WorksMathematics
The -Grassmannian, its minuscule Coxeter matroids, and the lines on a cubic surface
Abstract
We study the minuscule Coxeter matroids of type , whose ground set is the set of 27 lines on a smooth cubic surface. We characterize them by a tritangent exchange property and classify them into 230 orbits under the Weyl group. Every nonempty matroidal stratum of the -Grassmannian is an open subscheme of a torus, and all orbits but one are realizable over every infinite field. Finally, we compute the tropical -Grassmannian and its 60 classes of toric degenerations. This is an extended abstract for a forthcoming paper.
Introduction
The -Grassmannian is a remarkable variety that can be approached from several directions: through the exceptional Jordan algebra over the octonions (where it appears as the projective plane over the complexified octonions and is therefore also called the Cayley plane), through the representation theory of , and through the 27 lines on a smooth cubic surface. We show that the last of these gives an illuminating way to understand the –matroids and their strata in the -Grassmannian. Here is the maximal parabolic subgroup associated with the sixth simple root, numbered as in Bourbaki [2 Plate V]. The matroid axiom becomes an exchange property for tritangent planes, the defining quadrics of are signed sums over the tritangent planes through a line, and the analysis of each stratum is governed by how the tritangent planes through different lines meet.
The homogeneous coordinates of are indexed by the 27 weights of the minuscule representation of , whose convex hull is the Gosset polytope ; the –matroids are certain subsets of its vertices. Coxeter identified the vertices of with the 27 lines on a smooth cubic surface [5], and Lee realized as the convex hull of the classes of those lines in the Picard lattice [15]. We start from Lee’s description and project it to the orthogonal complement of the canonical class, which places in the weight lattice of . The inner product on then turns combinatorial statements about into statements about lines on a cubic surface, and this dictionary drives all of our results.
Partially supported by the AMS-Simons Research Enhancement Grant for PUI Faculty.
The combinatorics of –matroids
Using the inner product on , we characterize –matroids by a tritangent exchange property (Theorem 2.4): a set of lines is a matroid if and only if, whenever two lines of meet, the third line of their tritangent plane lies on a second tritangent plane whose other two lines also belong to . By the Gelfand–Serganova theorem, a set of lines is a matroid exactly when every edge of its convex hull is parallel to a root of ; testing this condition on every orbit of subsets of the 27 lines shows that there are 1 335 373 –matroids, forming 230 orbits under (Theorem 2.5).
Matroidal strata of the -Grassmannian
As for , the support of every point of is an –matroid, and the points with a given support form a locally closed subscheme defined over , the matroidal stratum of (§ 3). In a chart , the coordinates of the 16 lines skew to are free, and the coordinates of the 10 lines meeting are quadrics in them (Theorem 3.1); so each stratum is cut out of a torus by the vanishing of some of these quadrics. Our main geometric result (Theorem 3.6) is that these strata are as simple as possible. Up to symmetry, exactly one –matroid is realizable over no field. Every other stratum is isomorphic over to an open subscheme of an algebraic torus, so it is smooth, irreducible and rational, and the matroid is realizable over every infinite field. Both the non-realizable matroid and the one stratum that requires a separate argument are explained by a single linear relation among the chart quadrics.
Tropicalization and toric degenerations
Finally, we compute the tropicalization of the open stratum of (Theorem 4.1). It is a simplicial fan with 12 885 rays and 8 712 360 maximal cones, in 10 and 421 symmetry classes, and every cone induces a subdivision of into matroid polytopes; the classification of –matroids (Theorem 2.5) supplies the starting cone that the traversal algorithm needs. Exactly 60 classes of maximal cones are prime, and each gives a flat degeneration of to a projectively normal Gorenstein toric Fano variety whose polytope has 27 lattice points and normalized volume 78 (Theorem 4.3). One class is the Hibi degeneration [GL96], and two give the chain polytope of the minuscule poset.
(E6,P6)–matroids
Let be the lattice with symmetric bilinear form , , , and put , so that . This is the Picard lattice of a smooth cubic surface with its canonical class [7 Ch. 9]. With this connection, we use classical vocabulary below (lines, tritangent planes, double sixes). The orthogonal complement is the root lattice, whose 72 roots are the vectors with .
Definition 2.1. Let be the orthogonal projection, . Set
so that is a positive definite -invariant inner product on .
