Skip to main content

Section 6.3 Dedekind’s Problem via Containers

In Section 6.2, we used containers to count independent sets in regular graphs. The same basic idea is useful in many other settings: if the objects we want to count can be interpreted as independent sets in a graph or hypergraph, then a good container theorem often gives strong counting results. In this section, we prove a simple container theorem for antichains in the Boolean lattice and use it to prove the asymptotic solution to Dedekind’s problem on the number of antichains in \(2^{[n]}\text{.}\) The presentation here is inspired by the container approaches to Boolean-lattice problems in work of Balogh, Mycroft and Treglown [24] and Balogh, Treglown and Wagner [25].
described in detail following the image
An abstract illustration of containers in the Boolean lattice.
An antichain in \(2^{[n]}\) is the same thing as an independent set in the comparability graph of the Boolean lattice: this is the graph with vertex set \(2^{[n]}\text{,}\) where two distinct sets \(A,B\subseteq[n]\) are adjacent if one contains the other. We will also use comparability digraphs for \(2^{[n]}\text{,}\) which are directed graphs obtained by orienting each edge of the comparability graph in one direction or the other. We use these graphs only as a convenient way to talk about comparable pairs.
The first ingredient is a supersaturation theorem for the Boolean lattice. Its proof is a random-chain argument, in the same spirit as the proof of the LYM Inequality (Theorem 2.2.3). The main idea is that if a family is larger than a middle layer, then a random maximal chain should meet it in more than one set on average. If a chain meets a family in at least two sets, then it witnesses a comparable pair. We begin with a general theorem on counting comparisons relative to a random chain in a partially ordered set, or poset.

Proof.

We prove the result by induction on \(|\mathcal{F}|\text{.}\) The result is trivial when \(|\mathcal{F}|\leq M\text{,}\) since the right side is non-positive. For \(x\in\mathcal{F}\text{,}\) let
\begin{equation*} d^+_{\mathcal{F}}(x):=|\{y\in\mathcal{F}:(x,y)\in D\}|. \end{equation*}
If there exists \(x\in\mathcal{F}\) with \(d^+_{\mathcal{F}}(x)\geq L\text{,}\) then by induction applied to \(\mathcal{F}\setminus\{x\}\text{,}\)
\begin{equation*} |\{(u,v)\in D:u,v\in\mathcal{F}\setminus\{x\}\}| \geq L(|\mathcal{F}|-1-M). \end{equation*}
Adding the at least \(L\) arcs from \(x\) to elements of \(\mathcal{F}\) gives the desired bound.
Therefore, we may assume that
\begin{equation*} d^+_{\mathcal{F}}(x)<L \end{equation*}
for every \(x\in\mathcal{F}\text{.}\) Let \(X=|\mathcal{F}\cap\mathcal{C}|\text{.}\) Since every two elements of a chain are comparable, the number of directed arcs of \(D\) induced by \(\mathcal{F}\cap\mathcal{C}\) is \(\binom{X}{2}\text{.}\) Since \(X-\binom{X}{2}\leq 1\) for every nonnegative integer \(X\text{,}\) we have
\begin{equation*} 1\geq \mathbb{E}\left(X-\binom{X}{2}\right). \end{equation*}
Expanding this expectation gives
\begin{equation*} \begin{aligned} 1 &\geq \sum_{x\in\mathcal{F}}\mathbb{P}(x\in\mathcal{C}) - \sum_{\substack{(x,y)\in D\\ x,y\in\mathcal{F}}} \mathbb{P}(x,y\in\mathcal{C})\\ &= \sum_{x\in\mathcal{F}}\mathbb{P}(x\in\mathcal{C}) \left( 1- \sum_{\substack{y\in\mathcal{F}\\ (x,y)\in D}} \mathbb{P}(y\in\mathcal{C}\mid x\in\mathcal{C}) \right). \end{aligned} \end{equation*}
Since \(d^+_{\mathcal{F}}(x)<L\) and \(\mathbb{P}(y\in\mathcal{C}\mid x\in\mathcal{C})\leq 1/L\text{,}\) the expression in parentheses is positive for every \(x\text{.}\) Using the lower bound \(\mathbb{P}(x\in\mathcal{C})\geq1/M\) and the upper bound \(\mathbb{P}(y\in\mathcal{C}\mid x\in\mathcal{C})\leq1/L\text{,}\) we obtain
\begin{equation*} 1\geq \frac{1}{M} \sum_{x\in\mathcal{F}} \left( 1-\frac{d^+_{\mathcal{F}}(x)}{L} \right). \end{equation*}
Rearranging,
\begin{equation*} \sum_{x\in\mathcal{F}}d^+_{\mathcal{F}}(x) \geq L(|\mathcal{F}|-M). \end{equation*}
The left side is exactly \(|\{(x,y)\in D:x,y\in\mathcal{F}\}|\text{,}\) completing the proof.
We now apply this lemma to the Boolean lattice.

