Erdős-Rogers functions for arbitrary pairs of graphs

Abstract

Let $f_{F,G}(n)$ be the largest size of an induced $F$-free subgraph that every $n$-vertex $G$-free graph is guaranteed to contain. We prove that for any triangle-free graph $F$,

$$f_{F,K_{3}}(n)=f_{K_{2},K_{3}}(n)^{1+o(1)}=n^{\frac{1}{2}+o(1)}.$$

Along the way we give a slight improvement of a construction of Erdős-Frankl-Rödl for the Brown-Erdős-Sós $(3r-3,3)$-problem when $r$ is large.

In contrast to our result for $K_{3}$, for any $K_{4}$-free graph $F$ containing a cycle, we prove there exists $c_{F}>0$ such that

$$f_{F,K_{4}}(n)>f_{K_{2},K_{4}}(n)^{1+c_{F}}=n^{\frac{1}{3}+c_{F}+o(1)}.$$

For every graph $G$, we prove that there exists $\varepsilon_{G}>0$ such that whenever $F$ is a non-empty graph such that $G$ is not contained in any blowup of $F$, then $f_{F,G}(n)=O(n^{1-\varepsilon_{G}})$. On the other hand, for graph $G$ that is not a clique, and every $\varepsilon>0$, we exhibit a $G$-free graph $F$ such that $f_{F,G}(n)=\Omega(n^{1-\varepsilon})$.

1 Introduction

Say that a graph is $F$-free if it contains no subgraph isomorphic to $F$. Denote by $f_{F,G}(n)$ the maximum $m$ such that every $n$-vertex $G$-free graph contains an induced $F$-free subgraph on at least $m$ vertices. Hence the assertion $f_{F,G}(n)<b$ means that there exists an $n$ vertex $G$-free graph $H$ such that every vertex subset of $H$ of size $b$ contains a copy of $F$. The case $F=K_{s}$ and $G=K_{t}$ is the Erdős-Rogers [7] function $f_{s,t}(n)$. Classical results in Ramsey Theory [1, 10] give $r(3,t)=\Theta(t^{2}/\log t)$, which shows $f_{K_{2},K_{3}}(n)=\Theta(\sqrt{n\log n})$. We prove that roughly the same holds for $f_{F,K_{3}}(n)$ for any triangle-free graph $F$:

Theorem 1.

For any triangle-free graph $F$ containing at least one edge,

$$f_{F,K_{3}}(n)=n^{\frac{1}{2}+O(\sqrt{\frac{\log\log n}{\log n}})}=f_{K_{2},K_{3}}(n)^{1+o(1)}.$$

Our bound in Theorem 1 is much larger than $f_{K_{2},K_{3}}(n)=\Theta(\sqrt{n\log n})$, and therefore the following problem seems natural.

Problem 1.

Find a triangle free $F$ for which $f_{F,K_{3}}(n)/f_{K_{2},K_{3}}(n)\rightarrow\infty$.

A large pseudorandom triangle free graph with many edges seems an obvious choice for $F$ in Problem 1. Perhaps the simpler $F=K_{t,t}$ is another example. More generally, for each $s\geq 3$, one can ask whether there exists a $K_{s}$-free $F$ for which $f_{F,K_{s}}(n)/f_{K_{s-1},K_{s}}(n)\rightarrow\infty$.

Unlike the case of triangles, it appears that for $s\geq 4$, it is difficult to determine for each $K_{s}$-free graph $F$ a constant $c=c(F)$ such that $f_{F,K_{s}}(n)=n^{c+o(1)}$. The second author and Mattheus [14] proved $f_{K_{2},K_{4}}(n)=n^{1/3+o(1)}$ whereas it is well-known that $f_{K_{3},K_{4}}=n^{1/2+o(1)}$. We [16] recently proved $f_{K_{3},K_{4}}(n)=O(\sqrt{n}\log n)$ and the proof can be extended to prove that for every $K_{4}$-free graph $F$, we have $f_{F,K_{4}}(n)=O(\sqrt{n}\log n)$. Perhaps this can be improved for triangle-free $F$ as follows.

Problem 2.

Is it true that for every triangle-free graph $F$ there exists $\varepsilon=\varepsilon_{F}>0$ such that $f_{F,K_{4}}(n)<n^{1/2-\varepsilon}$?

Regardless of whether Problem 2 has an affirmative answer, one might suspect that there exists a sequence of triangle-free graphs where the exponent tends to $1/2$. We propose the following.

Problem 3.

Prove (or disprove) that $f_{K_{t,t},K_{4}}(n)=n^{1/2+o_{t}(1)}$.

The method of proof of Theorem 5 yields $f_{K_{t,t},K_{4}}(n)>n^{2/5-o_{t}(1)}$.

Our next result shows that for $s\geq 4$, we can find substantially larger $F$-free sets in $K_{s}$-free graphs than their conjectured [15] minimum independence number, which is $n^{1/(s-1)+o(1)}$.

Theorem 2.

Let $s\geq 4$ and let $F$ be any graph containing a cycle. Then there exists a constant $c_{F}>0$ such that

$$f_{F,K_{s}}(n)=\Omega(n^{\frac{1}{s-1}+c_{F}}).$$

If $F$ is a cycle, then this bound is almost tight for $K_{4}$, using the following proposition. Write $r(H,t)$ for the ramsey number of $H$ versus a clique on $t$ vertices.

Proposition 1.

For any graphs $F$ and $G$,

$$f_{F,G}(r(G,t)-1)<r(F,t).$$

Indeed, let $H$ be a $G$-free graph on $r(G,t)-1$ vertices with no independent set of size $t$. Then the maximum $F$-free subset of $H$ has size less than $m:=r(F,t)$ for any set of $m$ vertices in $H$ must contain either a copy of $F$ or an independent set of size $t$.

