An Eventown Result for Multisets

For a fixed coordinate i and a family \mathcal{F}, define

For i=1, we drop the superscript and just write T_e,T_o.

Case I: m is even

Suppose \delta=0, so m=2q. If q=0, the conclusion is immediate, otherwise, by Lemma 2.2, we have,

Now, we fix the first coordinate and replace \mathcal F by

This replacement is clearly an Eventown family. Moreover,

The selected first coordinate is now fixed, and all other coordinates remain unchanged.

Starting with \mathcal F, we perform this operation successively on all n coordinates. The resulting family \mathcal G contains at most one member, and at each iteration, the size of the family decreases by a factor of at most r. Consequently,

This proves the theorem when m is even.

Case II: m is odd

Suppose \delta=1, so m=2q+1. Define

both of size r+1.

We first transform the family \mathcal{F} into another Eventown family \mathcal{G}, without decreasing its size, such that the value in each coordinate is either from Q_e or Q_o.

Fix the first coordinate. First suppose that,

If |T_e|\ge |T_o|, replace \mathcal{F} by \{(s,x):s\in\{0,2,...,2q\},x\in T_e\}=V_e \times T_e; otherwise replace it by \{(t,y):t\in\{1,3,...,2q+1\},y\in T_o\}=V_o \times T_o. The replacement family is also an Eventown family, where the chosen first coordinate of each member lies in Q_e or Q_o. The size of the new family is r \max(T_e,T_o)\ge |\mathcal{F}|.

Now, suppose |\mathcal{F}|>r \max(|T_e|,|T_o|). This forces a+b >r. Therefore, by Lemma 2.2, we get,

and for every x\in T_e with a number of even extensions and any y \in T_o with b number of odd extensions, the even extensions come before the odd extensions. Therefore, selected tuples have exactly the following extensions:

for some l \in \{0\}\cup[q].

Now, every extension of a tuple in T_e precedes every extension of y. Consequently, each tuple in T_e has extensions at most 2l. Similarly, a tuple in T_o has all its extensions above 2l. Consequently, every even extension precedes every odd extension in \mathcal{F}.

Now, if |T_e|\ge |T_o| we replace \mathcal{F} by

otherwise by

Both of the families, \mathcal{F}_1^{(1)} and \mathcal{F}_2^{(1)} are Eventown families with the sizes r|T_e|+|T_o| and |T_e|+r|T_o|, respectively.

So, when |T_e|\ge |T_o| we have

and for |T_o|\ge|T_e|, we have

Thus the chosen replacement family, has size at least a|T_e|+b|T_o|\ge|\mathcal F|, preserves the Eventown condition and restricts the first coordinate to Q_e or Q_o.

We apply the above procedure successively to all coordinates, with each iteration modifying only the chosen coordinate of the family members. After all n iterations, we obtain an Eventown family \mathcal G satisfying

such that, in each coordinate, every member of \mathcal G takes values exclusively from either Q_e or Q_o. Hence by Lemma 2.3, we have,

Therefore,

If d=0, take I=K=\varnothing. Assume d>0, and let G be a d\times n matrix whose rows form a basis of C. For J\subseteq[n], let G_J denote the submatrix of G consisting of the columns indexed by J. Choose d independent columns, indexed by I. Since G_I is invertible, the restriction or projection map

is a bijection. Here, \mathbb{F}_2^I denotes the space of binary vectors indexed by I. We claim that restriction to the complementary coordinates is also injective. Indeed, suppose z\in C is supported on I, that is, z_i=0 for all i\notin I. By the self-orthogonality of C, we have

Since c|_I ranges over all of \mathbb{F}_2^I, this forces z|_I=0, and hence z=0.

Therefore restriction to I^c=[n]\setminus I is injective, and thus G_{I^c} has rank d. Choosing d independent columns of this matrix gives a set K\subseteq I^c, of size d, such that restriction to K is also injective.

If a_j=0 or b_j=0, the conclusion follows from the fact that \{0,1,...,m\} contains r even values and q+\delta \le r odd values.