Proof.

Define
\begin{equation*} M=\binom{n}{\lfloor n/2\rfloor} \qquad\text{and}\qquad L=\left\lceil\frac{n+1}{2}\right\rceil. \end{equation*}
First suppose that \(\emptyset\notin\mathcal{F}\text{.}\) Let \(\mathcal{P}=2^{[n]}\setminus\{\emptyset\}\text{.}\) Choose a uniformly random maximal chain in \(2^{[n]}\) and delete the empty set from it; call the resulting random chain \(\mathcal{C}\) in \(\mathcal{P}\text{.}\) For every \(A\in\mathcal{P}\text{,}\)
\begin{equation*} \mathbb{P}(A\in\mathcal{C})=\binom{n}{|A|}^{-1}\geq 1/M. \end{equation*}
We obtain a comparability digraph \(D\) for \(\mathcal{P}\) by orienting comparable pairs as follows. If \(A\) and \(B\) are comparable and \(|A|\neq |B|\text{,}\) orient the pair from the set whose cardinality is farther from \(n/2\) to the set whose cardinality is closer to \(n/2\text{.}\) In the case of a tie, so that \(A\subsetneq B\) and \(|A|=n-|B|\text{,}\) orient the pair from \(A\) to \(B\text{.}\)
We claim that, for every arc \((X,Y)\) in \(D\text{,}\)
\begin{equation*} \mathbb{P}(Y\in\mathcal{C}\mid X\in\mathcal{C})\leq 1/L. \end{equation*}
Suppose first that \(X\subsetneq Y\text{.}\) Then
\begin{equation*} \mathbb{P}(Y\in\mathcal{C}\mid X\in\mathcal{C}) = \binom{n-|X|}{|Y|-|X|}^{-1}. \end{equation*}
Since the arc is oriented from \(X\) to \(Y\text{,}\) the set \(X\) is at least as far from the middle level as \(Y\text{.}\) Thus \(|X|<n/2\text{.}\) Also, since \(X\neq\emptyset\text{,}\) the endpoint \(Y\) cannot be all of \([n]\text{;}\) otherwise \(Y\) would be farther from the middle than \(X\text{.}\) Hence
\begin{equation*} n-|X|\geq L \qquad\text{and}\qquad 1\leq |Y|-|X|<n-|X|. \end{equation*}
Therefore
\begin{equation*} \binom{n-|X|}{|Y|-|X|}\geq L. \end{equation*}
So
\begin{equation*} \mathbb{P}(Y\in\mathcal{C}\mid X\in\mathcal{C})\leq 1/L. \end{equation*}
The case \(Y\subsetneq X\) is similar. In this case,
\begin{equation*} \mathbb{P}(Y\in\mathcal{C}\mid X\in\mathcal{C}) = \binom{|X|}{|X|-|Y|}^{-1}. \end{equation*}
Since the arc is oriented from \(X\) to \(Y\text{,}\) the set \(X\) is at least as far from the middle level as \(Y\text{.}\) Thus \(|X|>n/2\text{.}\) Also \(Y\neq\emptyset\text{,}\) since \(\emptyset\notin\mathcal{P}\text{.}\) Therefore
\begin{equation*} |X|\geq L \qquad\text{and}\qquad 1\leq |X|-|Y|<|X|. \end{equation*}
It follows that
\begin{equation*} \binom{|X|}{|X|-|Y|}\geq L, \end{equation*}
as required.
Therefore, by Lemma 6.3.1,
\begin{equation*} |\{(A,B)\in\mathcal{F}^2:A\subsetneq B\}| \geq L(|\mathcal{F}|-M). \end{equation*}
This proves the result when \(\emptyset\notin\mathcal{F}\text{.}\)
Now suppose that \(\emptyset\in\mathcal{F}\text{.}\) Apply the previous case to \(\mathcal{F}'=\mathcal{F}\setminus\{\emptyset\}\text{.}\) The empty set is comparable to every other set in \(\mathcal{F}\text{,}\) so
\begin{equation*} \begin{aligned} |\{(A,B)\in\mathcal{F}^2:A\subsetneq B\}| &\geq L\left(|\mathcal{F}'|-M\right)+|\mathcal{F}'|\\ &= L\left(|\mathcal{F}|-M\right)+|\mathcal{F}'|-L. \end{aligned} \end{equation*}
If \(|\mathcal{F}'|\geq L\text{,}\) this is enough. If \(|\mathcal{F}'|<L\text{,}\) then \(|\mathcal{F}|\leq L\text{.}\) Since \(L\leq M\text{,}\) the quantity \(L(|\mathcal{F}|-M)\) is non-positive, and the desired inequality is trivial.
We will use the following compact notation. If \(X\) is a finite set and \(m\geq0\text{,}\) then
\begin{equation*} \binom{X}{\leq m}:=\{Y\subseteq X: |Y|\leq m\}. \end{equation*}
Thus,
\begin{equation*} \left|\binom{X}{\leq m}\right|=\sum_{j=0}^{\lfloor m\rfloor}\binom{|X|}{j}. \end{equation*}
The following container lemma is the abstract step which turns supersaturation into a small collection of containers.