When $F=C_{2k}$ or $F=C_{2k-1}$ we have $r(F,t)=O(t^{k/(k-1)}/(\log t)^{1/(k-1)})$ ([12, 20]). Moreover, recent results of [14], yield $r(K_{4},t)=\Omega(t^{3}/\log^{4}t)$. Putting these together in Proposition 1 yields

$$f_{F,K_{4}}(n)=O(n^{\frac{k}{3k-3}}(\log n)^{\frac{4k-3}{3k-3}})\qquad\hbox{ for }\qquad\hbox{ $F\in\{C_{2k},C_{2k-1}\}$}.$$ (1)

The constant in Theorem 2 satisfies $c_{F}=\Theta(1/k)$ for $F=C_{k}$ and with (1) this gives

$$f_{C_{k},K_{4}}(n)=f_{K_{2},K_{4}}(n)^{1+\Theta(\frac{1}{k})+o(1)}=n^{\frac{1}{3}+\Theta(\frac{1}{k})+o(1)}.$$ (2)

This shows that there are graphs $F$ for which $f_{F,K_{4}}(n)$ does not have the same exponent as $f_{K_{2},K_{4}}(n)$ or $f_{K_{3},K_{4}}(n)$, in contrast to the case of $f_{F,K_{3}}(n)=f_{K_{2},K_{3}}(n)^{1+o(1)}$ from Theorem 1. Using the graphs constructed in [14], and following the analysis along the lines of Janzer and Sudakov [9], Balogh et al. [2] improved the upper bound in (1) slightly in the case of even cycles, by showing

$$f_{C_{2k},K_{4}}(n)=O(n^{\frac{k}{3k-2}}(\log n)^{\frac{6k}{3k-2}}).$$

They also showed for complete multipartite graphs

$$f_{K_{s_{1},...,s_{r}},K_{r+2}}(n)=O(n^{\frac{2s-3}{4s-5}}(\log n)^{3}),$$

where $s=\sum s_{i}$. In the special case of 4-cycles this gives $f_{C_{4},K_{4}}(n)=O(n^{5/11})$.

We now address general Erdős-Rogers functions $f_{F,G}(n)$. For a given $G$, the first natural question is when $f_{F,G}(n)$ can be $n^{1-o(1)}$ as $|V(F)|\rightarrow\infty$. A blowup of a graph $F$ is obtained by replacing each vertex $v$ of $F$ with an independent set $I_{v}$ and adding all edges between $I_{u}$ and $I_{v}$ whenever $\{u,v\}\in E(F)$. The graph $F$ is a homomorphic image of $G$ if and only if some blowup of $F$ contains $G$. Consequently, we say that $F$ is $\mbox{hom}(G)$-free if no blowup of $F$ contains $G$. For instance, if $G$ is bipartite and $F$ contains at least one edge, then blowups of $F$ contain arbitrarily large complete bipartite graphs, and therefore $F$ is not $\mbox{hom}(G)$-free. This condition turns out to determine when Erdős-Rogers functions $f_{F,G}(n)$ can approach $n^{1-o(1)}$ as $|V(F)|\rightarrow\infty$:

Theorem 3.

For every graph $G$, there exists $\varepsilon_{G}>0$ such that if $F$ is any $\mbox{hom}(G)$-free graph containing at least one edge, then

$$f_{F,G}(n)=O(n^{1-\varepsilon_{G}}).$$

On the other hand, if $G$ is not a clique, then for any $\varepsilon>0$ there exists a $G$-free graph $F$ such that $f_{F,G}(n)=\Omega(n^{1-\varepsilon})$.

If $G$ is a clique, then every $G$-free graph is also $\mbox{hom}(G)$-free, hence the first part of Theorem 3 applies to all $G$-free graphs $F$ when $G$ is a clique. As mentioned earlier, in the case $G=K_{4}$, it turns out $f_{F,G}(n)=O(n^{1/2}\cdot\log n)$ due to our results in [16], so we may take $\varepsilon_{K_{4}}\geq 1/2$. It appears to be difficult to determine the largest possible value of $\varepsilon_{G}$ for each graph $G$ in Theorem 3.

2 Proof of Theorem 1

Ajtai, Komlós and Szemerédi [1] and Shearer [17] proved that $r(3,t)=O(t^{2}/\log t)$. Using the random triangle-free process, Kim [10] (see also Fiz Pontiveros, Griffths and Morris [8] and Bohman and Keevash [4]) showed $r(3,t)=\Omega(t^{2}/\log t)$, thereby determining the order of magnitude of $r(3,t)$. Consequently, for any non-empty graph $F$,

$$f_{F,K_{3}}(n)\geq f_{K_{2},K_{3}}(n)=\Theta(\sqrt{n\log n}).$$

To prove Theorem 1 we employ a construction of Erdős, Frankl and Rödl [6] of a linear triangle-free $R$-uniform $N$-vertex hypergraph. In the appendix, we give present a minor modification of their construction which gives a bound that is better than the bound from [6] when $R>\log N$; they prove a lower bound $N^{2}/e^{O(\log R\sqrt{\log N})}$ while our bound is $N^{2}/e^{O(\sqrt{\log R\log N})}$.

Theorem 4.

(Proposition A in Appendix) For any $R,N\geq 3$ and $N\geq R\geq\log N$, there exists an $N$-vertex $R$-uniform hypergraph $H$ with the following properties:

$\mathrm{(i)}$	$|E(H)|\geq N^{2}/R^{8\sqrt{\log_{R}N}}$
$\mathrm{(ii)}$	$H$ is linear, that is, for any distinct edges $e,f\in H$, $|e\cap f|\leq 1$.
$\mathrm{(iii)}$	$H$ is triangle-free, that is, for any three distinct edges $e,f,g\in H$, if $|e\cap f|=|f\cap g|=|g\cap e|=1$ then $|e\cap f\cap g|=1$.

Proof of Theorem 1. We are to prove that

$$f_{F,K_{3}}(n)=n^{\frac{1}{2}+O(\sqrt{\frac{\log\log n}{\log n}})}.$$

Let $t=|V(F)|$. We apply Theorem 4 with $R=\lceil 3t\log t\log N\rceil$, where $t=|V(F)|$. Then (i) yields

$$|E(H)|\geq\frac{N^{2}}{R^{8\sqrt{\log_{R}N}}}=N^{2-O(\sqrt{\frac{\log\log N}{\log N}})}.$$ (3)

Let $G$ be the graph whose vertex set is $E(H)$ and where $E(G)=\{e,f\in E(H):e\cap f\neq\emptyset\}$. For each vertex $v\in V(H)$, the set $K_{v}=\{e\in E(H):v\in e\}$ induces a clique in $G$. If for some distinct $v,w\in V(H)$ there exist distinct $e,f\in K_{v}\cap K_{w}$, then by definition $v,w\in e\cap f$, which contradicts that $H$ is linear. Therefore $|V(K_{v})\cap|V(K_{w})|\leq 1$ for all distinct $v,w\in V(H)$, and the cliques $K_{v}$ are edge-disjoint in $G$. Similarly, since $H$ is triangle-free, every triangle in $G$ is contained in a clique $K_{v}$ for some $v\in V(H)$.

