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

On tt-intersecting Hypergraphs with Minimum Positive Codegrees

Sam Spiro111Dept. of Mathematics, UCSD [email protected]. This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1650112.
Abstract

For a hypergraph \mathcal{H}, define the minimum positive codegree δi+()\delta_{i}^{+}(\mathcal{H}) to be the largest integer kk such that every ii-set which is contained in at least one edge of \mathcal{H} is contained in at least kk edges. For 1sk,t1\leq s\leq k,t and trt\leq r, we prove that for nn-vertex tt-intersecting rr-graphs \mathcal{H} with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}, the unique hypergraph with the maximum number of edges is the hypergraph \mathcal{H} consisting of every edge which intersects a set of size 2k2s+t2k-2s+t in at least ks+tk-s+t vertices provided nn is sufficiently large. This generalizes work of Balogh, Lemons, and Palmer who proved this for s=t=1s=t=1, as well as the Erdős-Ko-Rado theorem when k=sk=s.

1 Introduction

We say that a hypergraph \mathcal{H} is an rr-graph if every edge hh\in\mathcal{H} has size rr, and we say that \mathcal{H} is tt-intersecting if |hh|t|h\cap h^{\prime}|\geq t for any distinct h,hh,h^{\prime}\in\mathcal{H}. The central result concerning tt-intersecting rr-graphs is the famous Erdős-Ko-Rado theorem.

Theorem 1.1 ([4]).

For trt\leq r, if \mathcal{H} is an nn-vertex tt-intersecting rr-graph with the maximum number of edges, then there exists a set TT of size tt such that \mathcal{H} consists of every edge containing TT provided nn is sufficiently large.

There exist numerous extensions, variants, and applications of the Erdős-Ko-Rado theorem; see for example the book and survey by Frankl and Tokushige [8, 9] and the book by Godsil and Meagher [11]. Motivated by minimum degree variants of the Erdős-Ko-Rado theorem, Balogh, Lemons, and Palmer [3] considered a variant involving minimum positive codegrees.

Definition 1.

Given a hypergraph \mathcal{H} and integer ii, we define its minimum positive ii-degree δi+()\delta^{+}_{i}(\mathcal{H}) to be the largest integer kk such that if SS is a set of ii vertices contained in at least one edge, then SS is contained in at least kk edges. We adopt the convention that δi+()=\delta^{+}_{i}(\mathcal{H})=\infty if \mathcal{H} has no edges.

To state the main result of [3], we require one more definition.

Definition 2.

We say that an rr-graph \mathcal{H} is an (a,b)(a,b)-kernel system if there exists XV()X\subseteq V(\mathcal{H}) with |X|=a|X|=a such that hh\in\mathcal{H} if and only if |hX|b|h\cap X|\geq b.

For example, (t,t)(t,t)-kernel systems are exactly the extremal constructions appearing in the Erdős-Ko-Rado theorem. More generally, Ahlswede and Khachatrian [1] showed that every nn-vertex tt-intersecting rr-graph \mathcal{H} with the maximum number of edges is a (2i+t,i+t)(2i+t,i+t)-kernel system for some ii provided n>2rtn>2r-t. There are many other contexts where kernel systems appear as extremal constructions for variants of the Erdős-Ko-Rado theorem, especially for problems related to maximum degrees; see for example [6, 7, 14].

It is not hard to show that if rks+tr\geq k-s+t, then an rr-uniform (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system \mathcal{H} is tt-intersecting with δrs+()=(ks)\delta_{r-s}^{+}(\mathcal{H})={k\choose s}. The main result of Balogh, Lemons, and Palmer [3] shows that when s=t=1s=t=1, this is the unique rr-graph with these properties which has the maximum number of edges.

Theorem 1.2 ([3]).

Let k1k\geq 1 and let \mathcal{H} be an nn-vertex 1-intersecting rr-graph with δr1+()k\delta_{r-1}^{+}(\mathcal{H})\geq k. If \mathcal{H} has the maximum number of edges amongst hypergraphs with these properties, then \mathcal{H} is a (2k1,k)(2k-1,k)-kernel system if nn is sufficiently large in terms of rr.

The proof of Theorem 1.2 utilized the delta-system method. Using a similar approach together with some new ideas, we extend Theorem 1.2 to tt-intersecting \mathcal{H} which have bounded positive minimum (rs)(r-s)-degree for essentially all values of ss and tt.

Theorem 1.3.

Let k,r,s,tk,r,s,t be positive integers with sk,ts\leq k,t and trt\leq r, and let \mathcal{H} be an nn-vertex tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}. If \mathcal{H} has the maximum number of edges amongst hypergraphs with these properties, then \mathcal{H} is a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system if nn is sufficiently large in terms of rr.

This theorem shows a surprising phenomenon: despite only demanding δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s} in the hypothesis, the extremal constructions of Theorem 1.3 end up having δrs+()=(ks)\delta_{r-s}^{+}(\mathcal{H})={k\choose s}, which is a significantly stronger condition if s>1s>1. We note that the hypothesis δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s} in Theorem 1.3 is best possible, since if we only demanded δrs+()(k1s)\delta_{r-s}^{+}(\mathcal{H})\geq{k-1\choose s}, then a (2k22s+t,k1s+t)(2k-2-2s+t,k-1-s+t)-kernel system would satisfy the hypothesis and have significantly more edges.

Let us briefly discuss the range of parameters in Theorem 1.3. Observe that (k1s)=0{k-1\choose s}=0 for all ksk\leq s, so there is no loss in generality by considering ksk\geq s. If s>ts>t, then the problem is essentially trivial in view of Theorem 1.1. This is because a (t,t)(t,t)-kernel system will satisfy the positive minimum positive codegree condition if nn is sufficiently large, and (t,t)(t,t)-kernel systems have the maximum number of edges amongst tt-intersecting rr-graphs if nn is sufficiently large. If one considers r<tr<t, then any tt-intersecting rr-graph has at most one edge, so the problem becomes trivial. Thus Theorem 1.3 covers all of the non-trivial ranges of parameters that we could consider for this problem.

