Packing Designs with large block size

If n \leq \lambda, then the result is trivial. Otherwise, each (\lambda + 1)-set of blocks from \mathcal{B} can be a subset of \mathcal{B}_x for at most t-1 points x \in V, for otherwise t points would appear together in more than \lambda blocks. The number of (\lambda + 1)-subsets of \mathcal{B} is \binom{n}{\lambda+1} and, for each integer i \geq \lambda + 1, the number of (\lambda + 1)-subsets of \mathcal{B}_x for each point x \in V_i is \binom{i}{\lambda+1}. Thus \sum_{i=\lambda+1}^n \binom{i}{\lambda+1} N_i \leq (t-1)\binom{n}{\lambda+1} and the result follows.

Suppose T is a t-set of points in a \mathrm{PD}_{\lambda}(n;v,k,t). For each x \in T, there are n-r(x) blocks that x does not occur in and so in total there are at most tn-\sum_{x \in T}r(x) blocks that are not supersets of T. By the definition of a \mathrm{PD}_{\lambda}(n;v,k,t), at most \lambda blocks can be supersets of T and hence at least n-\lambda blocks are not. So tn-\sum_{x \in T}r(x) \geq n-\lambda and rearranging gives the result.

Consider a \mathrm{PD}_{\lambda}(d;v,k,t). For this packing design, by Lemma 2.3 we have

Since \sum_{i=1}^{d} N_i=v, \sum_{i=1}^{d} iN_i=dk and the function \binom{i}{\lambda+1} is convex in i, we have that the right hand side of (Proof) is minimized when N_{q+1}=r, N_q=v-r, and N_i=0 for all i \in \{1,\ldots,n\} \setminus \{q,q+1\}. Thus

where the equality follows by substituting \binom{q+1}{\lambda+1}=\binom{q}{\lambda+1}+\binom{q}{\lambda} and simplifying.

If v=k, then the multiset consisting of n copies of V forms the blocks of the required packing design, so we may assume v>k. Fix v and k, and proceed by induction on n. The result is trivial for n=0 and n=1. Otherwise, take a \mathrm{PD}_{\left\lceil nk/v \right\rceil}(n;v,k,t), (V,\mathcal{B}), such that r_{{\cal B}}(x) \in \left\{ \left\lfloor \frac{nk}{v} \right\rfloor, \left\lceil \frac{nk}{v} \right\rceil \right\} for all x \in V. Let \mathcal{B}' = \mathcal{B} \cup \{B\}, where B is a k-subset of V such that B \subseteq V_{\left\lfloor nk/v \right\rfloor} if N_{\left\lfloor nk/v \right\rfloor} \geq k, and V_{\left\lfloor nk/v \right\rfloor} \subseteq B if N_{\left\lfloor nk/v\right\rfloor} < k.

Then we have that |r_{{\cal B}'}(x)-r_{{\cal B}'}(y)| \leq 1 for all x,y \in V. Thus, since \sum_{x \in V}r_{{\cal B}'}(x)=(n+1)k, we have r_{{\cal B}'}(x) \in \{ \lfloor \frac{(n+1)k}{v} \rfloor, \lceil \frac{(n+1)k}{v} \rceil \}. Since any point is in at most \lceil \frac{(n+1)k}{v} \rceil blocks, the same is certainly true of any set of t points, and hence (V,\mathcal{B}') is a \mathrm{PD}_{\left\lceil (n+1)k/v \right\rceil}(n+1;v,k,t).

If n \leq \lambda, then taking \mathcal{B} to be any collection of n k-subsets of V gives a packing satisfying the required conditions. So we may assume that n \geq \lambda+1. Note that by combining our two hypotheses and simplifying we have v \geq (t-1)\binom{n}{\lambda+1} with equality only if both hypotheses hold with equality.

Let \mathcal{S} be the set of all (\lambda+1)-subsets of \{1, \ldots, n\}. Let U=\{u_{S,j} \mid S \in \mathcal{S}, j \in \{1,\ldots,t-1\}\} be a set of (t-1)\binom{n}{\lambda+1} elements indexed by \mathcal{S} \times \{1,\ldots,t-1\}, and let W be a set of v-(t-1)\binom{n}{\lambda+1} elements that is disjoint from U. If k = (t-1)\binom{n-1}{\lambda}, let W_1=\cdots=W_n=\emptyset. Otherwise, we claim that there is a \mathrm{PD}_{\lambda}\left(n;v-(t-1)\binom{n}{\lambda+1}, k-(t-1)\binom{n-1}{\lambda},t\right) on point set W with block set \mathcal{W}=\{W_1, W_2, \ldots, W_n\} such that |\mathcal{W}_x| \leq \lambda for each x \in W. Such a packing design exists by Lemma 3.2 because, using our hypothesis that \lambda v \geq nk-(t-1)\binom{n}{\lambda+1},

So the packing design (W,\mathcal{W}) does indeed exist.

