Radon partitions in convexity spaces∗

Boris Bukh†
arXiv:1009.2384v1 [math.CO] 13 Sep 2010

Abstract
Tverberg’s theorem asserts that every (k − 1)(d + 1) + 1 points in Rd can be
partitioned into k parts, so that the convex hulls of the parts have a common inter-
section. Calder and Eckhoff asked whether there is a purely combinatorial deduction
of Tverberg’s theorem from the special case k = 2. We dash the hopes of a purely
combinatorial deduction, but show that the case k = 2 does imply that every set of
O(k 2 log2 k) points admits a Tverberg partition into k parts.

Introduction
Radon’s lemma [Rad21] states that every set P of d + 2 points in Rd can be partitioned
into two classes P = P1 ∪ P2 so that the convex hulls of P1 and P2 intersect. Birch
[Bir59] (for d = 2) and Tverberg [Tve66] (for general d) extended Radon’s theorem to the
analogous statement for partitions of a set into more than two parts: For a set P ⊂ Rd of
|P | ≥ (k − 1)(d + 1) + 1 points there is a partition P = P1 ∪ · · · ∪ Pk into k parts, such that
the intersection of the convex hulls conv P1 ∩ · · · ∩ conv Pk is non-empty. The bound of
(k − 1)(d + 1) + 1 is sharp, as witnessed by any set of points in sufficiently general position.
Calder [Cal71] conjectured and Eckhoff [Eck79] speculated that Tverberg’s theorem
is a consequence of Radon’s theorem in the context of abstract convexity spaces. The
conjecture, which we now present, is commonly referred as “Eckhoff’s conjecture”, and we
will maintain this tradition to avoid additional confusion. If true, the conjecture would
have provided a purely combinatorial proof of Tverberg’s theorem. However, we will show
that the conjecture is false.
A convexity space on the ground set X is a family F ⊂ 2X of subsets of X, called
convex sets, such as both ∅ and X are convex, and intersection of any collection of convex
sets is convex. For example, the familiar convex sets in Rd form a convexity space on
Rd . Among the other examples are axis-parallel boxes in Rd , finite subsets on any ground
set, closed sets in any topological space (see the book [vdV93] for a through overview of
convexity spaces). If the ground set X in the convexity space (X, F) is clear from the
∗
The paper is in public domain, and is not protected by copyright.
†
[email protected]. Centre for Mathematical Sciences, Cambridge CB3 0WB, England and
Churchill College, Cambridge CB3 0DS, England.

1
context, we will speak simply of a convexity space F. The convex hull of a set P ⊂ X,
denoted conv P , is the intersection of all the convex sets containing P . We write convF P
if the convexity space is not clear from the context. The k-th Radon number of (X, F) is
the minimum natural number rk , if it exists, so that every set P ⊂ X of at least rk points
admits a partition P = P1 ∪ · · · ∪ Pk into k parts whose convex hulls have an element in
common. It is not hard to show1 that if r2 is finite, then so is rk . Eckhoff’s conjecture
states that rk ≤ (k − 1)(r2 − 1) + 1 in every convexity space. The conjecture has been
proved for r2 = 3 by Jamison [JW81], and for convexity space with at most 2r2 points by
Sierksma and Boland [SB83]. In section 4 we reproduce a version of Jamison’s proof.
The best bounds on rk are

rk1 k2 ≤ rk1 rk2 (due to Jamison [JW81]),

r2k+1 ≤ (r2 − 1)(rk+1 − 1) + rk + 1 (due to Eckhoff [Eck00]).

In particular,
rk ≤ k⌈log2 r2 ⌉ . (1)
The following result improves on (1).

Theorem 1. Let (X, F) be a convexity space, and assume that r2 is finite. Then

rk ≤ c(r2 )k2 log2 k,

where c(r2 ) is a constant that depends only on r2 .

Though this bound is not far from Eckhoff’s conjecture, the conjecture itself is false.

Theorem 2. For each k ≥ 3 there is a convexity space (X, F) such that r2 = 4, but
rk ≥ 3(k − 1) + 2.

