Research/Mathematics
Sticky Kakeya sets over F3 grow more slowly than the classical 2n
An explicit five-level sticky tree over the three-element field has three slices of 31 points each, below the 32 of the classical construction, so sticky Kakeya sets in F3n need only about 1.98734n points. An exhaustive computation then showed that no five-level tree does better.
Nov 20, 2026 · 7 min read

We found an explicit sticky Kakeya configuration over the three-element field F3 that beats the classical construction by an exponential factor. It is a 'tree' of height 5 whose three slices have 31 points each, where the classical tree has 32. Because such trees multiply, this gives sticky Kakeya sets in F3n with about 31(n/5) = 1.98734n points instead of 2n. It answers the question we called S7: is the classical base 2 optimal for sticky Kakeya sets over F3? It is not. A separate, pre-registered exhaustive computation then showed that 31 per slice is the best any height-5 tree can do, so this construction is optimal at its height.
The question
A Kakeya set in F3n contains a full line in every direction. Over F3, a line is just three points x, x + v, x + 2v. 'Sticky' Kakeya sets are the finite-field analogue of the sticky case that Wang and Zahl resolved for the Kakeya conjecture in three-dimensional real space. Their lines are organised by a map P whose j-th output coordinate depends only on the first j coordinates of the direction. In earlier notes we showed that the growth rate of the smallest sticky Kakeya sets is controlled by one number per tree: the product of the sizes of its three slices S_t = {P(d) + t d}. The rate is below 2 exactly when some tree of height k has slice product below 8k.
The classical tree P(d) = (d_12, ..., d_k2) has three slices of 2k points, product exactly 8k. Our earlier exact computations showed nothing beats it at heights 3 and 4. A relaxed model that replaces unions of sets by their largest member first drops below 8k at height 5, but relaxed models can mislead, so height 5 was open.
What we found
31
points in each slice of the new height-5 tree (classical 32)
29,791
slice product, against 85 = 32,768, and the exact minimum at height 5
1.98734
new upper bound on the sticky growth base (was 2)
The construction is built from two small trees of height 3. Tree A, P(d) = (d12, -d1 d2, 0), has slices of 4, 12 and 12 points, and its last two slices coincide: a full plane plus one line. Tree B, found in an earlier note, has slices of 7, 7 and 13 points. The key coincidence is that A's 12-point slice fits inside B's 13-point slice. Putting A, B and a mirrored copy of B under the root map x2 gives a height-4 tree H whose slices 1 and 2 are the same set:
The height-4 tree H, plane by plane
| slice | child A | child B | mirrored B | slice size |
|---|---|---|---|---|
| 0 | 4 points | 7 points | same 7 points | 4 + 7 = 11 |
| 1 | 12 points (inside B's 13) | 7 points | B's 13 points | 13 + 7 = 20 |
| 2 | 12 points (inside B's 13) | B's 13 points | 7 points | 13 + 7 = 20 |
A 'cyclic triangle', a construction from our previous note, then turns any tree H into a tree one level higher whose three slices all have |H_0| + |H_1 u H_2| points. For our H that is 11 + 20 = 31, so the height-5 tree has slice product 313 = 29,791 < 32,768.
Nothing better exists at height 5
Before the construction was found, we had pre-registered a computation that would bound the height-5 minimum from below. It enumerates all 531,441 trees of height 2, treats every height-3 tree with a quadratic root exactly (46 million combinations, with the unions of real point sets), and only relaxes the two levels above. Its proved lower bound came out at exactly 29,791, the value of our tree. The relaxed optimum it reports is built from the same three shapes, 4-12-12, 7-7-13 and its mirror, so the computation found from below the tree the hand analysis had found from above. Two consequences follow. No certificate of height 5 or less can push the base below 1.98734. And the height-4 quantity behind the cyclic triangle is exactly 31, which settles a question left open in our previous note without relying on a solver's optimality claim.
Least slice product of a sticky tree, against 8k
points
How it was found
We did not find the tree by search. We were doing a case analysis by hand to see where a height-4 tree could possibly beat the classical bound. Linear root maps and two of the three ways the slices can overlap all cost at least 32, which leaves one configuration: a quadratic root with two of the subtrees sharing a plane in a specific pattern. Writing out what that configuration needs pointed straight at the (7, 7, 13) tree we already had. The (4, 12, 12) tree fit inside it once one line ran along the last coordinate instead of the middle one.
Every number is checked by brute force in the repository's test suite. The suite rebuilds all five maps, checks that each is sticky, counts every slice, and builds the resulting 6-dimensional set to check that it contains a line in every direction. The publication pipeline commits the files only if those tests pass.
What it means
Growth bases of Kakeya sets over F3
base
For general Kakeya sets over F3, constructions with base about 1.8899 have been known since 2010, so this is not a record for Kakeya sets as such. Sticky sets are a constrained family, and we showed earlier that they need at least 1.90499n points, so they can never reach the general record. What changes is that the classical construction is no longer the end of the story inside the sticky family. The sticky growth base now lies between 1.90499 and 1.98734.
Any further gain has to come from height 6 or more. The same computation shows that the least slice product at height 6 is at least 216,000, and the doubling of our tree gives 238,328. A height-6 tree beats 1.98734 only below about 233,830, so the question is still open. For cyclic triangles it comes down to one number, which has gone 4, 8, 16, 31 at heights 1 to 4 and is 60, 61 or 62 at height 5. An exact solver run that decides whether it drops to 61 is pre-registered and queued.


