WorksMathematics
New bounds on simplex-avoiding sets in two dimensions
Abstract
We study extremal densities of measurable subsets of metric spaces avoiding prescribed finite geometric configurations. We prove two new upper bounds for three-point avoidance problems: a measurable subset of containing no orthonormal basis has normalized spherical measure at most , and a measurable subset of containing no unit equilateral triangle has upper Banach density at most . The first bound appears to be the first improvement over the trivial averaging bound for this problem, while the second improves the previously best known upper bound. The proofs are applications of the geometric fractional chromatic number method introduced by Ambrus et al. For this purpose, we formulate the method in a framework that applies both to compact homogeneous spaces and to Euclidean spaces with upper Banach density.
Introduction
We study the maximal density of measurable subsets of geometric spaces avoiding prescribed finite configurations. Following [5], we refer to such extremal quantities as independence densities, generalizing the independence ratio of a graph or hypergraph.
Our first problem concerns subsets of the unit sphere containing no orthonormal basis. Let denote the supremum of the normalized spherical measures of such sets. The elementary averaging argument gives . We prove the following improvement.
Theorem 1.1. We have
The problem has previously been considered in [4, 6, 7, 10], mostly in connection with lower bounds or higher-dimensional variants. To the best of our knowledge, our bound is the first explicitly established improvement over for this independence-density problem. We could deduce a weaker improvement from the work of Conway and Kochen [8]: their configuration of 40 orthonormal triples implies, by averaging, .
Our second problem concerns measurable subsets of containing no unit equilateral triangle. Let denote the supremum of their upper Banach densities. A construction of density 1/2 is conjectured to be optimal. The previously best known upper bound was 0.64355, obtained in [6]. We improve it as follows.
Theorem 1.2. We have
Both of our results are obtained using the geometric fractional chromatic number (GFCN) method, introduced in [15] and subsequently developed in [12, 14, 11]. The method considers random isometric placements of finite point configurations, called probes. The resulting distributions of intersection patterns satisfy linear constraints arising from the forbidden configurations and geometric congruences. Solving the corresponding linear programs yields upper bounds on independence densities.
Our bounds use a 15-point probe in the projective plane and a 14-point probe in the Euclidean plane. The probes were found by computer search, and the bounds are certified by exact rational dual solutions. The supplementary material [13] contains the certificates and an exact-arithmetic verifier.
Section 2 gives an informal overview of the GFCN method. In Sections 3 and 4, we formulate the method for compact homogeneous spaces and Euclidean spaces, respectively. Sections 5 and 6 establish the two bounds above.
The probe method: an informal overview
The geometric fractional chromatic number (GFCN) method was introduced in [15]. Here we describe its probabilistic, probe-based interpretation, introduced in [11]. The basic idea is to study a hypothetical configuration-avoiding set by testing it with randomly placed finite point configurations, which we call probes. Let be a Borel set avoiding the forbidden configurations, and let be a nonempty finite probe. We place a congruent copy of at random and record which points of land in . This produces a random subset , or equivalently a probability distribution on the subsets of . The placement rule is chosen to respect the symmetries of the ambient space. In Euclidean spaces, we initially work with periodic sets; the precise constructions are given in Sections 3 and 4.
There are two sources of constraints on the observation distribution. The first is combinatorial: an observed subset cannot contain a forbidden configuration, so for every such . Of course, all observation probabilities are nonnegative and sum to one. These restrictions depend only on which subsets of contain forbidden configurations.
The second source is geometric. For any , the probability that all points of land in , regardless of what happens to the remaining points of , is
If two subsets are congruent under the relevant group of isometries, their random placements have the same distribution, and therefore
These equalities apply to subconfigurations of any size, not just individual points or forbidden configurations. They use geometric information about that is not captured by its forbidden-configuration hypergraph. The strength of the method comes from combining these geometric consistency relations with the combinatorial restrictions on the possible observations.
Every point of the probe has the same probability of landing in , which equals the density of (or its periodic density in the Euclidean setting). We can therefore bound the density by maximizing the common singleton probability over all probability distributions on admissible subsets of satisfying the congruence equalities. This is a finite linear program. Every avoiding set under consideration yields a feasible observation distribution, so the LP optimum bounds its density.