Despite the failure of Eckhoff’s conjecture, we have been unable to rule out that the
convexity spaces with finite r2 might behave similarly to Euclidean spaces. It is conceivable
that rk is bounded by a linear function of k for each r2 . Moreover, it is possible that other
results from combinatorial convexity extend to such spaces. For instance, Radon proved
the lemma now bearing his name to give an alternative proof of Helly’s theorem that if
in some family of convex sets in Rd every d + 1 sets intersect, then all of them do. One
of the easy but startling consequences of Helly’s theorem is the centrepoint theorem. The
centrepoint theorem asserts that for every finite set P ⊂ Rd there is a point p ∈ Rd
d
(the “centrepoint”) such that every convex set containing more than d+1 |P | points of P
also contains p. Both the deduction of Helly’s theorem from Radon’s theorem, and the
deduction of centrepoint theorem from Helly’s theorem remain valid in the context of the
convexity spaces with finite r2 . This prompts the following question:
1
According to [Eck00] it was first shown by R.E.Jamison (1976). The first published proofs appear to
be in [DRS81] and [JW81].

2
Question 3 (Weak epsilon-nets). Suppose (X, F) is a convexity space with finite r2 . Let
ε > 0 be given. Let P ⊂ X be a set of points in the space. Is there a set N of |N | ≤ f (ǫ, r2 )
points such that every convex set S containing more than ε|P | points of P also contains
at least one point of N ?

The set N as in the question above is called a weak ε-net (with respect to convex sets)
for P . In Rd it is known that there are weak ε-nets of size only (1/ε)d logcd (1/ε). The
discussion above shows that the answer to the question is positive if ǫ > 1 − 1/(r2 − 1). It
is unclear whether the weak ε-nets of size f (ε, r2 ) exist for any ǫ < 1 − 1/(r2 − 1).
Bárány [Bár82] showed that if P is an n-point set in Rd , then there is a point p in
n n



cd d+1 of all the d+1 simplices spanned by P , where cd is a positive constant that
1
depends only on d. In R , it is immediate that c1 = 1/2 is admissible, and is best possible.
The situation for convexity spaces with bounded r2 is again unclear, except if r2 = 3:

Proposition 4 (Selection theorem). Let (X, F) be a space with r2 = 3. Let P ⊂ X be

point set. Then there is a point p ∈ X that is contained in at least 13 n2 + O(n) of all the

sets conv{x, y}.

Question 5. Does the preceding proposition hold with 1/2 in place of 1/3?

The standard greedy argument of Alon, Bárány, Füredi, Kleitman [ABFK92, Section 8]
shows that the selection theorem implies an affirmative answer to Question 3. In particular,
it gives f (ǫ, 3) ≤ O (1/ǫ)2 , which is probably not sharp.

The rest of the paper is organized as follows. In section 1 we introduce our only
technical tool, the nerves of convex sets. In lemma 7 we will show that the nerves encode
all the information about the convexity space that we need. In section 2 we present a
counterexample to Eckhoff’s conjecture. It is then followed in section 3 by the proof of
Theorem 1. We conclude the paper with a short discussion of convexity spaces with r2 = 3.

1 Nerves
Let P be a set of points in a some convexity space. We associate to P a collection N (P )
of subsets of 2P . A family F ⊂ 2P belongs to N (P ) if and only if the intersection
T
S∈F conv S is non-empty. In the conventional terminology one would say that the col-
lection N (P ) is the nerve of the family of convex sets {conv S : S ⊂ P }. Since we will
not use the nerves of any other families of sets, in this paper we abuse the language and
say that N (P ) is the nerve of P .

Proposition 6. If N = N (P ), then N satisfies the following properties:

(N1) N is a downset: if F ∈ N and F ′ ⊂ F, then F ′ ∈ N .

def
(N2) If F is in N , then so is F̂ = {S ′ : S ′ ⊃ S ∈ F}.

3
 def
(N3) For every p ∈ P the family F(p) = {S : p ∈ S} is in N .

(N4) The set P can be partitioned into k parts P = P1 ∪ · · · ∪ Pk so that (conv P1 ) ∩

· · · ∩ (conv Pk ) 6= ∅ if and only if there is a family {P1 , . . . , Pk } ∈ N consisting of k
disjoint sets.