2 Proof of Theorem 1.3

Our argument starts off nearly identical to that of [3]. We note that Theorem 1.3 implicitly says that if r<ks+tr<k-s+t, then any \mathcal{H} satisfying the hypothesis of Theorem 1.3 is empty (since (a,b)(a,b)-kernel systems are empty if r<br<b). The following confirms this is the case.

Lemma 2.1.

For 1sk,t1\leq s\leq k,t and trt\leq r, if \mathcal{H} is a non-empty tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}, then rks+tr\geq k-s+t.

Here and throughout the text we refer to sets II of size ii as ii-sets.

Proof.

The result is immediate if k=sk=s, so assume k>sk>s. Assume for contradiction that rks+t1r\leq k-s+t-1 and let hh\in\mathcal{H}. Because k>sk>s, the minimum positive codegree condition implies that there is another edge hhh^{\prime}\neq h in \mathcal{H}, and we will choose such an edge so that |hh||h\cap h^{\prime}| is as small as possible.

Observe that |hh|ts|h\cap h^{\prime}|\geq t\geq s since \mathcal{H} is tt-intersecting. Let ShhS\subseteq h\cap h^{\prime} be any ss-set. By the minimum positive codegree condition, the (rs)(r-s)-set hSh^{\prime}\setminus S is contained in more than (k1s)(rt+ss){k-1\choose s}\geq{r-t+s\choose s} edges. As h(hS)h\setminus(h^{\prime}\setminus S) has size at most rt+sr-t+s, we conclude that there exist some ss-set Sh(hS)S^{\prime}\not\subseteq h\setminus(h^{\prime}\setminus S) such that h′′:=(hS)Sh^{\prime\prime}:=(h^{\prime}\setminus S)\cup S^{\prime}\in\mathcal{H}. Observe that |hh′′|<|hh||h\cap h^{\prime\prime}|<|h\cap h^{\prime}| since h′′h^{\prime\prime} was obtained from hh^{\prime} be deleting an ss-subset of hh and adding an ss-set that was not contained entirely in hh. This contradicts us choosing |hh||h\cap h^{\prime}| as small as possible, a contradiction. ∎

The remainder of our proof relies heavily on sunflowers.

Definition 3.

We say that a hypergraph \mathcal{F} is a sunflower if there exists a set XX such that hh=Xh\cap h^{\prime}=X for any distinct h,h𝒮h,h^{\prime}\in\mathcal{S}. In this case we say that XX is the core of \mathcal{F} and that the sets hXh\setminus X with hh\in\mathcal{F} are the petals of \mathcal{F}. When \mathcal{F} consists of a single edge hh, we adopt the convention that hh is the core of \mathcal{F}.

The main result in the theory of sunflowers is the following result of Erdős and Rado.

Theorem 2.2 ([5]).

For every r,p1r,p\geq 1, there exists a constant f(r,p)r!(p1)rf(r,p)\leq r!(p-1)^{r} such that if \mathcal{H} is an rr-graph with more than f(r,p)f(r,p) edges, then \mathcal{H} contains a sunflower with at least pp petals.

Much stronger bounds for f(r,p)f(r,p) have been obtained in breakthrough work of Alweiss et. al. [2], but for our purposes we only need that f(r,p)f(r,p) is a constant. Theorem 2.2 does not give any control over the size of the core of a sunflower in \mathcal{H}, and for this we use a result of Mubayi and Zhao [15] which is based off of work of Füredi [10].

Proposition 2.3 ([15]).

If r>ks+tr>k-s+t and p1p\geq 1, then there exists a constant CC depending on r,pr,p such that if ||Cnrk+st1|\mathcal{H}|\geq Cn^{r-k+s-t-1}, then \mathcal{H} contains a sunflower with at least pp petals and core of size at most ks+tk-s+t.

Proposition 2.3 allows us to find sunflowers which have many petals and small cores. The next lemma shows that cores of sunflowers with many petals can not be too small.

Lemma 2.4.

For 1sk,t1\leq s\leq k,t, let \mathcal{H} be a tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}. If \mathcal{F} is a sunflower of \mathcal{H} with at least r+1r+1 petals and core YY, then |Y|ks+t|Y|\geq k-s+t.

Proof.

We first observe that every edge hh\in\mathcal{H} intersects YY in at least tt vertices. Indeed, because \mathcal{F} has at least r+1r+1 petals, there exists some petal PP of \mathcal{F} which is disjoint from hh, and having hh and PYP\cup Y as edges in \mathcal{H} implies that |hY|t|h\cap Y|\geq t.

Assume for contradiction that |Y|<ks+t|Y|<k-s+t, and let ZZ be a smallest set of vertices with the property that |hZ|t|h\cap Z|\geq t for all hHh\in H. Observe that |Z||Y|<ks+t|Z|\leq|Y|<k-s+t since YY is a set with this property. By the minimality of |Z||Z|, there exists some hh\in\mathcal{H} which intersects ZZ in exactly tt vertices {z1,,zt}\{z_{1},\ldots,z_{t}\}. Note that h{z1,,zs}h\setminus\{z_{1},\ldots,z_{s}\} is an (rs)(r-s)-set contained in an edge. Moreover, since every edge hh^{\prime} intersects ZZ in at least tt vertices, every edge hh^{\prime} containing this (rs)(r-s)-set must be of the form (h{z1,,zs})S(h\setminus\{z_{1},\ldots,z_{s}\})\cup S with SS an ss-subset of Z{zs+1,,zt}Z\setminus\{z_{s+1},\ldots,z_{t}\} since we have {zs+1,,zt}h{z1,,zs}\{z_{s+1},\ldots,z_{t}\}\subseteq h\setminus\{z_{1},\ldots,z_{s}\}. The number of choices of such ss-sets is exactly (|Z|t+ss)(k1s){|Z|-t+s\choose s}\leq{k-1\choose s}, a contradiction to this (rs)(r-s)-set being contained in more than (k1s){k-1\choose s} edges. We conclude the result. ∎

