Introduction
The exceptional group G 2 G_2 G 2 is an algebraic group of dimension 14, defined classically as the automorphism group of the octonions [37 §2.2–2.3] . We work with its split form over an arbitrary field K \mathbb{K} K . Following Aschbacher [3 ] , our starting point is the alternating trilinear form
Φ = e 123 + e 456 + e 147 + e 257 + e 367 ∈ ∧ 3 K 7 . (1) \Phi= e_{123} + e_{456} + e_{147} + e_{257} + e_{367} \in\wedge^3 \mathbb{K}^7. \tag*{(1)} Φ = e 123 + e 456 + e 147 + e 257 + e 367 ∈ ∧ 3 K 7 . (1)
Here, we identify K 7 \mathbb{K}^7 K 7 with its dual via the standard basis e 1 , … , e 7 e_1,\ldots,e_7 e 1 , … , e 7 , and we write e i j k = e i ∧ e j ∧ e k e_{ijk}=e_i\wedge e_j\wedge e_k e ij k = e i ∧ e j ∧ e k for the induced basis vectors of ∧ 3 K 7 \wedge^3\mathbb{K}^7 ∧ 3 K 7 . We define G 2 ⊂ S L 7 G_2\subset\mathrm{SL}_7 G 2 ⊂ SL 7 as the stabilizer of Φ \Phi Φ , so that its group of K \mathbb{K} K -points G 2 ( K ) G_2(\mathbb{K}) G 2 ( K ) consists of the 7 × 7 7\times7 7 × 7 -matrices in S L 7 ( K ) \mathrm{SL}_7(\mathbb{K}) SL 7 ( K ) that fix Φ \Phi Φ . Over every field K \mathbb{K} K , this is the split simple algebraic group of type G 2 G_2 G 2 , as shown in [3 §3] .
We find that G 2 G_2 G 2 is a subgroup of the special orthogonal group S O ( 7 , q ) \mathrm{SO}(7,q) SO ( 7 , q ) for
q ( x ) = x 1 x 4 + x 2 x 5 + x 3 x 6 − x 7 2 . (2) q(x) = x_1x_4 + x_2x_5 + x_3x_6 - x_7^2. \tag*{(2)} q ( x ) = x 1 x 4 + x 2 x 5 + x 3 x 6 − x 7 2 . (2)
This split quadratic form has signature ( 3 , 4 ) (3,4) ( 3 , 4 ) over R \mathbb{R} R . To derive q ( x ) q(x) q ( x ) , we write Φ \Phi Φ as a skew-symmetric 7 × 7 7\times7 7 × 7 matrix whose entries are linear forms in 7 variables x 1 , … , x 7 x_1,\ldots,x_7 x 1 , … , x 7 , namely
Φ ( x , ⋅ , ⋅ ) = x 1 ⋅ ( e 23 + e 47 ) − x 2 ⋅ ( e 13 − e 57 ) + x 3 ⋅ ( e 12 + e 67 ) − x 4 ⋅ ( e 17 − e 56 ) − x 5 ⋅ ( e 27 + e 46 ) − x 6 ⋅ ( e 37 − e 45 ) + x 7 ⋅ ( e 14 + e 25 + e 36 ) ∈ ∧ 2 K 7 . (3) \begin{aligned}
\Phi(x,\cdot,\cdot) ={}& x_1\cdot(e_{23}+e_{47})-x_2\cdot(e_{13}-e_{57})+x_3\cdot(e_{12}+e_{67})-x_4\cdot(e_{17}-e_{56})\\
&-x_5\cdot(e_{27}+e_{46})-x_6\cdot(e_{37}-e_{45})+x_7\cdot(e_{14}+e_{25}+e_{36})\in\wedge^2\mathbb{K}^7. \tag*{(3)}
\end{aligned} Φ ( x , ⋅ , ⋅ ) = x 1 ⋅ ( e 23 + e 47 ) − x 2 ⋅ ( e 13 − e 57 ) + x 3 ⋅ ( e 12 + e 67 ) − x 4 ⋅ ( e 17 − e 56 ) − x 5 ⋅ ( e 27 + e 46 ) − x 6 ⋅ ( e 37 − e 45 ) + x 7 ⋅ ( e 14 + e 25 + e 36 ) ∈ ∧ 2 K 7 . (3)
Up to sign, the 6 × 6 6\times6 6 × 6 Pfaffians of (3) are x 1 q ( x ) , x 2 q ( x ) , … , x 7 q ( x ) x_1q(x),x_2q(x),\ldots,x_7q(x) x 1 q ( x ) , x 2 q ( x ) , … , x 7 q ( x ) , so Φ \Phi Φ determines q ( x ) q(x) q ( x ) .
The group G 2 G_2 G 2 has an interesting history, described by Agricola [2 ] and Manivel [30 ] . It was discovered by Wilhelm Killing, who communicated it to Friedrich Engel in a letter dated May 23, 1887. Engel worked with Felix Klein and Sophus Lie, both professors at Leipzig. Engel presented G 2 G_2 G 2 in a lecture at the Saxonian Academy of Sciences on June 11, 1900.
According to Klein, geometries emerge from groups, and the lifeblood of a group is the spaces it acts on. Our group G 2 \mathrm{G}_2 G 2 acts naturally on three spaces: the quadric Q = V ( q ) ⊂ P 6 \mathcal{Q}=V(q)\subset\mathbb{P}^{6} Q = V ( q ) ⊂ P 6 , a Grassmannian G \mathcal{G} G of certain lines on Q \mathcal{Q} Q , and the flag variety F ⊂ Q × G \mathcal{F}\subset\mathcal{Q}\times\mathcal{G} F ⊂ Q × G which links them. Both Q \mathcal{Q} Q and G \mathcal{G} G have dimension 5 and F \mathcal{F} F has dimension 6. The G 2 \mathrm{G}_2 G 2 -Grassmannian G \mathcal{G} G is contained in the 7-dimensional orthogonal Grassmannian OGr ( 2 , 7 ) \operatorname{OGr}(2,7) OGr ( 2 , 7 ) of all lines on Q \mathcal{Q} Q , and hence in the 10-dimensional Grassmannian Gr ( 2 , 7 ) ⊂ P 20 \operatorname{Gr}(2,7)\subset\mathbb{P}^{20} Gr ( 2 , 7 ) ⊂ P 20 . In this paper we tropicalize these objects.
We now describe the polynomials that define our varieties. The ideal of the Grassmannian Gr ( 2 , 7 ) \operatorname{Gr}(2,7) Gr ( 2 , 7 ) in Plücker coordinates is generated by the 4 × 4 4\times4 4 × 4 subpfaffians p i j p k l − p i k p j l + p i l p j k p_{ij}p_{kl}-p_{ik}p_{jl}+p_{il}p_{jk} p ij p k l − p ik p j l + p i l p j k of a skew-symmetric 7 × 7 7\times7 7 × 7 matrix ( p i j ) (p_{ij}) ( p ij ) . The tropical Grassmannian trop ( Gr ( 2 , 7 ) ) \operatorname{trop}(\operatorname{Gr}(2,7)) trop ( Gr ( 2 , 7 )) is the space of phylogenetic trees with 7 leaves [29 Theorem 4.3.5] . We here also need the linear space
L = { p ∈ P 20 : p 23 + p 47 = p 13 − p 57 = p 12 + p 67 = p 17 − p 56 = p 27 + p 46 = p 37 − p 45 = p 14 + p 25 + p 36 = 0 } . (4) \mathcal{L}=\left\{p\in\mathbb{P}^{20}:\quad p_{23}+p_{47}=p_{13}-p_{57}=p_{12}+p_{67}=p_{17}-p_{56}
=p_{27}+p_{46}=p_{37}-p_{45}=p_{14}+p_{25}+p_{36}=0\right\}. \tag*{(4)} L = { p ∈ P 20 : p 23 + p 47 = p 13 − p 57 = p 12 + p 67 = p 17 − p 56 = p 27 + p 46 = p 37 − p 45 = p 14 + p 25 + p 36 = 0 } . (4)
These 7 linear equations are the coefficients of (3) when viewed as a linear form in x 1 , … , x 7 x_1,\ldots,x_7 x 1 , … , x 7 . The intersection in P 20 \mathbb{P}^{20} P 20 of the Grassmannian with this linear space is the G 2 \mathrm{G}_2 G 2 -Grassmannian:
G = Gr ( 2 , 7 ) ∩ L . (5) \mathcal{G}=\operatorname{Gr}(2,7)\cap\mathcal{L}. \tag*{(5)} G = Gr ( 2 , 7 ) ∩ L . (5)
We have dim ( G ) = 5 \dim(\mathcal{G})=5 dim ( G ) = 5 and deg ( G ) = 18 \deg(\mathcal{G})=18 deg ( G ) = 18 . Each point in G \mathcal{G} G is a line in P 6 \mathbb{P}^{6} P 6 , and the union of these lines is the quadric Q \mathcal{Q} Q . The flag variety F \mathcal{F} F for G 2 \mathrm{G}_2 G 2 is their incidence correspondence. This is a 6-dimensional variety, defined by 34 bilinear equations inside Q × G ⊂ P 6 × P 20 \mathcal{Q}\times\mathcal{G}\subset\mathbb{P}^{6}\times\mathbb{P}^{20} Q × G ⊂ P 6 × P 20 .
In Tits’ theory of buildings [38 ] , the triple ( Q , G , F ) (\mathcal{Q},\mathcal{G},\mathcal{F}) ( Q , G , F ) serves as a generalized hexagon [39 ] . In our coordinates, a distinguished hexagon is given by a cycle of six points and lines:
e 1 − e 15 − e 5 − e 35 − e 3 − e 34 − e 4 − e 24 − e 2 − e 26 − e 6 − e 16 − e 1 . (6) e_1-e_{15}-e_5-e_{35}-e_3-e_{34}-e_4-e_{24}-e_2-e_{26}-e_6-e_{16}-e_1. \tag*{(6)} e 1 − e 15 − e 5 − e 35 − e 3 − e 34 − e 4 − e 24 − e 2 − e 26 − e 6 − e 16 − e 1 . (6)
The group G 2 \mathrm{G}_2 G 2 acts on F \mathcal{F} F , and this action is transitive on the non-degenerate hexagons. The labels in Figure 1 show how the Weyl group D 12 \mathrm{D}_{12} D 12 acts on the 7 linear equations in (4) . Note that Φ \Phi Φ in (1) is the sum over two triangles and three diagonals that cross the hexagon.
Figure 1. Colorful redrawing of the hexagon (6) . Its symmetry group, the dihedral group D 12 \mathrm{D}_{12} D 12 of order 12, is the Weyl group of G 2 \mathrm{G}_2 G 2 . This group acts on our classical and tropical varieties.