(N5) If rt exists, then for every set of rt families F = {F1 , . . . , Frt } ⊂ N there is a
partition F = F 1 ∪ · · · ∪ F t of F into t parts so that ( F 1 ) ∪ · · · ∪ ( F t ) ∈ N .
T T

Proof. The first properties four properties are immediate from the definition of N (P ).
The final property is easy too: Suppose F = {F1 , . . . , Frt } is given. Let qi be any
T
point in S∈Fi conv S. The set Q = {q1 , . . . , qrt } of rt points can be partition into t parts
Q = Q1 ∪ · · · ∪ Qt so that (conv Q1 ) ∩ · · · ∩ (conv Qt ) is non-empty, thus containing some
point p. The partition Q = Q1 ∪· · ·∪Qt naturally induces the partition F = F 1 ∪· · · ∪F t .
T
It is easy to see that the point q belongs to S∈∩F i conv S for each i = 1, . . . , t.

Thanks to the following lemma we can avoid the convexity spaces in the rest of the
paper, and work exclusively with nerves.

Lemma 7. Let P be a set, and let N be a collection of subsets of 2P that satisfies the first
three properties in the Proposition 6. Then there are a ground set X ⊃ P and a convexity
space on X so that N (P ) = N .

Proof. For an arbitrary family F let C(F) = {F ′ ∈ N : F ⊂ F ′ }, and denote by C the

family of all the sets of the form C(F). Put X = N . We claim that C forms a desired
convexity space on X. It is clear that ∅, X ∈ C. Since C(F1 ) ∩ C(F2 ) = C(F1 ∪ F2 ), and
similarly for intersections of more than two sets, the collection C indeed forms a convexity
space on X. Define φ : P → X by φ(p) = F(p). The map φ is well-defined by property
(N3), and provides the embedding of P into X. We need to check that N (φ(P )) = φ(N )
Since F(p) ∈ C(F) if and only if F ⊂ F(p), it follows that {F(p1 ), . . . , F(pt )} ⊂ C(F)
precisely when F ⊂ F(pi ). Hence, if P ′ ⊂ P , then
T

 
\ \ \
conv C (φ(P ′ )) = C(F) = C(F) = C  F(p) .
T
φ(P ′ )⊂C(F ) F⊂ p∈P ′ F (p) p∈P ′

Hence, F ∈ convC (φ(P ′ )) if and only if {P : P ′ ⊂ P } ⊂ F. The intersection P ∈F ′ conv C (φ(P ))

T

is non-empty if and only if there is an F ∈ N so that F̂ ′ ⊂ F. Thus by the properties

(N1) and (N2) \
convC (φ(P )) 6= ∅ ⇐⇒ F ′ ∈ N .
P ∈F ′

Therefore N C (φ(P )) = φ(N ) as claimed.

4
2 Counterexample to Eckhoff’s conjecture
Proof of Theorem 2. We shall use the Lemma 7 to construct the requisite convexity space.
Let P = [3(k − 1) + 1]. Consider the three kinds of families:
 
 P
A[x] = {x} ∪ ,
4
   
  P P
B[xy : zw] = {x, y}, {z, w} ∪ S ∈ : {x, y, z, w} ∩ S 6= ∅ ∪ , distinct x, y, z, w
3 4
 
 P
C[xy] = {x, y} ∪ , x, y are distinct.
3

Here x, y, z, w are elements of P = [3(k − 1) + 1]. Let A[x], ˆ B̂[xy : zw] and Ĉ[xy] be
as in Proposition 6 property (N2). Let N consist of all the families, A[x], ˆ B̂[xy : zw]
and Ĉ[xy] and all their subfamilies. Let N 0 consist only of families Â[x], B̂[xy : zw]
and Ĉ[xy]. As N automatically satisfies properties (N1) and (N2) in Proposition 6 and
F(p) ⊂ Â[p], by Lemma 7 it is a nerve of some convexity space. As k ≥ 3, no family of
the form A[x], B[xy : zw] or C[xy] contains t disjoint sets. From that it follows that none
of Â[x], B̂[xy : zw] or Ĉ[xy] contain k disjoint sets either, and same holds for every family
in N . Therefore, to establish the theorem it remains to verify the property (N5) with
r2 = 4.
As Â-, B̂- and Ĉ-families are the maximal families in N , it suffices to show that
whenever F = {F1 , . . . , F4 } is a collections of four families in N 0 , then there is a partition
F = F 1 ∪ F 2 so that ( F 1 ) ∪ ( F 2 ) is contained in some F ∈ N 0 .
T T

To every family F we associate a subset

 
P
e(F) = F ∩ .
2

That is

e(Â[x]) = {x, y} : y ∈ P \ {x}

e(B̂[xy : zw]) = {x, y}, {z, w} ,

e(Ĉ[xy]) = {x, y} .

