ex(n, H) for some
bipartite H.
Moreover, a slight weakening of the second part of the conjecture was recently proved by
Bukh and Conlon. Recall from the homework that if H is a collection of graphs, then we
say that G is H-free if G contains no copy of any H ∈ H, and we write

ex(n, H) = max{e(G) : G is an n-vertex H-free graph}.

Theorem 4.11 (Bukh–Conlon 2018). For every rational α ∈ [1, 2), there exists some finite
collection H of bipartite graphs for which

ex(n, H) = Θ(nα ).

5 Extremal numbers of hypergraphs

It’s time for everything to get more hyper.
If we go back to bare basics, a graph is a collection V of vertices, plus a collection E of
edges, which are simply unordered pairs of vertices. Why restrict ourselves to pairs?

Definition 5.1. A k-uniform hypergraph (sometimes called an k-graph for short) consists
of a finite collection V of vertices, as well as a collection E of k-uniform hyperedges, which
are simply subsets of V of size k.
As with graphs, we say that one k-graph H is a subhypergraph (or simply subgraph) of
another k-graph G if we can obtain H from G by deleting some vertices and edges. We say
that G is H-free if G does not contain H as a subgraph (and we also say that G has no copy
of H).
As with graphs, we define the extremal number of H as

ex(n, H) = max{e(G) : G is an n-vertex H-free k-graph}.

In contrast to graphs (the case k = 2), we know extraordinarily little about ex(n, H) for
k-graphs H with k ≥ 3. For example, even the hypergraph analogue of Mantel’s theorem is
a famous open problem. To explain this formally, we make the following definition.
(k)
Definition 5.2. The complete k-graph on r vertices, denoted Kr , is the k-graph with r
vertices whose edge set consists of all subsets of size k.

13
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes
(k)
Then the amazing fact is that for any k > r ≥ 3, we do not know the value of ex(n, Kr ).
For literally no pair of (r, k)! This problem was proposed by Turán already in 1941, and he
made the following conjecture, which is a natural analogue of Mantel’s theorem.

Conjecture 5.3 (Turán 1941).

  
(3) 5 n
ex(n, K4 ) = + o(1) .
9 3
(3)
The reason for 5/9 is a specific construction of an n-vertex K4 -free 3-graph, which
Turán came up with, and which was the best he could come up with. You’ll see Turán’s
construction on the homework.
Erdős offered $500 for the resolution of Conjecture 5.3, and $1000 for a general formula
(k)
for ex(n, Kr ). So far, very little progress has been made on these questions. The best
(3)
known bound for ex(n, K4 ) is due to Razborov, who proved that
 
(3) n
ex(n, K4 ) ≤ (0.561666 + o(1)) .
3

Note that 5/9 = 0.555 . . . , so this is pretty close to Turán’s conjecture. Unfortunately,
Razborov’s technique is unlikely to yield the full resolution of Conjecture 5.3, because his
technique uses a computer to do complicated computations to what is essentially a “finite
approximation” to the problem.
In general, the best known lower bound is due to de Caen, who proved that
! 
1 n
ex(n, Kr(k) ) ≤ 1 − r−1
+ o(1) .
k−1
k

The best known general lower bound, due to Sidorenko, is

 k−1 ! 
k−1 n
ex(n, Kr(k) ) ≥ 1− + o(1) .
r−1 k

In the case of k = 3, this says that

 2 ! 
2 n
ex(n, Kr(3) ) ≥ 1− + o(1) ,
r−1 3

and this was conjectured to be optimal by Turán. You’ll see the construction in the home-
work.
Despite not knowing the hypergraph analogues of Mantel’s or Turán’s theorems, the
hypergraph analogue of the Kővári–Sós–Turán theorem is known, and is due to Erdős (1965).
To state this, we need to define the hypergraph analogue of a bipartite graph.

