Turán number for bushes

Zoltán Füredi∗ Alexandr Kostochka†

Abstract

Let a, b ∈ Z+ , r = a + b, and let T be a tree with parts U = {u1 , u2 , . . . , us } and V =

{v1 , v2 , . . . , vt }. Let U1 , . . . , Us and V1 , . . . , Vt be disjoint sets, such that |Ui | = a and |Vj | = b for
arXiv:2307.04932v1 [math.CO] 10 Jul 2023

all i, j. The (a, b)-blowup of T is the r-uniform hypergraph with edge set {Ui ∪ Vj : ui vj ∈ E(T )}.
We use the ∆-systems method to prove the following Turán-type result. Suppose a, b, t ∈ Z+ ,
r = a + b ≥ 3, a ≥ 2, and T is a fixed tree of diameter 4 in which the degree
of the center vertex
n
is t. Then there exists a C = C(r, t, T ) > 0 such that |H| ≤ (t − 1) r−1 + Cnr−2 for every
n-vertex r-uniform hypergraph H not containing an (a, b)-blowup of T . This is asymptotically
exact when t ≤ |V (T )|/2. A stability result is also presented.
Mathematics Subject Classification: 05D05, 05C65, 05C05.
Keywords: Hypergraph trees, extremal hypergraph theory, Delta-systems.

1 Introduction

1.1 Basic definitions and notation

An r-uniform hypergraph (an r-graph, for short), is a family of r-element subsets of a finite set. We
associate an r-graph H with its edge set and call its vertex set V (H). Often we take V (H) = [n],
where [n] := {1, 2, 3, . . . , n}. Given an r-graph F, let the Turán number of F, exr (n, F), denote
the maximum number of edges in an r-graph on n vertices that does not contain a copy of F.
Since a (graph) tree is connected and bipartite, it uniquely defines the parts in its bipartition. So,
we say a tree T is an (s, t)-tree if one part of V (T ) has s vertices and the other has t vertices.
Let s, t, a, b > 0 be integers, r = a + b, and let T = T (X, Y ) be an (s, t)-tree with parts U =
{u1 , u2 , . . . , us } and V = {v1 , v2 , . . . , vt }. Let U1 , . . . , Us and V1 , . . . , Vt be pairwise disjoint sets,
S
such that |Ui | = a and |Vj | = b for all i, j. So | Ui ∪ Vj | = as + bt. The (a, b)-blowup of T , denoted
by T (T, a, b), is the r-uniform hypergraph with edge set

T (T, a, b) := {Ui ∪ Vj : ui vj ∈ E(T )}.

The goal of this paper is to find the asymptotics of the Turán number for (a, b)-blowups of many
trees of radius 2 using the ∆-systems method. Earlier, (a, b)-blowups of different classes of trees
and different pairs (a, b) were considered in [7]. The main result in [7] is the following.
∗
Alfréd Rényi Institute of Mathematics, Budapest, Hungary. E-mail: z-furedi@illinois.edu. Research partially
supported by National Research, Development and Innovation Office NKFIH grants 132696 and 133819.
†
University of Illinois at Urbana–Champaign, Urbana, IL 61801 and Sobolev Institute of Mathematics, Novosibirsk
630090, Russia. E-mail: kostochk@illinois.edu. Research supported in part by NSF grant DMS-2153507 and NSF
RTG grant DMS-1937241.

1
Füredi and Kostochka: Bushes, July 4, 2023 2

u1 u2 us U Us
|U|=s 1
|Ui|=a
|V|=t V1 Vt |Vj |=b
v1 vt ⇒

Figure 1: An example of a (4, 2)-blowup.

Theorem 1 ([7]). Suppose r ≥ 3, s, t ≥ 2, a + b = r, b < a < r. Let T be an (s, t)-tree and let
T = T (T, a, b) be its (a, b)-blowup. Then (as n → ∞) any T -free n-vertex r-graph H satisfies
 
n
|H| ≤ (t − 1) + o(nr−1 ).
r−1

This is asymptotically sharp whenever t ≤ s.

This theorem asymptotically settles about a half of possible cases, but when t < s and a ≤ b,
it is expected that the asymptotics is different from the one in Theorem 1. More is known on
(a, b)-blowups of paths. Let Pℓ denote the (graph) path with ℓ edges.
The case of P2 was resolved asymptotically by Frankl [3] (for b = 1) and by Frankl and Füredi [4]
(for all 1 ≤ a ≤ r − 2 and b = r − a):
 
exr (n, T (P2 , a, b)) = Θ nmax{a−1,b} . (1)

The case of P3 was fully solved for large n by Füredi and Özkahya [8]. They showed that for fixed
1 ≤ a, b < r with r = a + b and for n > n0 (r),
 
n−1
exr (n, T (P3 , a, b)) = . (2)
r−1

For longer paths, the following was proved in [7].

Theorem 2 (Theorem 1 in [7]). Let a + b = r, a, b ≥ 1 and ℓ ≥ 3. Suppose further that (i) ℓ is
odd, or (ii) ℓ is even and a > b, or (iii) (ℓ, a, b) = (4, 1, 2).
Then   
ℓ−1 n
exr (n, T (Pℓ , a, b)) = + o(nr−1 ).
2 r−1

So, the situation with blowups of Pℓ is not resolved for the case when ℓ ≥ 4 is even and a ≤ b apart
from the case (ℓ, a, b) = (4, 1, 2).

In this paper, we first consider (a, b)-blowups of special trees of radius 2.