Independently for $v\in V(H)$, let $\chi_{v}:V(K_{v})\rightarrow V(F)$ be a random coloring of $K_{v}$. Next, we remove all edges $\{x,y\}$ of $G[K_{v}]$ such that $\chi_{v}(x)=\chi_{v}(y)$ or $\chi_{v}(x)\chi_{v}(y)\not\in E(F)$. In other words, we have placed a blowup of a copy of $F$ in each set $K_{v}$.

Since $F$ contains no triangle, the resulting graph $G^{*}$ is triangle-free. We now prove that $G^{*}$ has no $F$-free induced subgraph with at least $N$ vertices. To see this, fix a set $Z$ of $N$ vertices of $G^{*}$. The probability that $Z$ is an $F$-free set of $G^{*}$ is

$${\mathbb{P}}(Z)\leq\prod_{v\in V(H)}t\cdot\Bigl{(}1-\frac{1}{t}\Bigr{)}^{|K_{v}\cap Z|}.$$

Since $H$ is $R$-uniform,

$$\sum_{v\in V(H)}|K_{v}\cap Z|=\sum_{e\in Z}|e|=R|Z|=RN.$$

Using $(1-x)^{y}\leq e^{-xy}$ for $0\leq x\leq 1$ and $y\geq 1$,

$${\mathbb{P}}(Z)\leq t^{N}\Bigl{(}1-\frac{1}{t}\Bigr{)}^{RN}\leq e^{N\log t-RN/t}<N^{-2N}.$$

The number of sets of size $N$ in $G^{*}$ is no more than

$${N^{2}\choose N}\leq\frac{N^{2N}}{N!}.$$

Therefore the expected number of $F$-free sets $Z$ of size $N$ in $G^{*}$ is less than $1/N!$. We may therefore select $G^{*}$ so as to contain no $F$-free subgraph with at least $N$ vertices. Since $G^{*}$ is triangle-free, and $n:=|V(G^{*})|=|E(H)|$, the bound (3) gives

$$f_{F,K_{3}}(n)<N=n^{\frac{1}{2}+O(\sqrt{\frac{\log\log n}{\log n}})}.$$

This proves the theorem. ∎

3 Proof of Theorem 2: Large $C_{k}$-free subsets

To prove Theorem 2, it is sufficient to prove the following theorem:

Theorem 5.

For any graph $F$ containing a cycle $C_{k}$, there exists $\epsilon_{k}>1/100k$ such that

$$f_{F,K_{4}}(n)=\Omega(n^{\frac{1}{3}+\epsilon_{k}}).$$

To see that this implies Theorem 2, let $H$ be a $K_{s}$-free graph where $s\geq 5$. If $H$ has maximum degree $d$, then by Turán’s Theorem, $H$ has an independent set of size at least $n/(d+1)$, and the neighborhood of a vertex of degree $d$ induces a $K_{s-1}$-free subgraph. By induction, setting $\alpha_{k}(4)=1/3+\epsilon_{k}$, for $s\geq 5$, there exists $\alpha=\alpha_{k}(s-1)>1/(s-2)$ such that this $K_{s-1}$-free subgraph has an $F$-free subgraph with $\Omega(d^{\alpha_{k}(s-1)})$ vertices. Therefore we have an $F$-free subgraph of size at least

$$\Omega(\max\Bigl{\{}d^{\alpha_{k}(s-1)},\frac{n}{d+1}\Bigr{\}}).$$

Setting

$$\alpha_{k}(s)=1-\frac{1}{1+\alpha_{k}(s-1)},$$

since $\alpha_{k}(4)>1/3$ for all $k\geq 3$, by induction we have

$$\alpha_{k}(s)>1-\frac{1}{1+\frac{1}{s-2}}=\frac{1}{s-1}$$

as required. Moreover, if $\alpha_{k}(s-1)\geq 1/(s-2)+\epsilon$ where $\epsilon\leq 1$, the calculation above yields

$$\alpha_{k}(s)\geq 1-\frac{s-2}{s-1+\epsilon(s-2)} =\frac{1}{s-1}+\epsilon\left(\frac{(s-3)+\frac{1}{s-1}}{s-1+\epsilon(s-2)}\right)$$

$$>\frac{1}{s-1}+\epsilon\left(\frac{s-3}{2(s-1)}\right)$$

$$\geq\frac{1}{s-1}+\frac{\epsilon}{4}.$$

With $\alpha_{k}(4)>1/3+1/100k$, this gives $\alpha_{k}(s)=1/(s-1)+\Omega_{s}(1/k)$ as $k\rightarrow\infty$.

We will prove Theorem 5 as follows: a given $K_{4}$-free graph $H$ either has few $k$-cycles going through every vertex or has a vertex that lies in many $k$-cycles. In the former case, we apply standard results about hypergraph independent sets (Lemma 3) to obtain a large $C_{k}$-free subset. In the latter case, we show that $H$ contains a dense bipartite graph and then use the dependent random choice technique to extract from this a large independent set in one of the parts. These assertions are stated in the next three lemmas.

For sets $X,Y$ of vertices in a graph $G$, let $e(X,Y)$ denote the number of edges $\{x,y\}\in E(G)$ such that $x\in X$ and $y\in Y$.

Lemma 1.

Let $G$ be a graph of maximum degree $d$, and let $\delta>0$. Suppose the number of cycles of length $k$ containing a vertex $v_{0}\in V(G)$ is at least $\delta d^{k-1}$. Then there exist sets $X,Y\subseteq V(G)$ such that $e(X,Y)\geq\delta|X||Y|/(2\log_{2}d)^{k}$ and $|X|,|Y|\geq\delta d/(\log_{2}d)^{k-3}$.