This density formulation is equivalent to the usual GFCN formulation by a simple change of normalization. In the latter, one assigns nonnegative weights to admissible subsets, requires for every , imposes the same congruence relations, and minimizes the total weight . If a probability distribution has common singleton marginal , then has total weight . Conversely, normalizing such weights to have total mass one gives a probability distribution with singleton marginal . Thus the optimum of our density LP is the reciprocal of the corresponding GFCN. If we drop the congruence constraints for subsets of size at least two from the coloring formulation, retaining only the equal singleton marginals, then we recover the ordinary fractional chromatic number of the hypergraph on whose edges are the forbidden configurations (see [9]). Thus the additional geometric consistency constraints are precisely what distinguishes GFCN from standard fractional coloring. In this paper, we use the probability normalization because it directly expresses the desired density bound.
The practical challenge is to find probes for which this linear program gives strong bounds. In our applications, we search for such probes computationally and verify the resulting bounds using exact rational certificates. The search procedure is described in Appendix A.
A related use of probabilistic symmetrization appears in the recent proof by OpenAI [22] that the Euclidean plane is not five-colorable. Their proof averages over translations and rotations to construct invariant probability distributions on colorings, followed by a transfer to weak measurable colorings. Their arguments differ substantially from our finite-probe approach, but both methods exploit geometric symmetries through probabilistic averaging.
Probes in compact spaces
Let be a compact metric space equipped with a prescribed Borel probability measure . Let be a compact group acting continuously and transitively on by isometries, and assume that is -invariant. We write for normalized Haar measure on . The random placement procedure is to choose according to , and to send each point to .
Lemma 3.1. For , let be the map . The pushforward measure is equal to .
Proof. The measure is -invariant, since left translation preserves Haar measure. A compact transitive -space has a unique -invariant Borel probability measure: identify with and push Haar measure forward through the quotient map. Since is -invariant, the two measures are equal. □
Thus the Haar-random placement rule is compatible with the density measure: for every Borel set , is the probability that a randomly placed point lands in .
Let be a -invariant family of nonempty finite subsets of . Thus, if , then for every . Examples include the pairs of points at distance in a metric space, or the triples forming orthonormal bases on . A set is -independent if no member of is contained in . We write
For a probe , and for , define
and
This is the probability that a random placement of produces exactly the observed pattern . For a finite set , define
Since is continuous, these are probabilities of Borel events whenever is Borel.
For , define
Then
Indeed, the event is the disjoint union of the events over all .
Let
If is -independent, then for every , since otherwise a forbidden configuration inside would be placed inside . Also, if are congruent, meaning for some , then
because by right invariance of Haar measure. Finally, for every ,
Definition 3.2. Fix a probe and a point . Let be the optimum of the following linear program. The variables are , , and
Maximize , subject to
and
Since all singleton subsets of are congruent, the optimum does not depend on the choice of .
Theorem 3.3. For every probe ,
Proof. Every Borel -independent set gives a feasible point of the linear program above. Its objective value is
Taking the supremum over proves the theorem.
Remark 3.4. If one wants to take the supremum over all -measurable independent sets instead of Borel ones, the value is unchanged. Indeed, is a regular Borel probability measure on the compact metric space , and every measurable independent set contains compact independent subsets of arbitrarily close measure.
Probes in Euclidean spaces
We next record the corresponding form of the method for Euclidean spaces. Let be the Euclidean isometry group, and let be an -invariant family of nonempty finite subsets of . Thus, if , then for every and . A measurable set is -independent if no member of is contained in .
For a measurable , define its upper Banach density by
where denotes -dimensional Lebesgue measure. We write
First consider periodic independent sets. Let be a full-rank lattice, let
and let be the quotient map. If , where is Borel, its periodic density is
where is normalized Haar measure on .
The random placement procedure for an -periodic set is to choose
according to product Haar measure and to place at
For a probe , , and a periodic Borel set , define
For a finite set , define
These are probabilities of Borel events, since is continuous.
For , define by the same sum as in § 3. As before, , since the event that all points of land in is the disjoint union of the exact observations containing .
If is -independent, then whenever , because any placement observing would lift a forbidden configuration to for a lift of .
If are Euclidean-congruent, then
Indeed, if , then the change of variables preserves product Haar measure on . Finally,
for every , because is uniformly distributed on for every fixed .
For a Euclidean probe and , define by the linear program in § 3, with congruence interpreted as Euclidean congruence. The preceding arguments show that satisfies its constraints whenever is an -periodic -independent set.