Proof.

For \(A\subseteq\mathcal{P}\) and \(x\in A\text{,}\) let
\begin{equation*} N_A(x):=\{y\in A:y\neq x\text{ and }x,y\text{ are comparable}\}. \end{equation*}
Fix an arbitrary total order on \(\mathcal{P}\text{,}\) which will be used only to break ties.
Given an antichain \(\mathcal{I}\text{,}\) we run the following algorithm. Start with \(A=\mathcal{P}\) and \(T=\emptyset\text{.}\) While \(|A|>m\text{,}\) choose an element \(x\in A\) for which \(|N_A(x)|\) is as large as possible, breaking ties using the fixed total order. If \(x\in\mathcal{I}\text{,}\) add \(x\) to \(T\) and replace \(A\) by
\begin{equation*} A\setminus\big(\{x\}\cup N_A(x)\big). \end{equation*}
If \(x\notin\mathcal{I}\text{,}\) replace \(A\) by \(A\setminus\{x\}\text{.}\)
Whenever the loop runs, the current set \(A\) has size greater than \(m\text{.}\) By assumption, \(A\) contains at least \(d|A|\) comparable pairs. Hence
\begin{equation*} \sum_{x\in A}|N_A(x)|\geq 2d|A|. \end{equation*}
Therefore some element \(x\in A\) satisfies \(|N_A(x)|\geq 2d\text{.}\) Whenever an element is added to \(T\text{,}\) at least \(2d+1\) elements are deleted from \(A\text{.}\) Since initially \(|A|=|\mathcal{P}|\text{,}\) we obtain
\begin{equation*} |T|\leq\frac{|\mathcal{P}|}{2d+1}. \end{equation*}
When the algorithm stops, define \(f(T)\) to be the remaining set \(A\text{.}\) Then \(|f(T)|\leq m\text{.}\) Also, every element of \(\mathcal{I}\) is either added to \(T\) or remains in \(f(T)\text{,}\) since if an element of \(\mathcal{I}\) is deleted as a neighbour of a chosen element of \(\mathcal{I}\text{,}\) then two elements of the antichain would be comparable, which is impossible. Thus
\begin{equation*} \mathcal{I}\subseteq T\cup f(T). \end{equation*}
Moreover, \(T\cap f(T)=\emptyset\text{.}\)
Finally, this defines a deterministic function \(f\) on the sets \(T\in\binom{\mathcal{P}}{\leq |\mathcal{P}|/(2d+1)}\text{.}\) Indeed, to reconstruct \(f(T)\) from \(T\text{,}\) run the same algorithm and use membership in \(T\) to decide whether the currently chosen element should be treated as selected. This gives a deterministic function \(f\) with the desired properties.