A key point of tropical geometry is that tropicalization and intersection do not commute:
trop ( G ) = trop ( Gr ( 2 , 7 ) ∩ L ) ⊊ trop ( Gr ( 2 , 7 ) ) ∩ trop ( L ) . (7) \operatorname{trop}(\mathcal{G})=\operatorname{trop}(\operatorname{Gr}(2,7)\cap\mathcal{L})\subsetneq\operatorname{trop}(\operatorname{Gr}(2,7))\cap\operatorname{trop}(\mathcal{L}). \tag*{(7)} trop ( G ) = trop ( Gr ( 2 , 7 ) ∩ L ) ⊊ trop ( Gr ( 2 , 7 )) ∩ trop ( L ) . (7)
Both objects in (7) are polyhedral fans in R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 . Only that on the left is a tropical variety. Namely, it is the tropical G 2 \mathrm{G}_2 G 2 -Grassmannian. The object on the right is a tropical prevariety, called the G2 _2 2 -Dressian . This follows the nomenclature in [29 ] . We also study the tropical G2 _2 2 flag variety trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) , which is a tropical variety in R 7 / 1 × R 21 / 1 \mathbb{R}^{7}/\mathbf{1} \times\mathbb{R}^{21}/\mathbf{1} R 7 / 1 × R 21 / 1 . The G2 _2 2 flag Dressian is the tropical prevariety defined by the 78 generators of the ideal of F \mathcal{F} F listed in § 2 .
We now discuss the organization and main results in this paper. In § 2 we introduce G2 _2 2 -trees. These are phylogenetic trees on 7 leaves that lie on the G2 _2 2 -Dressian. The G2 _2 2 flag Dressian is the corresponding universal family. Its elements are G2 _2 2 -trees with a marked point. Complete descriptions of our two tropical prevarieties are given in Theorems 2.4 and 2.5 .
In § 3 we study the two tropical varieties associated with G2 _2 2 . Theorem 3.1 characterizes the Grassmannian trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) , and Theorem 3.3 characterizes the flag variety trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) . A tropical basis for G \mathcal{G} G is presented in Corollary 3.4 and explained in Figure 5 . These objects were found by algebraic computations over the field Q \mathbb{Q} Q of rational numbers. In § 4 we turn from characteristic 0 to characteristic p > 0 p > 0 p > 0 . The tropical Grassmannian and flag variety remain as in § 3 for p ≥ 3 p \ge3 p ≥ 3 . However, they have different tropicalizations in characteristic p = 2 p = 2 p = 2 . This is proved in Theorems 4.1 and 4.2 . Figure 6 offers a visualization.
Figure 5. We associate a G 2 G_2 G 2 -tree to each monomial in the cubics (11) . These are rays of trop 2 ( G ) \operatorname{trop}_{2}(\mathcal{G}) trop 2 ( G ) and hence shown in Figure 7 . From left to right, their orbits are E \mathcal{E} E , α 2 \alpha_{2} α 2 , α 4 \alpha_{4} α 4 , α 3 \alpha_{3} α 3 , α 3 \alpha_{3} α 3 .
Figure 6. The tetrahedron ε \varepsilon ε shows how the tropical G 2 \mathrm{G}_2 G 2 -Grassmannian depends on the field K K K .
§ 5 introduces the G2 _2 2 -forest. This is our gallery of G2 _2 2 -trees, one for each cell in trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . The rays are shown in Figure 7 . The maximal cells appear in Figure 8 and Figure 9 .
Figure 7. The 12 G 2 \mathrm{G}_{2} G 2 -trees for the rays of the tropical Grassmannian and Dressian.
Figure 8. The 20 types of maximal G 2 G_2 G 2 -trees that give toric degenerations in § 6 and § 7 .
Figure 9. The 15 types of maximal G 2 G_2 G 2 -trees that do not give toric degenerations.
In § 6 we examine toric degenerations of the Grassmannian G \mathcal{G} G and its flag variety F \mathcal{F} F . Such degenerations and their Newton–Okounkov polytopes are of considerable interest in the representation theory of G2 _2 2 . Specific instances appeared earlier in the articles [9 , 13 , 34 ] . Theorem 6.3 gives the complete classification. The census for G \mathcal{G} G is presented in Table 2 .
Table 2. Newton–Okounkov polytopes for all 20 toric degenerations of the G 2 \mathrm{G}_2 G 2 -Grassmannian over Q \mathbb{Q} Q and over F 2 \mathbb{F}_2 F 2 . The orbit sizes in Figure 8 sum to 147. The last two columns give the number of lattice points N N N in the polytope and the field characteristic. See [5 ] for multiplexes.
§ 7 introduces the G2 _2 2 -orchard. This is our gallery of decorated G2 _2 2 -trees, one for each cell in trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) . We focus on the fiber complexes. These are the preimages in trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) of individual maximal cells in trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . We also study the Newton–Okounkov polytopes for F \mathcal{F} F .
The present paper serves as a blueprint for importing reductive groups into tropical geometry. In § 8 we explore what it means to tropicalize a group and its homogeneous spaces. Theorem 8.1 shows how our approach to G2 _2 2 fits into the general theory. § 9 is motivated by the theory of buildings [38 , 39 ] . We study the 12-dimensional variety H \mathcal{H} H of all hexagons on F \mathcal{F} F , i.e., the closure of the orbit of (6) under the G2 _2 2 -action. Theorem 9.1 builds tropicalized hexagons. These are the loop trails for enjoying a stroll in the G2 _2 2 -orchard.
Our inspiration for this article was the theory of algebraic curves of genus 10. Their moduli space M 10 \mathcal{M}_{10} M 10 has dimension 27. It has a distinguished divisor K \mathcal{K} K , given by curves that lie on a K3 surface. According to Mukai [31 , 32 ] , the canonical model of a curve in K \mathcal{K} K is an intersection Λ ∩ G \Lambda\cap\mathcal{G} Λ ∩ G , where Λ \Lambda Λ is a linear space of codimension 4 in P 20 \mathbb{P}^{20} P 20 , so that Λ ∩ P 9 ≅ P 9 \Lambda\cap\mathbb{P}^{9} \cong\mathbb{P}^{9} Λ ∩ P 9 ≅ P 9 . In § 10 , we study intersections of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) with tropical linear spaces trop ( Λ ) \operatorname{trop}(\Lambda) trop ( Λ ) . Our tropical canonical curves are derived in Proposition 10.1 and shown in Figure 16 and Figure 17 .
Figure 16. A planar graph obtained with Method 1 (left). Its trace on trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) (right).
Figure 17. A nonplanar graph obtained with Method 2 (left). Its trace on trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) (right).
We now comment on our usage of AI tools. From the very start and throughout the entire project, tools based on large language models (mostly Claude Opus 5.5) were used for literature search, drawing pictures, writing code, drafting results and their proofs. Most importantly, Claude was invaluable in maintaining the overall consistency of this long paper. This is particularly important because several of our main results are based on nontrivial computations, mostly with the computer algebra system Oscar [12 , 33 ] . To check that our computational arguments are really complete, the reader is invited to use similar tools. Our Zenodo repository at 10.5281/zenodo.24104003 is organized in a way to facilitate linking the results in this article with Oscar output and computations; see Appendices § 12 and § 13 .
Trees and Dressians
The tropical Grassmannian trop ( Gr ( 2 , 7 ) ) \operatorname{trop}(\operatorname{Gr}(2,7)) trop ( Gr ( 2 , 7 )) is the set of points d = ( d i j ) 1 ≤ i < j ≤ 7 ∈ R 21 d=(d_{ij})_{1\leq i<j\leq7}\in\mathbb{R}^{21} d = ( d ij ) 1 ≤ i < j ≤ 7 ∈ R 21 such that
max ( d i j + d k l , d i k + d j l , d i l + d j k ) (8) \max(d_{ij}+d_{kl},d_{ik}+d_{jl},d_{il}+d_{jk}) \tag*{(8)} max ( d ij + d k l , d ik + d j l , d i l + d j k ) (8)
is attained at least twice [36 ] ; see also Theorem 10.47 of [24 ] and Lemma 4.3.6 of [29 ] . This is the four-point condition for phylogenetic trees. Hence d d d is a tree metric, provided its coordinates are positive and satisfy the triangle inequality. In algebraic geometry, we usually tropicalize with respect to the min-convention [29 ] . This is also Oscar’s default [33 ] . However, for our trees we pass to the max-convention by negating the valuation, so d i j = − val ( p i j ) d_{ij}=-\operatorname{val}(p_{ij}) d ij = − val ( p ij ) represents the tree distance between leaves i i i and j j j . This is because (8) is a max-tropical constraint.
A G 2 \mathrm{G}_2 G 2 -tree is a phylogenetic tree d d d with set of leaves [ 7 ] = { 1 , 2 , … , 7 } [7]=\{1,2,\ldots,7\} [ 7 ] = { 1 , 2 , … , 7 } that satisfies
d 17 = d 56 , d 27 = d 46 , d 37 = d 45 , d 47 = d 23 , d 57 = d 13 , d 67 = d 12 , (9) \begin{aligned}
d_{17}&=d_{56},\qquad d_{27}=d_{46},\qquad d_{37}=d_{45},\\
d_{47}&=d_{23},\qquad d_{57}=d_{13},\qquad d_{67}=d_{12}, \tag*{(9)}
\end{aligned} d 17 d 47 = d 56 , d 27 = d 46 , d 37 = d 45 , = d 23 , d 57 = d 13 , d 67 = d 12 , (9)
and max { d 14 , d 25 , d 36 } \max\{d_{14},d_{25},d_{36}\} max { d 14 , d 25 , d 36 } is attained at least twice.
The seven linear equations in (4) form a tropical basis, and (9) records their tropicalizations. These are the constraints that define trop ( L ) \operatorname{trop}(L) trop ( L ) , and hence the intersection on the right of (7) . In conclusion, the G 2 \mathrm{G}_2 G 2 -Dressian is the space of phylogenetic trees d d d that happen to be G 2 \mathrm{G}_2 G 2 -trees.
To make the notion of G 2 \mathrm{G}_2 G 2 -trees more tangible, we use the heart and the six colored shapes in Figure 1 to represent the seven leaves of each tree. The seven constraints in (9) are then:
the distance from any shape to the heart is the same as the distance between the two other shapes of the opposite color;
the maximum of the three distances between equal shapes is attained at least twice.
Example 2.1. Five G 2 \mathrm{G}_2 G 2 -trees are shown in Figure 2 . The rightmost tree gives d ∈ R 21 d\in\mathbb{R}^{21} d ∈ R 21 with
d 17 = d 56 = 7 , d 27 = d 46 = 7 , d 37 = d 45 = 6 , d 47 = d 23 = 7 , d 57 = d 13 = 7 , d 67 = d 12 = 6 , d_{17}=d_{56}=7,\quad d_{27}=d_{46}=7,\quad d_{37}=d_{45}=6,\quad d_{47}=d_{23}=7,\quad d_{57}=d_{13}=7,\quad d_{67}=d_{12}=6, d 17 = d 56 = 7 , d 27 = d 46 = 7 , d 37 = d 45 = 6 , d 47 = d 23 = 7 , d 57 = d 13 = 7 , d 67 = d 12 = 6 , d 36 = 6 < d 14 = d 25 = 8 , d 15 = 8 , d 16 = 1 , d 24 = 8 , d 26 = 5 , d 34 = 1 , d 35 = 5. d_{36}=6<d_{14}=d_{25}=8,\quad d_{15}=8,\quad d_{16}=1,\quad d_{24}=8,\quad d_{26}=5,\quad d_{34}=1,\quad d_{35}=5. d 36 = 6 < d 14 = d 25 = 8 , d 15 = 8 , d 16 = 1 , d 24 = 8 , d 26 = 5 , d 34 = 1 , d 35 = 5. The last six distances are the edges of our hexagon in Figure 1 . They do not appear in (9) .
Figure 2. From left to right, a G 2 \mathrm{G}_2 G 2 -tree with no inner edge, one inner edge, two inner edges, a snowflake G 2 \mathrm{G}_2 G 2 -tree, and a caterpillar G 2 \mathrm{G}_2 G 2 -tree. Check the conditions (i) and (ii) for each tree.