At this point in the analogous proof of [3] for s=t=1s=t=1 with222There is a small error in [3] where it is claimed that their argument works for rkr\geq k as opposed to just r>kr>k. However, a simple modification of their argument gives a correct proof for the r=kr=k case. r>kr>k, it is argued that if ||nrk1|\mathcal{H}|\gg n^{r-k-1}, then \mathcal{H} contains a set YZY\cup Z of size 2k12k-1 such that every kk-subset of YZY\cup Z is the core of a sunflower with at least r+1r+1 petals. From this observation one can quickly deduce that every edge intersects YZY\cup Z in at least kk vertices, which implies Theorem 1.2.

Essentially this same argument will work in our general setting if one assumes the stronger hypothesis δrs+()(ks)\delta_{r-s}^{+}(\mathcal{H})\geq{k\choose s}, but it fails when considering the weaker hypothesis δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}. For example, if s=t=2s=t=2 and k=4k=4, then one can consider the rr-graph \mathcal{H} which consists of every edge containing at least 4 vertices of {1,2,3,4,5,6}\{1,2,3,4,5,6\} except for the edges which contain {1,2,3,4}\{1,2,3,4\}. This construction has many edges and satisfies δr2+()=5>(k1s)\delta_{r-2}^{+}(\mathcal{H})=5>{k-1\choose s}, but it is not the case that there is a set of size 2k2s+t2k-2s+t such that every (ks+t)(k-s+t)-subset is the core of a sunflower with many edges. Thus from this point onwards we will have to deviate significantly from the approach of [3]. The key definition we need is the following.

Definition 4.

Given integers k,s,tk,s,t and a hypergraph \mathcal{H}, we say that a triple of vertex sets (h,Y,Z)(h,Y,Z) is a bad triple if the following conditions hold:

  • (1)

    The set hh is an edge of \mathcal{H} with |h(YZ)|<ks+t|h\cap(Y\cup Z)|<k-s+t.

  • (2)

    We have |Y|=|Z|=ks+t|Y|=|Z|=k-s+t and |YZ|=2k2s+t|Y\cup Z|=2k-2s+t (or equivalently, |YZ|=t|Y\cap Z|=t).

  • (3)

    The sets Y,ZY,Z are cores of sunflowers Y,Z\mathcal{F}_{Y},\mathcal{F}_{Z}. Moreover, every petal PP of Y\mathcal{F}_{Y} is disjoint from every edge of Z\mathcal{F}_{Z}, and every petal QQ of Z\mathcal{F}_{Z} is disjoint from every edge of Y\mathcal{F}_{Y}.

  • (4)

    For every edge hh^{\prime}\in\mathcal{H} there exist a petal PP of Y\mathcal{F}_{Y} and a petal QQ of Z\mathcal{F}_{Z} such that hP=hQ=h^{\prime}\cap P=h^{\prime}\cap Q=\emptyset.

  • (5)

    For any petal PP of Y\mathcal{F}_{Y}, define

    I(P)={Y:Y(YZ),PY}.I(P)=\{Y^{\prime}:Y^{\prime}\subseteq(Y\cup Z),\ P\cup Y^{\prime}\in\mathcal{H}\}.

    We have I(P)=I(P)I(P)=I(P^{\prime}) for all petals P,PP,P^{\prime} of Y\mathcal{F}_{Y}.

We note that if r=ks+tr=k-s+t, then all of the conditions except (1) are satisfied if there exist two edges Y,ZY,Z with |YZ|=t|Y\cap Z|=t (since one can take Y,Z\mathcal{F}_{Y},\mathcal{F}_{Z} to be sunflowers with 1 edge and the empty set as a petal). If r>ks+tr>k-s+t, then (4) will be satisfied provided Y,Z\mathcal{F}_{Y},\mathcal{F}_{Z} each have at least r+1r+1 edges.

Condition (5) will mostly be used as a technical convenience as follows: if there exists an edge PYP\cup Y^{\prime} with PP a petal of Y\mathcal{F}_{Y} and YYZY^{\prime}\subseteq Y\cup Z a set of size ks+tk-s+t, then (5) guarantees that PYP^{\prime}\cup Y^{\prime} will be an edge for any petal PP^{\prime} of Y\mathcal{F}_{Y}. We also note that (5) is the only condition which is asymmetric in YY and ZZ.

The following two results show that bad triples are the only obstruction to proving Theorem 1.3.

Lemma 2.5.

For 1st,k1\leq s\leq t,k and r=ks+tr=k-s+t, let \mathcal{H} be an nn-vertex tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}. If \mathcal{H} contains no bad triples, then \mathcal{H} is a subset of a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system.

Proof.

The result is trivial if \mathcal{H} is empty, so assume \mathcal{H} contains at least one edge. Let Y,ZY,Z be (possibly non-distinct) edges of \mathcal{H} such that |YZ||Y\cap Z| is as small as possible. We claim that |YZ|=t|Y\cap Z|=t. Indeed, we must have |YZ|t|Y\cap Z|\geq t since \mathcal{H} is tt-intersecting, so assume for contradiction that |YZ|t+1|Y\cap Z|\geq t+1. Let SZYZS_{Z}\subseteq Y\cap Z be an ss-set. Because

|Y(ZSZ)|ks+t(t+1s)=k1,|Y\setminus(Z\setminus S_{Z})|\leq k-s+t-(t+1-s)=k-1,

the positive minimum codegree condition implies there is an ss-set S~\tilde{S} such that Z~:=(ZSZ)S~\tilde{Z}:=(Z\setminus S_{Z})\cup\tilde{S} is an edge with S~Y(ZSZ)\tilde{S}\not\subseteq Y\setminus(Z\setminus S_{Z}), and in particular S~Y\tilde{S}\not\subseteq Y since S~\tilde{S} is disjoint from ZSZZ\setminus S_{Z}. Note that |YZ~|<|YZ||Y\cap\tilde{Z}|<|Y\cap Z| since Z~\tilde{Z} was formed from ZZ by deleting an ss-set SZYS_{Z}\subseteq Y and adding an ss-set S~Y\tilde{S}\not\subseteq Y. This contradicts us choosing ZZ to minimize |YZ||Y\cap Z|, so we conclude that |YZ|=t|Y\cap Z|=t.