Let \mathcal{B} = \{B_1, \ldots, B_n\}, where, for each i \in \{1, \ldots, n\}, B_i is the k-set

Note that |\mathcal{B}_x| = \lambda+1 for all x \in U and |\mathcal{B}_x| \leq \lambda for all x \in W. We claim (U \cup W,\mathcal{B}) is the required packing design. To see this, let T be a t-subset of U \cup W. If T \not\subseteq U, then there is some y \in T \cap W and we have |\bigcap_{x \in T}\mathcal{B}_x| \leq |\mathcal{B}_y| \leq \lambda. Otherwise, T \subseteq U, and we must have \{u_{S,j},u_{S',j'}\} \subseteq T for some S,S' \in \mathcal{S} and j,j' \in \{1,\ldots,t-1\} where S \neq S'. So,

First note that nk-(t-1)\binom{n}{\lambda+1} < (n+1)k- (t-1)\binom{n+1}{\lambda+1} implies

Since (Proof) and nk-(t-1)\binom{n}{\lambda+1} \leq \lambda v hold, there exists a \mathrm{PD}_\lambda(n;v,k,t) by Lemma 3.3.

We will complete the proof by showing that no \mathrm{PD}_\lambda(n+1;v,k,t) exists. By Lemma 3.1, it suffices to show that

where q and r are the integers such that (n+1)k=qv+r and 0 \leq r < v. To determine q and r, note that \lambda v < (n+1)k-\binom{n+1}{2} implies \lambda v < (n+1)k and hence q \geq \lambda. On the other hand, (Proof) yields (t-1)\binom{n}{\lambda+1}<\frac{n-\lambda}{\lambda+1}k. Substituting this into nk-(t-1)\binom{n}{\lambda+1} \leq \lambda v and simplifying produces (n+1)k<(\lambda+1)v. Hence q<\lambda+1. Thus q=\lambda and r=k(n+1)-\lambda v, so (Proof) becomes

which is equivalent to our hypothesis that \lambda v < k(n+1)-(t-1)\binom{n+1}{\lambda+1}.

Note that

by the definition of \ell. In view of (Proof) and the first inequality of (3.6), Lemma 3.3 guarantees the existence of a \mathrm{PDN}_{\lambda}(\ell;v,k,t). We will complete the proof by showing that Theorem 3.1 can be applied with q=\lambda+1 to show that no \mathrm{PDN}_{\lambda}(\ell+1;v,k,t) exists.

From (Proof) we have (t-1)\binom{\ell}{\lambda+1} \leq \frac{1}{\lambda+1}\ell k and substituting this into the first inequality of (3.6) and simplifying yields \ell k \leq (\lambda+1)v. Thus (\ell+1)k < (\lambda+2)v, since v > k follows from (Proof) and the first inequality of (3.6). Now (t-1)\binom{\ell}{\lambda} > k implies (t-1)\binom{\ell+1}{\lambda+1} > \frac{1}{\lambda+1}(\ell+1)k and substituting this into the second inequality of (3.6) and simplifying yields (\ell+1)k > (\lambda+1)v. So, applying Theorem 3.1 with q=\lambda+1 and r=(\ell+1)k-(\lambda+1)v, shows that no \mathrm{PDN}_{\lambda}(\ell+1;v,k,t) exists provided

Routine manipulation shows this is equivalent to the second inequality of (3.6).

We proceed by induction on v. The result is trivial when v=1. Now suppose (V,\mathcal{B}) is a \mathrm{PD}_2(v) with v \geq 2 and |\mathcal{B}_x| \leq 3 for each x \in V. Let a \in V, and let \mathcal{B}=\{B_1, \ldots, B_n\} and \mathcal{B}_a = \{B_1, \ldots, B_{t}\}, where t = r_{{\cal B}}(a) \in \{0,1,2,3\}. Let B'_i=B_i \setminus \{a\} for each i \in \{1, \ldots, n\}. By induction, there is a \mathrm{DPD}(v-1), (V\setminus\{a\}, \{T_1', \ldots, T_n'\}), such that T_i' is a permutation of B'_i for each i \in \{1, \ldots, n\}.