The lineality space of the tropical Grassmannian trop ( Gr ( 2 , 7 ) ) \operatorname{trop}(\operatorname{Gr}(2,7)) trop ( Gr ( 2 , 7 )) consists of all star trees d d d . It has dimension 6 in R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 . Namely, if s i s_i s i is the length of the i i i th pendant edge in the star tree then d i j = s i + s j d_{ij}=s_i+s_j d ij = s i + s j for all i , j i,j i , j . We now describe the lineality space of the G2 _2 2 -Dressian.
Lemma 2.2. The lineality space of the G2 _2 2 -Dressian has dimension 2 in R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 . It is the intersection of the lineality space of trop ( Gr ( 2 , 7 ) ) \operatorname{trop}(\operatorname{Gr}(2,7)) trop ( Gr ( 2 , 7 )) with the six hyperplanes in (9) . A basis is given by any two of the star trees in Figure 3 . Their edge lengths are the rows of the matrix
( 1 − 1 0 − 1 1 0 0 0 1 − 1 0 − 1 1 0 − 1 0 1 1 0 − 1 0 ) (10) \begin{pmatrix}
1 & -1 & 0 & -1 & 1 & 0 & 0 \\
0 & 1 & -1 & 0 & -1 & 1 & 0 \\
-1 & 0 & 1 & 1 & 0 & -1 & 0 \tag*{(10)}
\end{pmatrix} 1 0 − 1 − 1 1 0 0 − 1 1 − 1 0 1 1 − 1 0 0 1 − 1 0 0 0 (10)
Figure 3. These three stars are G2 _2 2 -trees. They span the lineality space of the G2 _2 2 -Dressian.
Proof. The G2 _2 2 -Dressian is the intersection in R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 of the two fans on the right in (7) . Its lineality space is therefore the intersection of the lineality spaces of the two intersectands. □ \square □
Our lineality space mirrors the root system of G2 _2 2 , which is shown in Figure 4 . Leaves of G2 _2 2 -trees are labeled by the six short roots, like α = ( 1 , 0 , − 1 ) \alpha=(1,0,-1) α = ( 1 , 0 , − 1 ) , or by the origin (heart). The six long roots include β = ( − 1 , − 1 , 2 ) \beta=(-1,-1,2) β = ( − 1 , − 1 , 2 ) . The roots are the degrees of our variables x i x_i x i and p i j p_{ij} p ij .
Figure 4. Root system of G2 _2 2 . Next to each root are the variables of that weight. Variables x 7 x_7 x 7 , p 14 p_{14} p 14 , p 25 p_{25} p 25 , p 36 p_{36} p 36 have weight 0. The maximal torus ( K ∗ ) 2 (\mathbb{K}^{*})^2 ( K ∗ ) 2 acts on Q , G , F \mathcal{Q},\mathcal{G},\mathcal{F} Q , G , F with these weights.
A G 2 \mathrm{G}_2 G 2 -tree is normalized if its edge lengths are relatively prime in N \mathbb{N} N , its shortest pendant edge has length 0 0 0 , and its vector s ∈ R 7 s \in\mathbb{R}^{7} s ∈ R 7 of pendant edge lengths is in the kernel of (10) . Two G 2 \mathrm{G}_2 G 2 -trees are equivalent if some positive multiples are congruent modulo lineality space.
Proposition 2.3. Every rational G 2 \mathrm{G}_2 G 2 -tree is equivalent to a unique normalized G 2 \mathrm{G}_2 G 2 -tree.
Proof. By Lemma 2.2, lineality translates the vector s s s of pendant edge lengths by the span of 1 = ( 1 , 1 , … , 1 ) \mathbf{1}=(1,1,\ldots,1) 1 = ( 1 , 1 , … , 1 ) and the rows of (10) , and changes nothing else. Exactly one translate of s s s is orthogonal to the plane spanned by the three rows. Their entries sum to zero, so adding a multiple of 1 \mathbf{1} 1 preserves orthogonality. Exactly one multiple makes the shortest pendant edge 0 0 0 . Both conditions are invariant under positive scaling, and the resulting tree d d d is rational and nonzero. Hence exactly one positive multiple has integer edge lengths with gcd 1 1 1 . □
The Weyl group of G 2 \mathrm{G}_2 G 2 is the dihedral group of order 12 12 12 , denoted D 12 \mathrm{D}_{12} D 12 . For us, D 12 \mathrm{D}_{12} D 12 is the permutation group generated by the rotation ( 1 5 3 4 2 6 ) (1\,5\,3\,4\,2\,6) ( 1 5 3 4 2 6 ) and the reflection ( 2 3 ) ( 5 6 ) (2\,3)(5\,6) ( 2 3 ) ( 5 6 ) . Geometrically, these are symmetries of the hexagons in Figure 1 and Figure 4 . It permutes the coordinates p i j p_{ij} p ij of P 20 \mathbb{P}^{20} P 20 , up to signs, and the coordinates d i j d_{ij} d ij of R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 . See also the proof of Theorem 8.1 (b). In this manner, D 12 \mathrm{D}_{12} D 12 acts on the varieties Q \mathcal{Q} Q , G \mathcal{G} G , F \mathcal{F} F and their tropical varieties and prevarieties. Equipped with this D 12 \mathrm{D}_{12} D 12 -action, we can now describe the G 2 \mathrm{G}_2 G 2 -Dressian.
Theorem 2.4. Modulo lineality, the G 2 \mathrm{G}_2 G 2 -Dressian is a 3 3 3 -dimensional polyhedral complex with f f f -vector ( 51 , 241 , 248 , 3 ) (51,241,248,3) ( 51 , 241 , 248 , 3 ) . Its 2 2 2 -dimensional cells are 245 245 245 triangles and three quadrilaterals, and its three 3 3 3 -cells are tetrahedra. Modulo the D 12 \mathrm{D}_{12} D 12 -action, the f f f -vector is ( 11 , 33 , 31 , 1 ) (11,33,31,1) ( 11 , 33 , 31 , 1 ) . The reduced integral homology of the G 2 \mathrm{G}_2 G 2 -Dressian is Z 54 \mathbb{Z}^{54} Z 54 , concentrated in dimension 2 2 2 .
Proof. This is a computation in the computer algebra system Oscar [12 , 33 ] . We compute the tropical prevariety of the ideal I I I given by the 35 35 35 Plücker relations and the 7 7 7 linear relations (4) . The output is a fan Σ \Sigma Σ over a polyhedral complex with f f f -vector ( 51 , 241 , 248 , 3 ) (51,241,248,3) ( 51 , 241 , 248 , 3 ) . That fan lies in a 15 15 15 -dimensional subspace cut out by the six linearly independent equations in (9) . Among the 248 248 248 polygons, we find 3 3 3 quadrilaterals and 245 245 245 triangles. After removing the 12 12 12 facets of the tetrahedra, we have 236 236 236 polygons that are maximal cells. In total, there are 239 239 239 maximal cells, namely 233 233 233 triangles, 3 3 3 quadrilaterals and 3 3 3 tetrahedra. Their cones span different linear spaces, so they give the coarsest fan structure on the support of Σ \Sigma Σ .
The action of D 12 \mathrm{D}_{12} D 12 permutes the 51 51 51 vertices and the 239 239 239 facets of the polyhedral complex underlying the fan Σ \Sigma Σ . The vertices come in 11 11 11 orbits. Their G 2 \mathrm{G}_2 G 2 -trees are shown in Figure 7 . The maximal triangles and quadrilaterals form 28 28 28 orbits. Their G 2 \mathrm{G}_2 G 2 -trees are the ones in Figure 8 and Figure 9 , except those labeled by “char 2 2 2 ” or “char ≠ 2 \ne2 = 2 ”. The tetrahedra form one orbit, represented by the G 2 \mathrm{G}_2 G 2 -tree labeled E \mathcal{E} E in Figure 7 . All trees in these figures are normalized as in Proposition 2.3 . The homology groups are also computed in Oscar. □
The G 2 \mathrm{G}_2 G 2 flag Dressian is a tropical prevariety in R 7 / 1 × R 21 / 1 \mathbb{R}^{7}/1 \times\mathbb{R}^{21}/1 R 7 /1 × R 21 /1 . We now state its definition in precise terms. The bihomogeneous prime ideal of the flag variety F ⊂ P 6 × P 20 \mathcal{F} \subset\mathbb{P}^{6} \times\mathbb{P}^{20} F ⊂ P 6 × P 20 lives in the polynomial ring K [ x , p ] \mathbb{K}[x,p] K [ x , p ] with 28 28 28 variables. We use the following 1 + 7 + ( 4 2 ) + ( 3 2 ) = 78 1+7+\binom{4}{2}+\binom{3}{2}=78 1 + 7 + ( 2 4 ) + ( 2 3 ) = 78 generators of this ideal. Namely, in bidegree ( 2 , 0 ) (2,0) ( 2 , 0 ) we have our quadratic q q q , in bidegree ( 0 , 1 ) (0,1) ( 0 , 1 ) we have the 7 7 7 linear forms in (4) , in bidegree ( 0 , 2 ) (0,2) ( 0 , 2 ) there are 35 35 35 Plücker quadrics, and in bidegree ( 1 , 1 ) (1,1) ( 1 , 1 ) we have the 35 35 35 incidence equations p i j x k − p i k x j + p j k x i = 0 p_{ij}x_k-p_{ik}x_j+p_{jk}x_i=0 p ij x k − p ik x j + p j k x i = 0 . These ensure that the point x x x lies on the line p p p . Modulo the products of the x i x_i x i with the 7 linear forms, only 34 of them are linearly independent. This is the number given in the Introduction. The G 2 \mathrm{G}_2 G 2 flag Dressian is the tropical prevariety defined by these 78 polynomials. Its lineality space is 2-dimensional, given by concatenations of each row vector in (10) with the associated star tree d d d . The former with the vector ( 1 , 1 , 1 , − 1 , − 1 , − 1 , 0 ) (1,1,1,-1,-1,-1,0) ( 1 , 1 , 1 , − 1 , − 1 , − 1 , 0 ) span the lineality of the 3-dimensional tropical hypersurface trop ( Q ) \operatorname{trop}(Q) trop ( Q ) and the latter span the lineality of the G 2 \mathrm{G}_2 G 2 -Dressian.
The tropical incidence relations are max ( d i j + d k 8 , d i k + d j 8 , d j k + d i 8 ) \max(d_{ij}+d_{k8},d_{ik}+d_{j8},d_{jk}+d_{i8}) max ( d ij + d k 8 , d ik + d j 8 , d j k + d i 8 ) , where d k 8 = − val ( x k ) d_{k8}=-\operatorname{val}(x_k) d k 8 = − val ( x k ) . They define the tropical Grassmannian trop Gr ( 2 , 8 ) \operatorname{trop}\operatorname{Gr}(2,8) trop Gr ( 2 , 8 ) over trop Gr ( 2 , 7 ) \operatorname{trop}\operatorname{Gr}(2,7) trop Gr ( 2 , 7 ) . Therefore, the points ( d i j , d k 8 ) (d_{ij},d_{k8}) ( d ij , d k 8 ) of the G 2 \mathrm{G}_2 G 2 flag Dressian are G 2 \mathrm{G}_2 G 2 -trees with one additional free leaf 8. The pendant edge at 8 can be normalized to 0. The resulting point in R 7 / 1 × R 21 / 1 \mathbb{R}^{7}/1\times\mathbb{R}^{21}/1 R 7 /1 × R 21 /1 is a decorated G 2 \mathrm{G}_2 G 2 -tree.
