Munim Technologies

Research/Mathematics

Smaller Kakeya sets over small finite fields

We give an explicit construction of Kakeya sets over every finite field whose growth rate is below every upper bound we found in the literature for fields with four or more elements, settle the growth rate for rank-r sets over the two-element field, and prove that the smallest Kakeya set in three-dimensional space over the three-element field has exactly 13 points.

Oct 14, 2026 · 7 min read

A Kakeya set is a set of points that contains a complete line pointing in every possible direction. Over a finite field, the natural question is how few points such a set can have. We prove that, for every finite field with q elements, the smallest Kakeya sets in n dimensions grow no faster than (q·2^(−(q−1)/q))^n, and we give the sets explicitly. For every q of 4 or more this is below every upper bound we could find in the literature: at q = 4 the growth base drops from 2.8284 to 2.3784, and at q = 8 from 6.3321 to 4.3620. We also pin down the exact growth rate of a higher-rank version of the problem over the two-element field, and prove that the smallest Kakeya set in the 27-point space F₃³ has exactly 13 points.

2.3784

new growth base at q = 4, down from 2.8284

4.3620

new growth base at q = 8, down from 6.3321

2^(3n/4)

points needed by every rank-2 Kakeya set in F₂ⁿ

13

points in the smallest Kakeya set in F₃³, proved

The question

In 1917 Kakeya asked how small a region of the plane can be if a needle of length one can be turned all the way around inside it. Besicovitch showed the area can be made as small as you like, and the question became one about the size of sets that contain a segment in every direction. The finite-field version replaces the plane with Fqⁿ, the n-dimensional space whose coordinates come from a field with q elements. A line is the q points a, a + d, a + 2d and so on, and a Kakeya set must contain one line in each direction d.

Dvir proved in 2008 that for a fixed dimension and large q a Kakeya set must fill a constant fraction of the space. A different regime stays open: fix a small field, say q = 3, and let the dimension n grow. Products of Kakeya sets are Kakeya sets, so the least size K(n, q) grows like βqⁿ for some growth base βq between q²/(2q − 1) and q. Kopparty, Lev, Saraf and Sudan (KLSS, 2010) asked for βq and called q = 3 the most intriguing case.

What we did

Earlier constructions start from one function f on the field, such as f(x) = x², and build lines whose points all lie in one symmetric set. That set has to be large enough for the worst point on every line. We cut the coordinates into blocks and shift each line's parameter separately on each block. The q points of a line then take turns being cheap on different blocks, so each point pays roughly the average cost rather than the worst. Counting points by how often each symbol appears turns that average into a Shannon entropy, and the entropy works out to log₂ q − (q − 1)/q per coordinate, which gives the bound. Membership in the set is decided by counting symbols, so the set is explicit; only the existence of good shifts uses a probabilistic argument, and for q = 3 we also give a deterministic rule.

The same averaging gives an exact formula for βq as an infimum over maps on small blocks. It shows that the new bound is the true value of βq exactly when no map on any block beats the coordinatewise squares. An exhaustive search over all 9⁹ maps on F₃² found none, so the first open case over F₃ is blocks of size three.

What we found

Growth base of the smallest Kakeya sets, by field size

Previous best upper boundThis workLower bound q²/(2q − 1)
02.557.510246810121416Field size qGrowth base (lower means smaller sets)
The previous bounds jump at q = 8 and q = 16 because the best earlier constructions for even q (Kyureghyan, Müller and Wang) differ from those for odd q. Source: kakeya-small-fields: output of `python kakeya_entropy.py table` (kakeya_entropy.py)

Upper bounds on the growth base βq

qLower boundPrevious bestThis workPrevious bound from
21.33331.41421.4142exact (difference bases, Banakh–Gavrylkiv)
31.80001.88991.8899, explicitKLSS Th. 8, random maps (explicit: 2)
42.28572.82842.3784KLSS Th. 8, random maps (explicit: 3)
52.77783.00002.8717KLSS Th. 6, (q + 1)/2
73.76924.00003.8643KLSS Th. 6, (q + 1)/2
84.26676.33214.3620Kyureghyan–Müller–Wang Th. 3.2
94.76475.00004.8603KLSS Th. 6, (q + 1)/2
115.76196.00005.8577KLSS Th. 6, (q + 1)/2
136.76007.00006.8560KLSS Th. 6, (q + 1)/2
168.258110.00008.3542Kyureghyan–Müller–Wang Th. 2.2
Source: python kakeya_entropy.py table in kakeya-small-fields; the repository's tests check every number. At q = 3 the value is not new: KLSS proved it with randomly chosen linear maps, and ours is the first explicit set we know of with that base.

The new bound sits within a factor of about 1 + 0.19/q of the lower bound, so for large q the remaining gap is small. For q = 2 nothing changes: the answer there, √2, was already known exactly.

Rank-r sets over F₂. A Kakeya set of rank r must contain a translate of every r-dimensional subspace, not just every line. KLSS asked for the growth base of these sets for even one case beyond ordinary lines, and called q = r = 2 non-trivial. Over the two-element field we prove, with a short entropy inequality, that every rank-r Kakeya set in F₂ⁿ has at least 2^((1 − 2^(−r))n) points. A construction of KLSS comes within a factor 2^(r+1) of that, so the growth base is exactly 2^(1 − 2^(−r)) for every r. For r = 2, every such set has at least 2^(3n/4) points. The same argument extends to cubes in any finite group.

Exact growth base of rank-r Kakeya sets over F₂

Rank rKLSS lower bound 2^(r+1)/(2^r + 1)Exact growth base 2^(1 − 2^(−r))
11.33331.4142 (known)
21.60001.6818
31.77781.8340
41.88241.9152
51.93941.9571
61.96921.9785
Source: python f2_rank.py table in kakeya-small-fields (f2_rank.py). The r = 1 case was already known.

Exact small cases. Growth rates say nothing about small dimensions, so in the companion repository kakeya-conjecture we also searched for the smallest sets directly. In F₃³ we prove that 13 points are needed and enough. After translating, a 12-point Kakeya set would have to contain one fixed line, and a bitmask search checks all 1,307,504 such 12-point sets in seconds without any solver; none works. Two independent integer-programming solvers also proved 13 optimal. The best construction we found before was Smyth's, with 15 points. In four other small spaces the solvers found sets smaller than any published construction we know of, without proving them optimal.

Smallest Kakeya sets found in small spaces

SpacePoints in spaceLower boundSmallest set foundSmyth constructionOptimal?
F₃³27131315Proved
F₃⁴8118 (Bukh–Chao)2731Not proved
F₃⁵24332 (Bukh–Chao)5663Not proved
F₄³6421 (Bukh–Chao)2834Not proved
F₄⁴25648 (Bukh–Chao)6496Not proved
Sources: kakeya-conjecture README section 7 and data/kakeya-min-v1/results.txt (solve.py run kakeya-min-v1, 2026-10-14). Every set is published as a coordinate list and re-checked line by line by the tests.

What it means

For fixed small fields, the best known Kakeya sets are now noticeably smaller for every q of 4 or more, and the formula for βq turns the remaining question into a finite one: a single map on a small block with lower entropy than the coordinatewise squares, which a computer can check, would beat our bound. Over F₂ the higher-rank question KLSS raised now has a complete answer. The gap between our upper bound and the lower bound q²/(2q − 1) remains open for every q of 3 or more, and we do not claim any improvement below 1.8899 at q = 3.

More research