Lower endpoint reported by Grace in a peer-reviewed 2026 article.
A line detector over three symbols
The Polynomial That Saw Every Line
A polynomial over three symbols detects every forbidden line, then becomes diagonal when none remain. Build a cap set, watch the rank certificate count its monomials, and minimise the live bound until its exponential base falls below three.
0 unless x+y+z=0, then 1
Each coordinate has only three values. A nonzero coordinate sum has square 1, so its factor vanishes. A zero coordinate sum has square 0, so its factor stays 1. The product sees the whole line at once.
Put the third point in
Choose two distinct points. The machine computes the one point that completes their line. Put it into the set and an off-diagonal tensor entry appears. Remove it and the entry is gone.
The set A
Every chip is one vector in F3n. Press a chip to include or exclude it.
Complete one line
0
9
0
The restricted tensor, one cell for each ordered pair (x,y)
A lit cell means the unique completion z is also in A. For a cap, only x=y=z survives, so the support is exactly the diagonal.
Count the slices, then move the base
Diagonal support has rank |A| in characteristic three. Expansion gives another description of the same tensor. Every term must have a low-degree x, y, or z block, so the number of possible slices puts a ceiling on |A|.
The exact monomial count
This is the proof cutoff. At least one of the three blocks must land here.
| coefficient convolution | 168 |
|---|---|
| exponent enumeration | 168 |
| 3mn | 504 |
| whole space 3n | 729 |
The finite bound, root by root
The purple path is (3mn)1/n. The floor in the cutoff makes it wobble. The horizontal gold line is the limiting generating-function base, and the grey line is 3.
An earlier bound of the form C3n/n1+epsilon has nth root tending to 3 for every fixed positive C and epsilon. No values for those unspecified constants are invented on this plot.
Every product term has total degree at most 2n. If its x, y, and z blocks all had degree above 2n/3, their sum would exceed 2n. One block must therefore enter a ledger.
10 distinct low blocks
10 distinct low blocks
1 distinct low block
All 343 expanded terms assigned. Each ledger uses no more than m3=10 low-block monomials.
Minimise the exponential envelope
The coefficient count obeys mn ≤ t-2n/3(1+t+t²)n for every 0<t<1. Move t. The curve itself tells you where the smallest base lives.
base = 2.801394247282882
Differentiation gives 4t²+t-2=0, so the positive root is t=(sqrt(33)-1)/8. The button keeps that exact JavaScript value. No stepped control rounds it.
The check
The green lane is recomputed in this browser. The other lanes name what entered from a source, what you were free to choose, and where the proof stops.
Recomputed now
The offline verifier also reconstructs the expanded polynomial from the grouped slices and exhausts every subset for n=1 and n=2. Run node research/cap-set-polynomial/verify-cap-set-polynomial.mjs.
Every free choice
- You choose the small-space dimension, the members of A, and the ordered pair to complete.
- Restore uses convenient constructed caps of sizes 2, 4, 9, and 16. The n=4 preset is not the published maximum 20.
- You choose n and the degree cutoff in the coefficient bench. Only
floor(2n/3)is the proof cutoff. - You choose the microscope dimension and t. The minimum button uses the exact positive root of
4t²+t-2=0. - The slice ledger assigns ties in x, then y, then z order. Another tie rule changes the ledger split, not the total ceiling of 3mn.
Vacuous and broken settings
- If A contains a line, the tensor is not diagonal. The diagonal-rank step no longer bounds |A|.
- Choosing x=y produces the unavoidable diagonal solution x=y=z. It is not a forbidden line.
- For some small n, 3mn is at least 3n. The displayed theorem is then true but no better than counting the whole space.
- The rank comparison uses output arithmetic in characteristic three. Over the real or complex numbers, this tensor does not give the same nontrivial rank bound.
Uncertainty and boundary
- The exact maxima 2, 4, 9, 20, 45, 112 are published context. This bound does not derive them.
- A peer-reviewed 2026 source says exact maxima are known through n=6. The exact value from n=7 onward remains unknown there.
- The true asymptotic growth base is not known. The polynomial gives the upper endpoint 2.755104613023633, rounded to 15 decimal places.
- The peer-reviewed lower endpoint reported in 2026 is 2.2202. A 2025 preprint claims 2.2203. The page does not merge their evidence levels.
What the certificate settled, and what it did not
Let M(n) be the largest cap in F3n. The method proved that limsup M(n)1/n is below 3. It did not determine M(n) in general, and it did not determine the true asymptotic base.
Published exact maxima, not outputs of the polynomial bound. The verifier independently recovers only the first two by exhaustive subset search.
Lower bound claimed by Zhai and coauthors in a 2025 preprint. Not peer-reviewed.
Upper base computed here from the minimiser and the published algebraic form. Rounded to 15 decimal places.
Source status and chronology
- Ellenberg and Gijswijt, Annals of Mathematics 185 (2017), 339-343. Peer-reviewed. This is the finite-field theorem, including |A| ≤ 3m and the q=3 exponential gap.
- Tao, 18 May 2016. Author exposition, not peer-reviewed. This supplies the symmetric product polynomial and characteristic-three rank account used here.
- Zeilberger, arXiv:1607.01804, v1, 6 July 2016. Preprint. Its title prints the exact radical and decimal for the upper base.
- Croot, Lev, and Pach, arXiv:1605.01506. It has two versions, v1 on 5 May 2016 and v2 on 21 May 2016, and concerns Z4n, not the finite-field cap-set theorem. Tao's symmetric account followed on 18 May. Ellenberg and Gijswijt, arXiv:1605.09223, has one version posted 30 May 2016 and records essentially identical arguments developed independently and simultaneously.
- Bateman and Katz, Journal of the American Mathematical Society 25 (2012), 585-613. Peer-reviewed source for the earlier bound with an unspecified positive constant and exponent. The page compares only its nth-root limit, not a fabricated finite curve.
- Gijswijt, STAtOR, June 2019, pages 12-13. Expository magazine source for the small-dimension table.
- Grace, Algebraic Capsets, 2026. Peer-reviewed, first published 21 May 2026. Source for the exact-status boundary and the 2.2202 to 2.756 interval.
- Zhai and coauthors, arXiv:2508.07932, v1, 11 August 2025. Preprint source for the claimed 2.2203 lower bound.