Theorem 2.5. Modulo lineality, the G 2 \mathrm{G}_2 G 2 flag Dressian is a 4-dimensional polyhedral complex with f f f -vector ( 141 , 1352 , 3584 , 2729 , 33 ) (141,1352,3584,2729,33) ( 141 , 1352 , 3584 , 2729 , 33 ) . The 3-cells are 2603 tetrahedra, 108 square pyramids and 18 triangular prisms. The 4-cells are 33 4-simplices. The reduced integral homology is Z 324 \mathbb{Z}^{324} Z 324 , concentrated in dimension 3. Modulo D 12 \mathrm{D}_{12} D 12 -action, the f f f -vector is ( 27 , 173 , 388 , 274 , 6 ) (27,173,388,274,6) ( 27 , 173 , 388 , 274 , 6 ) .
Proof. We used Oscar as well as Gfan [23 ] to compute the polyhedral fan. This fan has 2726 maximal cones. Exactly 99 pairs of them span the same linear space, and each pair has a convex union. Merging them removes six rays and gives 2627 maximal cones with pairwise distinct spans. The resulting fan is the coarsest, with f f f -vector ( 141 , 1352 , 3584 , 2729 , 33 ) (141,1352,3584,2729,33) ( 141 , 1352 , 3584 , 2729 , 33 ) .
Points in our flag Dressian are decorated G 2 \mathrm{G}_2 G 2 -trees. The lineality space is the diagonal: it is made of the star G 2 \mathrm{G}_2 G 2 -trees with decorated inner vertex. The rays are G 2 \mathrm{G}_2 G 2 -trees where one inner vertex is decorated. From Figure 7 , we count 2 ⋅ 20 + 3 ⋅ 30 + 4 = 134 2\cdot20+3\cdot30+4=134 2 ⋅ 20 + 3 ⋅ 30 + 4 = 134 rays. The 7 remaining rays lie over the lineality space of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) : the tree is the star tree, and the extra point 8 lies on a pendant edge. Only the index of the pendant edge matters: the supporting star tree can be moved freely in the lineality space. We similarly obtain maximal cells by decorating edges of the G 2 \mathrm{G}_2 G 2 -trees on the maximal cells of the G 2 \mathrm{G}_2 G 2 -Dressian. For example, the 33 4-simplices are obtained by decorating one of the 11 edges of the 3 tetrahedral G 2 \mathrm{G}_2 G 2 -trees. Decorating an edge of a 2-cell G 2 \mathrm{G}_2 G 2 -tree produces 2714 3-cells. The remaining 15 are obtained by decorating one of the 5 nodes of the 3 tetrahedral G 2 \mathrm{G}_2 G 2 -tree. The D 12 \mathrm{D}_{12} D 12 -action permutes rays and preserves maximal cells. The homology is computed with Oscar as in Theorem 2.4 . □
Grassmannian and Flag Variety
The tropical G 2 \mathrm{G}_2 G 2 -Grassmannian is contained in the G 2 \mathrm{G}_2 G 2 -Dressian. This inclusion, shown in (7) , must be strict by Theorem 2.4 . The cones over the three tetrahedra have dimension larger than dim ( G ) = 5 \dim(\mathcal{G})=5 dim ( G ) = 5 , whereas trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) is pure of dimension 5, by the Structure Theorem [29 Theorem 3.3.5] . Our next theorem shows that the tetrahedra are the only cells in the difference. For all results in this section, we assume that the field K has characteristic 0.
Theorem 3.1. Modulo lineality, the tropical G 2 \mathrm{G}_2 G 2 -Grassmannian trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) is a 2-dimensional polyhedral complex with f f f -vector ( 51 , 241 , 248 ) (51,241,248) ( 51 , 241 , 248 ) . Its 2-cells are 245 triangles and three quadrilaterals. The reduced integral homology is Z 57 \mathbb{Z}^{57} Z 57 , concentrated in dimension 2. Modulo the D 12 \mathrm{D}_{12} D 12 -action, the f f f -vector of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) is ( 11 , 33 , 31 ) (11,33,31) ( 11 , 33 , 31 ) . Each cell is encoded by a G 2 \mathrm{G}_2 G 2 -tree.
Proof. We perform the computation in Oscar [12 ] . The input is the same ideal I I I as in Theorem 2.4 . The output is the tropical variety, with its Gröbner fan structure. It has the same lineality space as the Dressian. Modulo lineality, the fan is simplicial and has f f f -vector ( 51 , 244 , 251 ) (51,244,251) ( 51 , 244 , 251 ) . The 251 maximal cones span 248 distinct linear spaces: three pairs have the same span. We merge each pair of triangles into a quadrilateral, by deleting its diagonal. This leaves 245 triangles and three quadrilaterals. The 248 maximal cones have pairwise distinct spans, so we have the coarsest fan structure. This gives the f f f -vector in Theorem 3.1 . The three quadrilaterals are precisely those of the G 2 \mathrm{G}_2 G 2 -Dressian. Thus, the Grassmannian is obtained from the Dressian by removing the three solid tetrahedra (cf. Theorem 2.4 ). The computation of orbits and of homology is analogous to those in the proof of Theorem 2.4 . □
The 11 types of rays are shown in Figure 7 and the 31 types of maximal cells are shown in Figure 8 and Figure 9 . These catalogues represent many refined aspects. Here is a first illustration.
Remark 3.2. All cones in the tropical Grassmannian trop ( Gr ( 2 , n ) ) \operatorname{trop}(\operatorname{Gr}(2,n)) trop ( Gr ( 2 , n )) have weight 1. This is not the case for trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . One of the 31 types of maximal cones has multiplicity 2. This is type 21, on the upper left in Figure 9 . See Example 6.2 for an algebraic explanation. All other cones have weights 1. With this, the weighted fan trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) is balanced [29 Theorem 3.3.5] .
We now turn to the flag variety F \mathcal{F} F . A direct Oscar computation of the tropical variety did not terminate on the ideal of F \mathcal{F} F . Instead we exploit symmetry with the following algorithm.
Algorithm 1 Tropical variety via symmetric pruning
Input: A homogeneous prime ideal J = ⟨ S ⟩ J=\langle S\rangle J = ⟨ S ⟩ of dimension d d d , the tropical prevariety Σ \Sigma Σ of the set S S S , and a group Γ \Gamma Γ of signed permutations of the variables that preserves J J J and Σ \Sigma Σ .
Output: The maximal cones of a fan with support trop ( V ( J ) ) \operatorname{trop}(V(J)) trop ( V ( J ))
for all Γ \Gamma Γ -orbits of cones σ ∈ Σ \sigma\in\Sigma σ ∈ Σ with dim σ ≥ d + 1 \dim\sigma\ge d+1 dim σ ≥ d + 1 do
Compute a Gröbner basis of J J J for a weight in relint ( σ ) \operatorname{relint}(\sigma) relint ( σ ) .
Find g ∈ J g\in J g ∈ J with in w ( g ) \operatorname{in}_{w}(g) in w ( g ) a monomial for all w ∈ relint ( σ ) w\in\operatorname{relint}(\sigma) w ∈ relint ( σ ) ; if none, return fail.
for all Γ \Gamma Γ -orbits of cones τ ∈ Σ \tau\in\Sigma τ ∈ Σ with dim τ = d \dim\tau=d dim τ = d do
Let w w w be the sum of the rays of τ \tau τ .
Keep the orbit of τ \tau τ if in w ( J ) \operatorname{in}_{w}(J) in w ( J ) contains no monomial.
return the kept cones
The correctness of Algorithm 1 follows from basics in tropical geometry, including the Structure Theorem [29 Theorem 3.3.5] and the Fundamental Theorem [29 Theorem 3.2.3] .
Theorem 3.3. Modulo lineality, the tropical G 2 \mathrm{G}_2 G 2 flag variety trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) is a 3-dimensional polyhedral complex with f f f -vector ( 141 , 1352 , 3584 , 2714 ) (141,1352,3584,2714) ( 141 , 1352 , 3584 , 2714 ) . Its 3-cells are 2588 tetrahedra, 108 square pyramids and 18 triangular prisms. The reduced integral homology is Z 342 \mathbb{Z}^{342} Z 342 , concentrated in dimension 3. Modulo the D 12 D_{12} D 12 -action, the f f f -vector is ( 27 , 173 , 388 , 270 ) (27,173,388,270) ( 27 , 173 , 388 , 270 ) . Each cell in trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) is encoded by a G 2 \mathrm{G}_2 G 2 -tree that is decorated as in § 7 .
Proof. We apply Algorithm 1 to the ideal J J J of F ⊂ Q × G \mathcal{F}\subset\mathcal{Q}\times\mathcal{G} F ⊂ Q × G , with S S S its 78 generators and Γ = D 12 \Gamma=D_{12} Γ = D 12 . For some vector w 0 w_0 w 0 , the initial ideal in w 0 ( J ) \operatorname{in}_{w_0}(J) in w 0 ( J ) is generated by 83 binomials and is
prime over Q \mathbb{Q} Q . Hence J J J is prime as the algorithm requires. The prevariety of S S S is the flag Dressian of Theorem 2.5 , having 2627 maximal cells, of dimensions 3 and 4 modulo lineality. The first loop removes the 33 4-simplices, in 6 orbits. This leaves the 2729 3-cells of the flag Dressian, in 274 orbits. The second loop keeps 270 of these orbits, with 2714 cells. The output is the coarsest and has f f f -vector ( 141 , 1352 , 3584 , 2714 ) (141,1352,3584,2714) ( 141 , 1352 , 3584 , 2714 ) , with 2588 tetrahedra, 108 square pyramids and 18 triangular prisms. It has the same lineality space as its Dressian. Projection to the second factor maps trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) onto trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . Each of the 2714 maximal cells maps onto one of the 248 maximal cells of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . Each maximal cell of trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) is a maximal G 2 \mathrm{G}_2 G 2 -tree decorated by a point in the relative interior of an edge. This removes the 15 tetrahedra of the G 2 \mathrm{G}_2 G 2 flag Dressian corresponding to node-decorated tetrahedral G 2 \mathrm{G}_2 G 2 -trees. The 141 rays are the same as those in the G 2 \mathrm{G}_2 G 2 flag Dressian. Namely, 134 lie over rays of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) , and all 51 of these occur. Each ray is a minimal G 2 \mathrm{G}_2 G 2 -tree with a marked point 8. The remaining 7 rays lie over the lineality space of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) : the tree is the star tree, with the point 8 on an edge.
Theorems Theorems 3.1 and 2.4 show that the Dressian and the tropical Grassmannian are different. Our G 2 \mathrm{G}_2 G 2 -theory is richer than the well-known identification of phylogenetic trees with points in trop ( Gr ( 2 , n ) ) \operatorname{trop}(\operatorname{Gr}(2,n)) trop ( Gr ( 2 , n )) . Namely, we learned that the 7 + 35 7+35 7 + 35 linear and quadratic generators for the prime ideal I I I of the Grassmannian G \mathcal{G} G do not form a tropical basis [29 Definition 2.6.3] . We now remedy this point by augmenting our given generators to a tropical basis for G \mathcal{G} G .