Proof.

Let $\mathcal{C}$ be the set of $k$-cycles containing $v_{0}$. For each $\sigma\in\mathcal{C}$, pick an ordering $(\sigma_{0},\sigma_{1},\dots,\sigma_{k-1},\sigma_{0})$ of the vertices of $\sigma$, where $\{\sigma_{i},\sigma_{i+1}\}\in E(\sigma)$ with subscripts modulo $k$. Let $X_{i}=\{\sigma_{i}:\sigma\in\mathcal{C}\}$ for $1\leq i\leq k-1$. Then for $2\leq i\leq k-2$ there exist sets $X_{i}^{\prime}\subseteq X_{i}$ and $a_{i}\in\{1,2,\dots,d\}$ such that every vertex of $X_{i}^{\prime}$ has at least $a_{i}/2$ and at most $a_{i}$ neighbors in $X_{i-1}^{\prime}$, and the number of cycles $\sigma\in\mathcal{C}$ with $\sigma_{i}\in X_{i}^{\prime}$ is at least $\delta d^{k-1}/(\log_{2}d)^{k-3}$. This can be done iteratively, starting by splitting $X_{2}$ into sets $X_{2j}$ such that every vertex of $X_{2j}$ has at least $d/2^{j+1}$ and at most $d/2^{j}$ neighbors in $X_{1}$, for $0\leq j\leq\log_{2}d$, and considering an $X_{2}^{\prime}=X_{2j}$ for which at least $|\mathcal{C}|/(\log_{2}d)$ of the cycles use an edge between $X_{1}$ and $X_{2}^{\prime}$. Call this collection of cycles $\mathcal{C}_{2}$. Then repeat the argument for the pair $X_{2}^{\prime}$ and $X_{3}$, with collection of cycles $\mathcal{C}_{2}$ so there exists $X_{3}^{\prime}\subseteq X_{3}$ and $\mathcal{C}_{3}\subset\mathcal{C}_{2}$ with $|\mathcal{C}_{3}|\geq|\mathcal{C}_{2}|/\log_{2}d$. We continue to obtain $X_{i}^{\prime}\subseteq X_{i}$ for all $i\leq k-2$ and set $\mathcal{C}^{\prime}:=\mathcal{C}_{k-2}$. Then

$$|\mathcal{C}^{\prime}|\geq\frac{|\mathcal{C}|}{(\log_{2}d)^{k-3}}\geq\frac{\delta d^{k-1}}{(\log_{2}d)^{k-3}}$$

and for every $\sigma\in\mathcal{C}^{\prime}$ we have $\sigma_{i}\in X_{i}^{\prime}$ for $i\leq k-2$.

Let $X=X_{k-2}^{\prime}$ and $Y=X_{k-1}^{\prime}$. The number of cycles in $\mathcal{C}^{\prime}$ containing an edge $\{x,y\}$ with $x\in X$ and $y\in Y$ is at most $a_{2}\cdots a_{k-2}\leq d^{k-3}$ as the maximum degree is $d$. Consequently,

$$|\mathcal{C}^{\prime}|\leq e(X,Y)\cdot a_{2}\cdots a_{k-2}\leq e(X,Y)\cdot d^{k-3}$$

and

$$d\cdot\min\{|X|,|Y|\}\geq e(X,Y)\geq\frac{|\mathcal{C}^{\prime}|}{d^{k-3}}\geq\frac{\delta d^{2}}{(\log_{2}d)^{k-3}}.$$

Therefore $\min\{|X|,|Y|\}\geq\delta d/(\log_{2}d)^{k-3}$.

Next we prove that $e(X,Y)\geq\delta|X||Y|/(2\log_{2}d)^{k}$. By construction, for $2\leq i\leq k-2$,

$$\frac{a_{i}}{2}|X_{i}^{\prime}|\leq e(X_{i}^{\prime},X_{i-1}^{\prime})\leq d|X_{i-1}^{\prime}|$$

and therefore $a_{i}\leq 2d|X_{i-1}^{\prime}|/|X_{i}^{\prime}|$. Since $|X_{1}^{\prime}|\leq|N(v_{0})|\leq d$ and $|Y|\leq|N(v_{0})|\leq d$,

$$a_{2}a_{3}\cdots a_{k-2}\leq\prod_{i=2}^{k-2}2d\frac{|X_{i-1}^{\prime}|}{|X_{i}^{\prime}|}=(2d)^{k-3}\frac{|X_{1}^{\prime}|}{|X_{k-2}^{\prime}|}=(2d)^{k-3}\frac{|X_{1}^{\prime}|}{|X|}\leq\frac{(2d)^{k-1}}{|X||Y|}.$$

Consequently,

$$e(X,Y)\geq\frac{|\mathcal{C}^{\prime}|}{a_{2}\dots a_{k-2}}\geq\frac{|\mathcal{C}^{\prime}||X||Y|}{(2d)^{k-1}}\geq\frac{\delta}{2^{k-1}(\log_{2}d)^{k-3}}|X||Y|$$

completing the proof. ∎

The following lemma is a standard consequence of the dependent random choice method and we omit the proof.

Lemma 2.

Let $\gamma\geq 0$, $s\geq 1$, and let $X$ and $Y$ be disjoint sets of vertices in a graph, such that $e(X,Y)\geq\gamma|X||Y|$. Then for any $s\geq 1$, there exists a set $Z\subseteq Y$ such that

$$|Z|\geq\frac{1}{2}\gamma^{s}|Y|$$

and every pair of vertices in $Z$ has at least $\gamma|X||Y|^{-1/s}$ neighbors in $X$.

Finally we need the following standard result about independent sets in hypergraphs first proved by Spencer [19].

Lemma 3.

For every $k\geq 2$, every $n$-vertex $k$-uniform hypergraph with average degree $d>0$ has an independent set of size at least $(1-1/k)n/d^{1/(k-1)}$.

We now have the necessary ingredients to prove Theorem 5.