A graph bush Bt,h is the radius 2 tree obtained from the star K1,t by joining each vertex of degree
one to other h new vertices. So Bt,h has 1 + t + th vertices. Let s = 1 + th. Then Bt,h is an
(s, t)-tree with s > t.
Füredi and Kostochka: Bushes, July 4, 2023 3

Suppose that a, b, t, h are positive integers, a + b = r and t ≥ 2. By (a, b, t, h)-bush, Bt,h (a, b), we
will call the (a, b)-blowup of Bt,h . This means the center vertex of Bt,h is replaced by an a-set A,
its neighbors by the b-sets B1 , . . . , Bt and its second neighbors by a-sets Ai,j , i ∈ [t], j ∈ [s].

In particular, the (a, b)-blowup of the path P4 is the (a, b, 2, 1)-bush B2,1 (a, b).

Since Bt,h (a, b) has t disjoint edges Bi ∪Ai,1 for i = 1, . . . , t, the example of the r-uniform hypergraph
with vertex set [n] in which every edge intersects the interval [t − 1] shows that
     
n n−t+1 n−1
exr (n, Bt,h (a, b)) ≥ − ∼ (t − 1) . (3)
r r r−1

We will use the ∆-systems approach to show that this is asymptotically correct in many cases.
Recall that for a > b ≥ 2 the asymptotic equality follows from Theorem 1. In this paper we deal
with all cases and also present a somewhat refined result by considering shadows of hypergraphs.
Recall that for an r-graph H the shadow, ∂H, is the collection of (r − 1)-sets that lie in some edge
of H. Our first main result is the following.

Theorem 3. Suppose that a, b, t, h are positive integers, r = a + b ≥ 3. Also suppose that in case
of (a, b) = (1, r − 1) we have h = 1. Then there exists a C = C(r, t, h) > 0 such that the following
holds. If H is an n-vertex r-uniform family not containing a bush Bt,h (a, b) then

|H| ≤ (t − 1)|∂H| + Cnr−2 . (4)

This implies that (3) is asymptotically exact in these cases as r, t, h are fixed and n → ∞. Note
that for (a, b) = (1, r − 1) we prove (4) only for h = 1. In fact, in this case (4) does not hold for
h ≥ 2. An example is this: V (H) = [n], A = [t − 1] and H = E1 ∪ E2 , where E1 is the set of
r-subsets of [n] with exactly one vertex in A and E2 is the Steiner system S1 (n − t + 1, r, r − 1) on
n
[n] − A. This example has asymptotically (t − 1 + 1r ) r−1


edges and does not contain Bt,2 (1, r − 1).
By increasing h, we can get examples without Bt,h (1, r − 1) that have even more edges.
McLennan [9] proved that in the graph case (a = b = 1) we have exr (n, Bt,h ) = 21 (t+th−1)n+O(1),
some extremal graphs are vertex-disjoint unions of complete graphs Kt+th . For the case t = 1 the
inequality (4) is rather weak, exr (n, B1,h (a, b)) = O(nr−2 ) is known for a ≥ 2. Even better bounds
were proved in [5]. So we usually suppose that t ≥ 2, r ≥ 3.
Since each tree of diameter 4 with the degree of the center equal to t is a subgraph of a graph bush
Bt,h for some h, Theorem 3 yields the following somewhat more general result.

Theorem 4. Suppose a, b, t ∈ Z+ , r = a + b ≥ 3 and a ≥ 2. Let T be a fixed tree of diameter 4

in which the degree of the center vertex is t. Then there exists a C = C(r, t, T ) > 0 such that the
following holds. If H is an n-vertex r-graph not containing an (a, b)-blowup of T , then
 
n
|H| ≤ (t − 1) + Cnr−2 . (5)
r−1

We also use the ∆-systems approach to show that for many a ≥ 2 and b ≥ 2 with a + b = r,
Füredi and Kostochka: Bushes, July 4, 2023 4

the r-uniform hypergraphs without Bt,h (a, b) of cardinality ”close” to extremal contain vertices of
”huge” degree.
Theorem 5. Suppose that a, b, t, h are positive integers, a, b ≥ 2 and r = a + b ≥ 5. Then for any
C0 > 0 there exist n0 > 0 and C1 > 0 such that the following holds. If n > n0 , H is an n-vertex
r-uniform family not containing a bush Bt,h (a, b) and |H| > (t − 1) n−1 r−2 , then there are


r−1 − C0 n
n−1 r−2


t − 1 vertices in [n] each of which is contained in at least r−1 − C1 n edges of H.

The structure of this paper is as follows. In the next section, we discuss the ∆-system method
and present a lemma by Füredi [6] from 1983 that will be our main tool. In Section 3 we describe
properties of so called intersection structures. It allows us to prove the main case of Theorem 3
(the case a ≥ 2) in Section 4 and the case of a = 1 and h = 1 in Section 5. Then in Section 6 we
prove Theorem 5.

2 Definitions for the ∆-system method and a lemma

A family of sets {F1 , . . . , Fs } is an s-star or a ∆-system or an s-sunflower of size s with kernel A,

if Fi ∩ Fj = A for all 1 ≤ i < j ≤ s. The sets Fi \ A are called petals.
For a member F of a family F, let the intersection structure of F relative to F be

I(F, F) = {F ∩ F ′ : F ′ ∈ F \ {F }}.

An r-uniform family F ⊆ [n]

r is r-partite if there exists a partition (X1 , . . . , Xr ) of the vertex set
[n] such that |F ∩ Xi | = 1 for each F ∈ F and each i ∈ [r].
For a partition (X1 , . . . , Xr ) of [n] and a set S ⊆ [n], the pattern Π(S) is the set {i ∈ [r] : S∩Xi ̸= ∅}.
Naturally, for a family L of subsets of [n],

Π(L) = {Π(S) : S ∈ L} ⊆ 2[r] .