14
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Definition 5.4. A k-graph H is called k-partite if its vertex set can be split into k parts,
so that every hyperedge of H contains exactly one vertex from each part.
(k)
The complete k-partite k-graph with parts of sizes s1 , . . . , sk , denoted Ks1 ,...,sk , is the k-
graph with parts of sizes s1 , . . . , sk , containing every edge with exactly one vertex from each
part.
Note that in case k = 2, this simply recovers the definition of a bipartite graph and a
complete bipartite graph. Because of this, the following result generalizes the Kővári–Sós–
Turán theorem.
Theorem 5.5 (Erdős 1965). Let s1 ≤ · · · ≤ sk be positive integers. Then
 k− 1 
(k)
ex(n, Ks1 ,...,sk ) ≤ O n s1 s2 ···sk−1
.

The most important thing here is that any n-vertex k-graph has at most nk = Θ(nk )

hyperedges. So this upper bound has a smaller exponent on n. This implies that if H is any
k-partite k-graph, then  
n
ex(n, H) = o(1) · .
k
The upper bound in Theorem 4.3 is still the best upper bound we have on extremal
numbers of k-partite k-graphs. Moreover, as in the case of graphs, it is known that the
bound in Theorem 5.5 is best possible if sk is sufficiently large with respect to s1 , . . . , sk−1 .
The proof of Theorem 5.5 very similar to that of Theorem 4.3, except that we combine
it with an induction on k. To keep the notation from getting too crazy, we will only prove
it in the case k = 3, which we will derive from the case k = 2, i.e. the Kővári–Sós–Turán
theorem. Also, we will only prove it in the case s1 = s2 = s3 = s, i.e. we will prove that
(3) 2
ex(n, Ks,s,s ) ≤ O(n3−1/s ). (2)

Hopefully you’ll believe me (or convince yourself that it’s true if you don’t!) that the general
result follows from the same technique, just with more bookkeeping.
2
Proof of (2). Let G be an n-vertex 3-graph with at least Cn3−1/s hyperedges, for some
(3)
constant C > 0 we will pick later. Suppose for contradiction that G is Ks,s,s -free. Let X be
(3)
the number of copies of K1,1,s in G. We bound X in two ways.
First, for a pair of distinct vertices v, w, let codeg(v, w) denote the number of hyperedges
containing both v and w. Then we first claim that
X
codeg(v, w) = 3e(G).
v,w∈V (G)
distinct

This is true for the same reason that the sum of the degrees in a graph equals twice the
number of edges. Namely, every hyperedge of G appears exactly three times in the sum on
the left-hand side.

15
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Using this, we see that by Jensen’s inequality,

X codeg(v, w) n n1 n 
P

codeg(v, w)
  
(2) v,w n 3e(G)/ 2
X= ≥ = .
s 2 s 2 s
v,w∈V (G)
distinct

2 2
Note too that since e(G) ≥ Cn3−1/s , we have that 3e(G)/ n2 ≥ Ω(n1−1/s ). This implies

that
n 3e(G)/ n2
 

2
X≥ ≥ cn2 · (Cn1−1/s )s = cC s n2+s−1/s
2 s
for some absolute constant c > 0, depending only on s.
On the other hand, we may upper-bound X by counting over s-sets of vertices which
(3)
can be the “outer” vertices of the K1,1,s . Namely, fix distinct u1 , . . . , us ∈ V (G). We define
a new graph (note: not a hypergraph, a graph) G(u1 , . . . , us ) as follows. The vertex set of
G(u1 , . . . , us ) is V (G) \ {u1 , . . . , us }. Moreover, given v, w ∈ V (G) \ {u1 , . . . , us }, we make
(3)
vw an edge of G(u1 , . . . , us ) if and only if {v, w, u1 , . . . , us } form a copy of K1,1,s .
Now, for every choice of u1 , . . . , us , we claim that G(u1 , . . . , us ) is a Ks,s -free graph.
(3)
Indeed, if we had a copy of Ks,s in G(u1 , . . . , us ), then we would find a copy of Ks,s,s in G,
which is a contradiction. So by the Kővári–Sós–Turán theorem, we know that

e(G(u1 , . . . , us )) ≤ O(n2−1/s )

for every choice of distinct u1 , . . . , us ∈ V (G).