Since is unimodular and is primitive, is the dual of with respect to . Thus is the weight lattice of ; it contains the root lattice with index , and is generated by the class of . The Weyl group , generated by the reflections in the roots, acts on . The vectors with and are the classes of the lines on the cubic surface,
and each projects to . We keep the names for these projections and write for the resulting set of vectors. Equivalently,
the set of weights of the -dimensional minuscule representation of (the other nontrivial coset of gives the weights of ). Following [15], the Gosset polytope is
a lattice polytope in centered at the origin, with as its full symmetry group. From we read off that two distinct elements satisfy , according as the corresponding lines are skew or meet, and that span an edge of exactly when , that is, when is a root.
Fix . The inner product partitions the remaining elements as
where is the set of neighbors of in , and the elements of are the vertices of the cross-polytope facet of opposite to . Two vertices of are antipodal in that cross-polytope if and only if , if and only if
More generally, a triple of distinct elements of is called a tritangent plane if in . In the cubic surface description, is the class of a hyperplane section of , so in (equivalently, the classes sum to in ) means that the three lines are coplanar: some plane cuts out .
So the five antipodal pairs of are the five tritangent planes through , and there are tritangent planes in all. We write for the antipode of in .
Example 2.2. For we get with antipodal pairs , and consists of and the ten with .
By the Gelfand–Serganova theorem [1], an –matroid is a nonempty subset such that every edge of the polytope is parallel to a root of . It is a strong –matroid if for all distinct there is a root whose reflecting hyperplane separates and and such that the reflections of and in both lie in .
Definition 2.3. A nonempty has the tritangent exchange property if, for every with , there are with , , and
Geometrically, if and are two meeting lines of and is the third line of their tritangent plane, the property asks for a second tritangent plane through with ; in other words, a tritangent plane through with two lines in can always be exchanged for another one. Since is injective on each level set of , the relation may be verified among the corresponding classes in .
Theorem 2.4. For a nonempty the following are equivalent.
(i) is an –matroid.
(ii) is a strong –matroid.
(iii) For every , the number of antipodal pairs of contained in is not equal to one.
(iv) has the tritangent exchange property.
Indication of proof. The equivalence of (i) and (ii) is [10 Thm. 4.22], and the equivalence of (ii) and (iii) is [4 Prop. 3.15]. For a direct proof of (i)(iii), if contains exactly one antipodal pair , an explicit linear functional on the cross-polytope shows that is an edge of , which is a face of because spans a facet; but , so is not parallel to a root. For (iii)(iv), if then lies in by the description of as the vectors of norm in a coset, and is an antipodal pair of ; conversely forces . □
In particular the adjective “strong” is redundant for and for all minuscule types by [10], but it fails for general pairs .
Theorem 2.5. There are exactly 1 335 373 –matroids, forming 230 orbits under . By the dimension of the orbits are distributed as follows.
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | |
| -orbits | 1 | 1 | 2 | 4 | 11 | 30 | 181 |
Table 2.5.
Indication of proof. Order and represent each orbit of subsets by its lexicographically minimal element. Every orbit of -subsets contains for some minimal representative of size , so the representatives can be grown one line at a time; those of size at least 14 are complements of those of size at most 13. This yields the 5 486 orbits of subsets of , and for each representative we test whether every edge of its convex hull is parallel to a root. □
Write for the graph on joining and when . Each edge of determines the line , the third line of their tritangent plane. One matroid plays a special role below:
Its complement is the union of three disjoint tritangent planes, namely , and , and every vertex of has degree 7.
Matroidal strata of the -Grassmannian
Equations and charts
Let and let be the skew-symmetric matrix with entries above the diagonal. The -invariant cubic on the 27-dimensional minuscule representation, which goes back to Cartan’s thesis [3] (see [16], and also [4 Prop. 4.15]), is a signed sum of 45 monomials , one for each tritangent plane, that is, each triple with in [8, 16]. With and it is
The 30 monomials of are the tritangent planes , and the 15 monomials of are the tritangent planes with . The -Grassmannian is the singular locus of the hypersurface defined by [13], so it is cut out by the partial derivatives of :
Differentiating the monomials shows that each quadric is a signed sum over the tritangent planes through :
(In the full version we match these coordinates with the model of inside the exceptional Jordan algebra over the octonions, under which becomes the determinant.)