Lemma 6 (The intersection semilattice lemma (Füredi [6])). For any positive integers s and r,
there exists a positive constant c(r, s) such that every family F ⊆ [n] ∗


r contains a subfamily F ⊆ F
satisfying
1. |F ∗ | ≥ c(r, s)|F|.
2. F ∗ is r-partite, together with an r-partition (X1 , . . . , Xr ).
3. There exists a family J of proper subsets of [r] such that Π(I(F, F ∗ )) = J holds for all F ∈ F ∗ .
4. J is closed under intersection, i.e., for all A, B ∈ J we have A ∩ B ∈ J , as well.
5. For any F ∈ F ∗ and each A ∈ I(F, F ∗ ), there is an s-star in F ∗ containing F with kernel A.

Remark 1. The proof of Lemma 6 in [6] yields that if F itself is r-partite with an r-partition
(X1 , . . . , Xr ), then the r-partition in the statement can be taken the same.

Remark 2. By definition, if for some M ⊂ [r] none of the members of the family J of proper
Füredi and Kostochka: Bushes, July 4, 2023 5

subsets of [r] in Lemma 6 contains M , then for any two sets F1 , F2 ∈ F ∗ , their
 intersections
m with
S ∗
Q n−(r−m)
j∈M Xj are distinct. It follows that if |M | = m, then |F | ≤ j∈M |Xj | ≤ m . Thus, if
m
|F ∗ | > n−r+m


m , then every m-element subset of [r] is contained in some B ∈ J .
Call a family J of proper subsets of [r] m-covering if every m-element subset of [r] is contained in
some B ∈ J . In these terms, Remark 2 says that

n−r+m m
if |F ∗ | >


m , then the corresponding J is m-covering. (6)

For k = 0, 1, . . . , r, define the family J (k) of proper subsets of [r] as follows. It contains
(a) the sets [r] − {i} for 1 ≤ i ≤ k,
(b) all (r − 2)-element subsets of [r] containing {1, 2, . . . , k}, and
(c) all the intersections of these subsets.
By definition, each J (k) is (r − 2)-covering. Moreover,

each (r − 2)-covering family of proper subsets of [r] closed under intersections contains
(7)
a subfamily isomorphic to some J (k) .

Indeed, if an (r − 2)-covering family J of proper subsets of [r] contains exactly k sets of size r − 1,
then it must contain as members all (r − 2)-element subsets of [r] not contained in these k sets, so
properties (a) and (b) of the definition hold. Part (c) follows since J is closed under intersections.

3 General claims on intersection structures.

Call a set B a (b, s)-kernel in a set system F if B is the kernel of size b in a sunflower with s petals
formed by members of F.
Lemma 7. If a + b = r and an r-uniform family H does not contain Bt,h (a, b), then there do not
exist disjoint sets A0 , B1 , B2 , . . . , Bt with |A0 | = a, |B1 | = . . . = |Bt | = b such that all B1 , . . . , Bt
are (b, thr)-kernels in H and the sets A0 ∪ B1 , . . . , A0 ∪ Bt are edges of H.

Proof. Suppose, there are such disjoint sets A0 , B1 , B2 , . . . , Bt . Let D0 = A0 ∪ tj=1 Bj . For
S

i = 1, . . . , t, do the following. Since Bi is a (b, thr)-kernel and |Di−1 \ Bi | = a + (t − 1)b + (i − 1)ha ≤

thr − h, there exist h petals Ai,j (1 ≤ j ≤ h) of a thr-sunflower with kernel Bi which are disjoint
from Di−1 . Let Di = Di−1 ∪j Yi,j . After t steps, we find a Bt,h (a, b) whose edges are A0 ∪ Bi and
Bi ∪ Ai,j for i = 1, . . . , t, j = 1, 2, . . . , h.

Suppose a + b = r and G ⊂ [n] 1

nr−2 does not contain Bt,h (a, b). By Lemma 6


r with |G| > c(r,thr)
and (6), there is G ∗ ⊆ G satisfying the lemma such that the corresponding family J of proper
subsets of [r] is (r − 2)-covering. Let (X1 , . . . , Xr ) be the corresponding partition.
Lemma 8. Family J does not contain disjoint members A and B such that |A| = a and |B| = b.

Proof. Suppose, it does. By renaming the elements of J , we may assume that A = {1, . . . , a} and
B = {a+1, . . . , r}. Let X = {x1 , . . . , xr } ∈ G ∗ , where xi ∈ Xi for all i. Since A ∈ J , {x1 , . . . , xa } is
Füredi and Kostochka: Bushes, July 4, 2023 6

an (a, thr)-kernel in G ∗ . Let B1 , . . . , Bt be some t petals in the sunflower with kernel {x1 , . . . , xa }.
Since [r] − [a] = B ∈ J , each of B1 , . . . , Bt is a (b, thr)-kernel in G ∗ , contradicting Lemma 7.

Lemma 9. If a + b = r and 2 ≤ a, b ≤ r − 2, then for each 0 ≤ k ≤ r − 2 and for k = r the family

J (k) has disjoint members A and B such that |A| = a and |B| = b, unless (r, a, b, k) = (4, 2, 2, 1).
If (a, b) = (1, r − 1) or (a, b) = (r − 1, 1) and r ≥ 3, then for each 1 ≤ k ≤ r − 2 and for k = r the
family J (k) has disjoint members A and B such that |A| = a and |B| = b.