We can use this to upper-bound X, as follows. Note that e(G(u1 , . . . , us )) is precisely
(3)
the number of copies of K1,1,s that have u1 , . . . , us as the outer vertices. This implies that
 
X X
2−1/s n
X= e(G(u1 , . . . , us )) ≤ O(n )= · O(n2−1/s ) = O(n2+s−1/s ).
s
u1 ,...,us ∈V (G) u1 ,...,us ∈V (G)
distinct distinct

Combining our upper and lower bounds on X, we see that

cC s n2+s−1/s ≤ O(n2+s−1/s ),

where both c and the implicit constant in the big-O depend only on s. Thus, if we pick C
(3)
sufficiently large, this is a contradiction, and we conclude that G has a copy of Ks,s,s .

16
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

6 Supersaturation
In this section, we discuss a special case of a very important phenomenon in extremal combi-
natorics, known as supersaturation. Roughly speaking, extremal combinatorics proves results
of the type “if some discrete structure is sufficiently ‘large’, then it contains at least one copy
of some other structure”. The example we’ve seen of this is Turán’s theorem: if a graph
(discrete structure) has sufficiently many edges (is large) then it contains a Kk subgraph (a
copy of some other structure). Supersaturation, in general, boosts this to a statement of the
type “if the discrete structure is just a bit larger, then it contains very many copies of the
other structure”. Specifically, we’ll prove the following supersaturation version of Turán’s
theorem. It was first explicitly stated by Erdős and Simonovits in 1983, but it can implicitly
be found in earlier works, e.g. of Erdős from 1971.

Theorem 6.1. For every integer k ≥ 3 and every real number ε > 0, there exists some δ > 0
so that the following holds for all sufficiently large n. If G is an n-vertex graph with
  
1 n
e(G) ≥ 1 − +ε ,
k−1 2

then G contains at least δ nk copies of Kk .

Note that G has at most nk copies of Kk , so this theorem is pretty remarkable: it says

that once we have just barely more edges than the Turán graph, we have not only one copy
of Kk , but a constant proportion of all possible copies of Kk . To prove this theorem, we need
the following useful lemma, which is stated in greater generality than we need.
For a graph G and a subset M ⊆ V (G), we denote by e(M ) the number of edges entirely
contained in M , or equivalently the number of edges in the induced subgraph G[M ].

Lemma 6.2. Let 0 < α < β < 1 be real numbers, let
m ≥ 2 be an integer, and let G be an
n-vertex graph with n ≥ m. Assume that e(G) ≥ β n2 .
Then the number of sets M ⊆ V (G)
m n


with |M | = m and e(M ) ≥ α 2 is at least (β − α) m .

Proof. The key identity which underlies this proof is

 
n−2 X
e(G) = e(M ).
m−2
M ⊆V (G)
|M |=m

This has a simple bijective proof. On the right-hand side, every edge uv is counted a number
of times, and that number of times is simply the number
of m-sets M which contain both u
n−2
and v. But the number of such m-sets is exactly m−2 , yielding the formula.
Now, let M0 denote the set of M with e(M ) < α 2 , and let M1 denote the set of M
m

with e(M ) ≥ α m2 . So our goal is to prove a lower bound on |M1 |. Continuing the identity

17
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

above, we can write

 
n−2 X X
e(G) = e(M ) + e(M )
m−2 M ∈M0 M ∈M1
X m X m
≤ α +
M ∈M
2 M ∈M1
2
 0
m
= (α|M0 | + |M1 |)
2

since every m-set in M0 has at most α m2 edges, and every m-set in M1 has at most m2

edges.
n n n




Note that |M0 | + |M1 | = m . Let x = |M1 |/ m , so that 1 − x = |M0 |/ m . Dividing by
n m



m 2
, the above inequality yields
n−2


m−2
n

m
e(G) ≤ α(1 − x) + x = α + (1 − α)x.
m 2

n


Now, we recall that e(G) ≥ β 2
, so
n−2 n−2 n




m−2 m−2 2
n

m
e(G) ≥ n

m

β.
m 2 m 2

n−2

n
n

m

The final step is another magic identity, which is that m−2 2
= m 2
; in other words,
the complicated fraction above is simply equal to 1. Indeed, both sides of this identity count
the same object, which is the number of ways of picking an m-set out of n objects, and then
picking 2 objects from the m-set.
Combining all these inequalities, we find that
β−α
β ≤ α + (1 − α)x ⇐⇒ x≥ ,
1−α
which implies that
     
n β−α n n
|M1 | = x ≥ ≥ (β − α) ,
m 1−α m m
as claimed.
With this lemma, we are ready to prove the supersaturation theorem, Theorem 6.1.
1
+ o(1)) m2 , where the o(1)