As noted before the lemma, if there existed an edge hh with |h(YZ)|<ks+t|h\cap(Y\cup Z)|<k-s+t, then (h,Y,Z)(h,Y,Z) would be a bad triple. Thus no such edge exists by hypothesis. We conclude that \mathcal{H} is a subset of a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system, namely the one consisting of every edge which intersects YZY\cup Z in at least ks+tk-s+t vertices. ∎

Proposition 2.6.

For 1st,k1\leq s\leq t,k and r>ks+tr>k-s+t, let \mathcal{H} be an nn-vertex tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}. There exists a constant C=C(r)C=C(r) such that if ||Cnrk+st1|\mathcal{H}|\geq Cn^{r-k+s-t-1} and \mathcal{H} contains no bad triples, then \mathcal{H} is a subset of a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system.

Proof.

Let \mathcal{H} be as in the proposition. By Proposition 2.3 and Lemma 2.4, we can guarantee, provided CC is sufficiently large in terms of rr, that \mathcal{H} contains a sunflower Y\mathcal{F}_{Y}^{\prime} with core Y={y1,,yks+t}Y=\{y_{1},\ldots,y_{k-s+t}\} and at least

p=(2r+1)f(s,r+1)ks+t+(r+1)222k2s+t+r(r+1)p=(2r+1)f(s,r+1)^{k-s+t}+(r+1)2^{2^{2k-2s+t}}+r(r+1)

petals, where f(s,r+1)f(s,r+1) is as in Theorem 2.2. Analogous to the proof of Lemma 2.5, we wish to show the following.

Claim 2.7.

There is a set of vertices Z={z1,,zks+t}Z=\{z_{1},\ldots,z_{k-s+t}\} such that |YZ|=t|Y\cap Z|=t and such that ZZ is the core of a sunflower Z\mathcal{F}_{Z}^{\prime} with at least 2r+12r+1 petals.

Proof.

For all i0i\geq 0, define g(i)=(2r+1)f(s,r+1)ig(i)=(2r+1)f(s,r+1)^{i}. We say that a (ks+t)(k-s+t)-set ZZ is good if there is a sunflower Z\mathcal{F}_{Z}^{\prime} with core ZZ and at least g(|YZ|)g(|Y\cap Z|) petals. Note that there exists a good set, namely YY. Let ZZ be a good set such that |YZ||Y\cap Z| is as small as possible. We claim that |YZ|=t|Y\cap Z|=t, from which the claim will follow since ZZ is contained in at least g(t)g(0)=2r+1g(t)\geq g(0)=2r+1 petals.

We first observe that |YZ|t|Y\cap Z|\geq t. Indeed if this was not the case, then since ZZ is the core of a sunflower with at least g(0)r+1g(0)\geq r+1 petals, one of these petals QQ must be disjoint from YY. Similarly one can find a petal PP of Y\mathcal{F}_{Y}^{\prime} which is disjoint from QZQ\cup Z, which means the edges QZQ\cup Z and PYP\cup Y intersect in less than tt vertices, a contradiction.

Assume for contradiction that |YZ|t+1|Y\cap Z|\geq t+1. Fix an ss-set SZYZS_{Z}\subseteq Y\cap Z and observe

|Y(ZSZ)|(ks+t)(t+1s)=k1.|Y\setminus(Z\setminus S_{Z})|\leq(k-s+t)-(t+1-s)=k-1.

This implies that for each petal QQ of Z\mathcal{F}_{Z}^{\prime}, there exists some ss-set S~Q\tilde{S}_{Q} such that S~QY\tilde{S}_{Q}\not\subseteq Y and Q(ZSZ)S~QQ\cup(Z\setminus S_{Z})\cup\tilde{S}_{Q}\in\mathcal{H}, as otherwise the (rs)(r-s)-set Q(ZSZ)Q\cup(Z\setminus S_{Z}) would be contained in at most (k1s){k-1\choose s} edges.

Consider the ss-graph 𝒮\mathcal{S} with edge set {S~Q:Q a petal of Z}\{\tilde{S}_{Q}:Q\textrm{ a petal of }\mathcal{F}_{Z}^{\prime}\}. We claim that 𝒮\mathcal{S} does not contain a sunflower with at least r+1r+1 petals. Indeed, assume 𝒮\mathcal{S} had distinct edges S~Q1,,S~Qr+1\tilde{S}_{Q_{1}},\ldots,\tilde{S}_{Q_{r+1}} which all have pairwise intersection WW. In this case \mathcal{H} would contain a sunflower \mathcal{F} with edges Qi(ZSZ)S~QiQ_{i}\cup(Z\setminus S_{Z})\cup\tilde{S}_{Q_{i}} and core (ZSZ)W(Z\setminus S_{Z})\cup W. Note that |W|<s|W|<s since it is the core of an ss-uniform sunflower with more than one edge, so \mathcal{F} is a sunflower in \mathcal{H} with at least r+1r+1 petals and a core of size less than ks+tk-s+t, a contradiction to Lemma 2.4.