Corollary 3.4. The 35 Plücker relations, the 7 linear relations (4) , and the 3 cubics
2 p 45 p 46 p 56 + p 23 p 36 p 56 − p 14 p 12 p 45 + p 26 p 13 p 45 − p 35 p 12 p 46 2 p 23 p 12 p 13 − p 46 p 14 p 13 + p 25 p 56 p 23 + p 16 p 45 p 23 − p 34 p 56 p 12 2 p 46 p 56 p 45 + p 12 p 25 p 45 − p 36 p 13 p 46 + p 15 p 23 p 46 − p 24 p 13 p 56 (11) \begin{aligned}
2p_{45}p_{46}p_{56} + p_{23}p_{36}p_{56} - p_{14}p_{12}p_{45} + p_{26}p_{13}p_{45} - p_{35}p_{12}p_{46} \\
2p_{23}p_{12}p_{13} - p_{46}p_{14}p_{13} + p_{25}p_{56}p_{23} + p_{16}p_{45}p_{23} - p_{34}p_{56}p_{12} \\
2p_{46}p_{56}p_{45} + p_{12}p_{25}p_{45} - p_{36}p_{13}p_{46} + p_{15}p_{23}p_{46} - p_{24}p_{13}p_{56} \tag*{(11)}
\end{aligned} 2 p 45 p 46 p 56 + p 23 p 36 p 56 − p 14 p 12 p 45 + p 26 p 13 p 45 − p 35 p 12 p 46 2 p 23 p 12 p 13 − p 46 p 14 p 13 + p 25 p 56 p 23 + p 16 p 45 p 23 − p 34 p 56 p 12 2 p 46 p 56 p 45 + p 12 p 25 p 45 − p 36 p 13 p 46 + p 15 p 23 p 46 − p 24 p 13 p 56 (11) form a tropical basis for the ideal I I I that defines the G 2 \mathrm{G}_2 G 2 -Grassmannian G \mathcal{G} G in P 20 \mathbb{P}^{20} P 20 .
Proof. One first checks that each of the three cubics in (11) lies in the ideal I I I of G \mathcal{G} G . To show that our 45 = 35 + 7 + 3 45=35+7+3 45 = 35 + 7 + 3 polynomials form a tropical basis, we must prove the following statement about an arbitrary element d d d in R 21 / 1 \mathbb{R}^{21}/\mathbf{1} R 21 / 1 . If d d d does not lie in trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) then our list of 45 candidates contains one polynomial f f f whose initial form in d ( f ) \operatorname{in}_{d}(f) in d ( f ) is a monomial. This claim clearly holds for all d d d that are not in the G 2 \mathrm{G}_2 G 2 -Dressian: here we can take f f f from among the 42 = 35 + 7 42=35+7 42 = 35 + 7 original generators. We may therefore assume that d d d lies in the G 2 \mathrm{G}_2 G 2 -Dressian.
From the description of the maximal cells in Theorem 2.4 , we know that d d d lies in the relative interior of one of the three special tetrahedra. To examine the tetrahedra in more detail, we showcase 15 special G 2 \mathrm{G}_2 G 2 -trees, one for each of the 5 + 5 + 5 5+5+5 5 + 5 + 5 monomials appearing in (11) . These G 2 \mathrm{G}_2 G 2 -trees are depicted in Figure 5 . Their edges have length 1, except for the double-edges, which have length 2. For instance, the trees in the second column are splits. For the last tree in the first row, the tree metric d d d satisfies d 13 = 3 d_{13}=3 d 13 = 3 , d 23 = 2 d_{23}=2 d 23 = 2 , d 12 = 1 d_{12}=1 d 12 = 1 , d 35 = 0 d_{35}=0 d 35 = 0 .
Each row in Figure 5 represents one of the three tetrahedra. The last four G 2 \mathrm{G}_2 G 2 -trees are the vertices of the tetrahedron, and the leftmost G 2 \mathrm{G}_2 G 2 -tree is in the interior of the tetrahedron. One checks the following property for the cubic polynomial f f f that is written above the trees. For each of the four trees d d d on the right, the initial form in d ( f ) \operatorname{in}_{d}(f) in d ( f ) is the sum of the four other monomials. This verifies the fact, already known from Theorem 3.1 , that d d d lies in trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . Consider the leftmost tree d d d . The initial form in d ( f ) \operatorname{in}_{d}(f) in d ( f ) is the monomial above it. This manifests d ∉ trop ( G ) d \notin\operatorname{trop}(\mathcal{G}) d ∈ / trop ( G ) . We now note that the initial form in d ( f ) \operatorname{in}_{d}(f) in d ( f ) remains unchanged if d d d gets replaced by any positive linear combination of the four G 2 G_{2} G 2 -trees on the right. For every d d d in the interior of the tetrahedron, in d ( f ) \operatorname{in}_{d}(f) in d ( f ) is a monomial. Hence the cubic f f f is a witness for the entire tetrahedron. This works for each of the three rows, and this completes the proof.
Characteristic Two
The integer 2 appears as the leading coefficient of the cubics (11) in our tropical basis. This suggests that the tropicalization of the Grassmannian G \mathcal{G} G depends on the field characteristic. In this section we show that this is indeed the case. In what follows, K \mathbb{K} K denotes an arbitrary field with trivial valuation. Since G \mathcal{G} G is defined over Z \mathbb{Z} Z , its tropicalization depends only on the characteristic p = char ( K ) p = \operatorname{char}(\mathbb{K}) p = char ( K ) , and we write trop p ( G ) \operatorname{trop}_{p}(\mathcal{G}) trop p ( G ) for this tropical variety in R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 . The tropical variety for p = 0 p = 0 p = 0 is the fan trop 0 ( G ) = trop ( G ) \operatorname{trop}_{0}(\mathcal{G}) = \operatorname{trop}(\mathcal{G}) trop 0 ( G ) = trop ( G ) , which is described in Theorem 3.1 .
Theorem 4.1. If p = char ( K ) p = \operatorname{char}(\mathbb{K}) p = char ( K ) is an odd prime then trop p ( G ) = trop ( G ) \operatorname{trop}_{p}(\mathcal{G}) = \operatorname{trop}(\mathcal{G}) trop p ( G ) = trop ( G ) . Now assume p = 2 p = 2 p = 2 . Modulo lineality, the tropical G 2 G_{2} G 2 -Grassmannian trop 2 ( G ) \operatorname{trop}_{2}(\mathcal{G}) trop 2 ( G ) is a 2-dimensional polyhedral complex with f f f -vector ( 54 , 253 , 254 ) (54, 253, 254) ( 54 , 253 , 254 ) . Its 2-cells are 251 triangles and three quadrilaterals. The reduced integral homology is Z 54 \mathbb{Z}^{54} Z 54 , concentrated in dimension 2. Modulo D 12 D_{12} D 12 -action, the f f f -vector of trop 2 ( G ) \operatorname{trop}_{2}(\mathcal{G}) trop 2 ( G ) is ( 12 , 36 , 32 ) (12, 36, 32) ( 12 , 36 , 32 ) . Each cell is encoded by a G 2 G_{2} G 2 -tree (see Figure 7 , Figure 8 , Figure 9 ).
The phenomenon of characteristic-dependence is familiar from [36 Section 7] , where it was shown for the tropicalization of Gr ( 3 , 7 ) \operatorname{Gr}(3,7) Gr ( 3 , 7 ) . In the classical case, by [36 Theorem 7.1] , trop ( Gr ( 2 , n ) ) \operatorname{trop}(\operatorname{Gr}(2,n)) trop ( Gr ( 2 , n )) is characteristic-free: every tree lifts to a line in P n − 1 \mathbb{P}^{n-1} P n − 1 . We here show that the situation is different on the locus in Gr ( 2 , 7 ) \operatorname{Gr}(2,7) Gr ( 2 , 7 ) that is relevant for the exceptional group G 2 \mathrm{G}_2 G 2 .
Before proving Theorem 4.1 , we explain its combinatorial essence. The characteristic 2 2 2 features emerge in the three tetrahedra of the G 2 \mathrm{G}_2 G 2 -Dressian. Each tetrahedron has a new ray, labeled ε \varepsilon ε in Figure 7 , that appears in its relative interior. For p = 2 p=2 p = 2 , the four facets of the tetrahedra are deleted and replaced by six triangles joining the center ε \varepsilon ε to the six edges. This is shown in Figure 6 . The vertices of the tetrahedron are G 2 \mathrm{G}_2 G 2 -trees of types α 2 \alpha_2 α 2 , α 4 \alpha_4 α 4 , α 3 \alpha_3 α 3 , α 3 \alpha_3 α 3 . These are shown in Figure 5 , and – more colorfully – in Figure 7 . The 4 + 6 4+6 4 + 6 triangles in the second row of Figure 6 are G 2 \mathrm{G}_2 G 2 -trees of types 3 2 \mathbf{3}_2 3 2 , 3 2 \mathbf{3}_2 3 2 , 3 3 \mathbf{3}_3 3 3 , 3 4 \mathbf{3}_4 3 4 for p ≠ 2 p\ne2 p = 2 and of types 1 8 \mathbf{1}_8 1 8 , 1 8 \mathbf{1}_8 1 8 , 1 9 \mathbf{1}_9 1 9 , 2 0 \mathbf{2}_0 2 0 , 3 5 \mathbf{3}_5 3 5 , 3 5 \mathbf{3}_5 3 5 for p = 2 p=2 p = 2 . Representatives for these orbits are shown in Figure 8 and Figure 9 .
Proof of Theorem 4.1 . Let I K I_K I K be the ideal in K [ p ] K[p] K [ p ] that is generated by the 35 Plücker quadrics and the seven linear forms in (4) . Since these generators have coefficients ± 1 \pm1 ± 1 , the tropical variety trop p ( G ) \operatorname{trop}_p(\mathcal{G}) trop p ( G ) is contained in the G 2 \mathrm{G}_2 G 2 -Dressian for every p p p . For odd primes p p p , the proof of Corollary 3.4 applies verbatim, as 2 ≠ 0 2\ne0 2 = 0 in K K K . This shows the inclusion trop p ( G ) ⊆ trop ( G ) \operatorname{trop}_p(\mathcal{G})\subseteq\operatorname{trop}(\mathcal{G}) trop p ( G ) ⊆ trop ( G ) .
For the reverse inclusion, we use the following criterion: if C C C is a cone and G G G a Gröbner basis of I K I_K I K for an order refining some w ∈ relint ( C ) w\in\operatorname{relint}(C) w ∈ relint ( C ) , with in v ( g ) = in w ( g ) \operatorname{in}_v(g)=\operatorname{in}_w(g) in v ( g ) = in w ( g ) for all g ∈ G g\in G g ∈ G and v ∈ relint ( C ) v\in\operatorname{relint}(C) v ∈ relint ( C ) , and the in w ( g ) \operatorname{in}_w(g) in w ( g ) have a common zero in the torus ( K ‾ ∗ ) 21 (\overline{K}^{*})^{21} ( K ∗ ) 21 , then C ⊆ trop p ( G ) C\subseteq\operatorname{trop}_p(\mathcal{G}) C ⊆ trop p ( G ) .