Note that e(F1 ∩ F2 ) = e(F1 ) ∩ e(F2 ). It is convenient think of e(F) as an edge of a

hypergraph on the ground set P2 .

Note that if F1 , F2 ∈ N 0 are two distinct families, then F1 ∩ F2 is contained in a

Ĉ-set. Moreover, if e(F1 ) ∩ e(F2 ) = ∅, then F1 ∩ F2 in contained in P3 .

Suppose F1 , . . . , F4 are four families in N 0 . If e(F1 ) ∩ e(F2 ) = ∅, then F1 ∩ F2 ⊂ P3

and F3 ∩ F4 ⊂ Ĉ[xy] for some x, y. Hence (F1 ∩ F2 ) ∪ (F3 ∩ F4 ) ⊂ P3 ∪ Ĉ[xy] = Ĉ[xy].

We may thus assume that e(F1 ) ∩ e(F2 ) is non-empty, and similarly for other pairs of sets
F1 , . . . , F4 .

5
 There are five cases according to the number of Â-families among the four families
F1 , . . . F4 .
There no Â-families: Since every two families meet, and e(F1 ), . . . , e(F4 ) contain
1 or 2 vertices each, it follows that e(F1 ), . . . , e(F4 ) must have a common vertex, say
{1, 2} ∈ P2 . Then (F1 ∩ F2 ) ∪ (F3 ∩ F4 ) ⊂ Ĉ[12].

There is a single Â-family F1 : As e(F2 ), e(F3 ) and e(F4 ) pairwise meet, they
either have a vertex in common, or F2 , F3 , F4 are B̂-families, and e(F2 ), e(F3 ), e(F4 ) form
a triangle. However, they cannot form the triangle because e(F1 ) would not meet each
of e(F2 ), e(F3 ) and e(F4 ). Thus, e(F1 ) ∩ · · · ∩ e(F4 ) is non-empty, and equals to say
{1, 2} ∈ P2 . Then (F1 ∩ F2 ) ∪ (F3 ∩ F4 ) ⊂ Ĉ[12].

There are two Â-families F1 and F2 : The intersection e(F3 ) ∩ e(F4 ) contains just
one element, say {x, y}. If F1 and F2 are just Â[x] and Â[y], then (F1 ∩ F2 ) ∪ (F3 ∩
F ) ⊂ Ĉ[xy]. If F1 = Â[z] and z 6∈ {x, y}, then it necessarily follows that e(F3 ) =
4 
{x, y}, {z, w3 } and e(F4 ) = {x, y}, {z, w4 } for some w3 and w4 . Thus F2 is either
Â[x] or Â[y]. In either case (F1 ∩ F3 ) ∪ (F2 ∩ F4 ) ⊂ B̂[xy : zw3 ].
There are three Â-families F1 , F2 and F3 : As e(F4 ) has to meet all of e(F1 ),
e(F2 ), e(F3 ), it must be that F4 is a B̂-family, implying that (F1 ∩ F2 ∩ F3 ) ∪ F4 = F4 .
All four families are Â-families: Say, F1 = Â[x], F2 = Â[y], F3 = Â[z] and
F4 = Â[w]. In that case (F1 ∩ F2 ) ∪ (F3 ∩ F4 ) ⊂ B̂[xy : zw].

3 Upper bound on Radon numbers

The main ingredient in the proof of theorem 1 is a version of Kruskal–Katona theorem
from [Buk10]. A d-dimensional r-uniform family is a d-tuple of r-element sets. In other
words, if we denote by Xr the family of all r-element subsets of X, then d-dimensional



d
d
r-uniform family is a subset of Xr . A shadow of such a family F ⊂ Xr is defined to
be
def 
∂F = (S1 \ {xi }, . . . , Sd \ {xd }) : (S1 , . . . , Sd ) ∈ F, and xi ∈ Si for i = 1, . . . , d .

d
Lemma 8 (Theorem 1 of [Buk10]). Suppose F ⊂ Xr is a d-dimensional r-uniform
family of size
 d
x
|F| = ,
r
where x ≥ r is a real number. Then
 d
x
|∂F| ≥ .
r−1
In addition to multidimensional Kruskal–Katona theorem, we shall need four lemmas.
The first two lemmas are a bound on Turán numbers of hypergraphs and a bound on
the independence numbers of graphs in which every subgraph have a large independence
number.