Fix and let . For we have , so by (2) the quadric contains the monomial ; write . The other four tritangent pairs of satisfy , hence lie in , so is a signed sum of four products .
Theorem 3.1 (Big cell). For every the morphism
which keeps the coordinates , , is an isomorphism.
Indication of proof. On , with , the equations say exactly that . It remains to check that the other 17 quadrics vanish identically after this substitution: becomes a quartic , and each , , becomes a relation that is linear in the . This is a straightforward computation in one chart, and permutes the charts transitively by signed permutations of the coordinates. □
Supports and strata
For a field and a -point of , let .
Proposition 3.2. The support of every point of is an –matroid.
Indeed, if contained exactly one tritangent pair, then by (2) the quadric would evaluate to a single nonzero monomial at . This is analogous to the fact that the support of a point of is a matroid, and it motivates the following definition.
Definition 3.3. The matroidal stratum of an –matroid is the locally closed subscheme
defined over . We say is realizable over if has a -point.
Thus , with permuting the strata, and the open stratum is the locus where all 27 coordinates are nonzero. These are analogous to the thin Schubert cells in [9] of the ordinary Grassmannian.
The equations of a stratum
Fix . Every point of lies in the chart , and setting the coordinates of to zero in Theorem 3.1 gives
where we keep writing after setting the coordinates of to zero, and is the set of for which is not the zero polynomial, so that is a genuine equation. Each is a signed sum of monomials over antipodal pairs of , which are disjoint, so no variable appears twice; and by Theorem 2.4(iii) is never a monomial. Hence each equation can be solved for any one of its variables. To organize the elimination, note that the monomial of comes from two meeting lines of with . So we define the exchange graph as the subgraph of on whose edges are the pairs with . Its edges are exactly the monomials of the equations, and the label partitions them into one matching per equation. A leaf order is an ordering of the matchings such that each one contains a vertex of degree one in the union of itself and its successors. The key combinatorial input is the following.
Lemma 3.4. Let be distinct and let be a tritangent pair of . Then meets in no vertex if , and in exactly one vertex otherwise.
This is a one-line computation, (both pairs sum to in ), and it implies that two matchings of never have the same vertex set, so every chart with at most two equations has a leaf order. Eliminating one variable per equation along a leaf order identifies with an open subscheme of a torus over .
Example 3.5. Let , the –matroid that supplies the starting cone in the proof of Theorem 4.1, and take . Then consists of , , , , , , , and of , , , , , , , , . Figure 1(a) shows , with on top, in the middle row and in the bottom row. Each of the 13 edges of on is labeled by the line ; for instance in . Figure 1(b) shows these edges with their labels. Seven labels lie in (dashed edges), and these edges give the coordinates
which are automatically units. The remaining six (solid) edges form , with matchings (blue), (orange) and (green), so and the equations are
With three equations Lemma 3.4 does not apply directly, but is the disjoint union of the paths , and , so each matching contains a leaf of the whole graph and every order is a leaf order. Solving each equation for a leaf that occurs in no other equation gives
Every coordinate is now a Laurent monomial in , so .

Figure 1. The graphs of Example 3.5.
Theorem 3.6. Let be an –matroid. If is -equivalent to
then is empty over every field. Otherwise is isomorphic over to an open subscheme of an algebraic torus with nonempty fiber over every prime. In
particular is realizable over every infinite field, and is smooth, irreducible and rational, of dimension for any chart admitting a leaf order.
Indication of proof. A chart with at most two equations has a leaf order by Lemma 3.4. A check over the 230 orbit representatives shows that every other than has a chart admitting a leaf order. In each chart of the three equations are the minors of a matrix, which cut out a translate of an 8-dimensional subtorus. For , a linear relation among the in the chart at forces , although both factors must be units. Finally, strong Gröbner bases over with leading coefficients show that the fibers over are nonempty. □
| 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | |
| orbits | 1 | 1 | 2 | 3 | 8 | 12 | 23 | 34 | 43 | 38 | 30 | 17 | 9 | 4 | 2 | 1 | 1 |
Table 1. Dimensions of the nonempty matroidal strata, up to .
Remark 3.7. The –matroids other than might not be realizable over certain finite fields. For instance itself is not realizable over . Indeed, a point with all coordinates nonzero would have all coordinates equal to , and then each quadric (2) would equal a sum of five terms , which is odd.