Proof of Theorem 6.1. Fix ε > 0. Recall that tk−1 (m) = (1 − k−1
term tends to 0 as m → ∞. This implies that there is some fixed m, depending only on ε,
so that   
1 ε m
tk−1 (m) ≤ 1 − + .
k−1 2 2

18
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Let this m be fixed, and let n ≥ m. Suppose that G is an n-vertex graph with at least
1
+ ε) n2 edges. We apply Lemma 6.2 with β = 1 − k−1 1 1
+ 2ε .


(1 − k−1 + ε and α = 1 − k−1
1
Then Lemma 6.2 tells us that the number of m-sets M ⊆ V (G) with e(M ) ≥ (1 − k−1 + 2ε )
n
is at least 2ε m


.
Every such m-set M has at least tk−1 (m) edges, so Turán’s theorem
implies that such an
ε n
M contains a copy of Kk . In other words, we’ve found at least 2 m copies of Kk , except
n−k


that we might have over-counted: each copy of Kk can be counted up to m−k times, since
n−k


the k vertices of the Kk can appear in m−k different m-sets M .
So in total, the number of Kk in G is at least
ε n n n



 
2 m ε m ε k
ε n
n−k

= · n−k
= · m = m
,
m−k
2 m−k 2 k
2 k k

where the

mmiddle
equality

uses the same magic identity as in the proof of Lemma 6.2, namely
n n n−k


that m k
= k m−k
.
To
conclude, we recall that m depends solely on ε and k. Therefore, if we define δ =
ε/(2 m
k
), then this will only depend on ε and k, and that yields the desired result.

7 Proof of the Erdős–Stone–Simonovits theorem

We are finally ready to prove the Erdős–Stone–Simonovits theorem. We begin by observing
a simple reduction, due to Erdős and Simonovits, which says that to prove the bound on
ex(n, H) for all H, it suffices to prove it for a very special class of H. Let Kk [s] denote the
complete k-partite graph with parts of size s. (Note that this is the same graph as the Turán
graph Tk (ks).)

Proposition 7.1. Suppose that for all positive integers k, s, we have that
  
1 n
ex(n, Kk [s]) = 1 − + o(1) .
k−1 2

Then   
1 n
ex(n, H) = 1 − + o(1)
χ(H) − 1 2
for every graph H.

Proof. We already proved the lower bound in the Erdős–Stone–Simonovits theorem, namely
that   
1 n
ex(n, H) ≥ tχ(H)−1 (n) = 1 − + o(1) .
χ(H) − 1 2
So it only suffices to prove the upper bound. Now, the key claim is that if H has chromatic
number k, then H is a subgraph of Kk [s] for some positive integer s. Indeed, if H has
chromatic number k, then we may split the vertices of H into k color classes, with the

19
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

property that no edge of H goes between two vertices in the same color class. If s is the
maximum size of one of the color classes, this precisely means that H is a subgraph of Kk [s].
But in that case, we see that
  
1 n
ex(n, H) ≤ ex(n, Kk [s]) = 1 − + o(1) ,
k−1 2

by assumption.
So it suffices to prove what is often called the Erdős–Stone theorem, namely the statement
1 n


that ex(n, Kk [s]) ≤ (1 − k−1 + o(1)) 2 for every k, s. This is what we now do.
Proof of the Erdős–Stone theorem. Fix some ε > 0. Our goal is to prove that if n is suffi-
ciently large in terms of ε, k, and s, and if G is an n-vertex graph with
  
1 n
e(G) ≥ 1 − +ε
k−1 2

edges, then G contains a copy of Kk [s].

