Stability and Erdős–Stone type results for F-free graphs with a fixed number of edges

# Stability and Erdős–Stone type results for F-free graphs with a fixed number of edges

Andrew Uzzell
Grinnell College
uzzellan@grinnell.edu
August 4, 2019
###### Abstract

A fundamental problem of extremal graph theory is to ask, “What is the maximum number of edges in an -free graph on vertices?” Recently Alon and Shikhelman proposed a more general, subgraph counting, version of this question. They considered the question of determining the maximum number of copies of a fixed graph in an -free graph on vertices.

In this more general context, where we are no longer counting edges, it is also natural to ask what is the maximum number of copies of in an -free graph with edges and no restriction on the number of vertices. Frohmader, in a different context, determined the answer when and are both complete graphs. We prove results for this problem analogous to the Erdős–Stone theorem, the Erdős–Simonovits theorem, and the stability theorem of Erdős–Simonovits.

## 1 Introduction

### 1.1 Extremal numbers and generalizations of extremal numbers

The fundamental problem of extremal graph theory is to compute the extremal number,

 ex(n,F)=max{e(G):G is an F-free graph on n % vertices}.

Recently Alon and Shikhelman [1] proposed a more general version of this problem. Rather than counting edges, they considered the problem of determining the maximum number of copies of a fixed graph in an -free graph on vertices. Letting be the number of copies of in , we define

 exT(n,F)=max{NT(G):G is an F-free graph% on n vertices}.

Turán’s theorem [16] states that , the number of edges in the Turán graph , the complete -partite graph on vertices with parts as equal in size as possible. Moreover the Turán graph is the unique extremal graph. The following result was proved by Zykov [17] (and has since been rediscovered many times).

###### Theorem 1.

For all , the maximum in the definition of is uniquely achieved by the Turán graph . In other words,

 exKt(n,Kr+1)=NKt(Tr(n)).

Now that we are counting copies of , rather than edges, it makes sense to shift away from our resource being a limited number of vertices we are allowed, and consider similar problems for the class of graphs with edges. We make a third parallel definition.

 mexT(m,F)=max{NT(G):G is an F-free % graph with m edges}.

It is important to note that this definition does not place any restriction on the number of vertices of .

### 1.2 Previous results about mexT(m,F)

Some results about are known (though not using that terminology). One can even think of the Kruskal–Katona theorem [11, 12] as proving a result in this direction. We start with a little background about that theorem.

Given , let denote the family of -sets of . The colexicographic or colex order on is defined as follows: for all , if and only if . For a family we define the shadow of on level to be the set

 ∂p(A)={B∈([n]p):∃A∈A s.t.\ B⊆A}.

The Kruskal–Katona theorem gives a bound for the minimum size of as a function of the size of .

###### Theorem 2.

If and is the colex initial segment of of size then for any we have

 |∂p(A)|≥|∂p(C)|.

One should also note is itself an initial segment in the colex order on .

It is an immediate corollary of Theorem 2 that for every and , the maximum number of copies of  in a graph with edges is achieved by the graph with vertex set  whose edge set consists of the first elements of  in colex order. We call this graph the colex graph with edges, and denote it . This is a slight abuse of notation, since we have not specified , but in our problems we only care that we have enough vertices, not how many there actually are.

Frohmader [10] determined the value of for all . His results were phrased in terms of simplicial complexes, so let us take a moment to recall the relevant definitions.

Let be a simplicial complex. If is a face of , then the dimension of  is . The dimension of  is . Let and, for each , , let denote the number of -dimensional faces in . Recall that the -vector of  is the -tuple . We say that a complex  is -colorable if there is a partion of its vertex set into parts such that each set in meets each part in at most one element.

A simplicial complex  is called a flag complex if every minimal non-face of  has two elements. This is equivalent to the notion of a “clique complex”: the clique complex of a graph  is the simplicial complex  whose vertex set is  and whose faces are the cliques of . It is easy to see that a flag complex is -colorable if and only if it is the clique complex of an -colorable graph. We say that a complex  is balanced if and is -colorable.