Proof. If k ≥ a, then we let A = [a], B = [r] − [a], and represent them as follows:
\ \ \
A= ([r] − {i, i′ }) ∩ ([r] − {i}), B = ([r] − {i}).
k+1≤i<i′ ≤r a+1≤i≤k 1≤i≤a

If k ≥ b, then we have a symmetric representation. In particular, this proves the claim for (a, b) =
(1, r − 1) or (a, b) = (r − 1, 1).
If k ≤ a − 2, then we again let A = [a], B = [r] − [a], but represent them as follows (using that
a ≤ r − 2 and k ≤ a − 2):
\ \ \
A= ([r] − {i, i′ }), B = ([r] − {i}) ∩ ([r] − {i, i′ }).
a+1≤i<i′ ≤r 1≤i≤k k+1≤i<i′ ≤a

By symmetry, the only remaining case is that k = a−1 = b−1. So r is even, and a = b = r/2 = k+1.
In this case, if r > 4, then we let A = [k − 1] ∪ {k + 1, k + 2}, B = ([r] − [k + 2]) ∪ {k}, and represent
them as follows:
\ \
A= ([r] − {i, i′ }) ∩ ([r] − {k}), B = ([r] − {i}) ∩ ([r] − {k + 1, k + 2}).
k+2≤i<i′ ≤r 1≤i≤k−1

4 Proof of the main Theorem

4.1 Basic procedure

Now we are ready to prove Theorem 3. Assume that an n-vertex r-uniform family H does not
contain Bt,h (a, b). Define C = C(r, t, h) := 1/c(r, thr), where c is from Lemma 6. For any r-uniform
family G, let G ∗ denote a family satisfying Lemma 6 and J (G ∗ ) ⊂ 2[r] denote the corresponding
intersection structure.
Do the following procedure. Let H1 = (H)∗ and J1 = J (H∗ ). For i = 1, 2, . . . , if |H \ ij=1 Hj | ≤
S

nr−2 /c(r, thr), then stop and let m := i and H0 = H\ ij=1 Hj ; otherwise, let Hi+1 := (H\ ij=1 Hj )∗
S S

and Ji+1 = J ((H \ ij=1 Hj )∗ ).

S

This procedure provides a partition of H, H = m

S Sm
i=0 Hi . Let H denote i=1 Hi . By definition,
b
r−2
|H0 | ≤ Cn , so we get
b + Cnr−2 ≥ |H|.
|H| (8)
Füredi and Kostochka: Bushes, July 4, 2023 7

We distinguish four cases, a, b ≥ 2 and r ≥ 5 (discussed in Subsection 4.2), (a, b) = (2, 2) and r = 4
(Subsection 4.3), (a, b) = (r − 1, 1) and r ≥ 3 (Subsection 4.4), and finally (a, b) = (1, r − 1), r ≥ 3
and h = 1 (Section 5).

4.2 Case of a, b ≥ 2 and r ≥ 5

Here 2 ≤ a, b ≤ r − 2 and (a, b) ̸= (2, 2), so Lemma 9 implies that for each 1 ≤ i ≤ m, J (Hi ) has
exactly r − 1 (r − 1)-subsets, it contains J (r−1) . Hence for each hyperedge E ∈ Hi ⊂ H b (1 ≤ i ≤ m)
there exists an element c(E) ∈ E such that each proper subset of E containing c(E) is a kernel of a
thr-star in Hi . Beware of the fact that although each Hi is k-partite, the partitions might differ for
different values of i. This does not cause any problem in our argument, we only need the existence
of the element c(E).
[n]


Define the function α on r−1 as follows: Given an (r − 1)-set Y , let α(Y ) be the number of edges
E ∈ H with Y = E \ {c(E)}.
b

Claim 10. α(Y ) ≤ t − 1, for all (r − 1)-subsets Y of [n].

Proof. Suppose to the contrary that there are t distinct E1 ∈ Hi1 , . . . , Et ∈ Hit such that Ei =
Y ∪ {c(Ei )}. If ij = ij ′ for some j ̸= j ′ , then the intersection structure of Hij would contain the
(r − 1)-set Y , a contradiction. Thus i1 , . . . , it are all distinct. By relabelling we may suppose that
Ei ∈ Hi .
We will find the t + 1 disjoint sets A0 , B1 , B2 , . . . , Bt contradicting Lemma 7 using induction as
follows. Let A0 ⊂ Y arbitrary, |A0 | = a and let D0 := ∪Ei , |D0 | = t + r − 1. We define the sets
Ei′ , Di , Bi step by step as follows. We will have Di := D0 ∪j≤i Ej′ and |Di | = t + r − 1 + i(r − 1 − a).
For i = 1, 2, . . . , t consider the family Hi and its member Ei in it. By the intersection structure of
Hi , the set A0 ∪ {c(Ei )} is an (a + 1, thr)-kernel in Hi . One of the thr petals of the sunflower in Hi
with kernel A0 ∪ {c(Ei )} should be disjoint from Di−1 ; let Ei′ be the corresponding set in Hi . Since
c(Ei ) ∈ Ei′ , and homogeneity gives c(Ei′ ) = c(Ei ), the set Bi := Ei′ − A0 is a (b, thr)-kernel.

Claim 10 implies X
(t − 1)|∂ H|
b ≥ α(Y ) = |H|.
b (9)
∀(r−1)-setY

This, together with (8), completes the proof of Theorem 3 in this case.

4.3 The case (a, b) = (2, 2)

Lemma 9 implies that for each 1 ≤ i ≤ m, either

— J (3) is contained in J (Hi ), it has exactly three 3-subsets, so [4] \ {1}, [4] \ {2}, and [4] \ {3} are
in J and J also contains all subsets containing the element 4 but it does not contain 123, or
— J (Hi ) is of J (1) type, it has a unique 3-subset, 234, and {1, 12, 13, 14} ⊂ J .
Call Hi (and its edges) type α if J (Hi ) has three 3-subsets. Each of these edges E ∈ Hi has an
element c(E) ∈ E such that each proper subset of E containing c(E) is a kernel of a 4 · t · h-star in
Hi . The union of the families Hi of type α is H bα . Call Hi (and its edges) type β if J (Hi ) has a
Füredi and Kostochka: Bushes, July 4, 2023 8

unique 3-subset. Each of these edges E ∈ Hi has an element b(E) ∈ E such that each set K ⊂ E
of the form {b(E), x} (x ∈ E \ {b(E)}) and the set E \ {b(E)} is a kernel of a 4 · t · h-star in Hi .
The union of the families Hi of type β is H
bβ .