Proof of Theorem 5. For $k\geq 3$, let

$$\epsilon_{k}=\frac{1}{100(k-1)}.$$

Let $H$ be an $n$-vertex $K_{4}$-free graph with maximum degree $d$. We will find a $C_{k}$-free subset of vertices in $H$ of size at least $n^{1/3+\epsilon_{k}}$. Suppose that $\triangle$ is the maximum number of copies of $C_{k}$ that a vertex is in. Define

$$\delta:=\frac{\triangle}{d^{k-1}}.$$

We now obtain two different bounds on the maximum $C_{k}$-free set.

Bound 1. Let $\mathcal{H}$ be the $k$-uniform hypergraph with $V(\mathcal{H})=V(H)$ and $E(\mathcal{H})=\{V(C_{k}):C_{k}\subseteq H\}$. Then $\mathcal{H}$ has maximum degree (and hence average degree) at most $\triangle$ and Lemma 3 implies that $H$ has an independent set of size at least

$$\Omega\left(\frac{n}{\triangle^{\frac{1}{k-1}}}\right)=\Omega\left(\frac{n}{\delta^{\frac{1}{k-1}}d}\right).$$

Bound 2. Let $v_{0}\in V(H)$ lie in $\triangle=\delta d^{k-1}$ copies of $C_{k}$. By Lemma 1, there exist sets $X,Y\subseteq V(H)$ such that

$$e(X,Y)\geq\frac{\delta}{(2\log_{2}d)^{k}}|X||Y|=:\gamma|X||Y|,$$

where $|X|\geq|Y|\geq\gamma d$. By Lemma 2 applied with $s=3$, there exists $Z\subseteq Y$ with

$$|Z|\geq\frac{1}{2}\gamma^{3}|Y|$$

such that every pair of vertices in $Z$ has at least $\gamma|X||Y|^{-1/3}\geq\gamma|Y|^{2/3}$ common neighbors in $X$. If $Z$ is not an independent set in $H$, then there exists $\{x,y\}\in E(H)$ with $x,y\in Z$. Since $H$ is $K_{4}$-free, $N(x)\cap N(y)$ is an independent set in $H$ of size at least $\gamma|Y|^{2/3}\geq\gamma^{5/3}d^{2/3}$. Otherwise, $Z$ is an independent set in $H$ of size at least $\frac{1}{2}\gamma^{3}|Y|\geq\frac{1}{2}\gamma^{4}d$. In particular, $H$ has a $C_{k}$-free induced subgraph of size at least

$$h(d,\gamma)=\min\left\{\gamma^{5/3}d^{2/3},\frac{1}{2}\gamma^{4}d\right\}.$$

It is also the case that $G$ always contains an independent set with at least $n/(d+1)$ vertices, by Turán’s Theorem. If $d\leq n^{2/3-\epsilon_{k}}$, this gives an independent set of size $n^{1/3+\epsilon_{k}}$ in $G$. If $d\geq n^{2/3+2\epsilon_{k}}$, then since the neighborhood of a vertex of degree $d$ induces a triangle-free graph, this neighborhood contains an independent set of size at least $d^{1/2}\geq n^{1/3+\epsilon_{k}}$ in $G$. Therefore we assume $n^{2/3-\epsilon_{k}}\leq d\leq n^{2/3+2\epsilon_{k}}$. In that case, by Bounds 1 and 2, we obtain a $C_{k}$-free set of size at least

$$\Omega\left(\max\left\{\frac{n}{\delta^{\frac{1}{k-1}}d},h(d,\gamma)\right\}\right).$$

If $\delta<n^{-1/25}$, then Bound 1 is at least

$$\Omega\left(\frac{n}{\delta^{\frac{1}{k-1}}d}\right)=\Omega\left(n^{\frac{1}{3}+\frac{1}{25(k-1)}-2\epsilon_{k}}\right)=\Omega(n^{\frac{1}{3}+\epsilon_{k}})$$

as $\epsilon_{k}<1/75(k-1)$. So we may assume that $\delta\geq n^{-1/25}$, and, as $n$ is sufficiently large, we may assume that $\gamma=\delta/(2\log_{2}d)^{k}>n^{-1/24}$. In this case, $\epsilon_{k}<1/100$ and $d>n^{2/3-\epsilon_{k}}$ yield

$$\gamma^{\frac{5}{3}}d^{\frac{2}{3}}>n^{\frac{-5}{72}+\frac{4}{9}-\frac{2\epsilon_{k}}{3}}>n^{\frac{1}{3}+\epsilon_{k}}\qquad\hbox{ and }\qquad\gamma^{4}d>n^{-\frac{1}{6}+\frac{2}{3}-\epsilon_{k}}>2n^{\frac{1}{3}+\epsilon_{k}}$$

and therefore $h(d,\gamma)>n^{1/3+\epsilon_{k}}$, completing the proof. ∎

4 Proof of Theorem 3

A sunflower is a collection of sets every pair of which have the same intersection, called the core. We need the well-known Erdos-Rado sunflower lemma in the form below.

Lemma 4.

Fix $t,m>0$. Every $t$-uniform hypergraph with more than $t!(m-1)^{t}$ edges has a sunflower of size $m$.

Proof of Theorem 3. Let $|V(G)|=k$. We may assume that $G$ is not acyclic, since otherwise $G$ would be contained in a blowup of $F$. Consider an $n$ by $n$ bipartite graph $H$ without cycles of length at most $2k$ and where every vertex has degree $d=n^{\frac{1}{3k}}$. Such graphs exist, for example the bipartite Ramanujan graphs of Lubotzky, Phillips and Sarnak [13], or even a random $d$-regular graph (if we are not fussy about the constant in the exponent). We now employ the methods of [15, 5]. Let $H^{\prime}$ be the restriction of the square of $H$ to one part of $H$, so that $H^{\prime}$ has $n$ vertices, and is a union of $n$ edge-disjoint cliques $K^{1},K^{2},\dots,K^{n}$ of order $d$. Since $G$ is not acyclic, every copy of $G$ in $H^{\prime}$ is contained in one of those cliques. In each of the cliques, take a random coloring with elements of $V(F)$, and put an edge between any two color classes corresponding to an edge of $F$. Since $F$ is $\mbox{hom}(G)$-free, this random graph $H^{*}$ is $G$-free. We claim (similarly to the proof of Theorem 1), that every set of at least $(2n|V(F)|\log|V(F)|)/d$ vertices of $H^{*}$ induces a copy of $F$. The probability that such a set $X$ does not induce a copy of $F$ is at most

$$\prod_{i=1}^{n}|V(F)|\cdot\Bigl{(}1-\frac{1}{|V(F)|}\Bigr{)}^{|X\cap V(K^{i})|}.$$