We will create a \mathrm{DPD}(v), (V,\mathcal{T}), with the required properties where \mathcal{T}=\{T_1, \ldots, T_n\}, T_i=T_i' for every i \in \{t+1, \ldots, n\} and for i \in \{1, \ldots, t\}, T_i is a permutation of B_i = B_i' \cup \{a\} such that T_i' is a subsequence of T_i. In other words, for i \in \{1, \ldots, t\}, T_i is obtained by inserting a into T_i'. Note that to confirm that (V,\mathcal{T}) is a \mathrm{DPD}(v), it will suffice to confirm that, for each point u \in V \setminus\{a\} such that |\mathcal{B}_u \cap \mathcal{B}_a| = 2, we have that the sequence (u,a) is a subsequence of one of the blocks in \mathcal{T}_u \cap \mathcal{T}_a and (a,u) is a subsequence of the other.

We specify how a should be inserted into T_1, \ldots, T_t according to the value of t. Note that if t=0 there is nothing to do, and if t=1, then we can take T_1 = (a) + T_1'. If t=2, we take T_1 = (a)+T_1' and T_2 = T_2'+(a). In each of these cases, it is easy to see that \mathcal{T} forms the blocks of a \mathrm{DPD}(v) with the required properties.

We now consider the case that t=3. Let X=B_1' \cap B_2', Y=B_1' \cap B_3' and Z=B_2' \cap B_3'. Note that X, Y and Z are pairwise disjoint because \mathcal{B}_a = \{B_1,B_2,B_3\} and (V,\mathcal{B}) is a \mathrm{PD}_2(v). Let p=|X|, q=|Y| and r=|Z| (noting that some or all of p, q and r may be zero) and say X=\{x_1, \ldots, x_p\}, Y=\{y_1, \ldots, y_p\} and Z=\{z_1, \ldots, z_p\}. Suppose without loss of generality that:

• (x_1, \ldots, x_p) and (y_1, \ldots, y_q) are subsequences of T_1',

• (x_p, \ldots, x_1) and (z_1, \ldots, z_r) are subsequences of T_2', and

• (y_q, \ldots, y_1) and (z_r, \ldots, z_1) are subsequences of T_3'.

Below, it will be convenient to use x_0, x_{p+1}, y_0, y_{q+1}, z_0, z_{r+1} as placeholders representing the beginning or end of certain sequences. For i \in \{0, \ldots, p+q\}, define j(i), k(i), R(i) and S(i) as follows.

• Let j(i) be the greatest element j of \{1, \ldots, p\} such that x_j is one of the first i entries of T_1'[X \cup Y] if such an element exists; otherwise, let j(i)=0.

• Let k(i) be the greatest element k of \{1, \ldots, q\} such that y_k is one of the first i entries of T_1'[X \cup Y] if such an element exists; otherwise, let k(i)=0.

• Let R(i) be the set of consecutive integers in \{0, \ldots, r+1\} such that, in (z_0,x_{p+1})+T_2'[X \cup Z] + (x_0,z_{r+1}), \min(R(i)) is the largest index of an element of Z that occurs before x_{j(i)+1} and \max(R(i)) is the smallest index of an element of Z that occurs after x_{j(i)}. So R(i) contains the set of all indices of elements of Z that lie between x_{j(i)} and x_{j(i)+1} in this block, as well as \min(R(i)) and \max(R(i)).

• Let S(i) be the set of consecutive integers in \{0, \ldots, r+1\} such that, in (z_{r+1},y_{q+1})+T_3'[Y \cup Z] + (y_0,z_{0}), \max(S(i)) is the smallest index of an element of Z that occurs before y_{k(i)+1} and \min(S(i)) is the largest index of an element of Z that occurs after y_{k(i)}. So S(i) contains the set of all indices of elements of Z that lie between y_{k(i)} and y_{k(i)+1} in this block, as well as \min(S(i)) and \max(S(i)).

Note that R(i) and S(i) are well defined because x_{j(i)} and x_{j(i+1)} occur between z_0 and z_{r+1} in (z_0,x_{p+1})+T_2'[X \cup Z] + (x_0,z_{r+1}) and y_{k(i)} and y_{k(i+1)} occur between z_{r+1} and z_0 in (z_{r+1},y_{q+1})+T_3'[Y \cup Z] + (y_0,z_{0}). We note the following properties.

1. (i) |R(i)| \geq 2 and |S(i)| \geq 2 for each i \in \{0, \ldots, p+q\}.

   This is because \min(R(i))<\max(R(i)) and \min(S(i))<\max(S(i)) by the definitions of R(i) and S(i).

2. (ii) j(0)=k(0)=0, so \max(R(0)) = r+1 and \min(S(0))=0.

   We have j(0)=k(0)=0 by definition. Thus z_{r+1} is the first element of Z after x_{j(0)} in (z_0,x_{p+1})+T_2'[X \cup Z] + (x_0,z_{r+1}) and z_0 is the first element of Z after y_{k(0)} in (z_{r+1},y_{q+1})+T_3'[Y \cup Z] + (y_0,z_{0}).

3. (iii) j(p+q)=p and k(p+q)=q, so \min(R(p+q))=0 and \max(S(p+q))=r+1.

