This paper was converted on www.awesomepapers.org from LaTeX by an anonymous user.
Want to know more? Visit the Converter page.

On stability of the Erdős-Rademacher Problem

József Balogh Department of Mathematics, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA, and Moscow Institute of Physics and Technology, Russian Federation.  and  Felix Christian Clemen Department of Mathematics, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA. [email protected], [email protected].
(Date: Received April 1, 2020)
Abstract.

Mantel’s theorem states that every nn-vertex graph with n24+t\lfloor\frac{n^{2}}{4}\rfloor+t edges, where t>0t>0, contains a triangle. The problem of determining the minimum number of triangles in such a graph is usually referred to as the Erdős-Rademacher problem. Lovász and Simonovits proved that there are at least tn/2t\lfloor n/2\rfloor triangles in each of those graphs. Katona and Xiao considered the same problem under the additional condition that there are no s1s-1 vertices covering all triangles. They settled the case t=1t=1 and s=2s=2. Solving their conjecture, we determine the minimum number of triangles for every fixed pair of ss and tt, when nn is sufficiently large. Additionally, solving another conjecture of Katona and Xiao, we extend the theory for considering cliques instead of triangles.

1991 Mathematics Subject Classification:
05C35
Research of the first author is partially supported by NSF Grant DMS-1764123, Arnold O. Beckman Research Award (UIUC Campus Research Board RB 18132) and the Langan Scholar Fund (UIUC)

1. Introduction

A classical result in extremal combinatorics is Mantel’s theorem [11]. It says that every nn-vertex graph with at least n24+1\lfloor\frac{n^{2}}{4}\rfloor+1 edges contains a triangle. There have been various extensions of Mantel’s theorem. Turán [13] generalized it for cliques: Every nn-vertex graph with at least tr(n)+1t_{r}(n)+1 edges contains a clique of size r+1r+1, where tr(n)t_{r}(n) is the number of edges of the complete balanced rr-partite graph on nn vertices. A result of Rademacher, see [3], says that every graph on nn vertices with n24+1\lfloor\frac{n^{2}}{4}\rfloor+1 edges contains at least n2\lfloor\frac{n}{2}\rfloor triangles and this is best possible. Erdős [3, 4] conjectured that an nn-vertex graph with n24+t\lfloor\frac{n^{2}}{4}\rfloor+t edges for t<n2t<\frac{n}{2} contains at least tn2t\lfloor\frac{n}{2}\rfloor triangles. This was proven by Lovász and Simonovits [10]. Xiao and Katona [14] proved a stability variant of the Lovász and Simonovits result for t=1t=1: If there is no vertex contained in all triangles, then there are at least n2n-2 triangles in GG. We will prove two extensions of this statement, one verifying a conjecture by Xiao and Katona [14] and the other proving a slight modification of another conjecture by Xiao and Katona [14].

For a graph GG denote e(G)e(G) the number of edges of GG and Tr(G)T_{r}(G) the number of copies of KrK_{r} in GG. A KrK_{r}-covering set in V(G)V(G) is a vertex set that contains at least one vertex of every copy of KrK_{r} in GG. The rr-clique covering number τr(G)\tau_{r}(G) is the size of the smallest KrK_{r}-covering set. A triangle covering set is a K3K_{3}-covering set and the triangle covering number is the 33-clique covering number. Given a vertex partition V(G)=V1V2VrV(G)=V_{1}\cup V_{2}\cup\ldots\cup V_{r}, we call an edge eE(G)e\in E(G) a class-edge if eE(G[Vi])e\in E(G[V_{i}]) for some ii and a cross-edge otherwise. A set of two vertices {x,y}\{x,y\} is a missing cross-edge if xyE(G)xy\not\in E(G) and xVi,yVjx\in V_{i},y\in V_{j} for some iji\neq j. Denote eG(V1,V2,,Vr)e_{G}(V_{1},V_{2},\ldots,V_{r}) the number of cross-edges and eGc(V1,V2,,Vr)e_{G}^{c}(V_{1},V_{2},\ldots,V_{r}) the number of missing cross-edges. We drop the index GG and just write e(V1,V2,,Vr)e(V_{1},V_{2},\ldots,V_{r}) and ec(V1,V2,,Vr)e^{c}(V_{1},V_{2},\ldots,V_{r}), respectively, if GG is clear from context.

Conjecture 1 (Xiao, Katona [14]).

Let t,st,s\in\mathbb{N} such that 0<t<s0<t<s. Suppose the graph GG has nn vertices and n24+t\left\lfloor\frac{n^{2}}{4}\right\rfloor+t edges satisfying τ3(G)s\tau_{3}(G)\geq s and nn(s,t)n\geq n(s,t) is large. Then GG contains at least

(s1)n2+n22(st)\displaystyle(s-1)\left\lfloor\frac{n}{2}\right\rfloor+\left\lceil\frac{n}{2}\right\rceil-2(s-t)

many triangles.

Xiao and Katona [14] proved their conjecture for t=1t=1 and s=2s=2 and that it is best possible. However, their conjecture in full generality holds only up to an additive constant. We will determine precisely the minimum number of triangles a graph on nn vertices with n2/4+t\lfloor n^{2}/4\rfloor+t edges and τ3(G)s\tau_{3}(G)\geq s can have. Depending on s,ts,t and nn one of the following two constructions will be an extremal example.
Construction 1: Let G1G_{1} be the graph on nn vertices, where V(G1)=ABV(G_{1})=A\cup B with

|A|=n2+a,|B|=n2a,\displaystyle|A|=\left\lceil\frac{n}{2}\right\rceil+a,\quad|B|=\left\lfloor\frac{n}{2}\right\rfloor-a,

where aa is chosen later to minimize the number of triangles in this construction. Pick 2(s1)2(s-1) vertices x1,y1,,xs1,ys1x_{1},y_{1},\ldots,x_{s-1},y_{s-1} in AA and two vertices u1,u2u_{1},u_{2} in BB. Add the edges {xi,yi}\{x_{i},y_{i}\} and {u1,u2}\{u_{1},u_{2}\} to K|A|,|B|K_{|A|,|B|} and delete the edges {u1,x1},,{u1,xα}\{u_{1},x_{1}\},\ldots,\{u_{1},x_{\alpha}\}, where α:=sta2𝟙{2n}a\alpha:=s-t-a^{2}-\mathds{1}_{\{2\nmid n\}}a. Then, G1G_{1} has n24+t\lfloor\frac{n^{2}}{4}\rfloor+t edges, triangle covering number τ3(G1)=s\tau_{3}(G_{1})=s and