With this we have |𝒮|f(s,r+1)|\mathcal{S}|\leq f(s,r+1), and hence there exists some S~𝒮\tilde{S}\in\mathcal{S} such that S~Q=S~\tilde{S}_{Q}=\tilde{S} for at least g(|YZ|)/f(s,r+1)=g(|YZ|1)g(|Y\cap Z|)/f(s,r+1)=g(|Y\cap Z|-1) petals QQ of Z\mathcal{F}_{Z}^{\prime}. Thus the set Z~:=(SSZ)S~\tilde{Z}:=(S\setminus S_{Z})\cup\tilde{S} is the core of a sunflower with at least g(|YZ~|1)g(|Y\cap\tilde{Z}|-1) petals. Note that |YZ~|<|YZ||Y\cap\tilde{Z}|<|Y\cap Z| because SZYS_{Z}\subseteq Y and S~Y\tilde{S}\subsetneq Y by definition of S~Q\tilde{S}_{Q}, a contradiction to us choosing ZZ to be good with |YZ||Y\cap Z| as small as possible. In total this implies |YZ|t|Y\cap Z|\leq t, completing the proof. ∎

With ZZ as in the claim, there are at least r+1r+1 petals Q1,,Qr+1Q_{1},\ldots,Q_{r+1} which are disjoint from YY, and we let Z\mathcal{F}_{Z} denote the sunflower with these petals. Amongst the petals of Y\mathcal{F}_{Y}^{\prime}, there are at least (r+1)222k2s+t(r+1)2^{2^{2k-2s+t}} petals which are disjoint from every edge of Z\mathcal{F}_{Z}, and amongst these petals there must exist at least r+1r+1 petals P1,,Pr+1P_{1},\ldots,P_{r+1} such that I(Pi)=I(Pj)I(P_{i})=I(P_{j}) for all i,ji,j (since there are at most 222k2s+t2^{2^{2k-2s+t}} possible sets that I(P)I(P) could be). Let Y\mathcal{F}_{Y} be the sunflower with core YY using these r+1r+1 petals. We must have |h(YZ)|ks+t|h\cap(Y\cup Z)|\geq k-s+t for all hh\in\mathcal{H}, as otherwise (h,Y,Z)(h,Y,Z) would be a bad triple. This means \mathcal{H} is a subset of a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system, namely the one consisting of every edge which intersects YZY\cup Z in at least ks+tk-s+t vertices. ∎

It remains to argue that hypergraphs as in Theorem 1.3 do not have bad triples.

Lemma 2.8.

For 1sk,t1\leq s\leq k,t, let \mathcal{H} be a tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}. If \mathcal{H} has a bad triple (h,Y,Z)(h,Y,Z) with |hZ|=t|h\cap Z|=t, then |hYZ|<s|h\cap Y\cap Z|<s.

Proof.

Assume for contradiction that there exists a bad triple (h,Y,Z)(h,Y,Z) as in the lemma statement and an ss-set ShYZS\subseteq h\cap Y\cap Z. Let QQ be a petal of Z\mathcal{F}_{Z} which is disjoint from hh, and consider the (rs)(r-s)-set Q(ZS)Q\cup(Z\setminus S). Observe that if h=Q(ZS)Sh^{\prime}=Q\cup(Z\setminus S)\cup S^{\prime} is an edge containing this (rs)(r-s)-set and PP is a petal of Y\mathcal{F}_{Y} disjoint from hh^{\prime}, then

|h(PY)|=|(ZS)Y|+|SY|=ts+|SY|,|h^{\prime}\cap(P\cup Y)|=|(Z\setminus S)\cap Y|+|S^{\prime}\cap Y|=t-s+|S^{\prime}\cap Y|,

so we must have |SY|=s|S^{\prime}\cap Y|=s, which implies SY(ZS):=YS^{\prime}\subseteq Y\setminus(Z\setminus S):=Y^{\prime} since SS^{\prime} must be disjoint from ZSZ\setminus S. As there are more than (k1s){k-1\choose s} edges containing Q(ZS)Q\cup(Z\setminus S), and since |Y|=ks+t(ts)=k|Y^{\prime}|=k-s+t-(t-s)=k, we conclude that for every yYy\in Y^{\prime} there exists an edge hyh_{y} containing yy and Q(ZS)Q\cup(Z\setminus S). Because (h,Y,Z)(h,Y,Z) is a bad triple, we have |h(YZ)|<ks+t=|Y||h\cap(Y\cup Z)|<k-s+t=|Y|, and hence there exists some yYhy\in Y\setminus h. With this we have |hhy|t1|h\cap h_{y}|\leq t-1 because |h(QZ)|=t|h\cap(Q\cup Z)|=t and we construct hyh_{y} by deleting from QZQ\cup Z an ss-set ShZS\subseteq h\cap Z and then adding an ss-set using some yhy\notin h. This contradicts \mathcal{H} being tt-intersecting, giving the result. ∎

Lemma 2.9.

For 1sk,t1\leq s\leq k,t, let \mathcal{H} be a tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}. If \mathcal{H} has a bad triple, then it has a bad triple of the form (h,Y,Z)(h,Y,Z) with |hZ|=t|h\cap Z|=t and |hYZ|ts+1|h\cap Y\cap Z|\geq t-s+1.

Proof.

Let (h,Y,Z)(h,Y,Z) be a bad triple with |hZ||h\cap Z| as small as possible, and assume for contradiction that |hZ|t+1|h\cap Z|\geq t+1. Let ShZS\subseteq h\cap Z be an arbitrary set of ss vertices. Observe that if hh^{\prime} is an edge containing hSh\setminus S and S:=h(hS)S^{\prime}:=h^{\prime}\setminus(h\setminus S), then SZ(hS)S^{\prime}\subseteq Z\setminus(h\setminus S), as otherwise (h,Y,Z)(h^{\prime},Y,Z) would be a bad triple with |hZ|<|hZ||h^{\prime}\cap Z|<|h\cap Z|, contradicting the minimality of (h,Y,Z)(h,Y,Z). We have |Z(hS)|(ks+t)(t+1s)=k1|Z\setminus(h\setminus S)|\leq(k-s+t)-(t+1-s)=k-1, so there are at most (k1s){k-1\choose s} edges containing hSh\setminus S, a contradiction to the minimum positive codegree condition. We conclude that |hZ|t|h\cap Z|\leq t, and we must have |hZ|t|h\cap Z|\geq t since \mathcal{H} is tt-intersecting and Z\mathcal{F}_{Z} contains a petal which is disjoint from hh. We conclude that |hZ|=t|h\cap Z|=t.

