Let a=(a1,a2,...,an)∈Rna = (a_1, a_2, ..., a_n) \in \mathbb{R}^n with ∣ai∣≥1|a_i| \ge 1. Let p(a)p(a) be the number of vectors c=(c1,c2,...,cn)∈Rnc = (c_1, c_2, ..., c_n) \in \mathbb{R}^n where ci=±1,∀i∈[n]c_i = \pm 1, \forall i \in [n], such that

−1≤<a,c>=∑i=1nciai<1 -1 \le \left< a , c \right> = \sum_{i = 1}^n c_i a_i < 1
  1. Prove that p(a)≤(n⌊n/2⌋)p(a) \le \binom{n}{\lfloor n / 2 \rfloor }.

  2. Find an a∈Rna \in \mathbb{R}^n with p(a)=(n⌊n/2⌋)p(a) = \binom{n}{\lfloor n / 2 \rfloor }.

Proof. By symmetry, we can assume that a>0a > 0 without lose of generality. Define

C={c=(c1,c2,...,cn)∈Rn∣ci=±1,i∈[n]}. C = \{ c = (c_1, c_2, ..., c_n) \in \mathbb{R}^n \mid c_i = \pm 1, i \in [n] \}.

Let [n][n] be the set of integers from 11 to nn, and P([n])\mathscr{P}([n]) be the power set of [n][n], i.e., the set of all subsets of [n][n]. For any subset s⊂[n]s \subset [n], we can associate to ss a vector in CC, which is denoted as

c(s)≐(c(s)1,c(s)2,...,c(s)n), c(s) \doteq (c(s)_1, c(s)_2, ..., c(s)_n),

where

{c(s)i=1,∀i∈sc(s)i=−1,∀i∉s \begin{cases} c(s)_i = 1, & \forall i \in s \\ c(s)_i = -1,& \forall i \notin s \end{cases}

It remains to show the set

S(a)≐{s⊂[n]:−1<<a,c(s)><1} S(a) \doteq \{ s \subset [n] : -1 < \left< a, c(s) \right> < 1 \}

has size at most (n⌊n/2⌋)\binom{n}{\lfloor n / 2 \rfloor }. The key observation is that

∀s,s′∈S(a)⇒s⊈s′. \forall s, s' \in S(a) \Rightarrow s \nsubseteq s'.

We call S(a)S(a) an antichain. To show this, suppose thta s′s' contains at least one element that is not in ss. Denote this element as jj. It follows that c(s′)j=1c(s')_j = 1 and c(s′)j=−1c(s')_j = -1 and that

<a,c(s′)>−<a,c(s)>≥c(s′)jaj−c(s)jaj=2aj≥2 \left< a, c(s') \right> - \left< a, c(s) \right> \ge c(s')_j a_j - c(s)_j a_j = 2 a_j \ge 2

We conclude that <a,c(s′)>≥<a,c(s)>+2≥1\left< a, c(s') \right> \ge \left< a, c(s) \right> + 2 \ge 1, a contradiction.

Now, consider an increasing sequence of nn subsets of [n][n] of the form:

∅⊂{x1}⊂{x1,x2}⊂{x1,x2,x3}⊂...⊂{x1,x2,...,xn} \emptyset \subset \{ x_1 \} \subset \{ x_1, x_2 \} \subset \{ x_1, x_2, x_3 \} \subset ... \subset \{ x_1, x_2, ..., x_n \}

where x1,...,xnx_1, ..., x_n are different integers in [n][n]. Each such sequence induces a permutation of integers in [n][n]. As a result, the total number of such sequence equals n!n!. The relation between S(a)S(a) and such sequence is established as follows:

  1. For each such sequence, there can be at most one set in S(a)S(a), because S(a)S(a) is a antichain.

  2. For each s={x1,x2,...,xk}∈S(a)s = \{x_1, x_2, ..., x_k \} \in S(a), where k=∣s∣≥1k = |s| \ge 1 is the size of ss, the number of sequences it appears in is given by

    ∣s∣!(n−∣s∣)!≥⌊n/2⌋!(n−⌊n/2⌋)! |s| ! (n - |s|)! \ge \lfloor n / 2 \rfloor ! (n - \lfloor n / 2 \rfloor)!

We obtain

∣S(a)∣⋅⌊n/2⌋!(n−⌊n/2⌋)!≤∑s∈S(a)∣s∣!(n−∣s∣)!≤n! |S(a) | \cdot \lfloor n / 2 \rfloor ! (n - \lfloor n / 2 \rfloor)! \le\sum_{s \in S(a) } |s| ! (n - |s|) ! \le n!

and

p(a)=∣S(a)∣=n!⌊n/2⌋!(n−⌊n/2⌋)!=(n⌊n/2⌋) p(a) = |S(a) | = \frac{n!}{ \lfloor n / 2 \rfloor ! (n - \lfloor n / 2 \rfloor)! } = \binom{n}{\lfloor n / 2 \rfloor }

This finishes the proof for problem 1.

Finally, ai=1,∀i∈[n]a_i = 1, \forall i \in [n] gives a tight example to problem 2.

■\blacksquare

Reference

[1] J. Matoušek and J. Nešetřil, Invitation to discrete mathematics, 2nd ed. Oxford ; New York: Oxford University Press, 2009.