Kalai (unpublished; see [15, p. 100]) and Eckhoff [4] conjectured that if is a flag complex, then there exists a balanced complex  with the same -vector as . Frohmader [10] proved the Kalai–Eckhoff conjecture. This is in fact sufficient to determine . For completeness we include a proof of this deduction below.

We will need to quote the “colored” version of the Kruskal–Katona theorem, proved by Frankl, Füredi, and Kalai [9]. The role played by the colex order in the Kruskal–Katona theorem is played here by the -partite colex order. This is colex order restricted to subsets of such that for all we have or .

Given and , the colex Turán graph  is the graph on vertex set whose edge set consists of the first edges in -partite colex order. (See Figure 1.) Note that if , then the unique non-trivial component of is isomorphic to .

###### Theorem 3 ([9]).

If is -colorable and is the initial segment of in the -partite colex order of size , then for any we have

 |∂p(A)|≥|∂p(C)|.

One should also note is itself an initial segment in the -partite colex order on .

In the next corollary and throughout the rest of the paper we write for . Also, given and , let and denote the number of copies of  in that contain and the number of copies of  in that contain , respectively.

###### Corollary 4.

If is an -partite graph with edges then

 kt(G)≤kt(CTr(m)).
###### Proof.

This is an immediate consequence of Theorem 3 and the definition of . ∎

From these results we can prove that achieves the maximum in the definition of .

For all we have

 mexKt(m,Kr+1)=kt(CTr(m)).
###### Proof.

Consider a -free graph having edges. Its clique complex is a flag complex, so, since the Kalai–Eckhoff conjecture is true, there is a balanced complex having the same -vector as . Since we know that is -colorable. By Corollary 4 we have . ∎

### 1.3 Structural supersaturation and stability

Turán’s theorem has inspired a great deal of research on the size and structure of extremal and near-extremal -free graphs. We mention several important theorems in this area in order to motivate our results.

Given a graph  and a positive integer , we let denote the -fold blowup of  (in which every vertex of is replaced by an independent set of size  and every edge by a copy of ). Erdős and Stone [8] showed that a graph with vertices and edges not only contains , but contains a sizable blow-up of .

###### Theorem 6.

For all , , and , there exists such that if and is a graph with vertices such that

 e(G)≥tr(n)+ϵn2,

then contains .

Erdős and Simonovits [7] observed that Theorem 6 implies that determines up to a error term.

###### Theorem 7.

Let be a graph. We have

 ex(n,F)=(χ(F)−2χ(F)−1+o(1))(n2).

Erdős and Simonovits [5, 14] also proved a stability result, which says that a -free graph with nearly the extremal number of edges has nearly extremal structure.

###### Theorem 8.

For all and all , there exist and such that if and is a -free graph with vertices such that

 e(G)≥(1−δ)tr(n),

then can be made -partite by deleting at most  edges.

### 1.4 Results

In Section 2 we prove analogues of the Erdős–Stone theorem (Theorem 6), the Erdős–Simonovits theorem (Theorem 7), and the Erdős–Simonovits stability theorem (Theorem 8) in the context of Frohmader’s theorem, Theorem 5. To be precise we prove the following results.

###### Theorem 9.

For all , all , and all , there exists such that if and is a graph with edges such that

 ks(G)≥mexKs(m,Kr+1)+ϵms/2,

then contains .

###### Theorem 10.

Let , let be a graph, and let . We have

 mexKs(m,F)=mexKs(m,Kr+1)+o(ms/2).

Theorem 10 follows from Theorem 9 in much the same way that Theorem 7 follows from Theorem 6, so we will omit the proof.

###### Theorem 11.

For all and all , there exist and such that if and is a -free graph with edges such that

 ks(G)≥(1−δ)mexKs(m,Kr+1),

then can be made -partite by deleting at most  edges.

###### Corollary 12.

For all , every graph  with chromatic number , and all , there exist and such that if and is an -free graph with edges such that

 ks(G)≥(1−δ)mexKs(m,F),

then can be made -partite by deleting at most  edges.