(1) T3(G1)\displaystyle T_{3}(G_{1}) =(s1)(n2a)+(n2+a)2(sta2𝟙{2n}a).\displaystyle=(s-1)\left(\left\lfloor\frac{n}{2}\right\rfloor-a\right)+\left(\left\lceil\frac{n}{2}\right\rceil+a\right)-2\left(s-t-a^{2}-\mathds{1}_{\{2\nmid n\}}a\right).

Now, choose a0a\geq 0 satisfying α0\alpha\geq 0 and minimizing (1). Note that G1G_{1} is similar to the construction of Xiao and Katona [14] resulting in Conjecture 1, except that the class sizes are different in some cases.
Construction 2: Let G2G_{2} be the graph on nn vertices, where V(G2)=ABV(G_{2})=A\cup B with

|A|=n2+a,|B|=n2a,\displaystyle|A|=\left\lceil\frac{n}{2}\right\rceil+a,\quad|B|=\left\lfloor\frac{n}{2}\right\rfloor-a,

where aa is chosen later to minimize the number of triangles in this construction. Pick 2s2s vertices x1,y1,,xs,ysAx_{1},y_{1},\ldots,x_{s},y_{s}\in A and one vertex uBu\in B. Add the edges {xi,yi}\{x_{i},y_{i}\} to K|A|,|B|K_{|A|,|B|} and delete the edges {u,x1},,{u,xα}\{u,x_{1}\},\ldots,\{u,x_{\alpha}\}, where α:=sta2𝟙{2n}a\alpha:=s-t-a^{2}-\mathds{1}_{\{2\nmid n\}}a. The graph G2G_{2} has n24+t\lfloor\frac{n^{2}}{4}\rfloor+t edges, triangle covering number τ3(G2)=s\tau_{3}(G_{2})=s and

(2) T3(G2)=s(n2a)(sta2𝟙{2n}a).\displaystyle T_{3}(G_{2})=s\left(\left\lfloor\frac{n}{2}\right\rfloor-a\right)-\left(s-t-a^{2}-\mathds{1}_{\{2\nmid n\}}a\right).

Now, choose a0a\geq 0 satisfying α0\alpha\geq 0 and minimizing (2). Our first main result is the following.

Theorem 1.

Let t,st,s\in\mathbb{N} such that 0<t<s0<t<s. Suppose the graph GG has nn vertices and n24+t\left\lfloor\frac{n^{2}}{4}\right\rfloor+t edges, τ3(G)s\tau_{3}(G)\geq s and nn(s,t)n\geq n(s,t) is large. Then GG contains at least min{T3(G1),T3(G2)}\min\{T_{3}(G_{1}),T_{3}(G_{2})\} triangles.

Note that both G1G_{1} and G2G_{2} achieve this minimum for some values of ss and tt. Xiao and Katona [14] also conjectured what the minimum number of copies of Kr+1K_{r+1} in a graph GG on nn vertices with tr1(n)+1t_{r-1}(n)+1 edges and τr+1(G)2\tau_{r+1}(G)\geq 2 could be. They conjectured the following graph to be an extremal example.
Construction 3: The graph G3G_{3} is constructed as follows from the complete rr-partite graph with vertex partition V(G3)=V1V2VrV(G_{3})=V_{1}\cup V_{2}\cup\ldots\cup V_{r} where |V1||V2||Vr||V_{1}|\geq|V_{2}|\geq\ldots\geq|V_{r}| and |V1||Vr|1|V_{1}|-|V_{r}|\leq 1. Let v1,v2V1v_{1},v_{2}\in V_{1} and u1,u2V2u_{1},u_{2}\in V_{2}. Remove the edge {v1,u1}\{v_{1},u_{1}\} and add the edges {v1,v2}\{v_{1},v_{2}\} and {u1,u2}\{u_{1},u_{2}\}. The graph G3G_{3} satisfies τr+1(G3)=2\tau_{r+1}(G_{3})=2, e(G3)=tr(n)+1e(G_{3})=t_{r}(n)+1 and

Tr+1(G3)=(|V1|+|V2|2)i=3r|Vi|.\displaystyle T_{r+1}(G_{3})=(|V_{1}|+|V_{2}|-2)\prod_{i=3}^{r}|V_{i}|.

We will verify Xiao and Katona’s [14] conjecture.

Theorem 2.

Let rr\in\mathbb{N} and nn(r)n\geq n(r) be a sufficiently large integer. If a graph GG on nn vertices has tr(n)+1t_{r}(n)+1 edges and the copies of Kr+1K_{r+1} have an empty intersection, then the number of copies of Kr+1K_{r+1} is at least as many as in G3G_{3}.

The key ingredient for proving Theorems 1 and 2 is the following general structural result.

Theorem 3.