Define the function α on [n]

3 as follows: Given an 3-set Y
let α(Y ) be the number of edges E ∈ Hα
b
[n]
with Y = E \ {c(E)}. Define the function β : H b→
r−1 as follows: Given a 3-set Y , let β(Y ) be
the number of edges E ∈ Hβ with b(E) ∈ Y ⊂ E.
b

Claim 11. α(Y ) + 13 β(Y ) ≤ t − 1, for all (r − 1)-subsets Y of [n].

Proof. Let E1 ∈ Hi1 , . . . , Eα ∈ Hiα be the α(Y ) distinct edges with E ∈ H bα and Y = E\{c(E)} and
let Eα+1 ∈ Hiα+1 , . . . , Eα+β ∈ Hiα+β be the β(Y ) distinct edges with E ∈ H bβ and b(E) ∈ Y ⊂ E.
′
If ij = ij ′ for some j ̸= j , then the intersection structure of Hij would contain the 3-set Y , a
contradiction. Thus i1 , i2 , . . . are all distinct. By relabelling we may suppose that Ei ∈ Hi .
Suppose to the contrary that α(Y ) + 31 β(Y ) > t − 1, so α + ⌈β/3⌉ ≥ t. Since |Y | = 3, one can find
an element y0 ∈ Y such that b(Ej ) = Y at least ⌈β/3⌉ times. So we may suppose that there are t
distinct Ei ∈ Hi such that the elements c(E1 ), . . . , c(Eα ) and d(Ej ) := Ej \ Y for α < j ≤ t are all
distinct and Ei = Y ∪ {c(Ei )}, b(Ej ) = y0 .
We will find the t + 1 disjoint sets A0 , B1 , B2 , . . . , Bt contradicting Lemma 7 using induction as
follows. Let A0 : Y \{y0 }, |A0 | = 2 and let D0 := ∪Ei , |D0 | = t+3. We define the sets Ei′ , Di , Bi step
by step as follows. We will have Di := D0 ∪j≤i Ej′ and |Di | = t + 3 + i. For i = 1, 2, . . . , t, consider
the family Hi and its member Ei in it. By the intersection structure of Hi , the set A0 ∪ {c(Ei )}
(and the set A0 ∪ {c(Ei )}) is an (a + 1, 4 · t · h)-kernel in Hi . One of the 4 · t · h petals of the sunflower
in Hi with kernel A0 ∪ {c(Ei )} should be disjoint from Di−1 ; let Ei′ be the corresponding set in Hi .
Since c(Ei ) ∈ Ei′ , the set Bi := Ei′ − A0 is an (b, 4 · t · h)-kernel.

Claim 11 implies
X  1

(t − 1)|∂ H|
b ≥ α(Y ) + β(Y ) = |H
b α | + |H
bβ |. (10)
3
∀3-setY

This, together with (8), completes the proof of Theorem 3 in this case.

4.4 The case (a, b) = (r − 1, 1)

Call Hi (as above) of type α if J (Hi ) has r − 1 (r − 1)-subsets. Each of these edges E ∈ Hi has an
element c(E) ∈ E such that each proper subset of E containing c(E) is a kernel of a thr-star in Hi .
The union of these Hi families is H bα . Call Hj (and its edges) type β if J (Hj ) has no (r − 1)-subset.
Note that each element y of an edge E ∈ Hj is a kernel of a thr-star in Hj . The union of these Hj
families is H bβ .
[n]


The definition of α on r−1 is the same as in the previous subsections, the definition of β : H b→
[n]


r−1 is even simpler: β(Y ) is the number of edges E ∈ H bβ with Y ⊂ E. If α(Y )+β(Y ) > t−1, then
taking A0 := Y and Bi := Ei \ Y one can see that each Bi is a kernel of a large star, contradicting
Lemma 7.
Füredi and Kostochka: Bushes, July 4, 2023 9

Finally, we complete the proof of this case as follows

X
(t − 1)|∂ H|
b ≥ (α(Y ) + β(Y )) = |H
bα | + r|H
bβ | ≥ |H|.
b (11)
∀(r−1)-setY

5 Hypergraphs without a bush Bt (1, r − 1)

Call an r-graph t-normal if it has no (r−1)-tuples of vertices whose codegree is positive but less than
t. For every edge Y in an r-graph H and any x ∈ Y , let QH (Y, x) = {y ∈ V (H)−x : Y −x+y ∈ H}.
In a t-normal r-graph H, |QH (Y, x)| ≥ t − 1 for every edge Y ∈ H and vertex x ∈ Y .
We will prove this case by induction on the number of edges. For |H| < α(r, t) the claim is trivial.
Suppose H is a counter-example with the fewest edges. If our H is not t-normal, then choose a
(r − 1)-tuple Y of vertices whose codegree is positive but less than t and let H′ be obtained from
H the edges containing Y . Then |H′ | − (t − 1)|∂H′ | ≤ |H| − (t − 1)|∂H|, so H′ satisfies (8) and is
Bt (1, r − 1)-free. This contradicts the minimality of H. Thus H is t-normal.
Let q ∗ be a huge number in terms of rt but small w.r.t. α(r, t)
By (8), there is a subfamily H∗ of H satisfying Lemma 6 for q = q ∗ with |H∗ | > nr−2 . Let J be
the corresponding intersection structure. By Lemmas 8 and 9, J contains a family isomorphic to
J (r−1) , or to J (0) . In both cases, J contains a singleton. So, H contains an element u and disjoint
(r − 1)-tuples U1 , . . . , Uq∗ such that Yi = Ui + u is an edge in H∗ for each 1 ≤ i ≤ q ∗ .
Since H is t-normal, for each 1 ≤ i ≤ q ∗ we can choose a (t − 1)-element subset Q′H (Yi , u) of
QH (Yi , u). Construct the auxiliary bigraph R with parts {Ui : 1 ≤ i ≤ q ∗ } and V (H) where
Ui v ∈ E(R) iff v ∈ Q′H (Yi , u).