Our results establish a number of parallels between and , with the colex Turán graph  playing the role in results about that the Turán graph  plays in results about . However, this correspondence is not perfect.

Let be a graph with a critical edge. Simonovits [14] used the stability method to determine (and the extremal graph) for all  sufficiently large.

###### Theorem 13.

Let be a graph with and suppose that contains an edge  such that . For all  sufficiently large, and is the unique extremal graph.

In contrast, if is as in the statement of Theorem 13 and , there exist infinitely many values of  such that is not an extremal graph for .

Given , let be the least integer such that . Let be the graph consisting of  and a vertex  that is joined to vertices of , distributed as evenly as possible among the classes of . Observe that if , then is -free but not -partite. Moreover,

 ks(G) =ks(Tr(n−1))+ks(v∗) =ks(Tr(n−1))+ks−1(Tr(m−tr(n−1))) >ks(Tr(n−1))+ks−1(Tr−1(m−tr(n−1))) =ks(CTr(m)).

Finally there are a number of very natural analogues of results concerning that are open for . In Section 3 we briefly discuss some of these open problems.

## 2 Proof of Theorems 9, 10, and 11

### 2.1 Preliminaries and notation

Let be a graph and let . Recall that for and , and denote the number of copies of  in that contain and the number of copies of  in that contain , respectively. The minimum values of these quantities are denoted

 δs(G) =min{ks(v):v∈V(G)} δ′s(G) =min{ks(e):e∈E(G)}.

In the extremal -free graph the average degree is a multiple of and the number of copies of is a multiple of . We define those constant multiples here. Given and , let

 βr=(2(r−1)r)1/2 (1)

and let

 cr,s=(rs)(r2)s/2. (2)

The following simple proposition collects some computations about .

###### Proposition 14.

If and , then and so

 ks(CTr(m))=ks(Tr(n))=(nr)s(rs)=cr,sms/2.

In particular in this case is -regular. Furthermore we have

 ks(CTr(m))=cr,sms/2+O(m(s−1)/2).
###### Proof.

Straightforward. ∎

We also record some properties of the constants and defined above.

###### Proposition 15.

For all and ,

 (r−1s−1)(r−1)s−1βs−2r =s2cr,s (3) and cr,s =2s/2s!⋅r(s)rs/2(r−1)s/2≤2s/2(r−2)s!(r−1). (4)
###### Proof.

If , then (3) holds trivially, as both sides equal . If , (1) and (2) imply that

 (r−1s−1)(r−1)s−1βs−2r =(r−1s−1)(r−1)s−1(2(r−1)r)s−22 =r2⋅(r−1)(s−1)(s−1)!⋅(2r(r−1))s/2 =s2⋅r(s)s!⋅(r2)−s/2 =s2cr,s.

Also, (4) is immediate from (2). ∎

We will need a consequence of the Kruskal–Katona theorem noted by Lovász [13, Exercise 13.31].

###### Theorem 16.

Let and let . If is a graph with edges, then .

###### Corollary 17.

Let . If is a graph with edges, then

 ks(G)≤(1+o(1))2s/2s!ms/2.
###### Proof.

Straightforward. ∎

We will also need the following result of Erdős, Frankl, and Rödl [6].

###### Theorem 18.

Let . For all and every graph  with chromatic number , there exists such that if is an -free graph of order , then can be made -free by removing at most  edges.

### 2.2 Proof of Theorem 9

###### Proof of Theorem 9.

First, we show that if is sufficiently large, then contains a subgraph  that has both positive edge density and many copies of  relative to .

Let be such that

 1>ρ>1−(s!2s/2(cr,s+ϵ3))2/s. (5)

(Proposition 15 implies that for all and , if is sufficiently small, then the right-hand side of (5) is positive.)

Let be sufficiently large. If , we do nothing. Otherwise, we let and, for each , if contains an edge  with , we set .

Suppose that we delete such edges and let denote the resulting subgraph. We have

 ks(G′) =ks(G)−⌊ρm⌋−1∑i=0ks(ei) ≥ks(G)−⌊ρm⌋−1∑i=02s−2ϵ2s−4(s−2)!(m−i)(s−2)/2 ≥ks(G)−2s−2ϵ2s−4(s−2)!(1+ϵ2)2s(ms/2−((1−ρ)m)s/2).