Now consider (h,Y,Z)(h,Y,Z) a bad triple with |hZ|=t|h\cap Z|=t and with |hYZ||h\cap Y\cap Z| as large as possible. Asume for contradiction that |hYZ|ts|h\cap Y\cap Z|\leq t-s, which implies there exists an ss-set S(YZ)hS\subseteq(Y\cap Z)\setminus h and also some z(Zh)Yz\in(Z\cap h)\setminus Y. Let PP be a petal of Y\mathcal{F}_{Y}, and observe that the (rs)(r-s)-set P(YS)P\cup(Y\setminus S) intersects any edge QZQ\cup Z with QQ a petal of Z\mathcal{F}_{Z} in tst-s vertices. Thus any edge containing P(YS)P\cup(Y\setminus S) must also contain an ss-set from the set Z(YS)Z\setminus(Y\setminus S). As |Z(YS)|=k|Z\setminus(Y\setminus S)|=k, and since the (rs)(r-s)-set is in more than (k1s){k-1\choose s} edges, there must exist a set SZ(YS)S^{\prime}\subseteq Z\setminus(Y\setminus S) such that P(YS)SP\cup(Y\setminus S)\cup S^{\prime} is an edge with zSz\in S^{\prime}. Define Y=(YS)SY^{\prime}=(Y\setminus S)\cup S^{\prime}, noting that |YZ|=t|Y^{\prime}\cap Z|=t. Observe that |hYZ|>|hYZ||h\cap Y^{\prime}\cap Z|>|h\cap Y\cap Z|, since we formed YY^{\prime} by deleting from YY an ss-set which is disjoint from hh and then added an ss-set containing a vertex in ZhZ\cap h. Also observe that due to condition (5) for bad triples, YY^{\prime} is the core of a sunflower whose petals are the same as Y\mathcal{F}_{Y} (i.e. since PYP\cup Y^{\prime} is an edge and YYZY^{\prime}\subseteq Y\cup Z, we have that PYP^{\prime}\cup Y^{\prime} is an edge for every petal PP^{\prime} of Y\mathcal{F}_{Y}). Further, YZ=YZY^{\prime}\cup Z=Y\cup Z, so (5) continues to hold and we conclude that (h,Y,Z)(h,Y^{\prime},Z) is a bad triple with |hZ|=t|h\cap Z|=t and |hYZ|>|hYZ||h\cap Y^{\prime}\cap Z|>|h\cap Y\cap Z|, a contradiction to our choice of triple. ∎

Note that these two lemmas immediately show that there are no bad triples if t2s1t\geq 2s-1, but we will need to work harder to get the result for all tt. The last tool we need is a particular case of the Kruskal-Katona theorem.

Lemma 2.10 ([12, 13]).

For an ss-graph 𝒮\mathcal{S} and 0is0\leq i\leq s, let i𝒮\partial^{i}\mathcal{S} be the ii-graph with edge set {S:SS,|S|=i}\{S^{\prime}:S^{\prime}\subseteq S\in\mathcal{H},\ |S^{\prime}|=i\}. If |𝒮|=(k1s)|\mathcal{S}|={k-1\choose s}, then |i𝒮|(k1i)|\partial^{i}\mathcal{S}|\geq{k-1\choose i}.

With this we can prove that bad triples do not exist.

Proposition 2.11.

For 1sk,t1\leq s\leq k,t, if \mathcal{H} is a tt-intersecting rr-graph with δrs+()>(k1s)\delta_{r-s}^{+}(\mathcal{H})>{k-1\choose s}, then \mathcal{H} does not contain a bad triple.

Proof.

Assume for contradiction that \mathcal{H} contains a bad triple, so by Lemma 2.9 there exists a bad triple (h,Y,Z)(h,Y,Z) with |hZ|=t|h\cap Z|=t and |hYZ|ts+1|h\cap Y\cap Z|\geq t-s+1. Let ThYZT\subseteq h\cap Y\cap Z be a set of tst-s vertices, and let Sh=(hZ)TS_{h}=(h\cap Z)\setminus T and SY=(YZ)TS_{Y}=(Y\cap Z)\setminus T. Note that Sh,SYS_{h},S_{Y} are ss-sets since h,Yh,Y both intersect ZZ in exactly tt vertices.

Claim 2.12.

Let PP be a petal of Y\mathcal{F}_{Y} which is disjoint from hh. Define

𝒮h={S:|S|=s,(hSh)S},𝒮Y={S:|S|=s,P(YSY)S}.\mathcal{S}_{h}=\{S:|S|=s,\ (h\setminus S_{h})\cup S\in\mathcal{H}\},\hskip 10.00002pt\mathcal{S}_{Y}=\{S:|S|=s,\ P\cup(Y\setminus S_{Y})\cup S\in\mathcal{H}\}.

The set systems 𝒮h,𝒮Y\mathcal{S}_{h},\mathcal{S}_{Y} have the following properties:

  • (a)

    We have |𝒮h|,|𝒮Y|>(k1s)|\mathcal{S}_{h}|,|\mathcal{S}_{Y}|>{k-1\choose s},

  • (b)

    Every S𝒮h𝒮YS\in\mathcal{S}_{h}\cup\mathcal{S}_{Y} is an ss-set which is a subset of the kk-set ZTZ\setminus T,

  • (c)

    The sets 𝒮h,𝒮Y\mathcal{S}_{h},\mathcal{S}_{Y} are disjoint, and

  • (d)

    Every Sh𝒮hS^{\prime}_{h}\in\mathcal{S}_{h} and SY𝒮YS^{\prime}_{Y}\in\mathcal{S}_{Y} satisfy |ShSY|2s+1k|S^{\prime}_{h}\cap S^{\prime}_{Y}|\geq 2s+1-k.