Case 1. R has a matching of size (r + 1)t, say M = {U1 , v1 , . . . , U(r+1)t v(r+1)t }. Since all Ui here
are disjoint and all vi are distinct, we can greedily find t disjoint sets Ui ∪ {vi }: start from M and
one by one take a set Ui ∪ {vi } and delete from the list all Ui′ ∪ {vi′ } with i′ ̸= i such that vi ∈ Ui′
or vi′ ∈ Ui′ . If our disjoint sets are U1 ∪ {v1 }, . . . , Ut ∪ {vt }, then we have Bt (1, r − 1) with the set
of edges {U1 ∪ {u}, . . . , Ut ∪ {u}, U1 ∪ {v1 }, . . . , Ut ∪ {vt }}

Case 2. R has no matching of size (r + 1)t. Then it has a vertex cover of size less than (r + 1)t.
So there is a set V ⊂ V (H) with |V | < (r + 1)t containing Q′H (Yi , u) for at least q ∗ − (r + 1)t
sets Yi . Then some (t − 1)-element set T = T (u) = {z1 , . . . , zt−1 } serves as Q′H (Yi , u) for at least

−1
(q ∗ − (r + 1)t) (r+1)t
t−1 sets Yi . We choose q ∗ so that this number is at least t − 1 and suppose
T (u) serves for Y1 , . . . , Yt−1 . If for some other Yi , Yi ∩ T = ∅ and Q′H (Yi , u) ̸= T , then we take an
element z ∈ Q′H (Yi , u) − T and construct a Bt (1, r − 1) with the set of edges

{Y1 ∪ {u}, . . . , Yt−1 ∪ {u}, Yi ∪ {u}} ∪ {Y1 ∪ {z1 }, . . . , Yt−1 ∪ {zt−1 }, Yi ∪ {z}}. (12)

So, T = Q′H (Yi , u) for all 1 ≤ i ≤ q ∗ such that Yi ∩ P = ∅. Hence if for at least one other edge Y
in H containing u and disjoint from T , QH (Yi , u) ̸= T , then we find a Bt (1, r − 1) similarly to (12)
Füredi and Kostochka: Bushes, July 4, 2023 10

because q ∗ is large. It follows that denoting T ′ (u) = T (u) ∪ {u},

the set of edges of H containing u and disjoint from T ′ (u) − z is contained in the set
(13)
of edges of H containing zj and disjoint from T ′ (u) − zj for each 1 ≤ j ≤ t − 1.

Case 2.1. J contains a family isomorphic to J (r−1) . We may assume that J contains all proper
subsets of [r] containing 1 and that {x1 , . . . , xr } ∈ H where xi ∈ Xi for i ∈ [r]. Since {1, 2} ∈ J ,
there are 2t disjoint sets Uh = {x3,h , x4,h , . . . , xr,h } such that Uh′ = Uh ∪ {x1 , x2 } ∈ H∗ for all
1 ≤ h ≤ 2t. At most t − 1 of them intersect T . So we may assume U1′ , . . . , Ut′ are disjoint
from T . Since [r] − {2} ∈ J , for every 1 ≤ h ≤ t, there are 2t elements x2,h,g ∈ X2 such that
Uh,g = Uh ∪ {x1 , x2,h,g } ∈ H∗ for all 1 ≤ h ≤ t and 1 ≤ g ≤ 2t. We may rename the elements
x2,h,g so that for all 1 ≤ h ≤ t and 1 ≤ g ≤ t, x2,h,g ∈ / T ∪ {x2 }. After that we rename them again
so that for all 1 ≤ h ≤ t the elements x2,h,1 are distinct. Now, letting zt = x1 , by (13) we have a
Bt (1, r − 1) with the set of edges {Uh ∪ {zh , x2 } : h ∈ [t]} ∪ {Uh ∪ {zh , x2,h,1 } : h ∈ [t]}.
Case 2.2. For each subfamily H∗ of H satisfying Lemma 6 for q = q ∗ with |H∗ | > nr−2 , the
intersection structure J = J (H∗ ) contains a family isomorphic to J (0) .
Do the following procedure. Let H1 = (H)∗ . For i = 1, 2, . . . , if |H − ij=1 Hj | ≤ C · nr−2 , then
S

stop, let m = i and H′ = ij=1 Hj and H0 = H − H′ ; otherwise, let Hi+1 = (H − ij=1 Hj )∗ .

S S

By the case, Hi contains a family isomorphic to J (0) for each i ≥ 1. By the structure of J (0) , each
v ∈ V (H′ can serve as vertex u in the argument at the beginning of Case 2. So, let u ∈ V (H′ ) and
T ′ = T ′ (u) be defined as above.
By (13), if for any zj ∈ T (u) and some edge Y in H containing zj and disjoint from T ′ (u) − zj ,
QH (Y, zj ) ̸= T ′ (u) − zj , then we again find a Bt (1, r − 1) similarly to (12) because q ∗ is large. It
follows that
there are disjoint t-element sets T1′ , . . . , of the form T ′ (u) covering V (H′ ) such that for
any two u, u′ ∈ Ti′ the set of edges of H containing u and disjoint from Ti − u equals (14)
the set of edges of H containing u′ and disjoint from Ti − u′ for every Ti .