Thus, using (5) twice, we have

 ks(G′) ≥ks(G)−2s−2ϵ2s−4(s−2)!ρms/2 >(cr,s+ϵ−2ϵ2s−4)ms/2 >(cr,s+2ϵ3)ms/2 >(2s/2s!(1−ρ)s/2+ϵ3)ms/2,

So, has a subgraph  with edges and vertices such that

 δ′s(G′)≥2s−2ϵ2s−4(s−2)!(m′)(s−2)/2.

We claim that

 ks(G′)≥(cr,s+ϵ)(m′)s/2. (6)

Indeed, given , suppose that . If is sufficiently small, then we have

 ks(Gi) ≥ks(Gi−1)−2s−2ϵ2s−4(s−2)!e(Gi−1)(s−2)/2 >(cr,s+ϵ)e(Gi−1)s/2−(cr,s+ϵ)s4e(Gi−1)(s−2)/2 >(cr,s+ϵ)(e(Gi−1)−1)s/2 =(cr,s+ϵ)e(Gi)s/2.

The claimed inequality (6) follows by induction on and our assumption on .

Observe that if and is an endpoint of , then . It follows that

 2s−2ϵ2s−4(s−2)!(m′)(s−2)/2≤δ′s(G′)≤(δ(G′)s−2)≤(2m′/n′s−2)≤(2m′/n′)s−2(s−2)!,

whence

 n′≤1ϵ2(m′)1/2. (7)

Suppose that does not contain a copy of . By the trivial bound , we may let be as large as we wish by taking to be sufficiently large. So, by (7) and Theorem 18, if is sufficiently large, then we can delete all copies of  in by removing at most  edges. This means that we remove at most  copies of . Let denote the resulting graph and let . By (6), Proposition 14, and Theorem 5, we have

 ks(G′′)≥ks(G′)−ϵ2s−3(m′)s/2 ≥(cr,s+ϵ−ϵ2s−3)(m′)s/2 ≥(cr,s(1−ϵ4s−7)s/2+ϵ2)(m′)s/2>mexKs(m′′,Kr+1),

### 2.3 Proof of Theorem 11

Proofs of stability results in extremal graph theory often begin by showing that a global density assumption on a graph  implies a minimum degree condition. This is frequently accomplished by iteratively deleting vertices of degree at most  (where is an appropriate constant) and showing that the density of  and the forbidden subgraph condition mean that only a small fraction of the vertices could have been deleted in this way.

However, in our case, if we delete vertices whose degree is too small as a function of the number of vertices of , then there is no reason to expect that the process will terminate quickly, for the simple reason that we do not know how many vertices has. In particular, we may end up deleting far more than  edges. Instead, letting denote the set of the vertices of  whose degree is too small as a function of the number of edges of , we will show that the vertices of  span only a small fraction of the edges of . We will then be able to show that has high minimum degree as a function of the number of vertices.

###### Lemma 19.

Given and , there exist and with the following property. If and is a -free graph with edges such that

 ks(G)≥(1−δ)mexKs(m,Kr+1),

then has a subgraph  with vertices and edges such that

 δ(G′)≥βr(1−2ϵ)(m′)1/2and % alsoδ(G′)≥(r−1r−4ϵ)n′.
###### Proof.

Let be sufficiently small and let

 δ=s(s−2)cr,s16ϵ2. (8)

Let . For each , if contains a vertex  with , set . Suppose that we delete edges in this way and that we delete edges incident to vertices. (To ensure that we delete exactly edges, if necessary we do not delete the final vertex , but instead delete the appropriate number of edges incident to it.) Let denote the resulting graph. We have

 ks(G′)≥ks(G)−i0−1∑i=0ks−1(Gi[NGi(vi)]).

Because is -free, for each , is -free. Hence, the number of copies of in that contain is at most . By Theorem 1, for all ,

 exKs−1(p,Kr)=ks−1(Tr−1(p))≤(pr−1)s−1(r−1s−1).