Let r,t,sr,t,s\in\mathbb{N} with 0<t<s0<t<s and let nn(s,t,r)n\geq n(s,t,r) be sufficiently large. Let GG be an nn-vertex graph with tr(n)+tt_{r}(n)+t edges satisfying τr+1(G)s\tau_{r+1}(G)\geq s. Denote \mathcal{H} the family of graphs HH on nn vertices with tr(n)+tt_{r}(n)+t edges such that there exists a vertex partition V(H)=V1V2VrV(H)=V_{1}\cup V_{2}\cup\ldots\cup V_{r} satisfying

  • the class-edges form a matching of size ss,

  • eHc(V1,,Vr)ste_{H}^{c}(V_{1},\ldots,V_{r})\leq s-t,

  • all missing cross-edges are incident to one of the class-edges,

  • if the class-edges are not all in the same class, then each missing cross-edge is incident to class-edges on both endpoints.

Then Tr+1(G)Tr+1(H)T_{r+1}(G)\geq T_{r+1}(H) for some HH\in\mathcal{H}.

We will observe (Lemma 4) that |Vi|=n/r+O(1)|V_{i}|=n/r+O(1) for all 1ir1\leq i\leq r. Hence, the family \mathcal{H} is small. An optimization argument will reduce the family \mathcal{H} to G1G_{1} and G2G_{2} in the case r=3r=3, and to G3G_{3} in the case r>3,s=2,t=1r>3,s=2,t=1. A similar cleaning argument could potentially be done in the general case, however the computational effort needed seems not worth the outcome. Also, it is likely that for some parameters r,t,sr,t,s, the extremal family realizing the minimum number of cliques of size r+1r+1 contains several graphs.

For tst\geq s we know from a result of Erdős [5] that the number of cliques of size r+1r+1 is at least (1+o(1))t(nr)r1(1+o(1))t\left(\frac{n}{r}\right)^{r-1} and the graph G4G_{4} constructed by adding a matching of size tt to one of the classes of the complete balanced rr-partite graph on nn vertices satisfies τr+1(G4)=t\tau_{r+1}(G_{4})=t and Tr+1(G4)=(1+o(1))t(nr)r1T_{r+1}(G_{4})=(1+o(1))t\left(\frac{n}{r}\right)^{r-1}. Hence this problem is interesting only when 0<t<s0<t<s.

Our result can be extended to consider ss and tt as functions of nn. In particular, in the triangle case r=3r=3, our proof allows ss to be linear in nn. However, the methods we use, especially our main tool Theorem 6, will not give us the entire range of ss where our theorem should hold. Thus, we do not attempt to optimize our proof with respect to the dependency of nn and ss.

Our main tool is a stability-supersaturation result by Balogh, Bushaw, Collares, Liu, Morris and Sharifzadeh [1], which extends on the Erdős-Simonovits stability theorem [6] and Füredi’s [7] proof’s of it.

We would like to point out that Liu and Mubayi [9] independently obtained similar results to ours.

Our paper is organized as follows. In Section 2 we prove Theorem 3, in Section 3 we conclude Theorem 1 and in Section 4 we conclude Theorem 2.

2. Proof of Theorem 3

The well-known Turán’s theorem [13] determines the maximum number of edges in a Kr+1K_{r+1}-free graph to be the number of edges tr(n)t_{r}(n) of the complete balanced rr-partite graph. We will make use of the following bound on tr(n)t_{r}(n).

(3) n22(11r)r2tr(n)n22(11r).\displaystyle\frac{n^{2}}{2}\left(1-\frac{1}{r}\right)-\frac{r}{2}\leq t_{r}(n)\leq\frac{n^{2}}{2}\left(1-\frac{1}{r}\right).

The classical Erdős-Simonovits stability theorem [6] asserts that any nn-vertex Kr+1K_{r+1}-free graph with almost tr(n)t_{r}(n) edges is close to the complete balanced rr-partite graph. There are many different quantitative versions [2, 8, 12] of this. A standard calculation shows that nn-vertex graphs with almost tr(n)t_{r}(n) edges and a vertex partition with few class-edges need to have almost balanced class sizes.

Lemma 4.

Let s,ns,n\in\mathbb{N}, tt\in\mathbb{Z} and GG a graph on nn vertices and tr(n)+tt_{r}(n)+t edges with vertex partition V(G)=V1V2VrV(G)=V_{1}\cup V_{2}\cup\ldots\cup V_{r} such that the number of class-edges is ss. Then |Vi|=nr+O(1)|V_{i}|=\frac{n}{r}+O(1) for all 1ir1\leq i\leq r.

Proof.

Without loss of generality we can assume that |V1||V2||Vr||V_{1}|\geq|V_{2}|\geq\ldots\geq|V_{r}|. Now, let x=|V1|nrx=|V_{1}|-\frac{n}{r}. Then,

i=1r|Vi|2\displaystyle\sum_{i=1}^{r}|V_{i}|^{2} =(nr+x)2+i=2r|Vi|2(nr+x)2+(i=2r|Vi|)2r1\displaystyle=\left(\frac{n}{r}+x\right)^{2}+\sum_{i=2}^{r}|V_{i}|^{2}\geq\left(\frac{n}{r}+x\right)^{2}+\frac{\left(\sum_{i=2}^{r}|V_{i}|\right)^{2}}{r-1}
(4) (nr+x)2+(n(11r)x)2r1n2r+x2.\displaystyle\geq\left(\frac{n}{r}+x\right)^{2}+\frac{\left(n\left(1-\frac{1}{r}\right)-x\right)^{2}}{r-1}\geq\frac{n^{2}}{r}+x^{2}.