6
Lemma 9 ([dC83]). If H is an s-uniform hypergraph on n vertices with fewer than
l−1
−1 n−l+1 |H|

s−1 n−s+1 s edges, then H contains an independent set on l vertices.

Lemma 10 (Special case of Theorem 2.1 from [AS07]). Let t < s ≤ 2s − 3, and let G be
a graph on n vertices. Suppose that every set of size s contains an independent set of size
t. Then G contains an independent set of size n − s + 1.

d
Our third lemma is purely computational. We say that a tuple (S1 , . . . , Sd ) ∈ Pa is
r-good if there are r pairwise disjoint sets Si1 , . . . , Sir among S’s.

Lemma 11. Let P be a finite set. There are at most

 d
2 d−r+1 |P |
C(d)(a /|P |)
a

P
d
r-bad tuples in a , where C(d) is a constant that depends only on d.

d
Proof. For S = (S1 , . . . , Sd ) ∈ Pa let G[S] be a graph on {1, . . . , d} with ij forming an
edge if Si ∩ Sj 6= ∅. A tuple S is r-bad if and only if the independence number of G[S]
is less than r. Suppose that the largest forest in G[S] has m edges, then by contracting
these edges we obtain an independent set of size d − m. Thus if a tuple S is r-bad, then
G[S] contains a forest F with d − r + 1 edges. We say that the forest F witnesses that S
is r-bad.
Fix a forest F . We shall bound the number of r-bad tuples S for which F is a witness
that S is r-bad. Let v1 , . . . , vd be a relabelling of {1, . . . , d} so that in F the vertex vi is
adjacent to at most one vertex vj with j < i. Pick S1 , . . . , Sd uniformly at random from
P

a . If vi is adjacent to some vj with j < i let Ei be the event that Si ∩ Sj 6= ∅. If vj is
adjacent to none vj with j < i let Ei be the event that holds with probability 1. Then
d
Y d
Y
Pr[F is a witness that S is r-bad] = Pr[Ei |E1 , . . . , Ei−1 ] = Pr[Ei ] ≤ (a2 /|P |)d−r+1 .
i=1 i=1

As the number of forests on d vertices depends only on d, the lemma follows by the union
bound.

Finally, the third lemma that we need is a restatement of Jamison’s upper bound
r2t ≤ r2t in terms of nerves. We include the proof for completeness.

Lemma 12. Suppose P a set in a convexity space, and N = N (P ) is its nerve. Then for
every set P ′ ⊂ P of size |P ′ | = r2t there is a family F ∈ N containing 2t disjoint subsets
of P ′ .

Proof. The proof is by induction on t. The base case t = 0 is trivial. Suppose t ≥ 1. Let
P ′ = P1′ ∪ · · · ∪ Pr′2 be a partition of P ′ into sets of size r2t−1 . By the induction hypothesis,

7
there are families F1 , . . . , Fr2 such that each Fi contains 2t−1 disjoint subsets of Pi′ . Let
these subsets be Ri,1 , . . . , Ri,2t−1 . By property (N5) of the proposition 6, there is a a set
I ⊂ [rk ] so that \ \
Fi ∪ Fi ∈ N .
i∈I i6∈I

By property (N2) the intersection Rj = i∈I Ri,j is in i∈I F̂i for each j = 1, . . . , 2t−1 .
T T

The sets Rj are 2t−1 disjoint subsets of i∈I Pi′ . Similarly one obtains 2t−1 disjoint subsets
S

of i6∈I Pi′ , for the total of 2t disjoint subsets of P ′ .