Remark 3.8. Although each stratum is smooth and irreducible, the strata do not form a stratification in the sense that the closure of a stratum is not necessarily the union of smaller strata. For example, for and we have that meets without containing it.
Tropicalization and toric degenerations
The tropical E-Grassmannian
Let be the open stratum, a closed subvariety of the dense torus of which is a smooth and irreducible very affine variety. Let . The initial form picks out the terms of of minimal -weight. Since is homogeneous, depends only on the class of in , and is the set of classes for which contains no monomial [17]. It is a pure 16-dimensional fan. Its lineality space is the 6-dimensional subspace given by the image of under , that is, the space of affine-linear functions on modulo constants.
Theorem 4.1. The restriction of the Gröbner fan of to is simplicial and each of its maximal cones has multiplicity one. Modulo its -vector is
Up to the action of it is .
For each in , the regular subdivision of induced by is a subdivision into –matroid polytopes.
Indication of proof. We computed the fan with gfan_tropicaltraverse [14], up to the symmetry of . This traversal needs a maximal cone to start from, and the heuristic for finding one, gfan_tropicalstartingcone, did not work for . Instead, we started from the height , the negative of the indicator vector of the matroid of Example 3.5. For the last statement, up to signs the quadrics are the strong exchange equations defining the Coxeter Dressian , so , and every point of the Dressian induces a matroidal subdivision by [4 Prop. 4.15] and [10 Thm. 4.22]. We also verified it directly on a representative of each of the 14 155 orbits of cones.
The ten orbits of rays have sizes 27, 36, 36, 216, 240, 270, 540, 720, 4 320 and 6 480; the two orbits of size 36 are spanned, modulo , by the indicator vectors of the 36 double sixes and of their complements. The subdivisions induced by maximal cones have between 51 and 69 maximal cells.
Proposition 4.2. The inclusion is strict.
Indication of proof. Let be the vector with for and otherwise. It induces a matroidal subdivision of into 27 cells, one of which is , so . A direct calculation of the initial ideal verified that .
Toric degenerations
A maximal cone of is prime if is a prime ideal for in its relative interior. In that case the Gröbner degeneration to is a flat degeneration to a variety which, having a dense torus orbit and a binomial ideal, is a projective toric variety ; here is the configuration of 27 lattice points obtained by dividing by the lattice of exponent differences of the binomials. Let .
Theorem 4.3. Exactly 60 of the 421 -orbits of maximal cones of are prime. For each prime cone , the special fiber is a projectively normal toric variety. Consequently it is Gorenstein with , and in particular Fano.
Indication of proof. For a representative of each maximal cone orbit we compute by a Gröbner basis computation; it is binomial for 388 orbits, and exactly 60 of these binomial ideals are prime. For these, Oscar [6] shows that is normal, so is arithmetically Cohen–Macaulay. The Hilbert series of is . Its numerator is palindromic, so is Gorenstein by [18 Thm. 4.4], and a toric computation gives .
The 60 polytopes share much of their structure. Each is a 16-dimensional lattice polytope whose only lattice points are its 27 vertices, with 297 edges; the 54 missing pairs are the diagonals of the 27 parallelograms recorded by the quadratic binomials of . Its -vector is . The orbits differ in finer invariants. For 53 of them is generated by the 27 quadratic binomials, while the other 7 need one, three or eight additional cubic binomials. The number of facets of ranges from 22 to 38 (Table 2), and the 60 polytopes realize 53 combinatorial types.
| facets of | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 34 | 38 |
| prime orbits | 1 | 1 | 8 | 3 | 7 | 7 | 7 | 3 | 12 | 2 | 3 | 5 | 1 |
Table 2. The 60 toric degenerations of by the number of facets of their polytopes.
Three of these degenerations have a known representation-theoretic origin. Ordering the 27 weights by when is a nonnegative combination of simple roots makes a distributive lattice, whose poset of join-irreducibles has 16 elements and 20 cover relations. Standard monomial theory degenerates flatly to the Hibi toric variety of this lattice [GL96], whose ideal is generated by the binomials for incomparable , , and whose polytope is the order polytope of [12]. Exactly one of our 60 orbits is this degeneration: up to and signs, its initial ideal is precisely the ideal of the 27 Hibi relations, and it is the unique orbit whose polytope has the minimal number 22 = 20 + 2 of facets. Exactly two further orbits give polytopes unimodularly equivalent to Stanley’s chain polytope of [Sta86], with 28 facets. We do not know a representation-theoretic description of the remaining 57 orbits; it would be interesting to realize them as marked poset polytopes or as Newton–Okounkov bodies of explicit valuations.