We apply this criterion to the 251 maximal cones C C C of the simplicial fan structure on trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) in the proof of Theorem 3.1 , with G G G a Gröbner basis over Z \mathbb{Z} Z , as in [1 §4] . This was computed in Oscar. We verify that in v ( g ) = in w ( g ) \operatorname{in}_v(g)=\operatorname{in}_w(g) in v ( g ) = in w ( g ) on relint ( C ) \operatorname{relint}(C) relint ( C ) , and that the leading coefficients appearing in G G G are 1 or 2. They are 1 unless C C C is a facet of a tetrahedron. For p p p not dividing them, G G G reduces modulo p p p to a Gröbner basis of I K I_K I K with the same initial forms. These have a common zero in { ± 1 } 21 \{\pm1\}^{21} { ± 1 } 21 for 237 cones. For the 12 facets of the tetrahedra and the two cones of type 21, we verify that the ideal generated by the in w ( g ) \operatorname{in}_w(g) in w ( g ) , by p i j − 1 p_{ij}-1 p ij − 1 for six coordinates i j ∈ S ij \in S ij ∈ S , and by y ∏ i ′ j ′ ∉ S p i ′ j ′ − 1 y\prod_{i'j'\notin S}p_{i'j'}-1 y ∏ i ′ j ′ ∈ / S p i ′ j ′ − 1 has a Gröbner basis over Z \mathbb{Z} Z without constant whose leading coefficients are powers of 2 2 2 , all 1 1 1 for type 21 \mathbf{21} 21 ; for p p p not dividing them, this ideal is proper over K K K , and its zeros are torus points. Hence C ⊆ trop p ( G ) C \subseteq\operatorname{trop}_p(\mathcal{G}) C ⊆ trop p ( G ) for odd p p p , which proves the first assertion. This also works for p = 2 p=2 p = 2 unless C C C is a facet of a tetrahedron.
Now let p = 2 p=2 p = 2 . Since trop 2 ( G ) \operatorname{trop}_2(\mathcal{G}) trop 2 ( G ) is contained in the Dressian, it remains to examine the three tetrahedra. Modulo 2 2 2 , the cubic f f f in (11) of a tetrahedron T T T has four terms, of equal weight at the G2 _2 2 -tree ε \varepsilon ε of T T T ; at each vertex of T T T , one term is strictly smaller than the other three, which are equal, and each term is the smaller one at exactly one vertex. Hence in w ( f ) \operatorname{in}_w(f) in w ( f ) is a monomial for all w ∈ T w \in T w ∈ T outside the six triangles spanned by ε \varepsilon ε and an edge of T T T , and trop 2 ( G ) ∩ T \operatorname{trop}_2(\mathcal{G}) \cap T trop 2 ( G ) ∩ T lies in these triangles. Conversely, each of these 18 triangles satisfies the criterion, with a Gröbner basis of I F 2 I_{\mathbb{F}_2} I F 2 and the common zero ( 1 , … , 1 ) (1,\ldots,1) ( 1 , … , 1 ) . Thus trop 2 ( G ) \operatorname{trop}_2(\mathcal{G}) trop 2 ( G ) consists of the 239 cones above and the 18 triangles. Its f f f -vector, orbits and homology are computed as in Theorem 3.1 , after merging three pairs of triangles into quadrilaterals.
An analogous result holds for the flag variety F \mathcal{F} F . We state this in a second theorem.
Theorem 4.2. If p = char ( K ) p=\operatorname{char}(\mathbb{K}) p = char ( K ) is odd then trop p ( F ) = trop ( F ) \operatorname{trop}_p(\mathcal{F})=\operatorname{trop}(\mathcal{F}) trop p ( F ) = trop ( F ) . If p = 2 p=2 p = 2 then, modulo lineality, the tropical G2 _2 2 flag variety trop 2 ( F ) \operatorname{trop}_2(\mathcal{F}) trop 2 ( F ) is a 3-dimensional polyhedral complex with f f f -vector ( 156 , 1445 , 3758 , 2792 ) (156,1445,3758,2792) ( 156 , 1445 , 3758 , 2792 ) . Its 3-cells are 2630 tetrahedra, 144 square pyramids and 18 triangular prisms. The reduced integral homology is Z 324 \mathbb{Z}^{324} Z 324 , concentrated in dimension 3. Modulo D12 _{12} 12 -action, trop 2 ( F ) \operatorname{trop}_2(\mathcal{F}) trop 2 ( F ) has f f f -vector ( 31 , 192 , 416 , 281 ) (31,192,416,281) ( 31 , 192 , 416 , 281 ) . Each cell is a decorated G2 _2 2 -tree.
Proof. We proceed as in the proof of Theorem 4.1 . Let p ≥ 3 p \ge3 p ≥ 3 . Since the generators of J J J have coefficients ± 1 \pm1 ± 1 , trop p ( F ) \operatorname{trop}_p(\mathcal{F}) trop p ( F ) lies in the flag Dressian. Outside trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) , it consists of the 33 open 4-simplices and 15 of their facets. On each 4-simplex, the initial form of a cubic in (11) is its term with coefficient 2. So trop p ( F ) ⊆ trop ( F ) \operatorname{trop}_p(\mathcal{F}) \subseteq\operatorname{trop}(\mathcal{F}) trop p ( F ) ⊆ trop ( F ) : the generators and the three cubics form a tropical basis for p ≠ 2 p \ne2 p = 2 . Conversely, we apply the criterion to one maximal cone in each of the 270 D12 _{12} 12 -orbits, as D12 _{12} 12 acts by signed permutations preserving J J J over Z \mathbb{Z} Z , or to the Gröbner cones in it if the initial forms vary (44 orbits). Leading coefficients are 1, 2 or 4; torus points are ± 1 \pm1 ± 1 -points or slices. For p = 2 p=2 p = 2 we apply Algorithm 1 over F 2 \mathbb{F}_2 F 2 to the 78 generators and 12 elements of a tropical basis of G \mathcal{G} G over F 2 \mathbb{F}_2 F 2 . Their prevalence’s 3038 cones merge into 2792 of dimension 8 in 281 orbits, all kept. We locate the output in the flag Dressian of Theorem 2.5 . Over each tetrahedron, a decorated version of Figure 6 emerges.
The G2 _2 2 -Forest
The G2 _2 2 -forest is composed of all G2 _2 2 -trees that correspond to rays and maximal cells of our G2 _2 2 -Grassmannians. Here, we include both the general version trop ( G ) = trop p ( G ) \operatorname{trop}(\mathcal{G})=\operatorname{trop}_p(\mathcal{G}) trop ( G ) = trop p ( G ) , p ≥ 3 p \ge3 p ≥ 3 , and the characteristic 2 version trop 2 ( G ) \operatorname{trop}_2(\mathcal{G}) trop 2 ( G ) . These are described in Theorems 3.1 and 4.1 . The D12 _{12} 12 -orbits of the G2 _2 2 -trees are shown in Figure 7 , Figure 8 , and Figure 9 . The diagrams are labeled by symbols in boldface, either Greek letters or numbers, followed by the orbit size in parentheses. The edge coloring and the green apples seen in Figure 8 will be explained in § 7 .
There are 12 types of rays, 11 for char ( K ) ≠ 2 \operatorname{char}(\mathbb{K}) \ne2 char ( K ) = 2 plus one for char ( K ) = 2 \operatorname{char}(\mathbb{K})=2 char ( K ) = 2 . Their G2 _2 2 -trees are shown in Figure 7 . The type α \alpha α represents splits, while β \beta β or γ \gamma γ represent caterpillars, with the heart either at an end or in the middle. Finally, δ \delta δ and ε \varepsilon ε represent snowflakes, where δ \delta δ is fixed under D 12 \mathrm{D}_{12} D 12 . Note that ε \varepsilon ε appears only in characteristic 2; see Figure 5 and Figure 6 .
We group the maximal cells into types 1–20 that yield a toric degeneration, in Figure 8 , and types 21–35 that do not, in Figure 9 . Toric degenerations will be explained in § 6 , where Table 2 matches Figure 8 . Types 18–20 and 35 appear only in characteristic 2 while types 32–34 disappear in characteristic 2. Type 30 is the orbit of the three quadrilaterals.
Each orbit size is easily found by looking at the G 2 \mathrm{G}_{2} G 2 -trees under the D 12 \mathrm{D}_{12} D 12 -action in Figure 1 . For instance, the split type α 2 \alpha_{2} α 2 has an orbit of size 3: its three G 2 \mathrm{G}_{2} G 2 -trees correspond to grouping the heart with two same shapes and splitting them from the other four leaves.
Table 1 lists which type of rays generates each type of maximal cells. This table also helps to identify pairs of G 2 \mathrm{G}_{2} G 2 -trees that form edges of our G 2 \mathrm{G}_{2} G 2 -Grassmannian: those are generated by two rays of this table. One necessary (but not sufficient) condition for several rays to generate a cell is for their associated trees to have compatible splits. For instance, δ \delta δ and α 4 \alpha_{4} α 4 are compatible, and they form an edge of the triangles 13 and 14, as listed in Table 1 .
Table 1. The ray orbits of the maximal cells in Figure 8 and Figure 9 , labeled as in Figure 7 .
Figure 10 describes all 4 + 4 + 1 4+4+1 4 + 4 + 1 G 2 G_2 G 2 -trees on a quadrilateral (type 30). Two vertices of the quadrilateral belong to the same orbit γ 1 \gamma_1 γ 1 . Inner edges of length zero were artificially added so the support trees agree. The G 2 G_2 G 2 -tree d d d on the quadrilateral is normalized and expressed as 1 6 ( γ 1 + γ 1 ′ − 2 s ) = γ 2 + γ 3 \frac{1}{6}(\gamma_1+\gamma'_1-2s)=\gamma_2+\gamma_3 6 1 ( γ 1 + γ 1 ′ − 2 s ) = γ 2 + γ 3 , where s s s is the star tree with all pendant edges of lengths 1. To be completely explicit, we have d 12 = d 16 = d 26 = d 34 = d 35 = d 36 = d 37 = d 45 = d 67 = 2 d_{12}=d_{16}=d_{26}=d_{34}=d_{35}=d_{36}=d_{37}=d_{45}=d_{67}=2 d 12 = d 16 = d 26 = d 34 = d 35 = d 36 = d 37 = d 45 = d 67 = 2 , d 13 = d 17 = d 23 = d 27 = d 46 = d 47 = d 56 = d 57 = 4 d_{13}=d_{17}=d_{23}=d_{27}=d_{46}=d_{47}=d_{56}=d_{57}=4 d 13 = d 17 = d 23 = d 27 = d 46 = d 47 = d 56 = d 57 = 4 , and d 14 = d 15 = d 24 = d 25 = 6 d_{14}=d_{15}=d_{24}=d_{25}=6 d 14 = d 15 = d 24 = d 25 = 6 for the tree metric d d d .
Figure 10. Detailed description of type 30 . This maximal cell of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) is a quadrilateral.
Toric Degenerations
Toric degeneration is a process which replaces a given projective variety by a toric variety. This process retains key invariants of the original variety, such as the dimension and the degree, and it offers new insights by means of polyhedral geometry and combinatorics. Associated with the toric variety is a lattice polytope, called the Newton–Okounkov polytope . The lattice points in this polytope encode elements in a Khovanskii basis for the given coordinate ring, and this in turn offers valuable algebraic information. This approach has been especially fruitful in representation theory, where one studies varieties with a group action, like Grassmannians and flag varieties. In what follows, we explore this for our group G 2 G_2 G 2 .