S

Proof of theorem 1. It suffices to show that for every nerve N on |P | = k2 log2 k points
there are k disjoint sets S1 , . . . , Sk ⊂ P and a family F that contains all of these sets.
For brevity we shall write r = r2 and t = 1 + ⌈log2 r⌉. Define a (2r − 3)-dimensional

2r−3
2r−3
family T ⊂ rPt as follows: A tuple (S1 , . . . , S2r−3 ) ∈ rPt is in T if there is a family
F ∈ N such that {S1 , . . . , S2r−3 } ⊂ F. Let P0 ⊂ P be any (2r − 3)r t -element subset of
P . Let P ′ ⊂ P0 be an arbitrary r t -element subset of P0 . By the preceding lemma there is
a family F that contains 2t disjoint subsets of P ′ . Since 2r − 3 ≤ 2t , by property (N2) it
follows that F contains 2r − 3 disjoint subsets of size r t each that partition P0 . In other
words, P0 gives rise to at least one tuple in T . Since P0 is an arbitrary (2r − 3)r t -element
subset of P , we conclude that
 2r−3 
(1 − c2 (r))|P | 2r−3
  
|P | (2r−3)r t |P |
|T | ≥ ≥ c1 (r)|P | ≥ −
(2r − 3)r t rt rt
for some positive constants c1 (r), c2 (r) that depend only on r.
P
2r−3
Let m = ⌈log k/c2 (r)⌉. Define a (2r − 3)-dimensional family T ′ ⊂ mr t in the
′
same way as T was defined: namely, S ∈ T if there is an F ∈ N such that S ⊂ F.
P
2r−3
Note that the property (N2) implies that if S ∈ mr t is not in T ′ , then neither is any
family obtained from S by removing some elements from each set in S. Lemma 8 applied
to the complement of T ′ yields

|P | 2r−3 (1 − c2 (r))|P | 2r−3

   
′
|T | ≥ − .
mr t+1 mr t
P
mr t ) be a (2r − 3)-uniform hypergraph on
Let H ⊂ (2r−3 P



mr t with edges

{S1 , . . . , S2r−3 } ∈ H ⇐⇒ (S1 , . . . , S2r−3 ) ∈ T ′ and (S1 , . . . , S2r−3 ) is r-good.

By Lemma 11, it follows that

 2r−3 !
1 ′ 2 2t r−2 |P |
|H| ≥ |T | − c3 (r)(m r /|P |)
(2r − 3)! mr t
 |P |  


mr t (2r−3)mr t 2 r−2
≥ 1 − (1 − c2 (r)) − c4 (r)(m /|P |)
2r − 3

8
Since m > log k/c2 (r), and |P | ≥ (9c4 (r))1/(r−2) m2 k2 it follows that the density of H is
 |P |

t
|H|/ mr t ≥ 1 − k(2r−3)r − (3k)−(2r−4) ≥ 1 − (2k)−(2r−4)
2r − 3

for k large enough.

P

By Lemma 9 the hypergraph H contains a clique on 2k vertices. Let S1 , . . . , S2k ∈ mr t

be the vertices of this clique. Since edges of H are r-good among every 2r − 3 of these 2k
sets there are r that are pairwise disjoint. Thus, by Lemma 10 there are k of them, say
S1 , . . . , Sk , that are pairwise disjoint.
We claim that for every I ⊂ [k] there is a family FI ∈ N that contains Si for every
i ∈ I. The proof is by induction on |I| starting with |I| = 2r − 3. If |I| = 2r − 3, then
the claim holds because {Si : i ∈ I} is an edge in H. Suppose |I| > 2r − 3. Pick any r
distinct |I|−1-element subsets I1 , . . . , Ir of I. Then by by property (N5) applied to families
FI1 , . . . , FIr it follows that there is a J ⊂ [r] so that F = ( j∈J FIj ) ∪ ( j6∈J FIj ) ∈ N .
T T

Since the family F contain Fi for every i ∈ I, we may put FI = F.

Finally, the family F[k] contains k disjoint sets S1 , . . . , Sk , as required.

4 Convexity spaces with r2 = 3

The space with r2 = 3 are especially nice because of the following lemma, which is implicit
in [JW81].

Lemma 13. Let P be a set in a convexity space with r2 = 3, and let N = N (P ) be its
nerve. Then there is a family Fp ∈ N for each p ∈ P , and these families satisfy

(J1) {p} ∈ Fp .

(J2) If p, q, r are any three points of P , then either {p, q} ∈ Fr or {p, r} ∈ Fq or {q, r} ∈
Fp .

(J3) If {q, r} ∈ Fp and {r, s} ∈ Fq , then {r, s} ∈ Fp .

Proof. Let Fp be a maximal family containing {p}. Then the other conditions follow from
the property (N5) applied to the triple of families Fp , Fq , Fr .