Proof.

Property (a) follows immediately since the (rs)(r-s)-sets hShh\setminus S_{h} and P(YSY)P\cup(Y\setminus S_{Y}) are contained in an edge. For (b), every element of 𝒮h𝒮Y\mathcal{S}_{h}\cup\mathcal{S}_{Y} is an ss-set by definition. Note that the (rs)(r-s)-set hShh\setminus S_{h} intersects ZZ in tst-s vertices. If SS is such that (hSh)S(h\setminus S_{h})\cup S is an edge, then one can choose a petal QQ of Z\mathcal{F}_{Z} which is disjoint from (hSh)S(h\setminus S_{h})\cup S, which means

|((hSh)S)(QZ)|=ts+|SZ|.|((h\setminus S_{h})\cup S)\cap(Q\cup Z)|=t-s+|S\cap Z|.

This implies that we must have |SZ|s|S\cap Z|\geq s, i.e. SZS\subseteq Z. Moreover we have SZTS\subseteq Z\setminus T since SS must be disjoint from (hSh)T(h\setminus S_{h})\supseteq T. An identical argument holds for 𝒮Y\mathcal{S}_{Y}, proving (b).

For (c), assume for contradiction that there existed S𝒮h𝒮YS\in\mathcal{S}_{h}\cap\mathcal{S}_{Y}. Let h=(hSh)Sh^{\prime}=(h\setminus S_{h})\cup S and Y:=(YSY)SY^{\prime}:=(Y\setminus S_{Y})\cup S. Because PYP\cup Y^{\prime} is an edge, condition (5) for bad triples implies YYZY^{\prime}\subseteq Y\cup Z is the core of a sunflower using the same petals as Y\mathcal{F}_{Y}. Thus (h,Y,Z)(h^{\prime},Y^{\prime},Z) is a bad triple with |hZ|=t|h^{\prime}\cap Z|=t and |hYZ||S|=s|h^{\prime}\cap Y^{\prime}\cap Z|\geq|S|=s, a contradiction to Lemma 2.8.

For (d), using Sh,SYZS_{h},S_{Y}\subseteq Z together with |hZ|=t|h\cap Z|=t and |h(YZ)|ks+t1|h\cap(Y\cup Z)|\leq k-s+t-1 implies

|((hSh)(YSY))Z|=|(hY)Z|ks1.|((h\setminus S_{h})\cap(Y\setminus S_{Y}))\setminus Z|=|(h\cap Y)\setminus Z|\leq k-s-1.

We also have |(hSh)(YSY)Z|=|T|=ts|(h\setminus S_{h})\cap(Y\setminus S_{Y})\cap Z|=|T|=t-s, so in total

|(hSh)(YSY)|k+t2s1.|(h\setminus S_{h})\cap(Y\setminus S_{Y})|\leq k+t-2s-1.

Observe that if Sh𝒮hS_{h}^{\prime}\in\mathcal{S}_{h}, then ShS_{h}^{\prime} is disjoint from P(YSY)P\cup(Y\setminus S_{Y}) since ShZTS_{h}^{\prime}\subseteq Z\setminus T. Similarly every SY𝒮YS_{Y}^{\prime}\in\mathcal{S}_{Y} is disjoint from hShh\setminus S_{h}. Thus if Sh𝒮hS_{h}^{\prime}\in\mathcal{S}_{h} and SY𝒮YS^{\prime}_{Y}\in\mathcal{S}_{Y}, for their corresponding edges to intersect in at least tt vertices we must have |ShSY|t(k+t2s1)=2s+1k|S^{\prime}_{h}\cap S^{\prime}_{Y}|\geq t-(k+t-2s-1)=2s+1-k, proving the result. ∎

It remains to show that it is impossible to construct set systems 𝒮h,𝒮Y\mathcal{S}_{h},\mathcal{S}_{Y} with the properties as in this claim. Observe that properties (b), (c), (a) imply that

(ks)|𝒮h𝒮Y|=|𝒮h|+|𝒮Y|2+2(k1s).{k\choose s}\geq|\mathcal{S}_{h}\cup\mathcal{S}_{Y}|=|\mathcal{S}_{h}|+|\mathcal{S}_{Y}|\geq 2+2{k-1\choose s}.

Note that this inequality is equivalent to (k1s1)2+(k1s){k-1\choose s-1}\geq 2+{k-1\choose s}, which is false for k2sk\geq 2s. Thus from now on we may assume k2s1k\leq 2s-1, which in particular means (d) is a non-trivial condition.

Observe that sk2s1s\leq k\leq 2s-1 implies 0kss10\leq k-s\leq s-1, so ks𝒮h\partial^{k-s}\mathcal{S}_{h} is well defined. Let 𝒮\mathcal{S}^{\prime} be the ss-graph with edge set {(ZT)S:Sks𝒮h}\{(Z\setminus T)\setminus S:S\in\partial^{k-s}\mathcal{S}_{h}\}. That is, 𝒮\mathcal{S}^{\prime} is the “complement” of ks𝒮h\partial^{k-s}\mathcal{S}_{h} in ZTZ\setminus T. For every S𝒮S^{\prime}\in\mathcal{S}^{\prime}, there exists some S𝒮hS\in\mathcal{S}_{h} such that |SS|=2sk|S\cap S^{\prime}|=2s-k; namely, this holds whenever SS^{\prime} is the complement of a (ks)(k-s)-subset of SS. By (d), it must be that 𝒮\mathcal{S}^{\prime} and 𝒮Y\mathcal{S}_{Y} are disjoint ss-graphs on ZTZ\setminus T. Using this together with (a), |𝒮|=|ks𝒮h||\mathcal{S}^{\prime}|=|\partial^{k-s}\mathcal{S}_{h}|, and Lemma 2.10 gives

(ks)|𝒮Y𝒮|=|𝒮Y|+|𝒮|>(k1s)+(k1ks)=(k1s)+(k1s1)=(ks).{k\choose s}\geq|\mathcal{S}_{Y}\cup\mathcal{S}^{\prime}|=|\mathcal{S}_{Y}|+|\mathcal{S}^{\prime}|>{k-1\choose s}+{k-1\choose k-s}={k-1\choose s}+{k-1\choose s-1}={k\choose s}.