By the supersaturation theorem, Theorem 6.1, we know that G has at least δ nk copies

of Kk , where δ > 0 depends only on ε and k. We define a k-uniform hypergraph G whose

vertex set is V (G), and we make a k-tuple of vertices a hyperedge of G if and only if the
k-tuple defines a copy of Kk in G. Then we have that
 
n
e(G) = #(copies of Kk in G) ≥ δ .
k

Recall that by Theorem 5.5, we have that

(k) k−1
ex(n, Ks,s,...,s ) ≤ Cnk−1/s

for some fixed constant C > 0. Now, if δ is fixed (which it is, since it only depends on ε and
k), and if n is sufficiently large, then
 
n k−1
δ > Cnk−1/s . (3)
k

This is because, as we’ve discussed previously, nk grows as Θ(nk ), and on the right-hand

side we have a smaller power of n. So as long as n is sufficiently large in terms of the other
parameters, we have that (3) holds.
(k)
Thus, for sufficiently large n, we have that e(G) > ex(n, Ks,s,...,s ), which implies that G
(k)
contains a copy of Ks,s,...,s . In other words, inside V (G), we can find k sets of s vertices each,
with the property that whenever we pick one vertex from each part, they yield a copy of Kk
in G. But that precisely means we have found a copy of Kk [s] in G, as claimed.

20
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

8 Topological problems in hypergraphs

The first result I mentioned in this class is that if G is an n-vertex graph with no cycle, then
G has at most n − 1 edges. There is a topological perspective on this: if we view a graph as
a topological object, then a cycle is simply a topological copy of the circle which is contained
in our graph.
Similarly, we can think of a 3-uniform hypergraph as a two-dimensional topological space.
So a natural analogue of this question for 3-graphs is the following: how many hyperedges
can we put in an n-vertex 3-graph without a topological copy of the sphere? Formally, we
make the following definition. Formally, we care about what are called triangulations of
the sphere, which are 3-graphs with the property that if we view the vertices and triangles
as geometric objects, we get something homeomorphic to a sphere (if you don’t like the
word homeomorphic, this just means that the resulting geometric object is an “ordinary”
polyhedron).
Let S be the family of all triangulations of the sphere. Then what we are interested in
is the 3-uniform extremal function ex(n, S). This question was studied, and resolved, by
Brown, Erdős, and Sós in 1973.
Theorem 8.1 (Brown–Erdős–Sos 1973).
ex(n, S) = Θ(n5/2 ).
As always, we need to prove both an upper and a lower bound. In this case, somewhat
surprisingly, the upper bound is a little easier than the lower bound, and closely follows the
proof of Theorem 5.5.
Proof of the upper bound. Let G be an n-vertex 3-graph with at least Cn5/2 edges, for some
big constant C > 0. We want to prove that G contains a triangulation of the sphere. In
fact, we will prove that G necessarily contains a double pyramid : this is a 3-graph consisting
of vertices x, y, z1 , . . . , z` and all edges of the form xzi zi+1 and yzi zi+1 , with addition mod
`. Here’s a picture of the double pyramid in case ` = 5, where every triangle you see is a
hyperedge.

(3) (3)
Let X be the number of copies of K1,1,2 in G, where K1,1,2 = . As always, we will
estimate X in two ways. First, as in the proof of Theorem 5.5, we know that
X codeg(v, w) n n1 n 
P

codeg(v, w)
  
(2) v,w n 3e(G)/
X= ≥ = 2 ≥ cC 2 n3 ,
2 2 s 2 2
v,w∈V (G)
distinct

21
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

where c > 0 is an absolute constant. On the other hand, for any pair u1 , u2 of distinct
vertices, let G(u1 , u2 ) be the graph with vertex set V (G) \ {u1 , u2 }, where a pair vw is an
(3)
edge if and only if {v, w, u1 , u2 } form a copy of K1,1,2 with v, w as the vertices that lie in
both hyperedges.
Now, the key claim is that if G(u1 , u2 ) contains a cycle for some u1 , u2 , then G contains a
double pyramid, and thus a triangulation of the sphere. Indeed, in the definition of a double
pyramid above, we can set x = u1 , y = u2 , and z1 , . . . , z` the vertices of a cycle in G(u1 , u2 ).
So if G has no triangulation of the sphere, then G(u1 , u2 ) has no cycle for all u1 , u2 . This
implies that
e(G(u1 , u2 )) ≤ v(G(u1 , u2 )) − 1 = n − 3.
Adding this all up, we conclude that
 
X n
X= e(G(u1 , u2 )) ≤ (n − 3) < n3 .
2
u1 ,u2 ∈V (G)
distinct

But if we let C be sufficiently large, this contradicts our lower bound X ≥ cC 2 n3 .