Now we use

$$\sum_{i=1}^{n}|X\cap V(K_{i})|=d|X|$$

and therefore the expected number of such $X$ is at most

$${n\choose|X|}\cdot|V(F)|^{n}\Bigl{(}1-\frac{1}{|V(F)|}\Bigr{)}^{d|X|}<e^{|X|\log n-d|X|/|V(F)|+n\log|V(F)|}.$$

This is vanishing since $d|X|/|V(F)|>2n\log|V(F)|$. Therefore

$$f_{F,G}(n)=O(n/d)=O(n^{1-\frac{1}{3k}})=O(n^{1-\frac{1}{3|V(G)|}})$$

and we may take $\varepsilon_{G}=1/3|V(G)|$ in Theorem 3.

We now prove the second statement of the theorem. Let $r:=|V(G)|-1$. If $G$ is acyclic, then any $n$-vertex $G$-free graph has an independent set $I$ of size linear in $n$, and $I$ is certainly $F$-free for any nonempty $F$ so we are done. If $G$ is not 2-connected, then let $F=K_{r}$ so that $F$ is clearly $G$-free. Suppose that $H$ is an $n$-vertex $G$-free graph. Then no two $r$-cliques in $H$ have a point in common, for otherwise the subgraph of $H$ induced by their union contains $G$. Indeed, we can pick some vertex in the intersection of the two cliques to be a cut vertex of $G$, and then easily embed $G$ in the union of the two cliques (the embedding is even easier if $G$ is not connected). Consequently, the $r$-cliques in $H$ are pairwise vertex disjoint. Then $H$ has a $K_{r}$-free induced subgraph of size at least $(1-1/r)n$ and we are done.

We may henceforth assume that $G$ is 2-connected. Since $G$ is not a clique, let $v,w$ be nonadjacent vertices in $G$. Let $G^{+}$ be the graph obtained from $G$ by adding all edges that are not already in $G$ between $\{v,w\}$ and $N_{G}(v)\cup N_{G}(w)$. So $G^{+}\supset G$, and $v$ and $w$ are clones in $G^{+}$. Let $G^{*}=G^{+}-\{w\}$ and let $G^{**}=G^{+}-\{v,w\}=G^{*}-\{v\}$. So $G^{*}$ has $r$ vertices and $G^{**}$ has $r-1$ vertices.

Assume that $t$ is sufficiently large in terms of $r$ and set $\delta=1/5r^{2}$. Apply Proposition B to obtain a $t$-vertex $r$-uniform hypergraph $F^{*}$ with girth larger than $r+1$, and the property that for every $s$-set $S$ with $t^{1-\delta}\leq s\leq t-1$, the number of edges in $F^{*}$ with exactly $r-1$ vertices in $S$ is at least

$$\frac{1}{10}{s\choose r-1}(t-s)t^{1-r+\frac{1}{2r}}>{s\choose r-1}(t-s)t^{1-r+\frac{1}{3r}}=:q_{s}.$$

Inside each hyperedge $e$ of $F^{*}$, place randomly a copy of $G^{*}$. More precisely, among all $r!$ ways to map the vertices of $G^{*}$ to $e$, we pick one with probability $1/r!$. Let $F$ be the resulting graph with $V(F)=V(F^{*})$ and $E(F)$ comprises the graph edges in all copies of $G^{*}$ that lie in edges of $F^{*}$.

As $G$ has $r+1$ vertices, and $F^{*}$ is $r$-uniform, there is no copy of $G$ in $F$ that lies entirely within an edge of $F^{*}$. If a copy of $G$ in $F$ has two vertices $x,y$ that do not lie in the same edge of $F^{*}$, then, since $G$ is 2-connected, there is a cycle in $G$ containing $x$ and $y$ and this cycle yields a hypergraph cycle in $F^{*}$ of length at most $r+1$ which does not exist by construction. We conclude that $F$ is $G$-free.

Furthermore, we claim that for any $s$-set $S$ in $F$, with $t^{1-\delta}\leq s\leq t-1$, there exists an edge $e$ of $F^{*}$ with $r-1$ vertices in $S$ and one vertex in $V(F^{*})-S$ such that

the copy of $$G^{*} e G^{**} e\cap S$$ placed inside induces a copy of within . (4)

Indeed, (4) follows from the following argument. For each of the $q_{s}$ edges $e$ of $F^{*}$ with exactly one vertex outside $S$, the probability that $e$ fails (4) is at most $1-1/r!$. Since any two such edges $e,e^{\prime}$ share at most one vertex by the girth property of $F^{*}$, the probability that all of these $q_{s}$ edges $e$ fail (4) is at most $(1-1/r!)^{q_{s}}$. Consequently, the probability that there exists an $s$-set for which there is no $e$ satisfying (4) is at most

$$\sum_{s=t^{1-\delta}}^{t-1}{t\choose s}e^{-q_{s}/r!}=\sum_{s=t^{1-\delta}}^{t-1}{t\choose t-s}e^{-q_{s}/r!}<t{t\choose t-s}e^{-q_{s}/r!}<e^{\log t+(t-s)\log t-q_{s}/r!}<1.$$

The final inequality holds as $\delta=1/5r^{2}$ implies

$$1-r+\frac{1}{3r}+(1-\delta)(r-1)>0.$$

Hence we may assume that for all $s$-sets $S$ with $t^{1-\delta}\leq s\leq t-1$ there exists an edge $e$ of $F^{*}$ with $r-1$ vertices in $S$ and one vertex in $V(F^{*})-S$ which satisfies (4).