Thus, we can conclude

tr(n)+t=e(G)=i=1re(G[Vi])+e(V1,V2,,Vr)s+1i<jr|Vi||Vj|\displaystyle\quad\ t_{r}(n)+t=e(G)=\sum_{i=1}^{r}e(G[V_{i}])+e(V_{1},V_{2},\ldots,V_{r})\leq s+\sum_{1\leq i<j\leq r}|V_{i}||V_{j}|
=s+12(n2i=1r|Vi|2)s+n22(11r)x22tr(n)+r2+sx2,\displaystyle=s+\frac{1}{2}\left(n^{2}-\sum_{i=1}^{r}|V_{i}|^{2}\right)\leq s+\frac{n^{2}}{2}\left(1-\frac{1}{r}\right)-\frac{x^{2}}{2}\leq t_{r}(n)+\frac{r}{2}+s-x^{2},

implying x=O(1)x=O(1). Note that in the second to last inequality we used (2) and in the last inequality we used (3). We conclude |V1|=nr+O(1)|V_{1}|=\frac{n}{r}+O(1). Similarly, we can conclude |Vr|=nr+O(1)|V_{r}|=\frac{n}{r}+O(1). ∎

Lemma 5.

Let s,ns,n\in\mathbb{N}, tt\in\mathbb{Z} and GG a graph on nn vertices and tr(n)+tt_{r}(n)+t edges with vertex partition V(G)=V1V2VrV(G)=V_{1}\cup V_{2}\cup\ldots\cup V_{r} such that the number of class-edges is ss. Then

Tr+1(G)=(1+o(1))s(nr)r1.\displaystyle T_{r+1}(G)=(1+o(1))s\left(\frac{n}{r}\right)^{r-1}.
Proof.

Any copy of Kr+1K_{r+1} needs to contain at least one class-edge. There are exactly (nr+O(1))r1(\frac{n}{r}+O(1))^{r-1} copies of Kr+1K_{r+1} containing one given class-edge but no others, since the other r1r-1 vertices need to be chosen from different classes. Since the number of missing cross-edges is O(1)O(1), they do not have an impact in this counting. There are ss class-edges and thus we conclude that the number of copies of Kr+1K_{r+1} containing exactly one class-edge is

(1+o(1))s(nr)r1.\displaystyle(1+o(1))s\left(\frac{n}{r}\right)^{r-1}.

The number of copies containing at least two class-edges is O(nr2)O(n^{r-2}), because there are O(1)O(1) ways to chose three vertices among the vertices being incident to the class-edges and at most O(nr2)O(n^{r-2}) ways to chose the remaining r2r-2 vertices. We conclude

Tr+1(G)=(1+o(1))s(nr)r1.T_{r+1}(G)=(1+o(1))s\left(\frac{n}{r}\right)^{r-1}.\qed

We say that a graph GG is xx-far from being rr-partite if χ(G)>r\chi(G^{\prime})>r for every subgraph GGG^{\prime}\subset G with e(G)>e(G)xe(G^{\prime})>e(G)-x. Balogh, Bushaw, Collares, Liu, Morris and Sharifzadeh [1] proved that graphs which are xx-far from being rr-partite, contain many cliques of size r+1r+1.

Theorem 6 ([1]).

For every n,x,rn,x,r\in\mathbb{N}, the following holds. Every graph GG on nn vertices which is xx-far from being rr-partite contains at least

nr1e2rr!(e(G)+x(11r)n22)\displaystyle\frac{n^{r-1}}{e^{2r}r!}\left(e(G)+x-\left(1-\frac{1}{r}\right)\frac{n^{2}}{2}\right)

copies of Kr+1K_{r+1}.

With this tool in hand, we now prove Theorem 3.

Proof of Theorem 3.

Let GG be a graph on nn vertices and tr(n)+tt_{r}(n)+t edges such that τr+1(G)s\tau_{r+1}(G)\geq s and nn is large enough. If GG is 2sr!e2r2sr!e^{2r}-far from being rr-partite, then by applying Theorem 6, the number of copies of Kr+1K_{r+1} in GG is at least

Tr+1(G)nr1r!e2r(tr(n)+t+2sr!e2r(11r)n22)snr1,\displaystyle T_{r+1}(G)\geq\frac{n^{r-1}}{r!e^{2r}}\left(t_{r}(n)+t+2sr!e^{2r}-\left(1-\frac{1}{r}\right)\frac{n^{2}}{2}\right)\geq sn^{r-1},

where we used (3) to lower bound tr(n)t_{r}(n). By Lemma 5, Tr+1(H)=(1+o(1))s(nr)r1T_{r+1}(H)=(1+o(1))s(\frac{n}{r})^{r-1} for every HH\in\mathcal{H}, and thus Tr+1(H)Tr+1(G)T_{r+1}(H)\leq T_{r+1}(G). Hence, we can assume that GG is not 2sr!e2r2sr!e^{2r}-far from being rr-partite. Then there exists a subgraph GGG^{\prime}\subset G with χ(G)r\chi(G^{\prime})\leq r and

e(G)>e(G)2sr!e2rs>tr(n)2sr!e2r.\displaystyle e(G^{\prime})>e(G)-2sr!e^{2r}s>t_{r}(n)-2sr!e^{2r}.

Let V(G)=V1V2VrV(G^{\prime})=V_{1}\cup V_{2}\cup\ldots\cup V_{r} be a partition such that V1,V2,,VrV_{1},V_{2},\ldots,V_{r} are independent sets in GG^{\prime}. By Lemma 4, the class sizes are roughly balanced, that is |Vi|=nr+O(1)|V_{i}|=\frac{n}{r}+O(1) for all 1ir1\leq i\leq r.