Proposition 4.1. If , with Borel, is -periodic and -independent, then
for every probe .
Proof. The actual observation probabilities form a feasible point of the linear program defining , and the objective value is
The following lemma gives a periodic approximation when the forbidden configurations have uniformly bounded diameter.
Write , with when .
Lemma 4.2. Assume . Let be Lebesgue measurable and -independent. For every , there are a full-rank lattice and an -periodic Borel -independent set such that
Proof. Choose . Since the boundary-strip proportion tends to zero as , the definition of lets us choose and such that
and such that the boundary strip
has measure at most . By inner regularity, choose a compact set
with
Let , and set
Then is -periodic and Borel. Distinct translates , , are separated by distance at least . Hence any member of contained in would have to lie in a single translate of . Translating it back would give a member of contained in , contradicting the -independence of .
The quotient map is injective on , and is compact in . Therefore and
Theorem 4.3. Assume . Then, for every probe ,
Proof. Apply Lemma 4.2 to a measurable -independent set , then apply Proposition 4.1 to the resulting periodic set, and let . □
Orthonormal basis-avoiding sets on the sphere
In this section we apply the GFCN machinery to prove Theorem 1.1, which states that .
The problem of avoiding orthonormal bases is closely tied to the Kochen–Specker theorem. The theorem states that there is no assignment of values 0 and 1 to all points of such that antipodal points receive the same value and every orthonormal basis of has exactly one vector assigned 1 [1]. The condition on orthonormal bases is equivalent to requiring that no two orthogonal points both receive 1 and that no orthonormal basis consists entirely of points receiving 0. Our problem retains only the latter restriction, requiring the measurable set under consideration to contain no orthonormal basis without requiring its complement to avoid orthogonal pairs.
Let denote normalized spherical measure. If contains no orthonormal basis, then at most two vectors of any orthonormal basis lie in . Averaging over uniformly random orthonormal bases, each vector lies in with probability , so . Therefore
For lower bounds, it is useful to consider the classical problem of avoiding orthogonal pairs. Witsenhausen asked for the largest measurable subset of containing no orthogonal pair, and the expected extremal example is the double cap: two antipodal open spherical caps of angular radius [2, 3]. On a double cap has measure . The best known upper bound for Witsenhausen’s problem is 0.2953, obtained in [11]. Since a single double cap contains no orthogonal pair, the union of two double caps with orthogonal axes contains no orthogonal triple; this already gives
Recently, Williams and Constantin [10] improved this lower bound by constructing a curved “scalloped belt” with no three mutually orthogonal directions and measure
Their construction supersedes earlier improvements by Perez [4]. They conjecture it to be optimal. This construction arose from the study of subsets of admitting assignments of 0 and 1 that agree at antipodal points, assign 1 to no orthogonal pair, and assign exactly one 1 to every orthonormal basis contained in the subset. The set is disjoint from the double cap centered at the north and south poles, and their union admits such an assignment, with value 0 on and value 1 on the double cap.
Since changing the signs of the vectors in an orthonormal basis independently preserves orthonormality, the set is basis-avoiding whenever is, and . We may therefore restrict to antipodally symmetric sets and pass to the projective plane by identifying antipodal points on the sphere. Let be the antipodal quotient, and endow with the quotient round metric and the probability measure . Let be the family of projective orthonormal triples
where ranges over orthonormal bases of . For a probe , write .
Lemma 5.1. We have
Proof. Let be measurable and contain no orthonormal basis. For every , choose a compact set with . The set is compact and -independent, since a projective orthonormal triple in would lift to an orthonormal basis contained in . As , we have
Taking the supremum over and letting gives . Conversely, if is Borel and -independent, then is Borel, contains no orthonormal basis, and satisfies . This gives the reverse inequality. □
The space is compact, and the natural action of is continuous, transitive, and isometric. Both and are -invariant. Thus Theorem 3.3 applies to projective probes. Their random placement is obtained by applying a Haar-uniform element of to all the projective points. Here two projective subconfigurations are congruent precisely when a single element of maps one onto the other.
We now present a fifteen-point probe . The fifteen points of are represented by the nonzero integer vectors listed in Table 1. Figure 1 shows the corresponding unit representatives chosen in the closed northern hemisphere. The left panel gives an oblique view of the hemisphere, while the right panel shows its orthogonal projection onto the plane . Quarter great-circle arcs indicate orthogonal pairs among the chosen representatives. The labels in both panels match the row order of Table 1.