Kaveh and Manon [28 ] proved that the toric degenerations correspond to the maximal cones in the tropical variety whose associated initial ideal is prime. From this data, one can then read off the toric variety and the Newton–Okounkov polytope. We shall explain this process for the G 2 G_2 G 2 -Grassmannian, thereby classifying all toric degenerations of G \mathcal{G} G inside P 20 \mathbb{P}^{20} P 20 . Our description is for fields of every characteristic, and hence for both trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) and trop 2 ( G ) \operatorname{trop}_2(\mathcal{G}) trop 2 ( G ) .
Let C C C be a maximal cone of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . We call C C C a prime cone if the initial ideal in C ( G ) \operatorname{in}_C(\mathcal{G}) in C ( G ) is prime over the algebraic closure K ‾ \overline{\mathbb{K}} K of K \mathbb{K} K . Here we mean the ideal in d ( G ) \operatorname{in}_d(\mathcal{G}) in d ( G ) for any d ∈ relint ( C ) d \in\operatorname{relint}(C) d ∈ relint ( C ) . If C C C is a prime cone then in C ( G ) \operatorname{in}_C(\mathcal{G}) in C ( G ) is a toric ideal , and we also write in C ( G ) \operatorname{in}_C(\mathcal{G}) in C ( G ) for the corresponding projective toric variety in P 20 \mathbb{P}^{20} P 20 . The Newton–Okounkov polytope P C P_C P C , which is associated to this toric variety, has dimension 5 = dim ( G ) 5 = \dim(\mathcal{G}) 5 = dim ( G ) . Its normalized volume is 18 = degree ( G ) 18 = \operatorname{degree}(\mathcal{G}) 18 = degree ( G ) . We compute P C P_C P C as follows. Consider any 5 × 21 5 \times21 5 × 21 matrix whose rows form a Z \mathbb{Z} Z -basis for the lattice points in the linear span of C C C . Then P C P_C P C is the convex hull in R 5 \mathbb{R}^5 R 5 of the 21 21 21 columns.
Example 6.1 (Type 1). The first tree in Figure 8 equals the rightmost tree in Figure 2 . Its metric d d d is shown explicitly in Example 2.1. The initial ideal in d ( G ) \operatorname{in}_d(\mathcal{G}) in d ( G ) is toric. Its generators are
p 37 − p 45 , p 12 + p 67 , p 27 + p 46 , p 23 + p 47 , p 17 − p 56 , p 13 − p 57 , p 14 + p 25 , p 26 p 57 + p 36 p 67 , p 36 p 45 − p 35 p 46 , p 45 p 56 − p 36 p 57 , p 36 p 56 + p 57 p 67 , p 45 p 46 − p 36 p 47 , p 36 p 46 + p 47 p 67 , p 25 p 35 + p 45 p 57 , p 24 p 35 + p 45 p 47 , p 25 p 26 − p 46 p 67 , p 15 p 26 + p 56 p 67 , p 47 p 56 − p 46 p 57 , p 46 p 56 − p 25 p 67 , p 25 p 45 + p 47 p 57 , p 25 p 36 + p 46 p 57 , p 24 p 36 + p 46 p 47 , p 15 p 36 − p 56 p 57 , p 25 p 47 − p 24 p 57 , p 15 p 47 + p 25 p 57 , p 25 p 46 − p 24 p 56 , p 15 p 46 + p 25 p 56 , p 25 2 + p 15 p 24 , p 36 2 + p 45 p 67 , p 45 2 − p 35 p 47 , p 46 2 − p 24 p 67 , p 47 2 + p 24 p 45 , p 56 2 + p 15 p 67 , p 57 2 − p 15 p 45 , p 67 2 − p 26 p 56 . \begin{aligned}
&p_{37}-p_{45},\ p_{12}+p_{67},\ p_{27}+p_{46},\ p_{23}+p_{47},\ p_{17}-p_{56},\ p_{13}-p_{57},\ p_{14}+p_{25},\\
&p_{26}p_{57}+p_{36}p_{67},\ p_{36}p_{45}-p_{35}p_{46},\ p_{45}p_{56}-p_{36}p_{57},\ p_{36}p_{56}+p_{57}p_{67},\\
&p_{45}p_{46}-p_{36}p_{47},\ p_{36}p_{46}+p_{47}p_{67},\ p_{25}p_{35}+p_{45}p_{57},\ p_{24}p_{35}+p_{45}p_{47},\\
&p_{25}p_{26}-p_{46}p_{67},\ p_{15}p_{26}+p_{56}p_{67},\ p_{47}p_{56}-p_{46}p_{57},\ p_{46}p_{56}-p_{25}p_{67},\\
&p_{25}p_{45}+p_{47}p_{57},\ p_{25}p_{36}+p_{46}p_{57},\ p_{24}p_{36}+p_{46}p_{47},\ p_{15}p_{36}-p_{56}p_{57},\\
&p_{25}p_{47}-p_{24}p_{57},\ p_{15}p_{47}+p_{25}p_{57},\ p_{25}p_{46}-p_{24}p_{56},\ p_{15}p_{46}+p_{25}p_{56},\\
&p_{25}^{2}+p_{15}p_{24},\ p_{36}^{2}+p_{45}p_{67},\ p_{45}^{2}-p_{35}p_{47},\ p_{46}^{2}-p_{24}p_{67},\\
&p_{47}^{2}+p_{24}p_{45},\ p_{56}^{2}+p_{15}p_{67},\ p_{57}^{2}-p_{15}p_{45},\ p_{67}^{2}-p_{26}p_{56}.
\end{aligned} p 37 − p 45 , p 12 + p 67 , p 27 + p 46 , p 23 + p 47 , p 17 − p 56 , p 13 − p 57 , p 14 + p 25 , p 26 p 57 + p 36 p 67 , p 36 p 45 − p 35 p 46 , p 45 p 56 − p 36 p 57 , p 36 p 56 + p 57 p 67 , p 45 p 46 − p 36 p 47 , p 36 p 46 + p 47 p 67 , p 25 p 35 + p 45 p 57 , p 24 p 35 + p 45 p 47 , p 25 p 26 − p 46 p 67 , p 15 p 26 + p 56 p 67 , p 47 p 56 − p 46 p 57 , p 46 p 56 − p 25 p 67 , p 25 p 45 + p 47 p 57 , p 25 p 36 + p 46 p 57 , p 24 p 36 + p 46 p 47 , p 15 p 36 − p 56 p 57 , p 25 p 47 − p 24 p 57 , p 15 p 47 + p 25 p 57 , p 25 p 46 − p 24 p 56 , p 15 p 46 + p 25 p 56 , p 25 2 + p 15 p 24 , p 36 2 + p 45 p 67 , p 45 2 − p 35 p 47 , p 46 2 − p 24 p 67 , p 47 2 + p 24 p 45 , p 56 2 + p 15 p 67 , p 57 2 − p 15 p 45 , p 67 2 − p 26 p 56 .
The 7 7 7 linear binomials say that P C P_C P C is the convex hull of 14 = 21 − 7 14 = 21 - 7 14 = 21 − 7 distinct lattice points. The last 8 8 8 quadrics show that only 6 = 14 − 8 6 = 14 - 8 6 = 14 − 8 of these are vertices. Hence P C P_C P C is a 5 5 5 -simplex. See the first row of Table 2 . This is the Chevalley polytope studied by Spacek and Wang [34 ] .
Example 6.2 (Type 21). The first tree in Figure 9 equals that on the triangle in Figure 12 . We see that d 15 = d 26 = d 34 = 0 d_{15}=d_{26}=d_{34}=0 d 15 = d 26 = d 34 = 0 and d i j = 2 d_{ij}=2 d ij = 2 for all other i , j i,j i , j . The initial ideal in d ( G ) \operatorname{in}_d(\mathcal{G}) in d ( G ) is prime over Q \mathbb{Q} Q , so Oscar calls this a prime ideal. However, the ideal is not prime over C \mathbb{C} C . To see this, we check that ( p 12 2 p 13 ) 2 + ( p 12 2 p 13 ) ( p 14 2 p 16 ) + ( p 14 2 p 16 ) 2 (p_{12}^{2}p_{13})^{2}+(p_{12}^{2}p_{13})(p_{14}^{2}p_{16})+(p_{14}^{2}p_{16})^{2} ( p 12 2 p 13 ) 2 + ( p 12 2 p 13 ) ( p 14 2 p 16 ) + ( p 14 2 p 16 ) 2 lies in in d ( G ) \operatorname{in}_d(\mathcal{G}) in d ( G ) . This polynomial factors over C \mathbb{C} C , but neither factor lies in in d ( G ) \operatorname{in}_d(\mathcal{G}) in d ( G ) . In particular, in d ( G ) \operatorname{in}_d(\mathcal{G}) in d ( G ) is not a toric ideal. We conclude the cones of type 21 21 21 are not prime cones of the tropical Grassmannian trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) .
Figure 12. The polyhedral complex of type A 1 \mathfrak{A}_1 A 1 projecting to the triangle of type 21 21 21 .
We compute the initial ideal for every maximal cone C C C of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) , we test whether it is toric, and – if yes – we compute P C P_C P C as described above. We then redo the same computation for the flag variety F \mathcal{F} F . Using this approach, we obtain the following classification result.
Theorem 6.3. In any characteristic, the G 2 \mathrm{G}_2 G 2 -Grassmannian has 135 toric degenerations in 17 orbits. In characteristic 2, there are 12 further toric degenerations in 3 orbits. All 20 orbits are listed in Table 2 . In any characteristic, the G 2 \mathrm{G}_2 G 2 flag variety has 807 toric degenerations in 81 orbits. In characteristic 2, there are 18 further toric degenerations in 2 orbits.
Remark 6.4. A toric degeneration fails to be normal when ∣ P ∩ Z 6 ∣ \lvert P \cap\mathbb{Z}^{6}\rvert ∣ P ∩ Z 6 ∣ is 15 or 16: a lattice point of P P P that is not among the 14 points of the generators gives a monomial in the integral closure of the toric ring that is not in the ring. For instance, at a cell of type 6, the toric ideal contains p 45 p 56 2 − p 57 2 p 67 p_{45}p_{56}^{2}-p_{57}^{2}p_{67} p 45 p 56 2 − p 57 2 p 67 . So u = p 57 p 67 / p 56 u=p_{57}p_{67}/p_{56} u = p 57 p 67 / p 56 is integral, as u 2 = p 45 p 67 u^{2}=p_{45}p_{67} u 2 = p 45 p 67 , but it is not in the toric ring: its lattice point, the midpoint of those of p 45 p_{45} p 45 and p 67 p_{67} p 67 , is not that of any p i j p_{ij} p ij .
Several of the Newton–Okounkov polytopes in Table 2 already appeared in the literature. Type 1 is the Chevalley polytope of Spacek and Wang in [34 ] , Section A.4. Christophersen and Ilten [9 Theorem 3.1] degenerate G \mathcal{G} G to the Stanley–Reisner schemes P ( K ∗ Δ 2 ) \mathbb{P}(K * \Delta_{2}) P ( K ∗ Δ 2 ) of simplicial 2-spheres K K K with 11 vertices, one for each of their term orders [9 Fig. 3] . For two of them K K K is the sphere T 11 T_{11} T 11 in [9 Fig. 1] . Each of these degenerations factors through our toric degenerations among types 7, 10, 11, 12 and 17: for suitable cells of these types. The canonical weight lies in the closed Gröbner cone of their term order. Hence these Newton–Okounkov polytopes admit a regular unimodular triangulation of the form T 11 ∗ Δ 2 T_{11} * \Delta_{2} T 11 ∗ Δ 2 . In particular, they are normal. Cutting by general hyperplanes [9 Corollary 3.3] degenerates the Fano threefold V 18 V_{18} V 18 to P ( T 11 ∗ Δ 0 ) \mathbb{P}(T_{11} * \Delta_{0}) P ( T 11 ∗ Δ 0 ) and the K3 surface of genus 10 to P ( T 11 ) \mathbb{P}(T_{11}) P ( T 11 ) , a connection which is important for our Mukai curves in § 10 . One of these toric degenerations, namely type 7, also appeared in work of Feigin, Fourier and Littelmann [13 Definition 11.5] .