Proof of Proposition 4. Let I = {(p, q, r) : p ∈ conv{q, r}}. Since there are n3 triples

{p, q, r}, each of which contributes at least at least one element I, the proposition follows
by the pigeonhole principle.

Since Jamison’s proof of Eckhoff’s conjecture is especially short in the language of

nerves, we include it:

Theorem 14. If r2 = 3, then rk ≤ 2(k − 1) + 1.

9
Proof. Suppose |P | = 2(k − 1) + 1. We shall show that one of Fp contains k pairwise
disjoint sets. We claim that there is a pair of elements p, q ∈ P so that {p, q} ∈ Fr for
every r 6= p, q. Indeed, it is true if |P | ≤ 3. If |P | ≥ 4, and s is any element of Fp ,
then by induction there is a p, q ∈ P \ {s} so that {p, q} ∈ Fr for every r 6= p, q, s. If in
addition {p, q} ∈ Fs , then we are done. Otherwise by property (J2) either {p, s} ∈ Fq or
{q, s} ∈ Fp . Say {p, s} ∈ Fq . Then by property (J3) applied to {p, q} ∈ Fr and either
{p, s} ∈ Fq we conclude that {p, s} is in every Fr , r 6= p, s. The claim is proved.
Let p, q be a pair of element so that {p, q} ∈ Fr for r 6= p, q. By the induction
hypothesis applied to P \ {p, q} there is r ∈ {p, q} so that Fr contains k − 1 disjoint sets
that are also disjoint from {p, q}. Together with {p, q} these form a desired family of
disjoint sets.

References
[ABFK92] Noga Alon, Imre Bárány, Zoltán Füredi, and Daniel J. Kleitman. Point selec-
tions and weak ǫ-nets for convex hulls. Combin. Probab. Comput., 1(3):189–200,
1992. http://www.tau.ac.il/~nogaa/PDFS/abfk3.pdf.

[AS07] Noga Alon and Benny Sudakov. On graphs with subgraphs having
large independence numbers. J. Graph Theory, 56(2):149–157, 2007.
http://www.math.ucla.edu/~bsudakov/erdos-hajnal.pdf.

[Bár82] Imre Bárány. A generalization of Carathéodory’s theorem. Discrete Math.,

40(2-3):141–152, 1982.

[Bir59] B. J. Birch. On 3N points in a plane. Proc. Cambridge Philos. Soc., 55:289–

293, 1959.

[Buk10] Boris Bukh. Multidimensional Kruskal–Katona theorem. arXiv, Sep 2010.

[Cal71] J. R. Calder. Some elementary properties of interval convexities. J. London

Math. Soc. (2), 3:422–428, 1971.

[dC83] D. de Caen. Extension of a theorem of Moon and Moser on complete subgraphs.

Ars Combin., 16:5–10, 1983.

[DRS81] Jean-Paul Doignon, John R. Reay, and Gerard Sierksma. A Tverberg-type

generalization of the Helly number of a convexity space. J. Geom., 16(2):117–
125, 1981.

[Eck79] Jürgen Eckhoff. Radon’s theorem revisited. In Contributions to geometry

(Proc. Geom. Sympos., Siegen, 1978), pages 164–185. Birkhäuser, Basel, 1979.

[Eck00] Jürgen Eckhoff. The partition conjecture. Discrete Math., 221(1-3):61–78,

2000. Selected papers in honor of Ludwig Danzer.

10
[JW81] Robert E. Jamison-Waldner. Partition numbers for trees
and ordered sets. Pacific J. Math., 96(1):115–140, 1981.
http://projecteuclid.org/getRecord?id=euclid.pjm/1102734951.

[Rad21] Johann Radon. Mengen konvexer Körper, die einen gemeinsamen Punkt en-
thalten. Math. Ann., 83(1-2):113–115, 1921.

[SB83] Gerard Sierksma and Jan Ch. Boland. On Eckhoff’s conjecture for Radon
numbers; or how far the proof is still away. J. Geom., 20(2):116–121, 1983.

[Tve66] H. Tverberg. A generalization of Radon’s theorem. J. London Math. Soc.,

41:123–128, 1966.

[vdV93] M. L. J. van de Vel. Theory of convex structures, volume 50 of North-Holland

Mathematical Library. North-Holland Publishing Co., Amsterdam, 1993.

11