Now let $H$ be any $G$-free $n$-vertex graph. We are to find an $F$-free set of size $\Omega(n^{1-\varepsilon})$. Let $R=r!+1$ and

$$T=t!\left(R{t-1\choose r-1}-1\right)^{t}.$$

Set $b=t^{1-\delta}$. We claim the number of copies of $F$ in $H$ is at most

$$T{n\choose b}.$$ (5)

If (5) holds, then there are at most $O(n^{b})$ copies of $F$ in $H$ and we finish the proof as follows. Consider the $t$-uniform hypergraph $\mathcal{H}$ with $V(\mathcal{H})=V(H)$ and $E(\mathcal{H})=\{V(F):F\subseteq H\}$. The average degree of $\mathcal{H}$ is $O(n^{b-1})$. By Lemma 3, $\mathcal{H}$ contains an independent set $I$ of size $\Omega(n^{1-(b-1)/(t-1)})$. Since $(b-1)/(t-1)<t^{1-\delta}/(t-1)<\varepsilon$ for large $t$, we conclude that $I$ is an $F$-free set of size at least $\Omega(n^{1-\varepsilon})$.

We now prove (5). Assume to the contrary. Then to each copy of $F$ we may associate any $b$-subset of its vertices. By pigeonhole, there exists a set $C$ of $b$ vertices in $H$ and at least $T$ copies of $F$, say $F_{1},F_{2},\dots,F_{T}$ for which $V(F_{i})\cap V(F_{j})\supseteq C$. Amongst these sets of size $t$, Lemma 4 gives a sunflower of size $R{t-1\choose r-1}$ with core $S\supseteq C$. As $t^{1-\delta}\leq|S|\leq t-1$, by (4), for each of these $R{t-1\choose r-1}$ copies $A$ of $F$, there is a vertex $v_{A}$ outside $S$ that forms an edge $e_{A}$ in $A$ with $r-1$ vertices in $S$ and $v_{A}$ plays the role of vertex $v$ in the copy of $G^{*}$ within $e_{A}$ (in other words, $e^{\prime}_{A}=e_{A}-\{v_{A}\}$ induces a copy of $G^{**}$). By pigeonhole, there exists a set $e^{\prime}\subseteq S$ of size $r-1$ and vertices $v_{1},v_{2},\ldots,v_{R}\not\in S$ such that $e_{i}=e^{\prime}\cup\{v_{i}\}$ is an edge of $F^{*}$ for all $i\in[R]$ and $v_{i}$ plays the role of $v$ in the copy of $G^{*}$ in $e_{i}$ (in other words, $e^{\prime}$ induces a copy of $G^{**}$). Since $R>r!$, we can find vertices, say $v_{1},v_{2}$, such that the copies of $G^{**}$ within $e^{\prime}$ for both $v_{1}$ and $v_{2}$ are identical. This copy of $G^{**}$ together with $v_{1}$ and $v_{2}$ is a copy of $G^{+}$. We conclude $H\supseteq G^{+}\supseteq G$, a contradiction. ∎

5 Appendix

Proposition A.

(Erdős-Frankl-Rödl) For any $N,R\geq 3$ such that $N\geq R\geq\log N$, there exists a linear triangle-free $N$-vertex $R$-uniform hypergraph $H$ with

$$|E(H)|\geq\frac{N^{2}}{R^{8\sqrt{\log_{R}N}}}.$$

Proof.

The construction is based on the construction of Behrend [3] of a dense subset of $\{1,2,\dots,n\}$ with no three-term arithmetic progression. For completeness, we describe this construction here, which is slightly better than the construction of Erdős, Frankl and Rödl [6]. Let $A$ be the set of positive integer points on the sphere of radius $r$ in $\mathbb{R}^{d}$. For any choice of positive integers $x_{1},x_{2},\dots,x_{d-4}\leq r/\sqrt{d}$, there exist positive integers $x_{d-3},x_{d-2},x_{d-1},x_{d}$ such that $x_{1}^{2}+x_{2}^{2}+\dots+x_{d}^{2}=r^{2}$ by Lagrange’s four squares theorem. Therefore

$$|A|\geq\Bigl{(}\frac{r}{\sqrt{d}}\Bigr{)}^{d-4}.$$

Let $X_{i}=[ir]^{d}$. Then define an $R$-uniform $R$-partite hypergraph $H$ where $V(H)$ consists of $X_{1}\cup X_{2}\cup\dots\cup X_{R}$ and let $E(H)=\{x,x+a,x+2a,\dots,x+(R-1)a\}$ where $a\in A$ and $x\in X_{1}$. Then

$$|V(H)|=N\leq R^{d+1}r^{d}\quad\quad\mbox{ and }\quad\quad|E(H)|=|A||X_{1}|\geq\frac{r^{2d-4}}{\sqrt{d}^{d-4}}.$$

Put $d=\lfloor\sqrt{\log_{R}N}\rfloor<\log N\leq R$ and $r=R^{d}$. Then $r^{4}\leq R^{4d}$ and $d^{d}<R^{d}$ and hence

$$|E(H)|\geq\frac{N^{2}}{R^{2d+2}r^{4}d^{\frac{d-4}{2}}}\geq\frac{N^{2}}{R^{8d}}>\frac{N^{2}}{R^{8\sqrt{\log_{R}N}}}.$$

This establishes (i). If $e=\{x,x+a,x+2a,\dots,x+(R-1)a\}$ and $f=\{y,y+b,y+2b,\dots,y+(R-1)b\}$ intersect in two vertices of $H$, say $x+ia=y+ib$ and $x+ja=y+jb$, then $x=y$ and $a=b$, establishing (ii). If $e,f$ and $g=\{z,z+c,z+2c,\dots,z+(R-1)c\}$ have $|e\cap f|=|f\cap g|=|g\cap e|=1$, then we may assume $x+ia=y+ib$ and $y+jb=z+jc$ and $z+kc=x+ka$ for some distinct $i,j,k\in\{0,1,2,\dots,R-1\}$ and $a,b,c\in A$. This implies $i(b-a)+j(c-b)+k(a-c)=0$ which means $(k-i)a+(i-j)b+(j-k)c=0$. Since the sphere is strictly convex, $a,b,c$ cannot all lie in a line, and hence we conclude two of $i,j,k$ are identical, a contradiction. This proves (iii). ∎

For the next proposition, we need some definitions. A cycle of length two in a hypergraph is a set of two edges that share at least two vertices. A cycle of length $\ell>2$ is a collection of $\ell$ distinct vertices $v_{1},v_{2},\ldots,v_{\ell}$ and $\ell$ distinct edges $e_{1},\ldots,e_{\ell}$ where $e_{i}\cap e_{i+1}=\{v_{i+1}\}$ (indices modulo $\ell$) and $e_{i}\cap e_{j}=\emptyset$ otherwise. So an $\ell$-cycle in an $r$-uniform hypergraph ($\ell>2$) has $\ell$ edges and $\ell(r-1)$ vertices (these are often called loose cycles). Say that a hypergraph $H$ has girth $g$ if the length of the shortest cycle in $H$ is $g$.