Figure 1. The fifteen-point projective probe , represented by unit vectors in the closed northern hemisphere of . Both panels show the same labeled representatives and quarter great-circle arcs joining orthogonal pairs.
For this probe, the finite linear program gives . Combining Lemma 5.1 and Theorem 3.3, we obtain
as asserted in Theorem 1.1.
Unit equilateral triangles in the plane
The seminal paper by Erdős et al. [17], which introduced Euclidean Ramsey theory, begins with a simple illustrative example:
We are concerned in this paper with problems of the type illustrated as follows: Is it true that for any partition of the Euclidean plane into two classes (we say that the plane is
| 1 | 1 | ||
| 2 | 1 | ||
| 3 | 1 | 1 | |
| 4 | 2 | 1 | |
| 5 | 0 | 1 | |
| 6 | 0 | 0 | 1 |
| 7 | 0 | 1 | 1 |
| 8 | 0 | 1 | 0 |
| 9 | 1 | 0 | |
| 10 | 1 | 1 | |
| 11 | 1 | 1 | 1 |
| 12 | 1 | 0 | 1 |
| 13 | 2 | 1 | 1 |
| 14 | 2 | 1 | 0 |
| 15 | 1 | 0 | 0 |
caption: Table 1. Integer vectors representing the projective points of . The indices match the labels in Figure 1.
:::
two-colored), there exists a set of three points all in the same class forming the vertices of an equilateral triangle of side length 1? (We call such a set monochromatic.)
In this example the answer is “no,” as can be seen by dividing the points into two classes according to the parity of .
The two color classes in this construction both avoid unit equilateral triangles and have upper Banach density . A different family of “zebra-stripe” constructions was given by Jelínek et al. [19], achieving the same density. In contrast, Currier, Moore, and Yip [21] showed that every two-coloring of the plane contains a monochromatic congruent copy of any prescribed three-term arithmetic progression.
Let be the vertex set of a unit equilateral triangle, and let
We write
Thus . Although the independence density problem imposes no restriction on the complement of a -independent set, no such set with upper Banach density greater than is known to us. We conjecture that .
To describe the two probes used below, we identify with the complex plane and set and .
We first establish an elementary upper bound using a variation of the construction of Leo and William Moser [16].
Theorem 6.1. .
Proof. Adding one vertex to the Moser spindle gives the eight-vertex unit-distance graph shown in Figure 2.
Figure 2. The eight-point probe and its unit-distance graph , obtained by adding vertex 8 to the Moser spindle.
It is easy to verify that any six vertices of induce a subgraph containing a triangle. By Lemma 4.2, it suffices to prove the bound for -periodic Borel triangle-free sets . By this property of , every placement of in the plane has at most five vertices in . By linearity of expectation, this implies that .
The bound uses only that every triangle-free subset of has at most five points and that each point of has the same inclusion probability under a random placement. The GFCN linear program also imposes congruence constraints for subconfigurations containing more than one point. We apply these additional constraints to a fourteen-point probe .
For , let , using and as above. Figure 3 depicts this probe, and Table 2 lists the integer coefficients .
Figure 3. The fourteen-point planar probe , with edges joining pairs of points at unit distance.
| 1 | 3 | 3 | 1 | 0 |
| 2 | 2 | 4 | 1 | 0 |
| 3 | 0 | 3 | 4 | 0 |
| 4 | 0 | 4 | 4 | 0 |
| 5 | 2 | 2 | 2 | 1 |
| 6 | 1 | 3 | 2 | 1 |
| 7 | 2 | 3 | 2 | 1 |
| 8 | 3 | 3 | 2 | 1 |
| 9 | 1 | 4 | 2 | 1 |
| 10 | 4 | 1 | 0 | 2 |
| 11 | 1 | 1 | 3 | 2 |
| 12 | 3 | 0 | 1 | 3 |
| 13 | 3 | 1 | 1 | 3 |
| 14 | 2 | 2 | 1 | 3 |
Table 2. Integer coefficients specifying the points of . The indices match the labels in Figure 3.
Proof of Theorem 1.2. Every member of has diameter 1. Theorem 4.3 then gives . The linear program gives .
Open Problems
Several natural questions remain open concerning both extremal densities and the effectiveness of the GFCN method itself.