Since GGG^{\prime}\subset G, and e(G)>e(G)2sr!e2re(G^{\prime})>e(G)-2sr!e^{2r}, the vertex partition V(G)=V1V2VrV(G)=V_{1}\cup V_{2}\cup\ldots\cup V_{r} contains at most 2sr!e2r2sr!e^{2r} class-edges in GG. If the number of class-edges in GG is less than ss, then there is a Kr+1K_{r+1}-covering set of size at most s1s-1, contradicting τr+1(G)s\tau_{r+1}(G)\geq s. If the number of class-edges is more than ss, then by Lemma 5

Tr+1(G)=(1+o(1))(s+1)(nr)r1(1+o(1))s(nr)r1=Tr+1(H)\displaystyle T_{r+1}(G)=(1+o(1))(s+1)\left(\frac{n}{r}\right)^{r-1}\geq(1+o(1))s\left(\frac{n}{r}\right)^{r-1}=T_{r+1}(H)

for every HH\in\mathcal{H}. Therefore, we can assume that the number of class-edges in GG is exactly ss. These ss edges need to form a matching, otherwise there is a Kr+1K_{r+1}-covering set of size s1s-1, contradicting τr+1(G)s\tau_{r+1}(G)\geq s. Further, ec(V1,,Vr)ste^{c}(V_{1},\ldots,V_{r})\leq s-t, because otherwise

e(G)\displaystyle e(G) =e(V1,V2,,Vr)+i=1re(G[Vi])tr+1(n)ec(V1,V2,,Vr)+s\displaystyle=e(V_{1},V_{2},\ldots,V_{r})+\sum_{i=1}^{r}e(G[V_{i}])\leq t_{r+1}(n)-e^{c}(V_{1},V_{2},\ldots,V_{r})+s
<tr+1(n)+t=e(G).\displaystyle<t_{r+1}(n)+t=e(G).

Next, we assume that there is a missing cross-edge uvE(G)uv\not\in E(G) not incident to any of the class-edges. Take a cross-edge xyE(G)xy\in E(G) which is incident to exactly one class-edge. We will show that Tr+1(G+uvxy)<Tr+1(G)T_{r+1}(G+uv-xy)<T_{r+1}(G). Adding uvuv to GG increases the number of cliques of size r+1r+1 by at most snr3sn^{r-3}, because for each of the ss class-edges ee the number of cliques of size r+1r+1 containing ee and uvuv is increased by at most nr3n^{r-3}. Removing xyxy from G+uvG+uv decreases the number of cliques of size r+1r+1 by at least

min(|Vi|s)r2(nr+O(1)s)r2=Ω(nr2),\displaystyle\min(|V_{i}|-s)^{r-2}\geq\left(\frac{n}{r}+O(1)-s\right)^{r-2}=\Omega(n^{r-2}),

where we used that in each class ViV_{i} the number of vertices incident to no missing cross-edge is at least |Vi|s|V_{i}|-s. Thus, performing this edge flip decreases the number of cliques of size r+1r+1, i.e. Tr+1(G+uvxy)<Tr+1(G)T_{r+1}(G+uv-xy)<T_{r+1}(G). We repeat this process until we end up with a graph G¯\bar{G} such that Tr+1(G¯)Tr+1(G)T_{r+1}(\bar{G})\leq T_{r+1}(G) and every missing cross-edge is incident to a class-edge.
If G¯\bar{G} has all class-edges in the same class, then G¯\bar{G}\in\mathcal{H}. Thus, we can assume that G¯\bar{G} has the property that not all class-edges are in the same class and there is a missing cross-edge not adjacent to class-edges on both ends. Denote GG^{\prime} the graph constructed from G¯\bar{G} by adding all missing cross-edges. This graph satisfies

Tr+1(G)\displaystyle T_{r+1}(G^{\prime}) Tr+1(G¯)+(2eG¯c(V1,V2,,Vr)1)(nr+O(1))r2\displaystyle\leq T_{r+1}(\bar{G})+\left(2e_{\bar{G}}^{c}(V_{1},V_{2},\ldots,V_{r})-1\right)\left(\frac{n}{r}+O(1)\right)^{r-2}
+seG¯c(V1,V2,,Vr)nr3\displaystyle+se_{\bar{G}}^{c}(V_{1},V_{2},\ldots,V_{r})n^{r-3}
Tr+1(G)+(1+o(1))(2eG¯c(V1,V2,,Vr)1)(nr)r2,\displaystyle\leq T_{r+1}(G)+(1+o(1))\left(2e_{\bar{G}}^{c}(V_{1},V_{2},\ldots,V_{r})-1\right)\left(\frac{n}{r}\right)^{r-2},

because for each class-edge ee and each added edge ee^{\prime}, the number of cliques of size r+1r+1 containing ee and ee^{\prime} increases by at most (nr+O(1))r2\left(\frac{n}{r}+O(1)\right)^{r-2} if ee and ee^{\prime} share a vertex and by at most nr3n^{r-3} if ee and ee^{\prime} are disjoint.

There is a set MM of eG¯c(V1,V2,,Vr)s1e_{\bar{G}}^{c}(V_{1},V_{2},\ldots,V_{r})\leq s-1 cross-edges such that each eMe\in M is incident to class-edges on both endpoints and every class-edge is incident to edges from MM on at most one endpoint. Denote HH the graph constructed from GG^{\prime} by removing those edges. This graph satisfies HH\in\mathcal{H} and

Tr+1(H)\displaystyle T_{r+1}(H) Tr+1(G)(2+o(1))eG¯c(V1,V2,,Vr)(nr+O(1)s)r2\displaystyle\leq T_{r+1}(G^{\prime})-(2+o(1))e_{\bar{G}}^{c}(V_{1},V_{2},\ldots,V_{r})\left(\frac{n}{r}+O(1)-s\right)^{r-2}
Tr+1(G),\displaystyle\leq T_{r+1}(G),

completing the proof. ∎

3. Proof of Theorem 1

Proof of Theorem 1.