Code
The code and data for the computations in this abstract are available at: https://dcorey2814.github.io/E6P6CoxeterMatroids/
AI disclosure
The author used Gemini (Gemini 3.1), ChatGPT (Sol) and Claude (Opus 5.5) for brainstorming, for finding new proofs of results that the author had obtained independently, for drafting this paper, and for adapting the author’s code into the companion GitHub repository and website. Notably, Gemini discovered the tritangent exchange property, Claude formulated the leaf order proof of Theorem 3.6 and found the representation-theoretic origin of three of the toric degenerations. The author checked all claims made by these tools.
Department of Mathematics, Embry-Riddle Aeronautical University, Daytona Beach, FL, USA
Email address: [email protected]
References
- [1][BGW03] A. V. Borovik, I. M. Gelfand, and N. White, Coxeter matroids, Progress in Mathematics, vol. 216, Birkhäuser Boston, Boston, MA, 2003. ↑2
- [2][Bou02] N. Bourbaki, Lie groups and Lie algebras. Chapters 4–6, Elements of Mathematics, Springer-Verlag, Berlin, 2002. ↑1
- [3][Car94] É. Cartan, Sur la structure des groupes de transformations finis et continus, Ph.D. Thesis, Paris, 1894. ↑3.1
- [4][CDFS25] K. Calvert, A. Dermenjian, A. Fink, and B. Smith, Quadratic exchange equations for Coxeter matroids, 2025. arXiv:2511.13498. ↑2, 3.1, 4.1arxiv.org/abs/2511.13498
- [5][Cox40] H. S. M. Coxeter, The polytope 221, whose twenty-seven vertices correspond to the lines on the general cubic surface, Amer. J. Math. 62 (1940), 457–486. ↑1
- [6][DEF+25] W. Decker, C. Eder, C. Fieker, M. Horn, and M. Joswig (eds.), The computer algebra system OSCAR: Algorithms and examples, Algorithms and Computation in Mathematics, vol. 32, Springer, 2025. ↑4.2DOI
- [7][Dol26] I. V. Dolgachev, Classical algebraic geometry: A modern view, 2nd ed., Vol. 2, Cambridge University Press, Cambridge, 2026. ↑2
- [8][Fau01] J. R. Faulkner, Generalized quadrangles and cubic forms, Comm. Algebra 29 (2001), no. 10, 4641–4653. ↑3.1DOI
- [9][GGMS87] I. M. Gelfand, R. M. Goresky, R. D. MacPherson, and V. V. Serganova, Combinatorial geometries, convex polyhedra, and Schubert cells, Adv. Math. 63 (1987), no. 3, 301–316. ↑3.2DOI
- [10]A. Gross, K. Kühn, and D. Luber, Minuscule Coxeter Dressians, 2025. arXiv:2512.09703. ↑2, 4.1arxiv.org/abs/2512.09703
- [12]T. Hibi, Distributive lattices, affine semigroup rings and algebras with straightening laws, Commutative algebra and combinatorics (Kyoto, 1985), 1987, pp. 93–109. ↑4.2
- [13]A. Iliev and L. Manivel, On cubic hypersurfaces of dimensions 7 and 8, Proc. Lond. Math. Soc. (3) 108 (2014), no. 2, 517–540. ↑3.1
- [14]A. N. Jensen, Gfan, a software system for Gröbner fans and tropical varieties. Available at https://users-math.au.dk/jensen/software/gfan/gfan.html. ↑4.1
- [15]J.-H. Lee, Gosset polytopes in Picard groups of del Pezzo surfaces, Canad. J. Math. 64 (2012), no. 1, 123–150. ↑1, 2arxiv.org/abs/0905.3025
- [16]L. Manivel, Configurations of lines and models of Lie algebras, J. Algebra 304 (2006), no. 1, 457–486. ↑3.1
- [17]D. Maclagan and B. Sturmfels, Introduction to tropical geometry, Graduate Studies in Mathematics, vol. 161, American Mathematical Society, Providence, RI, 2015. ↑4.1
- [18]R. P. Stanley, Hilbert functions of graded algebras, Adv. Math. 28 (1978), no. 1, 57–83. ↑4.2DOI