The most immediate question is whether the density construction described in § 6 is optimal. For unit equilateral triangle avoidance, our upper bound of is still above the conjectured optimal value of .
Curiously, we were unable to obtain a better upper bound even under the stronger assumption that both a measurable set and its complement avoid unit equilateral triangles.
The corresponding problem on the sphere also remains open. Williams and Constantin [10] conjectured that their scalloped-belt construction is optimal, which would give
We establish the upper bound . Closing this gap is a natural direction for further work.
A more general question concerns the effectiveness of the GFCN method itself. For a family of forbidden configurations in a space , Theorems 3.3 and 4.3 establish, under their respective assumptions, that
For which spaces and forbidden families does equality hold? Equivalently, can the optimal independence density always be approximated arbitrarily closely by bounds obtained from finite probes? For unit-distance avoidance in the plane, the results of [18] and [20] imply equality, since the LP bound for every finite probe is at most its independence ratio. The question remains open in most other cases.
Use of generative AI tools. Generative AI tools, including GPT-6 and Claude Fable 5.2, were used to assist with writing the computer code, drafting the supplementary documentation, and improving the exposition and organization of the manuscript. All AI-assisted material was reviewed by the authors, who take full responsibility for the mathematical results, computational verification, and final content of the paper.
Acknowledgements
D. C., Á. D., D. V., and P. Zs. were supported by the Hungarian Ministry of Innovation and Technology NRDI Office within the framework of the Artificial Intelligence National Laboratory (RRF-2.3.1-21-2022-00004). Á. D. and D. V. were supported by the grant NKFIH-153165. D. C. and P. M. were supported by the grant NKFIH-154121. We thank András Szabó for his help in preparing the supplementary material, and Gergely Ambrus, Máté Matolcsi, and Kenneth Moore for helpful discussions and valuable insights.
Appendix A. Computational search for the probes
The probes and were obtained by computer-assisted searches over finite geometric configurations, following the general approach of [15, 12, 11]. The search software generates candidate configurations by geometric extensions and ranks them using numerical solutions of their GFCN linear programs. Both beam search and Monte Carlo tree search are implemented for exploring the resulting space of configurations.
For the planar problem, points are represented by integer quadruples encoding , where and are defined in § 6. Candidate extensions include adjoining points at unit distance from existing points, completing equilateral triangles, and completing parallelograms. Configurations are canonicalized under translations and symmetries of the Moser lattice to avoid duplicate candidates. The fourteen integer quadruples for the resulting probe are listed in Table 2.
For the spherical problem, the search works with projective directions, identifying a unit vector with . The extension operation takes two nonparallel representatives and adjoins the projective direction represented by , which is orthogonal to both. The same cross-product extension rule is used in the spherical search described in [11]. The exact coordinates of are listed in Table 1.
For each candidate , its graph has edges between pairs at unit distance in the planar problem or between orthogonal pairs in the spherical problem. The triangle-free vertex subsets are precisely the admissible observation patterns, and the linear program also imposes the congruence constraints of § 3 and § 4. The search evaluates the LP numerically, using a smaller value of or as a better score. For the final probes, has 23 unit-distance pairs and 6 unit equilateral triangles, while has 25 orthogonal pairs and 6 orthonormal triples. These counts can be checked directly from the exact coordinates, whereas the LP congruence constraints also use geometric information not contained in the graphs. The exact LP values and are verified by rational primal and dual certificates independently of the numerical search.
Domonkos Czifra, HUN-REN Alfréd Rényi Institute of Mathematics, Reáltanoda utca 13-15, 1053 Budapest, Hungary Email address: [email protected]
Ákos Dúcz, HUN-REN Alfréd Rényi Institute of Mathematics, Reáltanoda utca 13-15, 1053 Budapest, Hungary Email address: [email protected]
Péter Madarasi, HUN-REN Alfréd Rényi Institute of Mathematics, Reáltanoda utca 13-15, 1053 Budapest, Hungary; and Department of Operations Research, ELTE Eötvös Loránd University, Pázmány P. s. 1/C, 1117 Budapest, Hungary Email address: [email protected]
Dániel Varga, HUN-REN Alfréd Rényi Institute of Mathematics, Reáltanoda utca 13-15, 1053 Budapest, Hungary Email address: [email protected]
Pál Zsámboki, HUN-REN Alfréd Rényi Institute of Mathematics, Reáltanoda utca 13-15, 1053 Budapest, Hungary Email address: [email protected]
References
- [1]S. Kochen and E. P. Specker, The problem of hidden variables in quantum mechanics, J. Math. Mech. 17 (1967), 59–87.