By Theorem 3, we can assume that GG has a vertex partition V(G)=ABV(G)=A\cup B where |A||B||A|\geq|B| with exactly ss class-edges forming a matching. Let |A|=n/2+a,|B|=n/2a|A|=\lceil n/2\rceil+a,|B|=\lfloor n/2\rfloor-a for some a0a\geq 0. We have

n24+t\displaystyle\left\lfloor\frac{n^{2}}{4}\right\rfloor+t =e(G)=|A||B|ec(A,B)+s\displaystyle=e(G)=|A||B|-e^{c}(A,B)+s
=(n2+a)(n2a)ec(A,B)+s,\displaystyle=\left(\left\lceil\frac{n}{2}\right\rceil+a\right)\left(\left\lfloor\frac{n}{2}\right\rfloor-a\right)-e^{c}(A,B)+s,

and thus the number of missing cross-edges is

ec(A,B)=sta2𝟙{2n}.\displaystyle e^{c}(A,B)=s-t-a^{2}-\mathds{1}_{\{2\nmid n\}}.

Let MAM_{A} be the matching of class-edges inside AA and MBM_{B} be the matching of class-edges inside BB. Denote GG^{\prime} the graph constructed from GG by adding all missing cross-edges ec(A,B)e^{c}(A,B). The number of triangles in this graph is |MA||B|+|MB||A||M_{A}||B|+|M_{B}||A|.
Case 1: |MB|>0|M_{B}|>0.
The number of triangles in GG is at least |MA||B|+|MB||A|2ec(A,B)|M_{A}||B|+|M_{B}||A|-2e^{c}(A,B), because each missing cross-edge is in at most two triangles in GG^{\prime}. Since |A||B||A|\geq|B|, we get

T3(G)\displaystyle T_{3}(G) (s1)(n2a)+(n2+a)2ec(A,B)\displaystyle\geq(s-1)\left(\left\lfloor\frac{n}{2}\right\rfloor-a\right)+\left(\left\lceil\frac{n}{2}\right\rceil+a\right)-2e^{c}(A,B)
=(s1)(n2a)+(n2+a)2(sta2𝟙{2n}a)\displaystyle=(s-1)\left(\left\lfloor\frac{n}{2}\right\rfloor-a\right)+\left(\left\lceil\frac{n}{2}\right\rceil+a\right)-2\left(s-t-a^{2}-\mathds{1}_{\{2\nmid n\}}a\right)
T3(G1).\displaystyle\geq T_{3}(G_{1}).

Case 2: |MB|=0|M_{B}|=0.
The number of triangles in GG is at least |MA||B|ec(A,B)|M_{A}||B|-e^{c}(A,B), because each missing cross-edge is in at most one triangle in GG^{\prime}. Therefore

T3(G)\displaystyle T_{3}(G) |MA||B|ec(A,B)=s(n2a)(sta2𝟙{2n}a)\displaystyle\geq|M_{A}||B|-e^{c}(A,B)=s\left(\left\lfloor\frac{n}{2}\right\rfloor-a\right)-\left(s-t-a^{2}-\mathds{1}_{\{2\nmid n\}a}\right)
T3(G2).\displaystyle\geq T_{3}(G_{2}).\hfill\qed

4. Proof of Theorem 2

Let =nmodr\ell=n\mod r and mm be an integer such that n=rm+n=rm+\ell. The number of cliques of size r+1r+1 in G3G_{3} is

Tr+1(G3)\displaystyle T_{r+1}(G_{3}) =(|V1|+|V2|2)i=3r|Vi|={(2m2)mr2, iff =0,(2m1)mr2, iff =1,2m(m+1)2mr, otherwise.\displaystyle=(|V_{1}|+|V_{2}|-2)\prod_{i=3}^{r}|V_{i}|=\begin{cases}\left(2m-2\right)m^{r-2}&,\text{ iff }\ell=0,\\ \left(2m-1\right)m^{r-2}&,\text{ iff }\ell=1,\\ 2m(m+1)^{\ell-2}m^{r-\ell}&,\text{ otherwise.}\\ \end{cases}
Proof of Theorem 2.

Let GG be an nn-vertex graph with tr(n)+1t_{r}(n)+1 edges satisfying τr+1(G)2\tau_{r+1}(G)\geq 2. By applying Theorem 3 for s=2s=2 and t=1t=1, we can assume without loss of generality that GG\in\mathcal{H} and thus has a vertex partition V(G)=V1V2VrV(G)=V_{1}\cup V_{2}\cup\ldots\cup V_{r} where |V1||V2||Vr||V_{1}|\geq|V_{2}|\geq\ldots\geq|V_{r}| with exactly two class-edges e1,e2e_{1},e_{2} and at most one missing cross-edge. We have

(5) 1i<jr|Vi||Vj|ec(V1,V2,,Vr)+2=e(G)=tr(n)+1.\displaystyle\sum_{1\leq i<j\leq r}|V_{i}||V_{j}|-e^{c}(V_{1},V_{2},\ldots,V_{r})+2=e(G)=t_{r}(n)+1.

This means we have two cases:

ec(V1,V2,,Vr)=0orec(V1,V2,,Vr)=1.\displaystyle e^{c}(V_{1},V_{2},\ldots,V_{r})=0\quad\quad\quad\text{or}\quad\quad\quad e^{c}(V_{1},V_{2},\ldots,V_{r})=1.

Case 1: ec(V1,V2,,Vr)=0e^{c}(V_{1},V_{2},\ldots,V_{r})=0.
In this case

1i<jr|Vi||Vj|=tr(n)1\displaystyle\sum_{1\leq i<j\leq r}|V_{i}||V_{j}|=t_{r}(n)-1

by (5). Note that |V1||Vr|=2|V_{1}|-|V_{r}|=2, because if |V1||Vr|3|V_{1}|-|V_{r}|\geq 3, then moving one vertex from V1V_{1} to VrV_{r} increases the sum 1i<jr|Vi||Vj|\sum_{1\leq i<j\leq r}|V_{i}||V_{j}| by at least two and thus

(6) 1i<jr|Vi||Vj|tr(n)2,\displaystyle\sum_{1\leq i<j\leq r}|V_{i}||V_{j}|\leq t_{r}(n)-2,

a contradiction. If |V1||Vr|1,|V_{1}|-|V_{r}|\leq 1, then 1i<jr|Vi||Vj|=tr(n)\sum_{1\leq i<j\leq r}|V_{i}||V_{j}|=t_{r}(n), giving a contradiction as well. Thus, we have |V1||Vr|=2|V_{1}|-|V_{r}|=2. Similarly, |V2||Vr1|1|V_{2}|-|V_{r-1}|\leq 1, as otherwise (6) holds. Conclusively, there are only the following types of combination of class sizes. We list them depending on =nmodr\ell=n\mod r. If {0,1}\ell\in\{0,1\}, then

  • |V1|==|V+1|=m+1|V_{1}|=\ldots=|V_{\ell+1}|=m+1, |V+2|==|Vr1|=m|V_{\ell+2}|=\ldots=|V_{r-1}|=m, |Vr|=m1|V_{r}|=m-1.