The lower bound, just as Klein’s construction of a C4 -free graph with many edges, uses
a clever input from finite fields. But first, there is a key observation. Let Pk be the single
pyramid hypergraph with a base of size k. Here are drawings of P3 , P4 , P5 , where the shaded
triangles are the hyperedges.

P3 = P4 = P5 =

Lemma 8.2. Let H ∈ S be a triangulation of the sphere. Then H contains a copy of Pk for
some k ∈ {3, 4, 5}.

Proof. The key fact we need here, which is a simple consequence of Euler’s formula, is that
every planar graph has a vertex of degree at most 5. Said differently, in any triangulation
of the sphere, one of the vertices is incident to at most 5 (two-dimensional) edges of the
polyhedron. But that means that that vertex plus its neighbors forms a copy of Pk for some
k ∈ {3, 4, 5}.
To construct the lower bound in Theorem 8.1, Brown, Erdős, and Sós constructed an
n-vertex 3-graph with Ω(n5/2 ) edges and no copy of P3 , P4 , or P5 . Thus, this 3-graph is
S-free, by Lemma 8.2.
Before describing the construction, let’s briefly return to Klein’s construction of a C4 -free
graph. Recall that we defined it as follows. We let P be the set of points in F2p , and L the set
of (non-vertical) lines in F2p , and the edges were given by incidence: some p ∈ P is adjacent

22
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

to some ` ∈ L if and only if p ∈ `. Said differently, we identify P with pairs (x, y) where
x, y ∈ Fp , and we identify L with pairs (m, b) with m, b ∈ Fp , and incidence is given by

(x, y) ∼ (m, b) ⇐⇒ y = mx + b.

By changing variables, we see that this is more or less the same construction as the following.
Define a bipartite graph with parts A, B. Vertices of A are pairs (a1 , a2 ) with a1 , a2 ∈ Fp \{0},
vertices of B are pairs (b1 , b2 ) with b1 , b2 ∈ Fp \ {0}, and adjacency is given by

(a1 , a2 ) ∼ (b1 , b2 ) ⇐⇒ a1 b1 + a2 b2 = 1.

It is not hard to check that this graph, like the earlier construction we defined, has n =
2(p − 1)2 = Θ(p2 ) vertices, Θ(p3 ) = Θ(n3/2 ) edges, and no copy of C4 . The proof is the
same as before: assume there is a C4 , which yields four linear equations, and by tracing
through these linear equations we find that we can’t have two distinct vertices from A and
two distinct vertices from B.
In fact, we can generalize this construction even further. If we fix non-zero c1 , c2 ∈ Fp ,
then we can define an alternate bipartite graph G(c1 , c2 ) with vertex set A∪B, with adjacency
given by
(a1 , a2 ) ∼ (b1 , b2 ) ⇐⇒ c1 · a1 b1 + c2 · a2 b2 = 1.
Then again, G is a C4 -free graph with n = 2(p − 1)2 vertices and Θ(n3/2 ) edges. Again, the
point is that we can trace through the linear equations, and the non-zero multipliers c1 , c2
don’t matter.
With this, we can pretty easily prove the lower bound in Theorem 8.1. As discussed
above, thanks to Lemma 8.2, it suffices to prove the following result.
Proposition 8.3 (Brown–Erdős–Sós 1973). For all n, there exists an n-vertex 3-graph G
with Θ(n5/2 ) edges and no copy of P3 , P4 , or P5 .
Proof. We assume for simplicity that n = 3(p − 1)2 for some prime p; as in the proof of
Theorem 4.4, this assumption can be lifted thanks to Bertrand’s postulate. Let G be the
3-partite 3-graph with parts A, B, C, where

A = {(a1 , a2 ) : a1 , a2 ∈ Fp \ {0}}
B = {(b1 , b2 ) : b1 , b2 ∈ Fp \ {0}}
C = {(c1 , c2 ) : c1 , c2 ∈ Fp \ {0}},

and we make a triple {(a1 , a2 ), (b1 , b2 ), (c1 , c2 )} a hyperedge of G if and only if

a1 b1 c1 + a2 b2 c2 = 1.