- [2]H. S. Witsenhausen, Spherical sets without orthogonal point pairs, Amer. Math. Monthly 81 (1974), no. 10, 1101–1102.DOI
- [3]G. Kalai, Some old and new problems in combinatorial geometry I: around Borsuk’s problem, in Surveys in Combinatorics 2015, London Math. Soc. Lecture Note Ser. 424, Cambridge Univ. Press, Cambridge, 2015, pp. 147–174.arxiv.org/abs/1505.04952
- [4]A.-M. Perez, On generalizations of the double cap conjecture, Research Science Institute report, 2019.
- [5]D. Castro-Silva, Geometrical sets with forbidden configurations, Forum Math. Sigma 11 (2023), e44.
- [6]D. Castro-Silva, F. M. de Oliveira Filho, L. Slot, and F. Vallentin, A recursive Lovász theta number for simplex-avoiding sets, Proc. Amer. Math. Soc. 150 (2022), no. 8, 3307–3322.
- [7]D. Zakharov, Spherical sets avoiding orthonormal bases, C. R. Math. 363 (2025), 1351–1362.arxiv.org/abs/2310.06821
- [8]J. H. Conway and S. Kochen, The free will theorem, Foundations of Physics 36 (2006), 1441–1473.DOI
- [9]E. R. Scheinerman and D. H. Ullman, Fractional Graph Theory: A Rational Approach to the Theory of Graphs, Wiley, New York, 1997.
- [10]T. Williams and A. Constantin, Maximal non-Kochen-Specker sets and a lower bound on the size of Kochen-Specker sets, Phys. Rev. A 111 (2025), no. 1, 012223.DOI
- [11]D. Czifra, Á. Dúcz, M. Matolcsi, D. Varga, and P. Zsámboki, Improved bounds for the double cap conjecture, preprint, https://arxiv.org/abs/2605.28709.
- [12]M. Matolcsi, I. Z. Ruzsa, D. Varga, and P. Zsámboki, The fractional chromatic number of the plane is at least 4, preprint, https://arxiv.org/abs/2311.10069.
- [13]Supplementary material for this paper (verification code and data). Available at https://users.renyi.hu/~akos/triangles/triangle_reproduction/site/.
- [14]V. F. Lev, M. Matolcsi, P. P. Pach, and D. Varga, On the density of Kravitz sets, preprint, https://arxiv.org/abs/2510.16522.
- [15]G. Ambrus, A. Csizsárik, M. Matolcsi, D. Varga, and P. Zsámboki, The density of planar sets avoiding unit distances, Math. Program. 207 (2024), 303–327.DOI
- [16]L. Moser and W. Moser, Solution to problem 10, Canad. Math. Bull. 4 (1961), 187–189.
- [17]P. Erdős, R. L. Graham, P. Montgomery, B. L. Rothschild, J. Spencer, and E. G. Straus, Euclidean Ramsey theorems. I, Journal of Combinatorial Theory, Series A 14 (1973), no. 3, 341–363.DOI
- [18]S. Li and Y. Su, Every n points in the plane contain (m₁ + o(1))n points with no unit distance, preprint, arXiv:2610.15424 (2026).arxiv.org/abs/2610.15424
- [19]V. Jelínek, J. Kynčl, R. Stolař, and T. Valla, Monochromatic triangles in two-colored plane, Combinatorica 29 (2009), 699–718.DOI
- [20]Á. Dúcz, Measurable Independence Density Equals the Finite Independence-Ratio Infimum in the Euclidean Plane, Hexagon preprint, hexagon:2610.00171v1 (2026). https://hexagonmath.org/2610.00171v1.
- [21]G. Currier, K. Moore, and C. H. Yip, Any two-coloring of the plane contains monochromatic 3-term arithmetic progressions, Combinatorica 44 (2024), 1367–1380.DOI
- [22]OpenAI, The Euclidean plane is not five-colorable, OpenAI Math Release preprint (2026). https://github.com/openai/math/blob/main/preprints/The-Euclidean-plane-is-not-five-colorable-September-23-2026/paper.pdf.