If =2\ell=2, then

  • Type 1: |V1|=m+2|V_{1}|=m+2, |V2|==|Vr|=m|V_{2}|=\ldots=|V_{r}|=m,

  • Type 2: |V1|=|V2|=|V3|=m+1|V_{1}|=|V_{2}|=|V_{3}|=m+1, |V4|==|Vr1|=m|V_{4}|=\ldots=|V_{r-1}|=m, |Vr|=m1|V_{r}|=m-1.

If 3r33\leq\ell\leq r-3, then

  • Type 1: |V1|=m+2|V_{1}|=m+2, |V2|==|V1|=m+1|V_{2}|=\ldots=|V_{\ell-1}|=m+1, |V|==|Vr|=m|V_{\ell}|=\ldots=|V_{r}|=m,

  • Type 2: |V1|==|V+1|=m+1|V_{1}|=\ldots=|V_{\ell+1}|=m+1, |V+2|==|Vr1|=m|V_{\ell+2}|=\ldots=|V_{r-1}|=m, |Vr|=m1|V_{r}|=m-1.

If =r2\ell=r-2, then

  • Type 1: |V1|=m+2|V_{1}|=m+2, |V2|==|Vr3|=m+1|V_{2}|=\ldots=|V_{r-3}|=m+1, |Vr2|==|Vr|=m|V_{r-2}|=\ldots=|V_{r}|=m,

  • Type 2: |V1|==|Vr1|=m+1|V_{1}|=\ldots=|V_{r-1}|=m+1, |Vr|=m1|V_{r}|=m-1.

If =r1\ell=r-1, then

  • |V1|=m+2|V_{1}|=m+2, |V2|==|Vr2|=m+1|V_{2}|=\ldots=|V_{r-2}|=m+1, |Vr1|=|Vr|=m|V_{r-1}|=|V_{r}|=m.

Denote the two class-edges e1,e2e_{1},e_{2}. Let 1αβr1\leq\alpha\leq\beta\leq r such that e1Vαe_{1}\subset V_{\alpha} and e2Vβe_{2}\subset V_{\beta}. The number of copies of Kr+1K_{r+1} containing e1e_{1} but not e2e_{2} is iα|Vi|\prod_{i\neq\alpha}|V_{i}|. Similarly, the number of copies of Kr+1K_{r+1} containing e2e_{2} but not e1e_{1} is iβ|Vi|\prod_{i\neq\beta}|V_{i}|. Thus, the total number of copies of Kr+1K_{r+1} in GG is at least

(7) Tr+1(G)iα|Vi|+iβ|Vi|2i=2r|Vi|.\displaystyle T_{r+1}(G)\geq\prod_{i\neq\alpha}|V_{i}|+\prod_{i\neq\beta}|V_{i}|\geq 2\prod_{i=2}^{r}|V_{i}|.

We will now check that for any possible combination of class sizes the right hand side in (7) is at least Tr+1(G3)T_{r+1}(G_{3}). We distinguish cases depending on =nmodr\ell=n\mod r. If =0\ell=0, then

2i=2r|Vi|=2mr2(m1)(2m2)mr2=Tr+1(G3).\displaystyle 2\prod_{i=2}^{r}|V_{i}|=2m^{r-2}(m-1)\geq(2m-2)m^{r-2}=T_{r+1}(G_{3}).

If =1\ell=1, then

2i=2r|Vi|=2(m+1)mr3(m1)(2m1)mr2=Tr+1(G3).\displaystyle 2\prod_{i=2}^{r}|V_{i}|=2(m+1)m^{r-3}(m-1)\geq(2m-1)m^{r-2}=T_{r+1}(G_{3}).

If =2\ell=2, then

2i=2r|Vi|={2mr1,for Type 12(m+1)2mr4(m1),for Type 22mr1=Tr+1(G3).\displaystyle 2\prod_{i=2}^{r}|V_{i}|=\begin{cases}2m^{r-1},&\text{for Type 1}\\ 2(m+1)^{2}m^{r-4}(m-1),&\text{for Type 2}\end{cases}\geq 2m^{r-1}=T_{r+1}(G_{3}).

If 3r33\leq\ell\leq r-3, then

2i=2r|Vi|\displaystyle 2\prod_{i=2}^{r}|V_{i}| ={2(m+1)2mr+1,for Type 12(m+1)mr2(m1),for Type 22mr+1(m+1)2\displaystyle=\begin{cases}2(m+1)^{\ell-2}m^{r-\ell+1},&\text{for Type 1}\\ 2(m+1)^{\ell}m^{r-\ell-2}(m-1),&\text{for Type 2}\end{cases}\geq 2m^{r-\ell+1}(m+1)^{\ell-2}
=Tr+1(G3).\displaystyle=T_{r+1}(G_{3}).

