Write B=bs for the largest column offset. Summing the separation
inequalities gives ai′−ai>(i′−i)B whenever i′>i; in particular
ai′−ai>2B whenever i′≥i+2. For (i,j)∈S the
multiples of pij are
Mij(t)=z−ai+tpij=z+tq+(t−1)ai+tbj(t∈Z).
Call Mij(t) the layer-t location of the cell. The layer 0
location Ri=z−ai depends only on the row and the layer 1
location Cj=z+q+bj only on the column. Write Eij=Mij(−1)
and Dij=Mij(2)=z+2q+ai+2bj.
Clusters. Since all offsets are nonnegative and at most ar and
B, the layer locations of the cells of S lie in
t=−1t=1t=3: [z−q−2ar−B, z−q],: [z+q, z+q+B],: [z+3q, z+3q+2ar+3B].t=0t=2: [z−ar, z],: [z+2q, z+2q+ar+2B],
These five intervals are pairwise disjoint because 2ar+3B<q.
Every layer t≤−2 lies at or below z−2q and every layer t≥4
at or above z+4q.
Multiplicities. (i) Layer −1 is injective on S: an equality
2ai+bj=2ai′+bj′ with i<i′ gives
bj−bj′=2(ai′−ai)>2B, impossible; with i=i′ it forces
j=j′. (ii) Each value of layer 2 is taken by at most two cells: if
ai+2bj=ai′+2bj′ with i<i′, then
ai′−ai=2(bj−bj′)≤2B, so i′=i+1; and within one row the
value determines the column. So the cells sharing a value lie in at most
two adjacent rows, one in each. (iii) The same holds for layer 3: an
equality 2ai+3bj=2ai′+3bj′ with i′≥i+2 would give
2(ai′−ai)=3(bj−bj′)≤3B, while 2(ai′−ai)>4B≥3B;
and within one row the value again determines the column.
In particular the set TS of layer 2 offsets ai+2bj has
∣TS∣≥⌈u/2⌉.
Blocks retaining the row and column locations. Let J be a block of
K consecutive integers containing every Ri and Cj of S. Fix
(i,j)∈S. The integers strictly between Eij=Ri−pij and
Dij=Ri+2pij number 3pij−1<3P≤K, so J, which contains
Ri, also contains Eij or Dij. Let Tin be the set of
t∈TS with z+2q+t∈J, and choose for each
t∈TS∖Tin one cell with ai+2bj=t. Its layer 2
location is not in J, so its layer −1 location is. Distinct values
t give distinct cells and, by (i), distinct layer −1 locations.
By the cluster disjointness, J therefore contains at least
rS+sS+∣Tin∣+∣TS∖Tin∣=rS+sS+∣TS∣
distinct multiples.
For J0=[z−q+1,z−q+K]: it contains every Ri because ar<q;
every Cj because K≥3P≥3q; and every Dij because
P=q+maxS(ai+bj) gives K≥3P≥3q+ai+2bj for each cell of
S. It contains no Eij, which is at most z−q, and no layer
3 location, which is at least z+3q>z−q+K because K<4q. The
remaining layers are farther away. So the multiples in J0 are exactly
the rS+sS+∣TS∣ locations of layers 0, 1 and 2.
Blocks near z. Let J be a block of K consecutive integers inside
[z−2q+1,z+4q−1]. It contains no location of a layer t≤−2 or
t≥4. Each cell has at least three consecutive multiples in J,
since K≥3pij, and three consecutive layers among
−1,0,1,2,3 include at least one of the outer layers −1, 2 and 3.
Let nt be the number of cells whose layer-t location is in J,
for t∈{−1,2,3}; then n−1+n2+n3≥u. By (i) to (iii) these
locations take at least ⌈nt/2⌉ distinct values in each
outer layer, and the clusters are disjoint, so J contains at least
⌈n−1/2⌉+⌈n2/2⌉+⌈n3/2⌉ ≥ ⌈u/2⌉
distinct multiples. This proves the window bound.
Replacing z by z+n∏(i,j)∈Spij keeps the congruences
and translates every block statement, so the canonical block may be
taken to consist of positive integers.
The sharp full-grid bound. For $S={1,\ldots,r}\times
{1,\ldots,s}$, only adjacent rows share third-layer values, and no
value belongs to three rows. Rows i and i+1 share one value for
each pair of columns with
bj−bj′=(ai+1−ai)/2=:di>B/2.
No column participates in two of these pairs: a two-edge chain would
have span 2di>B, and each column has at most one predecessor and one
successor at distance di. Thus the pairs form a matching and number
at most ⌊s/2⌋. With B={b1,…,bs},
counting the overlaps gives
∣TS∣=rs−i=1∑r−1∣(ai+2⋅B)∩(ai+1+2⋅B)∣≥rs−(r−1)⌊s/2⌋.
Here 2⋅B is the dilate, not a two-fold sumset.
For an equality pattern take columns bj=j−1 and rows
ai=2(i−1)⌈s/2⌉. The row gap exceeds s−1 and adjacent
third-layer rows share exactly s−⌈s/2⌉=⌊s/2⌋
values. Fix 3≤α<4 and apply L202 to the cells (−ai,bj).
Their differences ai+bj are distinct. The proof of L202 supplies
integers h,q with all q+h(ai+bj) prime and
(4−α)q≥16hΩ+1, where
Ω≥max(1,ar,bs). Consequently
2har+3hbs<q and
K=⌊αi,jmax(q+hai+hbj)⌋≤α(q+2hΩ)<4q.
The scaled grid therefore meets all the hypotheses and preserves every
projection coincidence. Its canonical positive block has exactly
r+s+rs−(r−1)⌊s/2⌋ counted integers, which is the
core-retaining minimum already proved.