Proof.

Let
\begin{equation*} M=\binom{n}{\lfloor n/2\rfloor}, \qquad L=\left\lceil\frac{n+1}{2}\right\rceil, \qquad m=(1+\alpha)M. \end{equation*}
If \(A\subseteq2^{[n]}\) has \(|A|>m\text{,}\) then by Lemma 6.3.2,
\begin{equation*} |\{(B,C)\in A^2:B\subsetneq C\}| \geq L(|A|-M). \end{equation*}
Since \(|A|>(1+\alpha)M\text{,}\) we have
\begin{equation*} |A|-M>\frac{\alpha}{1+\alpha}|A|\geq \frac{\alpha}{2}|A|. \end{equation*}
Hence \(A\) contains at least
\begin{equation*} \frac{\alpha L}{2}|A| \end{equation*}
comparable pairs. For sufficiently large \(n\text{,}\) let
\begin{equation*} d=\left\lfloor\frac{\alpha L}{3}\right\rfloor. \end{equation*}
Then \(d\geq1\text{,}\) and every subset \(A\subseteq2^{[n]}\) with \(|A|>m\) contains at least \(d|A|\) comparable pairs.
Applying Lemma 6.3.3 with \(\mathcal{P}=2^{[n]}\text{,}\) this value of \(d\text{,}\) and this value of \(m\text{,}\) gives the desired function \(f\text{.}\) Since
\begin{equation*} \frac{|2^{[n]}|}{2d+1} \leq K(\alpha)\left(\frac{2^n}{n}\right), \end{equation*}
where \(K=K(\alpha)\) is chosen sufficiently large with respect to \(\alpha\text{.}\)
We now turn to the classical problem of counting all antichains in \(2^{[n]}\text{.}\) This is known as Dedekind’s problem, since antichains in \(2^{[n]}\) are in bijection with monotone Boolean functions on \(n\) variables.

Definition 6.3.5.

Let \(M(n)\) denote the number of antichains in \(2^{[n]}\text{.}\) The number \(M(n)\) is called the Dedekind number.

Proof.

The lower bound follows by taking arbitrary subfamilies of the middle layer:
\begin{equation*} M(n)\geq 2^{\binom{n}{\lfloor n/2\rfloor}}. \end{equation*}
For the upper bound, fix \(0<\alpha<1\) and apply Corollary 6.3.4. Let
\begin{equation*} M=\binom{n}{\lfloor n/2\rfloor}. \end{equation*}
Every antichain \(\mathcal{I}\) is contained in a set of the form \(T\cup f(T)\text{,}\) where
\begin{equation*} T\in\binom{2^{[n]}}{\leq K2^n/n} \qquad\text{and}\qquad |f(T)|\leq(1+\alpha)M. \end{equation*}
Since
\begin{equation*} M=(1+o(1))\sqrt{\frac{2}{\pi n}}\,2^n, \end{equation*}
we have
\begin{equation*} \frac{2^n}{n}=o(M). \end{equation*}
Therefore the number of possible choices for \(T\) is
\begin{equation*} \left|\binom{2^{[n]}}{\leq K2^n/n}\right| = \sum_{j\leq K2^n/n}\binom{2^n}{j} = 2^{o(M)}. \end{equation*}
For each fixed \(T\text{,}\) the container \(f(T)\) is determined, and the number of possible antichains contained in \(T\cup f(T)\) is at most
\begin{equation*} 2^{|T|+|f(T)|} \leq 2^{o(M)}2^{(1+\alpha)M}. \end{equation*}
Hence
\begin{equation*} M(n)\leq 2^{(1+\alpha+o(1))M}. \end{equation*}
Since \(\alpha>0\) was arbitrary, this gives
\begin{equation*} M(n)\leq 2^{(1+o(1))\binom{n}{\lfloor n/2\rfloor}}. \end{equation*}
Together with the lower bound, this proves the theorem.