This is a contradiction, proving the result. ∎

Proof of Theorem 1.3.

Let \mathcal{H} be a hypergraph satisfying the hypothesis of the theorem with the maximum number of edges, and observe that \mathcal{H} has no bad triples by Proposition 2.11. If r<ks+tr<k-s+t, then \mathcal{H} is empty by Lemma 2.1. If r=ks+tr=k-s+t, then \mathcal{H} is contained in a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system by Lemma 2.5, and since \mathcal{H} has the maximum number of edges it must in fact be such a kernel system. Thus we may assume r>ks+tr>k-s+t.

Because (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel systems satisfy the hypothesis of the theorem and have Ω(nrk+st)\Omega(n^{r-k+s-t}) edges, we must have ||=Ω(nrk+st)|\mathcal{H}|=\Omega(n^{r-k+s-t}) as well. Proposition 2.6 implies \mathcal{H} is contained in a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system if nn is sufficiently large, and because \mathcal{H} has the maximum number of edges, it must be such a kernel system, proving the result. ∎

Our proof in fact gives a stability result: if \mathcal{H} is as in Theorem 1.3 with ||nrk+st1|\mathcal{H}|\gg n^{r-k+s-t-1}, then \mathcal{H} is a subset of a (2k2s+t,ks+t)(2k-2s+t,k-s+t)-kernel system. In the special case of r=ks+tr=k-s+t, this conclusion holds regardless of the size of \mathcal{H}.

We made no attempt at optimizing how large nn must be for Theorem 1.3 to hold. A careful analysis shows that our argument works provided nrrs(ks+t)+2r22k2s+tn\approx r^{rs(k-s+t)}+2^{r2^{2k-2s+t}} due to demanding a sunflower with sufficiently many petals to find a sunflower on ZZ, as well as to utilize condition (5) for bad triples. It is not too difficult to provide an alternative proof by replacing (5) with the condition that if r>ks+tr>k-s+t, then the sunflower Y\mathcal{F}_{Y} contains at least (2r)k(s+1|hYZ|)(2r)^{k(s+1-|h\cap Y\cap Z|)} petals. This ultimately gives a bound that works for nrrs(ks+t)n\approx r^{rs(k-s+t)}, which matches the bound of nrrkn\approx r^{rk} implicitly proven in [3] when s=t=1s=t=1. However, this alternative proof is slightly more involved, so we elected to use the current argument at the cost of weaker (implicit) bounds.

As in [3], one may be able to prove that Theorem 1.3 holds for nrs(ks+t)n\approx r^{s(k-s+t)} if kk is small, and in particular one may be able to adapt the ideas of [3] to prove such a bound when k=s+1k=s+1.

Acknowledgments. The author thanks Jason O’Neill for engaging conversations and comments on an earlier draft, and Cory Palmer for clarifying some of the points from [3].

References

  • [1] R. Ahlswede and L. H. Khachatrian. The complete intersection theorem for systems of finite sets. European Journal of Combinatorics, 18(2):125–136, 1997.
  • [2] R. Alweiss, S. Lovett, K. Wu, and J. Zhang. Improved bounds for the sunflower lemma. In Proceedings of the 52nd Annual ACM SIGACT Symposium on Theory of Computing, pages 624–630, 2020.
  • [3] J. Balogh, N. Lemons, and C. Palmer. Maximum size intersecting families of bounded minimum positive co-degree. SIAM Journal on Discrete Mathematics, 35(3):1525–1535, 2021.
  • [4] P. Erdős, C. Ko, and R. Rado. Intersection theorems for systems of finite sets. The Quarterly Journal of Mathematics, 12(1):313–320, 1961.
  • [5] P. Erdős and R. Rado. Intersection theorems for systems of sets. Journal of the London Mathematical Society, 1(1):85–90, 1960.
  • [6] P. Frankl. On intersecting families of finite sets. Journal of Combinatorial Theory, Series A, 24(2):146–161, 1978.
  • [7] P. Frankl. Erdös-Ko-Rado theorem with conditions on the maximal degree. Journal of Combinatorial Theory, Series A, 46(2):252–263, 1987.
  • [8] P. Frankl and N. Tokushige. Invitation to intersection problems for finite sets. Journal of Combinatorial Theory, Series A, 144:157–211, 2016.
  • [9] P. Frankl and N. Tokushige. Extremal problems for finite sets, volume 86. American Mathematical Soc., 2018.
  • [10] Z. Füredi. On finite set-systems whose every intersection is a kernel of a star. Discrete mathematics, 47:129–132, 1983.
  • [11] C. Godsil and K. Meagher. Erdos-Ko-Rado theorems: algebraic approaches. Number 149. Cambridge University Press, 2016.
  • [12] G. Katona. A theorem of finite sets. In Classic Papers in Combinatorics, pages 381–401. Springer, 2009.
  • [13] J. B. Kruskal. The number of simplices in a complex. Mathematical optimization techniques, 10:251–278, 1963.
  • [14] N. Lemons and C. Palmer. The unbalance of set systems. Graphs and Combinatorics, 24(4):361–365, 2008.
  • [15] D. Mubayi and Y. Zhao. Forbidding complete hypergraphs as traces. Graphs and Combinatorics, 23(6):667–679, 2007.