Since J (0) does not contain sets of size r − 1, each (r − 1)-element set Y ∈ ∂Hi is only in one set in
Hi , thus |Hi | ≤ |∂Hi |/r. Moreover, since |H0 | < α(r, t)nr−2 , by (8) some (r − 1)-tuple Y0 belongs
to at least r(t − 1) families Hi , say to H1 , . . . , Hr(t−1) . For i ∈ [r(t − 1)], let yi be the vertex such
that Y0 ∪ {yi } ∈ Hi . Let Y ′ = {y1 , . . . , yr(t−1) }. For every y ∈ Y0 , |T ′ (y) ∩ Y ′ | ≤ t − 1; thus there
is yj ∈ Y ′ \ y∈Y0 T ′ (y), say j = 1. Let T (y1 ) = {z1 , . . . , zt−1 }. Since |Y ′ | > t, we may assume
S

y2 ∈/ T ′ (y1 ).
Since J (H1 ) contains all singletons, there are disjoint (r − 1)-tuples U1 , . . . , Uq∗ such that Ui + y1
is an edge in H1 for each 1 ≤ i ≤ q ∗ . Since q ∗ is large, we can choose among them t − 1 sets, say
U1 , . . . , Ut−1 disjoint from Y0 ∪ Y ′ ∪ T (y1 ). Then we have a Bt (1, r − 1) with the set of edges

{Uh ∪ {y1 } : h ∈ [t − 1]} ∪ {Uh ∪ {zh } : h ∈ [t − 1]} ∪ {Y0 ∪ {y1 }, Y0 ∪ {y2 }},

a contradiction.
Füredi and Kostochka: Bushes, July 4, 2023 11

6 Stability: Proof of Theorem 5

Recall the Lovász form of the Kruskal-Katona Theorem:

[n]
x

x


Theorem 12. If x is a positive real, 1 ≤ k < n, G ⊆ k and |G| = k , then |∂G| ≥ k−1 .

Choose n0 so that

nr−1
   
n0 − 1 n0 − 1
> 3(C + C0 )nr−2
0 and 0
<2 . (15)
r−1 (r − 1)! r−1

Let n > n0 and H be an n-vertex r-uniform family not containing a bush Bt,h (a, b) with |H| >
(t − 1) n−1 r−2 . Define C, m, H , . . . , H and H


r−1 − C0 n 0 m
b as in Subsection 4.1. By (8),
 
n−1
|H| ≥ (t − 1)
b − (C + C0 )nr−2 . (16)
r−1

As in Subsection 4.2, for each 1 ≤ i ≤ m, the intersection structure J (Hi ) contains J (r−1) . So,
again for each hyperedge E ∈ Hi ⊂ H b (1 ≤ i ≤ m) there is an element c(E) ∈ E such that each
[n]


proper subset of E containing c(E) is a kernel of a thr-star in Hi . For every Y ∈ r−2 , let U(Y )
be the set of vertices v ∈ [n] − Y such that there is an edge E ∈ H containing Y and such that
b
v = c(E).
[n]
, v, v ′ ∈ U(Y ), v ̸= v ′ and edges E, E ′ ∈ H


Claim 13. Suppose Y ∈ r−2 b are such that v = c(E),
′ ′ ′ ′
v = c(E ) and Y ⊆ E ∩ E . If E ∈ Hi and E ∈ Hi′ , then i ̸= i . ′

Proof. Suppose v ̸= v ′ , but i = i′ . We may assume that the partition of [n] corresponding to Hi is
(X1 , . . . , Xr ) and v, v ′ ∈ Xr . By symmetry, we also may assume that Y ⊂ X1 ∪ . . . ∪ Xr−2 . Since
Y ⊆ E ∩ E ′ , [r − 2] ∈ Ji or [r − 1] ∈ Ji . Since J (r−1) ⊂ Ji and 2 ≤ a, b ≤ r − 2, [r] − [a] ∈ Ji
and [a] ∪ {r} ∈ Ji . Since Ji is intersecting, [a] = ([a] ∪ {r}) ∩ [r − 1] = ([a] ∪ {r}) ∩ [r − 1] ∈ Ji .
Together with [r] − [a] ∈ Ji , this contradicts Lemma 8.

Similarly to Claim 10, the following holds.

Claim 14. For every (r − 2)-subset Y of [n], |U(Y )| ≤ t − 1.

Proof. Suppose to the contrary that there are t distinct v1 , . . . , vt ∈ [n] and distinct E1 , . . . , Et ∈ H
b
such that Y ⊆ E1 ∩ . . . ∩ Et and vi = c(Ei ) for i = 1, . . . , t. Let E1 ∈ Hi1 , . . . , Et ∈ Hit . By
Claim 13, i1 , . . . , it are all distinct. By relabelling we may suppose that Ei ∈ Hi .
We will find the t + 1 disjoint sets A0 , B1 , B2 , . . . , Bt contradicting Lemma 7 using induction as
follows. Fix any subset A0 of Y with |A0 | = a and let D0 := ∪Ei . Then |D0 | ≤ 2t + r − 2.
We define the sets Ei′ , Di , Bi step by step as follows. We will have Di := D0 ∪ j≤i Ej′ and
S

|Di | ≤ 2t + r − 2 + i(r − 1 − a). For i = 1, 2, . . . , t consider the family Hi and its member Ei in
it. By the intersection structure of Hi , the set A0 ∪ {c(Ei )} is an (a + 1, thr)-kernel in Hi . One
of the thr petals of the sunflower in Hi with kernel A0 ∪ {c(Ei )} should be disjoint from Di−1 ; let
Füredi and Kostochka: Bushes, July 4, 2023 12

Ei′ be the corresponding set in Hi . Since c(Ei ) ∈ Ei′ , and homogeneity gives c(Ei′ ) = c(Ei ), the set
Bi := Ei′ − A0 is a (b, thr)-kernel.