Proposition B.

Fix $r\geq 2$ and $\delta=1/5r^{2}$. For $t$ sufficiently large, there exists a $t$-vertex $r$-uniform hypergraph $F^{*}$ with girth at least $r+2$ such that for every $s$-subset $S$ with $t^{1-\delta}<s<t$, the number of edges with exactly one vertex outside $S$ is at least

$$\frac{1}{10}{s\choose r-1}(t-s)t^{1-r+\frac{1}{2r}}.$$ (6)

Proof.

Consider the binomial random $r$-uniform hypergraph $H\sim H^{(r)}(t,p)$ with $t$ vertices where each edge appears independently with probability $p=t^{1-r+\frac{1}{2r}}$. For each $2\leq\ell\leq r+1$, Let $C_{\ell}$ denote the cycle of length $\ell$ (this is unique except for $\ell=2$) and let $\mathcal{B_{\ell}}$ denote a maximal collection of edge-disjoint copies of $C_{\ell}$ in $H$. Form $F^{*}$ by starting with $H$ and deleting all $\ell$ edges from every copy of $C_{\ell}$ in $\mathcal{B_{\ell}}$ for all $2\leq\ell\leq r+1$. Then, by the maximality of $\mathcal{B_{\ell}}$, the remaining hypergraph $F^{*}$ has girth at least $r+2$. We will now show that with high probability $F^{*}$ has the required property.

Pick $S\subset V(F^{*})$ of size $s$ where $t^{1-\delta}<s<t$. Call an edge in $H$ with exactly one vertex outside $S$ an $S$-edge. Let $X=X_{S}$ be the number of $S$-edges, let $Y_{\ell}=Y_{S,\ell}$ be the number of copies of $C_{\ell}$ that contain at least one $S$-edge and let $Z_{\ell}=Z_{S,\ell}$ be the maximal number of pairwise edge-disjoint copies of $C_{\ell}$, each containing at least one $S$-edge. Obviously, $Z_{\ell}\leq Y_{\ell}$. Define the event

$$A_{\ell}=A_{S,\ell}=\{X>10r\ell\,Z_{\ell}\}.$$

We note that if $A_{S,\ell}$ holds for every appropriate $S$, and $2\leq\ell\leq r+1$, then the number of $S$-edges in $F^{*}$ is at least

$$X-\sum_{\ell=2}^{r+1}\ell\,Z_{\ell}\geq|X|-\sum_{\ell=2}^{r+1}\frac{|X|}{10r}>(0.9)|X|.$$

Moreover, ${\mathbb{E}}(X)={s\choose r-1}(t-s)p$, so if it is also the case that $X>{\mathbb{E}}(X)/2$, then the number of $S$-edges in $F^{*}$ is at least $(0.4){\mathbb{E}}(X)$ and $S$ satisfies (6).

We see that

$${\mathbb{E}}(Y_{\ell})<{s\choose r-1}(t-s)t^{\ell(r-1)-r}p^{\ell}.$$

As $p=t^{1-r+\frac{1}{2r}}$ and $\ell\leq r+1$, we have $p^{\ell}t^{\ell(r-1)-r}\ll p$. Therefore ${\mathbb{E}}(Y_{\ell})\ll{\mathbb{E}}(X)$. Now

$${\mathbb{P}}(\overline{A_{\ell}})={\mathbb{P}}(X\leq 10r\ell\,Z_{\ell})\leq{\mathbb{P}}\left(X\leq\frac{{\mathbb{E}}(X)}{2}\right)+{\mathbb{P}}\left(Z_{\ell}\geq\frac{{\mathbb{E}}(X)}{20r\ell}\right).$$

Krivelevich [11, Claim 1] proved that in this setup, for any constant $c>0$,

$${\mathbb{P}}(Z_{\ell}\geq c\,{\mathbb{E}}(Y_{\ell}))<e^{-c\,(\log c-1){\mathbb{E}}(Y_{\ell})}.$$

Using this and ${\mathbb{E}}(Y_{\ell})\ll{\mathbb{E}}(X)$ we have

$${\mathbb{P}}\left(Z_{\ell}\geq\frac{{\mathbb{E}}(X)}{20r\ell}\right)={\mathbb{P}}\left(Z_{\ell}\geq\frac{{\mathbb{E}}(X)}{20r\ell\,{\mathbb{E}}(Y_{\ell})}{\mathbb{E}}(Y_{\ell})\right)<e^{\frac{-{\mathbb{E}}(X)}{20r\ell}(\log(\frac{{\mathbb{E}}(X)}{20r\ell\,{\mathbb{E}}(Y_{\ell})})-1)}<e^{-{\mathbb{E}}(X)}.$$

The standard Chernoff bound gives ${\mathbb{P}}(X\leq{\mathbb{E}}(X)/2)<e^{-{\mathbb{E}}(X)/8}$ so altogether we obtain ${\mathbb{P}}(\overline{A_{\ell}})<e^{-{\mathbb{E}}(X)/9}$. Using the union bound, the probability that there exists an $S$ that fails (6) is at most

$$\sum_{s=t^{1-\delta}}^{t-1}{t\choose s}e^{-{s\choose r-1}(t-s)p/9}<e^{\log t+(t-s)\log t-{s\choose r-1}(t-s)p/9}.$$

The power of $t$ in $s^{r-1}p$ is at least $1-r+1/2r+(1-\delta)(r-1)>0$ as $\delta<1/2r^{2}$ and hence the quantity above vanishes for large $t$. We conclude that (6) holds in $F^{*}$ with high probability. ∎