Research/Mathematics
No sticky Kakeya tree over F3 beats the classical one below height 5
A pre-registered, exhaustive computation proves that every sticky Kakeya tree of height 4 over the three-element field has slice product at least 84 = 4096. A relaxation had allowed 3971; real trees cannot realise it. Any finite proof that sticky Kakeya sets over F3 grow slower than 2n must now have height 5 or more.
Nov 20, 2026 · 6 min read

MV2's open question S7 asks whether sticky Kakeya sets over the three-element field F3 can grow with base below 2, that is, whether the classical construction can be beaten by an exponential factor. Note a188 reduced S7 to a finite search: it holds if and only if some sticky tree of some height k has slice product below 8k. We have now proved that no tree of height 4 or less does: the least slice product at height 4 is exactly 4096 = 84. A relaxation used in note a190 had allowed 3971, which left height 4 open. An exhaustive, pre-registered computation of real slice unions shows that this cheaper shape does not exist.
The question
A Kakeya set over a finite field contains a line in every direction. Over F3, a set is sticky if the line in direction d passes through a point whose j-th coordinate depends only on the first j coordinates of d. The smallest such sets are governed by sticky maps P from F3k to itself, called full trees of height k. Each has three slices, S_t = {P(d) + t d} for t = 0, 1, 2. Write pi_k for the least value of |S_0| |S_1| |S_2|. Theorem 78 in a188 shows that the growth base of sticky Kakeya sets equals the cube root of the limit of pi_k(1/k). The classical product map, x maps to x2 in every coordinate, gives exactly 8k. So S7 is true precisely when some tree has pi_k below 8k, and a single explicit tree would settle it.
4096
least slice product at height 4, now proved exact
3971
what the earlier relaxation allowed
46,016,815
exact height-3 combinations checked
What we did
A tree is built from a root map and three subtrees. In each slice, subtrees whose planes coincide contribute the union of their slice sets. Earlier notes replaced every union by its largest member, a relaxation we call the game. The game is easy to compute, but it can be too optimistic. It found a height-4 shape with product 3971 = 11 x 19 x 19, and that shape needs exact coincidences of slice sets at height 3. The job sticky-h3-v1 removed the relaxation at height 3. It enumerated all 531,441 height-2 trees as bit masks and computed the real unions for every quadratic root at height 3. The only reductions it used are exact: entries dominated by subsets, Pareto pruning, and symmetry under the 108 lower-triangular affine maps of F32. It relaxed only height 4. The method, the self-checks and the prediction (55% for this outcome) were fixed in notes/prereg-sticky-h3.md before the run. The run took 80 seconds, its unit tests passed and no self-check failed.
Least slice product of sticky trees over F3
| Height k | pi_k (proved) | Game relaxation | 8k (classical) |
|---|---|---|---|
| 1 | 8 | 8 | 8 |
| 2 | 64 | 64 | 64 |
| 3 | 512 | 512 | 512 |
| 4 | 4096 (new) | 3971 | 4096 |
| 5 | 27,000 to 32,768 | 27,000 | 32,768 |
What we found
Only six height-3 slice-size vectors from the game can take part in a height-4 tree below 4096: the permutations of (4,12,12) and of (7,7,12). The exact computation shows that real trees realise the first three but not the second three. Every real height-3 vector in the relevant range dominates a permutation of (4,12,12). Combined at a quadratic root, these give at best (4+12)(4+12)(4+12) = 4096. A root with an affine map costs at least 9 x 512 = 4608 (Lemma 162). So Proposition 88: every full sticky tree of height 4 over F3 has slice product at least 4096, and the classical tree attains it.
Lower bounds on the height-4 slice product
product
The realised shape has a short formula: P(d1, d2, d3) = (d12, 0, -d1 d3) has slices of sizes 4, 12 and 12, which can be checked by hand. The game shape (7,7,12) needs two subtrees to have identical slice sets inside a shared plane. The exhaustive search shows that no choice of real height-2 subtrees produces it.
Height-3 shapes that could enter a height-4 tree below 4096
| Shape | Product | Realised by a real tree? |
|---|---|---|
| (4,12,12) and permutations | 576 | yes |
| (7,7,12) and permutations | 588 | no (game only) |
What it means
Together with a189, this gives pi_k = 8k exactly for every height up to 4. A tree that answers S7 must have height at least 5. The question stays open. Note a185 proved that the game itself eventually falls below 8k by an exponential factor, and at height 5 it allows 27,000 against the classical 32,768. Two outcomes remain, and either would be new. If exact unions keep raising the game's cheap shapes height after height, that points towards sticky Kakeya sets over F3 having growth base exactly 2. If a real tree of height 5 or more realises a cheap shape, sticky sets beat the classical construction exponentially. The next pre-registered job keeps height 3 exact on its whole range and relaxes heights 4 and 5. It either proves pi_5 = 32,768 or names the height-4 shapes that an exact check must test next.


