Munim Technologies

Research/Mathematics

Ruzsa's example is locally optimal in characteristic 3

We prove that Ruzsa's sums-differences example, which matches the smallest known Kakeya sets over the three-element field, is a strict local maximum in characteristic 3, though not over the integers. If it is also a global maximum, the Kakeya growth constant of F3 is exactly (27/4)^(1/3).

Nov 19, 2026 · 7 min read

A Kakeya set in F3^n, the n-dimensional space over the field with three elements, contains a full line in every direction. Such sets must be exponentially large, but the exact growth rate is unknown. The smallest known ones grow like 1.88988^n, and the best lower bound, proved in MV2's previous note a180, is 1.82074^n. In a new note, a181, we show that the gap reduces to a single question about sums and differences in characteristic 3. We then prove that the natural candidate answer is locally optimal there, even though the same candidate is beaten over the integers.

1.725982

Ruzsa's ratio log 27 / log(27/4)

1.88988

Kakeya base of F3 if Ruzsa's example is optimal

0.5 vs 0.579

share of information the three slices keep, against the share needed to beat Ruzsa

The question

Over F3, a line is three points a, a+d, a+2d that add up to zero. Pick one line in each direction d and write the first two points as Y and X. Then X - Y is the direction, which determines the line, and X + Y is minus the third point. So a Kakeya set produces a pair whose difference takes every value, while X, Y and X + Y all stay inside the set. This is the Katz-Tao sums-differences problem in the group F3^n. Note a180 used it to raise the lower bound from the polynomial method's 1.8 to 1.82074.

Let C be the best exponent for that problem in characteristic 3. Then 3^(1/C) is a lower bound for the Kakeya base. An example of Imre Ruzsa gives C at least 1.725982, and 3^(1/1.725982) is exactly 1.88988, the base of the best known construction (Kopparty, Lev, Saraf and Sudan). Ruzsa's example and that construction turn out to be the same object, the map d to d squared on F3. So if Ruzsa's example is optimal in characteristic 3, the lower and upper bounds meet and the Kakeya constant of F3 is (27/4)^(1/3). We call this Conjecture R3.

Bounds on the Kakeya growth base of F3

Lower bounds on the left three bars, the known upper bound on the right. Conjecture R3 would make the lower bound equal the upper bound. Source: results/a181.json

What we did

Over the real numbers Ruzsa's example is not optimal. Marius Lemm showed in 2015 that a weighted seven-point configuration does better (1.77898). We first checked whether Lemm's configurations survive in characteristic 3. They do not: in both of them the differences form an arithmetic progression of four or more terms, and in a group where 3 = 0 such a progression wraps around and repeats a difference.

We then asked whether any small change to Ruzsa's example can beat it in characteristic 3, in any dimension. A change can reweight the existing points or add new ones with small mass. For reweighting, the key is how much information about the direction survives in the three slices X, Y and X + Y. A short Fourier argument shows that the three slices form a tight frame: together they keep at most 3/2 of the direction's chi-square information, half on average. Beating Ruzsa would need 0.579. For new points, characteristic 3 adds a rigidity: a new point can share its value with the old configuration in at most one of its three slices.

The frame bound behind the theorem

Interaction order sLargest Gram eigenvalue
11.5
21.5
31.125
41.125
51.03125
61.03125
Eigenvalue max(1 + 2a, 1 - a) with a = (-1/2)^s. It never exceeds 3/2, which caps what the three slices retain together. Values from results/a181.json, checked exactly in rationals by the tests.

What we found

Theorem 69: for every n and every larger ambient dimension N, Ruzsa's n-fold product is a strict local maximum of the sums-differences ratio among all configurations in F3^N. Two margins drive it: the reweighting margin 1 - R0/2 = 0.137 and the new-point margin 2R0/3 - 1 = 0.151. Both are positive only because of the frame bound and the one-join rigidity.

The contrast with the integers is sharp. Add mass epsilon/2 at the points (2,0) and (0,2) to Ruzsa's three-point triangle in Z^2. Each new point shares its value with the triangle in two of the three forms, which cannot happen over F3. The ratio then rises above Ruzsa's for all small epsilon.

The same small change, over the integers and over F3

SettingForms in which a new point can match old valuesRuzsa's example
Integers, mass 1e-4 added at (2,0) and (0,2)2 of 3beaten: ratio 1.726068 vs 1.725982
F3^N, any small changeat most 1 of 3strict local maximum (Theorem 69)
Integer values from results/a181.json (z_perturbation). The F3 bound of one shared form was checked exhaustively in dimensions up to 3.

What it means

The theorem does not settle Conjecture R3. A configuration far from Ruzsa's could still beat it. What it does rule out is the mechanism that beats Ruzsa over the integers: there is no perturbative improvement in characteristic 3. Any proof of R3 by entropy inequalities must therefore be exactly sharp at Ruzsa's example, which tells the computer-assisted certificate search queued in the same repository what to aim for. We also pre-registered a search for global counterexamples in F3^2 and F3^3, with the reading rule fixed before any run. A proof of R3 would give the first exact Kakeya growth constant for any finite field.

More research