Combinatorics, Probability and Computing (2003) 12, 89–93.

c 2003 Cambridge University Press

DOI: 10.1017/S0963548302005230 Printed in the United Kingdom

An Average Set Size Theorem

DAVID REIMER
Department of Mathematics,
The College of New Jersey, Ewing, NJ 086280718, USA
(e-mail: [email protected])

Received 26 August 2001; revised 30 January 2002

In this paper we prove that, if a collection of sets A is union-closed, then the average set
size of A is at least 12 log2 (|A|).

1. Introduction
In what follows, all logs will be base two. For n ∈ N define [n] ≡ {1, 2, . . . , n}. Let P([n]) be
the power set of [n], and for A ⊆ B ⊆ [n] define [A, B] as [A, B] ≡ {C ⊆ [n] : A ⊆ C ⊆ B}.
For A ⊆ P([n]), we say A is union-closed if A, B ∈ A implies A ∪ B ∈ A.
The Union-Closed Sets Conjecture, attributed to P. Frankl (see [4, p. 525]), states that,
if A ⊆ P([n]) is union-closed, then some element of [n] belongs to at least half of the
sets in A. This paper will prove a theorem that puts a lower bound on the average set
size of a collection of union-closed sets. The theorem, conjectured by Jeff Kahn [3], was
motivated by, and can be shown to follow from, Frankl’s conjecture.

Theorem 1.1. If n ∈ N and A ⊆ P([n]), where A is union-closed, then

P
A∈A |A| 1
> log(|A|).
|A| 2

When A = P([n]), both sides of the equation become 2n and hence the theorem is
sharp. Later we will show that the only sharp cases are the families A = P(M), where
M ⊆ [n].
We will prove Theorem 1.1 using two lemmas. The first lemma gives a strong result in
the special case that A is a filter.

Definition. F ⊆ P([n]) is a filter if and only if i ∈ [n] and F ∈ F implies F ∪ {i} ∈ F.

90 D. Reimer

The following definition treats P([n]) as a digraph whose vertices are the sets, and
whose edges are pairs of sets that differ by exactly one element, directed from the smaller
set to the larger set.

Definition. Let A, B ∈ P([n]). Then A ⊂· B if and only if A ∪ {i} = B, for some i ∈ [n],

V (P([n])) ≡ P([n]),
E(P([n])) ≡ {(A, B) ∈ P([n])2 : A ⊂· B}.

Define the edge boundary of filter F as

EB(F) ≡ {(A, B) ∈ E(P([n])) : A ∈

/ F, B ∈ F}.

Lemma 1.2. For all filters F,

X
2 |F| = n|F| + |EB(F)|.
F∈F

It is not hard to demonstrate that this lemma implies the well-known fact that Theo-
rem 1.1 holds for filters. Simply divide both sides by 2|F| and notice that n > log(|F|).
The second lemma establishes a special bijection between a union-closed family and an
associated filter.

Lemma 1.3. Let A ⊆ P([n]) be closed under unions. Then there exist a filter F ⊆ P([n])
and a bijection from A to F, A → FA such that:
(i) A ⊆ FA ;
(ii) A, B ∈ A, A 6= B ⇒ [A, FA ] ∩ [B, FB ] = φ;
P
(iii) A∈A |FA − A| 6 |A|n − |A| log(|A|)

In Section 2 we will prove that Lemmas 1.2 and 1.3 imply Theorem 1.1, and in Section 3
we will prove these two lemmas.

2. Proof of the main theorem

Proof of Theorem 1.1. Let A be a subset of P([n]) that is closed under unions and apply
Lemma 1.3 to obtain filter F. For each A ∈ A, define edge set EA ≡ {(B, FA ) ∈ E(P([n])) :
B = FA −{i}, i ∈ FA −A}. Clearly |EA | = |FA −A|. Also note that EA ⊆ E([A, FA ])∩EB(F)
P P
and the sets EA are disjoint. Hence A∈A |EA | 6 |EB(F)|, or equivalently A∈A |FA −A| 6
|EB(F)|. Apply Lemma 1.2 to F. By part (i) of Lemma 1.3, we know A ⊆ FA , so
X X
2 |A| + 2 |FA − A| = |A|n + |EB(F)|,
A∈A A∈A
X X
2 |A| + |FA − A| > |A|n.
A∈A A∈A
 An Average Set Size Theorem 91

By part (iii) we get

X
2 |A| + |A|n − |A| log(|A|) > |A|n,
A∈A
P
A∈A |A| 1
> log(|A|).
|A| 2

3. Proofs of the lemmas

Proof of Lemma 1.2.
X XX X X X X
2 |F| = 1+ 1+ 1
F∈F F∈F G⊂·F F∈F G∈F:G⊂·F ·F
F∈F G∈F:G⊂
/
XX XX
= 1+ 1 + |EB(F)|
F∈F G⊂·F ·G
G∈F F⊃
X
= n + |EB(F)|
F∈F

= n|F| + |EB(F)|

We can now prove Lemma 1.3, following some preliminaries. We first define the i-rising
function RS,i (S). This is in fact a well-known operator, appearing, for example, in [1] and
[2]. Intuitively, RS,i (S) adds i to all sets of S that do not contain i, unless such a set
already exists in S. To simplify our notation, we will often refer to RS,i (S) as Ri (S).

Definition. Given a family of sets S, define


S : S ∪ {i} ∈ S,
Ri (S) ≡
S ∪ {i} : S ∪ {i} ∈
/ S.
Then the i-rising of S is
Ri (S) ≡ {Ri (S) : S ∈ S}.