   The first p+q elements of T_1'[X\cup Y] comprise all of X and Y, so j(p+q)=p and k(p+q)=q follow. Thus \min(R(p+q)) is the index of the last element of Z before x_{p+1} in (z_0,x_{p+1})+T_2'[X \cup Z] + (x_0,z_{r+1}) and \max(S(p+q)) is the index of the last element of Z before y_{q+1} in (z_{r+1},y_{q+1})+T_3'[Y \cup Z] + (y_0,z_{0}).

4. (iv) If the (i+1)st entry of T_1'[X \cup Y] is in X, then j(i+1)=j(i)+1, k(i+1)=k(i), \max(R(i+1))=\min(R(i))+1 and S(i+1)=S(i).

   Since the (i+1)st entry of T_1'[X \cup Y] is in X, it is clear that j(i+1)=j(i)+1. Since the (i+1)st entry is not in Y, k(i+1)=k(i), and it follows that S(i+1)=S(i).

   Also, \max(R(i+1))=\min(R(i))+1 because x_{j(i)+1}=x_{j(i+1)} and, in (z_0,x_{p+1})+T_2'[X \cup Z] + (x_0,z_{r+1}), the indices of the last element of Z before x_{j(i+1)} and the first element of Z after x_{j(i+1)} are consecutive.

5. (v) If the (i+1)st entry of T_1'[X \cup Y] is in Y, then j(i+1)=j(i), k(i+1)=k(i)+1, R(i+1)=R(i), and \min(S(i+1))=\max(S(i))-1.

   This follows similarly to (iv), with the roles of X and Y interchanged, and considering the block T_3' in place of T_2'.

Let \ell be the least element of \{0, \ldots, p+q\} such that \min(R(\ell)) < \max(S(\ell)). (Note that such an \ell exists by (iii).)

Claim. |R(\ell) \cap S(\ell)| \geq 2. 1ex

Proof of claim. Note that \ell \neq 0 since \min(R(0)) \geq 0 = \max(S(0)) by (ii). Thus \ell-1 \geq 0 and by minimality of \ell, \min(R(\ell-1)) \geq \max(S(\ell-1)).

First, suppose that the \ellth entry of T_1'[X \cup Y] is in X. Then by (iv),

Thus, since \min(R(\ell)) < \max(S(\ell)), it follows that

As the elements of R(\ell) are consecutive, this means that R(\ell) must contain both \max(S(\ell))-1 and \max(S(\ell)). But from (i), |S(\ell)| \geq 2, and the elements of S(\ell) are consecutive, so \{\max(S(\ell))-1, \max(S(\ell))\} \subseteq S(\ell). Thus |R(\ell) \cap S(\ell)| \geq 2.

Otherwise, the \ellth entry of T_1'[X \cup Y] is in Y. A similar argument using (v) shows that |R(\ell) \cap S(\ell)| \geq 2. This concludes the proof of the claim. 1ex

It follows that for some m, \{m, m+1\} \subseteq R(\ell) \cap S(\ell).

• Let T_1 be the permutation of B_1 obtained by inserting a into T_1' immediately after the \ellth entry (and at the beginning of the sequence if \ell=0). Note that, in (x_0,y_0) + T_1 + (x_{p+1},y_{q+1}), a will appear between x_{j(\ell)} and x_{j(\ell)+1} and also between y_{k(\ell)} and y_{k(\ell)+1}.

• Let T_2 be a permutation of B_2 obtained by inserting a into T_2' so that, in (z_0,x_{p+1})+T_2+(x_0,z_{r+1}), a appears between x_{j(\ell)+1} and x_{j(\ell)} and also between z_m and z_{m+1}. A suitable position to insert a exists because \{m,m+1\} \subseteq R(\ell).

• Let T_3 be a permutation of B_3 obtained by inserting a into T_3' so that, in (z_{r+1},y_{q+1})+T_3'+(y_0,z_0), a appears between y_{k(\ell)+1} and y_{k(\ell)} and also between z_{m+1} and z_m. A suitable position to insert a exists because \{m,m+1\}\subseteq S(\ell).

From the above, it is easy to confirm that, for each point u \in V \setminus \{a\} such that |\mathcal{B}_u \cap \mathcal{B}_a|=2 we have that (u,a) is in one of the blocks in \mathcal{T}_u \cap \mathcal{T}_a and (a,u) is in the other.

The upper bound follows from Lemmas 1.5 and 3.1. By Lemma 3.3, there exists a \mathrm{PD}_2(n;v,k), (V,\mathcal{B}), such that |\mathcal{B}_x| \leq 3 for each x \in V; existence of a \mathrm{DPD}(n;v,k) now follows from Lemma 4.1.