Otherwise, choose u,v such that |E_j(u)|=a_j and |O_j(v)|=b_j. Let E_j(u)=\{s_1<s_2<...<s_{a_j}\} and O_j(v)=\{t_1<t_2<...<t_{b_j}\}. Then, we have

Thus all \min(s_i,t_{i'}) have same parity. Since s_i’s are even and t_{i'}’s are odd, either s_1<...<s_{a_j}<t_1<...<t_{b_j} or t_1<...<t_{b_j}<s_1<...<s_{a_j}. So, either every element of E_j(u) lies below every element of O_j(v), or conversely.

In the first case, write \max E_j(u)=2l. There are at most l+1 even values at most 2l, and q+\delta -l odd values greater than 2l. Hence

so a_j+b_j\le r+\delta.

In the second case, write \max O_j(v)=2l-1. Then,

and therefore a_j+b_j \le r\le r+\delta. This proves the lemma.

Observe that the identity \min(s,t)\bmod2=(s\bmod2)(t\bmod2) holds within each of Q_e,Q_o. For x \in \mathcal{G}, define its parity vector as

Using the above identity we get for x,y\in \mathcal{G},

Let C=\text{span}_{\mathbb{F}_2}\{\pi(x):x \in \mathcal{G}\}. Then C is a self-orthogonal subspace of \mathbb{F}_2^n. Therefore if dimension of C is d, then d\le k.

By Lemma 2.1, there exist disjoint coordinate sets I,J\subseteq[n], each of size d, such that either restriction determines a vector of C. Put L=[n]\setminus(I\cup J), so |L|=n-2d. Note that these restrictions determine the parity vector, not the actual multiplicities.

In each coordinate, the allowed multiplicities have parity counts (r,1) or (1,r). Thus, one parity has r multiplicities, and the other has one. Let P be the set of parity vectors occurring in \mathcal{G}. For each c \in P, define

Once c is fixed, each coordinate in L has at most r multiplicity choices. Hence the number of members of \mathcal{G} with parity vector c is at most a_cb_cr^{n-2d}. Consequently,

Next, we shall bound \sum_{c \in P} a_c^2. For every binary pattern u on I, let A(u) be the number of allowed multiplicity assignments with that pattern. Since I contains d coordinates, there are 2^d possible patterns for u. If u selects the parity with r choices in exactly p coordinates, then

Since l' coordinates are chosen from d coordinates, there are exactly \binom{d}{p} such patterns. Consequently,

Since I determines the complete parity vector, distinct c \in P provide distinct restrictions c_I. Therefore, we have,

Similarly from J we have

Hence

Now we characterize the extremal families. Since |\mathcal{G}|=r^{n-2k}(1+r^2)^k, every inequality above is an equality. Consequently, we have the following

1. d=k,

2. P=C, and

3. a_c=b_c for all c \in C.

Also, if a parity pattern occurs in P, then every allowed choice of multiplicity vector realizing that pattern belongs to \mathcal{G}.

If n is odd, there is one coordinate outside I \cup J. This coordinate must contribute exactly r choices for each parity pattern that occurs. Since 0\in C=P, the even parity must provide r choices. Hence, the multiplicity assigned to this coordinate must belong to Q_e. In contrast, the odd parity provides only one choice, so no member of the family can have odd multiplicity in this coordinate. Consequently, the remaining coordinate takes values precisely from V_e, independently of the other coordinates.

Now since I,J both determines C, their restrictions have the form

where T=\pi_J \circ \pi_I^{-1} is an invertible linear map. Define \alpha \in \mathbb{F}_2^{|I|} as follows

Similarly, define \beta\in \mathbb{F}_2^{|J|} on J. Thus, a coordinate of I gives r choices precisely when u_i=\alpha_i. Therefore,

Similarly,

where B(Tu) is the analogous multiplicity assignment count on J. By using a_c=b_c for c\in C and r>1, we get,

Taking u=\alpha we have T\alpha=\beta. Substituting u=v+\alpha into the preceding equation, we obtain

Each standard basis vector has Hamming weight 1. Since T preserves Hamming weight, it maps every standard basis vector to a standard basis vector. The invertibility of T ensures that these images are distinct. Hence, T is a coordinate permutation, inducing a pairing between the coordinates of I and those of J. Within each pair, the two coordinates have the same parity, and both take values exclusively from either Q_e or Q_o.

A pair of Q_e-coordinates, therefore, gives

and a pair of Q_o-coordinates gives

Since all parity patterns and all permitted assignments occur, we have,

after a coordinate permutation. The converse is immediate.