For v ∈ [n], let H(v) = {E ∈ Hb : v = c(E)} and G(v) = {E − v : E ∈ H}. b Let G = S

v∈[n] G(v). By
Claim 13, |G(v)| = |H(v)|
b b ≥ (t − 1)nr−1 − (C +
for each v ∈ [n]; in particular, by (16), |G| = |H|
C0 )nr−2 .
In these terms, Claim 14 can be restated as follows:

Each (r − 2)-subset Y of [n] is in the shadow of at most t − 1 families G(v). (17)

xj

For every j ∈ [n], there a real xj such that |G(j)| = r−1 . Reorder the elements of [n] so that
x1 ≥ x2 ≥ . . . ≥ xn . Let b = 3(r − 1)!(C + C0 ).

Claim 15. For every 1 ≤ i ≤ t − 1,

n − xi ≤ b. (18)

Proof. Suppose the claim fails and 1 ≤ i ≤ t − 1 is the smallest index for which (15) does not hold.
Then
Since for each real 0 < x < n,
x x

 

r−1 n−x r−2
n

= 1− n ,

(19)
r−1
n−r+2 r−2

Theorem 12 yields that for every i ≤ j ≤ n,

  n

xj
n

 
xj b
|∂G(j)| ≥ = r−2n

 r−1
≥ r−2
n |G(j)| 1

+ ,
r−2 r−1 1−
n−xj
r−1
n−r+2
n−r+2

and for every 1 ≤ j ≤ i − 1,

n  n
 


xj r−2 xj r−2
|∂G(j)| ≥ ≥ n

≥ n |G(j)|.


r−2 r−1
r−1 r−1

Hence
n

 b

n 1+ n i−1 n i−1


X r−2 n−r+2 X X r−2
X
|∂G(j)| ≥ n

|G(j)| and |∂G(j)| ≥ n

|G(j)|. (20)
j=i r−1 j=i j=1 r−1 j=1

By (17) and (20),

n i−1 n n

 n n

X X X r−2
X b X
(t − 1)|∂G| ≥ |∂G(j)| = |∂G(j)| + |∂G(j)| ≥ n

 |G(j)| + |G(j)| .
r−1
n−r+2
j=1 j=1 j=i j=1 j=i
Füredi and Kostochka: Bushes, July 4, 2023 13

n Pn n−1
− (C + C0 )nr−2 , this implies



Since |∂G| ≤ r−2 and by (16), j=1 |G(j)| = |G| = |H|
b ≥ (t − 1)
r−1

    n
n n−1 b X
(t − 1) ≥ (t − 1) − (C + C0 )nr−2 + |G(j)|. (21)
r−1 r−1 n−r+2
j=i

Again by (16), nj=i |G(j)| ≥ |H|b − (i − 1) n−1 ≥ (t − i) n−1 − (C + C0 )nr−2 . Since t − i ≥ 1,

P


r−1 r−1
plugging this into (21) and rearranging we get
  
b n−1
(C + C0 )nr−2 ≥ − (C + C0 )nr−2 . (22)
n−r+2 r−1

Using the definition of b and (15), inequality (22) implies

   
r−2 3(r − 1)!(C + C0 ) n−1 1 n−1
(C + C0 )n ≥ − ,
n−r+2 r−1 3 r−1

which in turn again using (15) yields

 
r−2 n−1
n (n − r + 2) ≥ 2(r − 1)! ≥ nr−1 ,
r−1

a contradiction.

Now we are ready to finish the proof of the theorem. By Claim 15, for every 1 ≤ i ≤ t − 1,
      r−2   
xi n−b n Y n−b−j n (r − 1)b
|H(i)|
b = ≥ = ≥ 1−
r−1 r−1 r−1 n−j r−1 n−r+2
j=0
     
n n n b
= −b ≥ − nr−2 .
r−1 r−2 r−1 (r − 1)!
b
This proves Theorem 5 for C1 = (r−1)! = 3(C + C0 ).

References

[1] M. Deza, P. Erdős and P. Frankl, Intersection properties of systems of finite sets, Proc. London Math.
Soc. (3) 36 (1978), 369–384.
[2] P. Erdős, and T. Gallai, On maximal paths and circuits of graphs. Acta Math. Acad. Sci. Hungar. 10
(1959), 337–356.
[3] P. Frankl, On families of finite sets no two of which intersect in a singleton. Bull. Austral. Math. Soc.
17 (1977), 125–134.
[4] P. Frankl, and Z. Füredi, Forbidding just one intersection. J. Combin. Th., Ser. A 39 (1985), 160–176.
[5] P. Frankl, and Z. Füredi, Exact solution of some Turán-type problems. J. Combin. Th., Ser. A 45 (1987),
226–262.
Füredi and Kostochka: Bushes, July 4, 2023 14

[6] Z. Füredi, On finite set-systems whose every intersection is a kernel of a star, Discrete Math. 47 (1983),
129–132.
[7] Z. Füredi, T. Jiang, A. Kostochka, D. Mubayi, and J. Verstraëte, Extremal problems for hypergraph
blowups of trees. To appear in SIDMA, 17 pp.
[8] Z. Füredi, and L. Özkahya, Unavoidable subhypergraphs: a-clusters. J. Combin. Th., Ser. A 118 (2011),
2246–2256.
[9] A. McLennan, The Erdős-Sós conjecture for trees of diameter four. J. Graph Theory 49 (2005), no. 4,
291–301.