Note that Ri (S) is one to one. We will now raise A in each of the n dimensions. For
our family A, define

A : if i = 0,
Ai ≡
Ri (Ai−1 ) : if i > 0,


Ai ≡ RAi−1 ,i ◦ RAi−2 ,i−1 · · · ◦ RA0 ,1 (A).
If Ai 6= Ai−1 we say Ai−1 rises, and note that Ai−1 ∈ Ai−1 implies that Ai ∪ {i} ∈ Ai .
We will need the following lemma.

Lemma 3.1. If Ai−1 is closed under unions and Ci−1 , Di−1 ∈ Ai−1 are such that Di−1 =
Ci−1 ∪ X for some X ⊆ [n], then Ci ∪ X ∈ Ai . Furthermore Ci−1 6= Ci−1 ∪ X implies
Ci 6= Ci ∪ X.
92 D. Reimer

Proof. Ci−1 ∪ X ∈ Ai−1 implies Ci−1 ∪ X ∪ {i} ∈ Ai . So, if Ci ∪ X ∈ / Ai then Ci ,

Ci−1 , X, and Di−1 do not contain i. Since Ci−1 does not rise, Ci−1 ∪ {i} ∈ Ai−1 , and
by closure (Ci−1 ∪ {i}) ∪ Di−1 = Di−1 ∪ {i} ∈ Ai−1 , and hence Di−1 does not rise. So
Ci ∪ X = Ci−1 ∪ X = Di−1 = Di ∈ Ai .
If Ci−1 6= Ci−1 ∪ X and Ci = Ci ∪ X, then Ci−1 rises. But Ci−1 ∪ {i} = Ci−1 ∪ X = Di−1 ∈
Ai−1 , so Ci−1 does not rise, leading to a contradiction.

Lemma 3.2. For all i ∈ {0, 1, . . . , n}, Ai is closed under unions.

Proof. We will prove this by induction on i. Pick any Ai , Bi ∈ Ai . We need to show

Ai ∪ Bi ∈ Ai . Let Ci−1 = Ai−1 , and Di−1 = Ai−1 ∪ Bi−1 ∈ Ai−1 and apply Lemma 3.1. Then
Ai ∪ Di−1 ∈ Ai . Ai ∪ Bi is either Ai ∪ Di−1 or Ai ∪ Di−1 ∪ {i}. But Ai ∪ Di−1 ∈ Ai implies
Ai ∪ Di−1 ∪ {i} ∈ Ai .

Lemma 3.3. Let i, j ∈ [n] where i < j. Let Ci , Di ∈ Ai such that Di = Ci ∪ X for some
X ⊆ [n]. Then Cj ∪ X ∈ Aj . Furthermore, Ci 6= Ci ∪ X implies Cj 6= Cj ∪ X.

Proof. This result is obtained by repeated application of Lemma 3.1.

Lemma 3.4. An is a filter.

Proof. We need to show, for all A ∈ A, that An ∪ {i} ∈ An . Pick i ∈ [n]. We know
Ai ∈ Ai implies Ai ∪ {i} ∈ Ai . Let Ci = Ai , X = {i} and j = n and apply Lemma 3.3.
Then An ∪ {i} ∈ An .

Define FA ≡ An . Then A → FA is a bijection from A to filter F ≡ An , since it is the

composition of bijections RAi−1 ,i . We now need to prove parts (i)–(iii) of Lemma 1.3.

Proof of Lemma 1.3. (i) For every i ∈ [n], Ai−1 ⊆ Ai , so A = A0 ⊆ An = FA .

(ii) Let A, B ∈ A, and assume for a contradiction that there exists Y ∈ [A, FA ] ∩ [B, FB ].
We may assume one of A and B contains the other. For suppose A 6⊆ B and B 6⊆ A.
Since A ⊆ Y and B ⊆ Y , we know A ∪ B ⊆ Y , and hence A ∪ B ∈ [A, FA ]. By closure,
A ∪ B ∈ A and [A, FA ] ∩ [A ∪ B, FA∪B ] 6= φ. We know A ∪ B 6= A since B 6⊆ A. So, if we
replace B with A ∪ B, then A ⊂ B.
Let C ≡ A, X = B, i = 0 and j = n and apply Lemma 3.3. Since C 6= C ∪ X, we get
An 6= An ∪ B or equivalently FA 6= FA ∪ B, which implies B 6⊆ FA and B ∈ / [A, FA ], which
is a contradiction.
(iii) Since A ⊆ FA , we know |[A, FA ]| = 2|FA −A| . These subcubes are disjoint so their total
size must be less than the whole cube, giving
X X
2|FA −A| = |[A, FA ]| 6 2n .
A∈A A∈A
P
We will use this inequality to obtain an upper bound for A∈A |FA − A| using Jensen’s
 An Average Set Size Theorem 93

inequality, which essentially says that the maximum sum is obtained when all the subcubes
[A, FA ] are the same size. By Jensen’s inequality we have
" #
X X
exp2 |A|−1 |FA − A| 6 |A|−1 2|FA −A| 6 2n |A|−1
A∈A A∈A
or
X
|FA − A| 6 n|A| − |A| log |A|.
A∈A

The inequalities in the above proof are only sharp if the subcubes [A, FA ] fill the cube
and are all the same size. Without loss of generality, we may assume the maximal set
M ∈ A is [n], and therefore does not rise. Hence |[M, FM ]| = 1 and, if the inequality is
sharp, |[A, FA ]| = 1, for all A ∈ A, which implies that A = P([n]). Therefore the only
sharp cases are those with A = P(M), for some M ⊆ [n], as claimed.

References
[1] Alon, N. (1983) On the density of sets of vectors. Discrete Math. 46 199–202.
[2] Frankl, P. (1983) On the trace of finite sets. J. Combin. Theory Ser. A 34 41–45.
[3] Kahn, J. Personal communication.
[4] Rival, I. (1985) Graphs and Order, Reidel, Dordrecht/Boston.