If =r2\ell=r-2, then

2i=2r|Vi|\displaystyle 2\prod_{i=2}^{r}|V_{i}| ={2(m+1)r4m3,for Type 12(m+1)r2(m1),for Type 22m3(m+1)r4\displaystyle=\begin{cases}2(m+1)^{r-4}m^{3},&\text{for Type 1}\\ 2(m+1)^{r-2}(m-1),&\text{for Type 2}\end{cases}\geq 2m^{3}(m+1)^{r-4}
=Tr+1(G3).\displaystyle=T_{r+1}(G_{3}).

Finally, if =r1\ell=r-1, then

2i=2r|Vi|=2(m+1)r3m2=Tr+1(G3).\displaystyle 2\prod_{i=2}^{r}|V_{i}|=2(m+1)^{r-3}m^{2}=T_{r+1}(G_{3}).

Thus, in Case 1, Tr+1(G)Tr+1(G3)T_{r+1}(G)\geq T_{r+1}(G_{3}).

Case 2: ec(V1,V2,,Vr)=1e^{c}(V_{1},V_{2},\ldots,V_{r})=1.
Observe that in this case the class sizes of GG and G3G_{3} are the same. Further, the number of missing cross-edges is exactly one. Now, we distinguish two cases depending on where e1e_{1} and e2e_{2} are. Case 22a is when both class-edges are inside the same class. In Case 22b, e1e_{1} and e2e_{2} are in different classes.
Case 2a: e1,e2Vαe_{1},e_{2}\subseteq V_{\alpha} for some 1αr.1\leq\alpha\leq r.
The missing edge is incident to e1e_{1} or e2e_{2}. The second endpoint of the missing edge is in VβV_{\beta} for some βα\beta\neq\alpha. Then

Tr+1(G)\displaystyle T_{r+1}(G) =2iα|Vi|iα,β|Vi|=(2|Vβ|1)iα,β|Vi|\displaystyle=2\prod_{i\neq\alpha}|V_{i}|-\prod_{i\neq\alpha,\beta}|V_{i}|=(2|V_{\beta}|-1)\prod_{i\neq\alpha,\beta}|V_{i}|
(|Vα|+|Vβ|2)iα,β|Vi|(|V1|+|V2|2)i=3r|Vi|=Tr+1(G3),\displaystyle\geq(|V_{\alpha}|+|V_{\beta}|-2)\prod_{i\neq\alpha,\beta}|V_{i}|\geq(|V_{1}|+|V_{2}|-2)\prod_{i=3}^{r}|V_{i}|=T_{r+1}(G_{3}),

where the sum of the factors are the same in the two products, hence the product is larger when the factors are ‘closer’ to each other.
Case 2b: e1Vα,e2Vβe_{1}\subseteq V_{\alpha},e_{2}\subseteq V_{\beta} for some αβ\alpha\neq\beta.
Since the missing edge is incident to both class-edges,

Tr+1(G)\displaystyle T_{r+1}(G) =(|Vβ|1)iα,β|Vi|+(|Vα|1)iα,β|Vi|\displaystyle=(|V_{\beta}|-1)\prod_{i\neq\alpha,\beta}|V_{i}|+(|V_{\alpha}|-1)\prod_{i\neq\alpha,\beta}|V_{i}|
=(|Vα|+|Vβ|2)iα,β|Vi|(|V1|+|V2|2)i=3r|Vi|=Tr+1(G3),\displaystyle=(|V_{\alpha}|+|V_{\beta}|-2)\prod_{i\neq\alpha,\beta}|V_{i}|\geq(|V_{1}|+|V_{2}|-2)\prod_{i=3}^{r}|V_{i}|=T_{r+1}(G_{3}),

where the last inequality holds by the same reasoning as in Case 22a.

Acknowledgement

We would like to thank an anonymous referee for helpful comments, in particular, for pointing out an error in an earlier version of this paper.

References

  • [1] J. Balogh, N. Bushaw, M. Collares, H. Liu, R. Morris, and M. Sharifzadeh. The typical structure of graphs with no large cliques. Combinatorica, 37(4):617–632, 2017.
  • [2] J. Balogh, F. C. Clemen, M. Lavrov, B. Lidický, and F. Pfender. Making Kr+1{K}_{r+1}-free graphs rr-partite. Combinatorics, Probability and Computing, pages 1–10, 2020.
  • [3] P. Erdős. Some theorems on graphs. Riveon Lematematika, 9:13–17, 1955.
  • [4] P. Erdős. On a theorem of Rademacher-Turán. Illinois J. Math., 6:122–127, 1962.
  • [5] P. Erdős. On the number of complete subgraphs and circuits contained in graphs. Časopis Pěst. Mat., 94:290–296, 1969.
  • [6] P. Erdős and M. Simonovits. A limit theorem in graph theory. Studia Sci. Math. Hungar, 1:51–57, 1966.
  • [7] Z. Füredi. A proof of the stability of extremal graphs, Simonovits’ stability from Szemerédi’s regularity. J. Combin. Theory Ser. B, 115:66–71, 2015.
  • [8] D. Korándi, A. Roberts, and A. Scott. Exact stability for Turán’s theorem. arXiv:2004.10685, 2020.
  • [9] X. Liu and D. Mubayi. On a generalized Erdős-Rademacher problem. arXiv:2005.07224, 2020.
  • [10] L. Lovász and M. Simonovits. On the number of complete subgraphs of a graph. II. In Studies in pure mathematics, pages 459–495. Birkhäuser, Basel, 1983.
  • [11] W. Mantel. Problem 28. Wiskundige Opgaven, 10:60–61, 1907.
  • [12] A. Roberts and A. Scott. Stability results for graphs with a critical edge. arXiv:1610.08389, 2018.
  • [13] P. Turán. Eine Extremalaufgabe aus der Graphentheorie. Mat. Fiz. Lapok, 48:436–452, 1941.
  • [14] C. Xiao and G. O. H. Katona. The number of triangles is more when they have no common vertex. arXiv:2003.04450, 2020.