So, we have

 ks(G′)≥ks(G)−i0−1∑i=0(dGi(vi)r−1)s−1(r−1s−1). (9)

By assumption, for all , . Moreover, by the definition of ,

 i0−1∑i=0dGi(vi)<ϵm+dGi0−1(vi0−1)≤ϵm+βr(1−2ϵ)m1/2.

Combining this bound with (9) gives

 ks(G)−ks(G′) <⌈ϵm+βr(1−2ϵ)m1/2βr(1−2ϵ)(m−ϵm)1/2⌉(r−1s−1)(r−1)s−1βs−1r(1−2ϵ)s−1m(s−1)/2 <ϵ(r−1s−1)(r−1)s−1(βr(1−ϵ))s−2ms/2. (10)

On the other hand, if is sufficiently large, then Theorem 5, Proposition 14, and our assumption on imply that

 ks(G)−ks(G′) ≥ks(G)−mexKs((1−ϵ)m,Kr+1) ≥(cr,s−δ)ms/2−(1+ϵ3)cr,s(1−ϵ)s/2ms/2 ≥(cr,s−δ−cr,s(1−sϵ2+3s(s−2)ϵ216))ms/2.

By (8) and Proposition 15,

 ks(G)−ks(G′) ≥(−δ+sϵ2cr,s−3s(s−2)ϵ216cr,s)ms/2 =(sϵ2cr,s−s(s−2)ϵ24cr,s)ms/2 =ϵ(r−1s−1)(r−1)s−1βs−2r(1−s−22ϵ)ms/2 >ϵ(r−1s−1)(r−1)s−1(βr(1−ϵ))s−2ms/2,

So, has a subgraph  with vertices and edges such that

 δ(G′)≥βr(1−2ϵ)(m′)1/2.

On the other hand, , so

 n′≤2βr(1−2ϵ)(m′)1/2. (11)

It follows from (11) and (1) that

 δ(G′)≥β2r(1−2ϵ)2n′2>β2r(1−4ϵ)n′2>(r−1r−4ϵ)n′.

This completes the proof. ∎

Now we are ready to prove Theorem 11. The argument is similar to the proof of the -free case of the Erdős–Simonovits stability theorem, Theorem 8.

###### Proof of Theorem 11.

Given , let

 ϵ′=ϵ16r+1. (12)

Let be the graph obtained by inputting , , and into Lemma 19 and let .

By Lemma 19,

 e(G′)≥(r−1r−4ϵ′)(n′)22.

So, if is sufficiently small, Turán’s theorem implies that contains a copy of  with vertex set . Because is -free, every vertex of  has at most  neighbors in . Let and let . By definition,

 e(U,V(G′)∖U)≤(r−1)|A|+(r−2)|B|=(r−1)(n′−r−|B|)+(r−2)|B|.

On the other hand, by Lemma 19,

 e(U,V(G′)∖U)≥r(r−1r−4ϵ′)n′−(r2)=(r−1−4ϵ′r)n′−(r)2.

It follows that

 |B|≤(r−1)(n′−r)−(r−1−4ϵ′r)n′+(r2)=4ϵ′rn′−(r2)<4ϵ′rn′. (13)

For , …, , let . It is easy to see that the partition and that each is an independent set. So, if we delete all of the vertices of  from , the resulting graph is -partite.

It remains to show that deleting the vertices of  from removes only a small number of edges. By Lemma 19, we have , which means that

 n′≤2βr(1−2ϵ)(m′)1/2

(just as in (11)). This bound, (13), and (1) imply that if is sufficiently large, then the number of edges of  incident to a vertex of  is at most

 |B|(n−|B|)+(|B|2)≤4ϵ′r(n′)2+8ϵ′2r2(n′)2≤16ϵ′rβ2r(1−2ϵ′)2m′+32ϵ′2r2β2r(1−2ϵ′)2m′<16ϵ′rm.

It follows from (12) that we have deleted at most  edges of . This completes the proof. ∎

Let and