Note that if we fix (c1 , c2 ) ∈ C, then the graph G(c1 , c2 ) of pairs (a1 , a2 ) ∈ A, (b1 , b2 ) ∈ B
with {(a1 , a2 ), (b1 , b2 ), (c1 , c2 )} a hyperedge is precisely the graph G(c1 , c2 ) defined above.
In particular, we see that there are Θ(p3 ) hyperedges containing every (c1 , c2 ) ∈ C, which
implies that there are Θ(p5 ) = Θ(n5/2 ) hyperedges in G.

23
Mathcamp 2022 Extremal graph theory (Yuval) Lecture notes

Additionally, note that for every (c1 , c2 ) ∈ C, the graph G(c1 , c2 ) is bipartite and C4 -free.
So in particular, G(c1 , c2 ) does not contain any copy of C3 , C4 , or C5 . This implies that there
is no copy of P3 , P4 , or P5 centered at any (c1 , c2 ) ∈ C. But the hypergraph is symmetric,
so the same argument applies if we try to center the Pk in A or in B. So we conclude that
G is {P3 , P4 , P5 }-free, as claimed.
Given the Brown–Erdős–Sós theorem, it is natural to ask about triangulations of other
surfaces. Namely, for an integer g ≥ 1, let Sg denote the set of all 3-graphs which are
triangulations of a genus-g surface. (Recall that a genus-g surface is the same as a g-hole
torus.) Then it is natural to ask for the value of ex(n, Sg ). For many years, there was no
real progress on this problem, but it was very recently resolved.

Theorem 8.4 (Kupavskii–Polyanskii–Tomon–Zakharov 2021). For every g ≥ 1, we have

that
ex(n, Sg ) = Θ(n5/2 ).

Note that the implicit constant in the big-Θ above depends on g. To prove this, they
had to prove both upper and lower bounds. The lower bound is totally different from
above, essentially because there is not as simple of a “local” obstruction to the presence
of a triangulation of a genus-g surface. Instead of the lower bound above, they use the
probabilistic method, plus a powerful result of Gao on the number of n-vertex triangulations
of a genus-g surface.
The upper bound uses two separate, pretty cool ideas. The first is a much trickier
version of the double pyramid argument of Brown, Erdős, and Sós, which allows one to build
a triangulated torus in a 3-graph by finding and gluing together very many double pyramids
along a long cycle. The second is a simple probabilistic argument that shows that we can
upper-bound ex(n, Sg+1 ) in terms of ex(n, Sg ). Basically, the idea is that we can randomly
partition the edges of G into two parts. In one part, we can find a triangulation of a genus-
g surface, using what we already know about ex(n, Sg ). In the other part, we can find a
triangulation of the torus, using what we know about ex(n, S1 ). Moreover, by exploiting the
randomness, we can ensure that we can glue these together to get a triangulated genus-(g+1)
surface. This can be used to show inductively that ex(n, Sg+1 ) ≤ O(n5/2 ).
What about non-orientable surfaces? It turns out that every non-orientable surface has
a non-orientable genus k ≥ 1, and a non-orientable surface of genus k can be obtained by
gluing together k copies of the real projective plane. Let S−k denote the set of triangulations
of the genus-k non-orientable surface. Then the same probabilistic lower bound implies that
ex(n, S−
k ) ≥ Ω(n
5/2
) for all n. Moreover, the same gluing argument for the upper bound
shows that it suffices to understand the case of the real projective plane, i.e. to upper-bound
ex(n, S−
1 ). This was extremely recently resolved by Maya Sankar.

Theorem 8.5 (Sankar 2022+). ex(n, S−

1 ) = Θ(n
5/2
). As a consequence, ex(n, S−
k ) = Θ(n
5/2
)
for all k ≥ 1.

24