The G2 _2 2 -Orchard
The G2 _2 2 -orchard consists of G2 _2 2 -trees decorated by a fruit, namely the green apple in Figure 11 . This represents the leaf 8 discussed prior to Theorem 2.5 . The distance from the apple to leaf k k k equals d k 8 = − val ( x k ) d_{k8}=-\operatorname{val}(x_k) d k 8 = − val ( x k ) . Every decorated G2 _2 2 -tree is a point of the G2 _2 2 flag Dressian.
Figure 11. Two types of vertices and a 3-cell of trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) .
The green apples explain Theorem 3.3 . They can grow on an edge or on an inner vertex of a G2 _2 2 -tree. The 141 vertices of trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) are 7 star trees with an apple on an edge and 134 decorated G2 _2 2 -trees with an apple on a node in one of the 11 trees in Figure 7 . Two of them are on the left in Figure 11 . The 2714 maximal cells of trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) are the edge-decorated G2 _2 2 -trees of type 1–35 (Figure 8 and Figure 9 ). One of them is shown on the right in Figure 11 .
This gives a way of identifying each decorated G2 _2 2 -tree, but it does not say much about its surroundings in trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) . We next provide charts for guiding the reader through the orchard. This rests on the projection trop ( F ) → trop ( G ) \operatorname{trop}(\mathcal{F}) \to\operatorname{trop}(\mathcal{G}) trop ( F ) → trop ( G ) , which comes from picking the apple.
The fiber over the lineality space of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) is given by the 7 decorated star trees. These appear over each cell of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . Fix a maximal (i.e., 2-dimensional) cell C C C of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . Let F C \mathcal{F}_{C} F C denote the set of all cells in trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) that map to C C C . This is a full-dimensional subcomplex of trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) . We call it the fiber complex over C C C . The tropical G2 _2 2 flag variety trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) is the union of the fiber complexes F C \mathcal{F}_{C} F C , where C C C runs over the maximal cells of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) .
The cells of F C \mathcal{F}_{C} F C are all possible ways of decorating the G2 _2 2 -trees given by the faces of C C C . The fiber complex F C \mathcal{F}_{C} F C has either 10 or 11 maximal cells, of dimension 3. Indeed, an apple can grow from any of the edges of the G2 _2 2 -tree of C C C . The 3-cells are glued along facets. These 2-cells correspond to apples growing on a node. They all biject to C C C . Every fiber complex is attached to the 7 vertices corresponding to the edge-decorated star trees. The remaining vertices are obtained by decorating the nodes of the G 2 G_2 G 2 -trees associated to the vertices of C C C . The following description of the G 2 G_2 G 2 -orchard was obtained by direct computations.
Proposition 7.1. There are 23 combinatorial types of fiber complexes F C \mathcal{F}_C F C over the trees C C C in Figure 8 and Figure 9 . These are listed in Table 3 . Six distinct f f f -vectors are labeled by A , … , F \mathfrak{A}, \ldots, \mathfrak{F} A , … , F .
Table 3. The 23 combinatorial types of fiber complexes F C \mathcal{F}_C F C . These cover trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) .
We now describe the 23 fiber complexes in more detail. Pictures will be helpful to appreciate the beauty of the G 2 G_2 G 2 -orchard. For instance, Figure 12 shows the complex F A 1 \mathcal{F}_{\mathfrak{A}_1} F A 1 . Here A 1 \mathfrak{A}_1 A 1 refers to the first line in Table 3 . The triangle-tetrahedron incidence graph of this 3-dimensional complex is the inner-edge subdivision of the G 2 G_2 G 2 -tree it projects to. We see this in Figure 12 : each edge of the tree gives a tetrahedron, and each inner node a triangle.
The vertices of each fiber complex are labeled by our 7 shapes in Figure 1 for the edge-decorated star trees, and by letters A , B , … A, B, \ldots A , B , … for internal nodes. For instance, in Figure 12 , we have six nodes A , B , C , D , E , F A, B, C, D, E, F A , B , C , D , E , F in the three trees appearing in the triangle of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) underlying A 1 \mathfrak{A}_1 A 1 . The maximal cells of F A 1 \mathcal{F}_{\mathfrak{A}_1} F A 1 are tetrahedra, specified by the 4 vertex labels. Likewise, triangles are given by 3 vertex labels. For instance, the three tetrahedra A B D E ABDE A B D E , B C D E BCDE B C D E and B D E F BDEF B D E F share the triangle B D E BDE B D E . The tetrahedron over an edge e e e has as vertices the images of the endpoints of e e e under the relevant contractions of the tree. For instance, for the tetrahedron A B D E ABDE A B D E , one contracts the edge A B AB A B and retains the nodes D D D and E E E . For each pendant edge, one of the vertices of the tetrahedron is the associated colorful shape.
In Figure 13 , we present the fiber complex F D \mathcal{F}_{\mathcal{D}} F D over the triangle of type 5 in Figure 8 . The fiber complex F D \mathcal{F}_{\mathcal{D}} F D is not a simplicial complex. One of its 3-cells is a triangular prism, labeled by B C D E G H BCDEGH B C D E G H . This appears because the blue square and the blue triangle form a cherry in each of the three vertex G 2 \mathrm{G}_2 G 2 -trees. Hence the corresponding edges B C BC B C , D E DE D E and G H GH G H are consistent. The two facets C D G CDG C D G and B E H BEH B E H of the triangular prism are also in F D \mathcal{F}_{\mathcal{D}} F D .
Figure 13. The polyhedral complex of type D \mathcal{D} D projecting to the triangle of type 5.
Table 3 summarizes the fiber complexes for all types 1–35. Only the first two rows yield simplicial complexes. Pyramids and prisms appear in the other fiber complexes.
We close this section with a discussion of the toric degenerations of the G 2 \mathrm{G}_2 G 2 flag variety F \mathcal{F} F . Each of these induces a toric degeneration of the G 2 \mathrm{G}_2 G 2 -Grassmannian G \mathcal{G} G . We know from Theorem 6.3 that trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) has 807 prime cones. Each of these 807 projects onto one of the 135 prime cones in trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . Hence, the 81 types of toric degenerations of F \mathcal{F} F map to the 17 types of toric degenerations of G \mathcal{G} G . There is an analogous correspondence in characteristic 2.
We examined this map for the types 1–20. For each 3-cell C C C of each fiber complex in Table 3 , we computed the initial ideal in C ( F ) \operatorname{in}_{C}(\mathcal{F}) in C ( F ) , and we recorded whether it is prime over Q \mathbb{Q} Q . The resulting prime cells are marked by green edges in Figure 8 . The green number next to the green apple records the number of green edges. By multiplying each green number with the parenthesized orbit size, and by adding up all 19 products, we arrive at the total count:
807 + 18 = ( 6 × 4 + 6 × 4 + 12 × 6 + ⋯ + 3 × 11 + ⋯ + 12 × 8 ) + ( 6 × 2 + 3 × 2 ) . 807 + 18 = (6 \times4 + 6 \times4 + 12 \times6 + \cdots+ 3 \times11 + \cdots+ 12 \times8) + (6 \times2 + 3 \times2). 807 + 18 = ( 6 × 4 + 6 × 4 + 12 × 6 + ⋯ + 3 × 11 + ⋯ + 12 × 8 ) + ( 6 × 2 + 3 × 2 ) .
The edges that are colored in red in Figure 8 do not give any toric degeneration of F \mathcal{F} F . We note that the characteristic 2 type 19 has no prime cone in its fiber complex inside trop 2 ( F ) \operatorname{trop}_2(\mathcal{F}) trop 2 ( F ) .
Remark 7.2. Our discussion matches what is in the literature on G 2 \mathrm{G}_2 G 2 . The toric degeneration of F \mathcal{F} F due to Spacek and Wang [34 ] is obtained by decorating either of the two circle leaves of type 1. The toric degeneration given by Feigin, Fourier and Littelmann [13 Definition 11.5] is obtained by decorating one of the two edges supporting the cherries in type 7.
Tropicalizing Homogeneous Spaces
Tropicalization depends on coordinate choices. What does this mean for projective varieties that are homogeneous spaces? When the representation of interest is minuscule [7 , 18 , 11 ] , a coordinate system is handed to us. This was the case when the tropical Grassmannian first arose. The authors of [36 ] did not think about choices: they simply used Plücker coordinates; see also [4 ] for other Grassmannians. On the other hand, simple (or, more generally, reductive) groups can be tropicalized without coordinates, through their buildings [10 , 15 ] . The present paper is one of the first to address this dichotomy as a practical issue.
The group G 2 \mathrm{G}_2 G 2 has three natural projective homogeneous spaces, namely the quotients by its two maximal parabolic subgroups P 1 P_1 P 1 and P 2 P_2 P 2 , and by its Borel subgroup B = P 1 ∩ P 2 B = P_1 \cap P_2 B = P 1 ∩ P 2 . Tropicalizing G 2 / P 1 ≅ Q \mathrm{G}_2/P_1 \cong Q G 2 / P 1 ≅ Q is easy; the weights of K 7 \mathbb{K}^7 K 7 have multiplicity one. However, for G 2 / P 2 ≅ G \mathrm{G}_2/P_2 \cong\mathcal{G} G 2 / P 2 ≅ G , the representation has a zero weight plane; so a choice of coordinates seems necessary. Coordinates for both then give coordinates for G 2 / B ≅ F ⊂ Q × G \mathrm{G}_2/B \cong\mathcal{F} \subset Q \times\mathcal{G} G 2 / B ≅ F ⊂ Q × G .
Forests and orchards grew in § 2 –§ 7 . Their trees came from embedding G \mathcal{G} G into Gr ( 2 , 7 ) \operatorname{Gr}(2,7) Gr ( 2 , 7 ) . The group only entered through the linear forms in (4) , the grading of Figure 4 , the star trees that span the lineality space, and the Weyl group D 12 \mathrm{D}_{12} D 12 . We will see that all of these come from the root system R R R in Figure 4 , through a maximal torus T T T of G 2 \mathrm{G}_2 G 2 . In other words, we actually had no choice. A torus acting on a representation whose weight spaces are lines forces coordinates: its weight vectors, which are unique up to scaling. After that forcing, the group still acts, but only through its torus and its Weyl group. These make up the tropical reductive group G 2 t r o p \mathrm{G}_2^{\mathrm{trop}} G 2 trop of [19 §1.5] . This is what tropicalization retains of the action of G 2 \mathrm{G}_2 G 2 on the three spaces. The following theorem is our main result in this section.
Theorem 8.1. Let L ⊃ K \mathbb{L} \supset\mathbb{K} L ⊃ K be algebraically closed fields with a valuation onto R \mathbb{R} R , trivial on K \mathbb{K} K .
(a) The coordinates are forced: x 1 , … , x 7 x_1,\ldots,x_7 x 1 , … , x 7 by T T T on K 7 \mathbb{K}^7 K 7 , and the Plücker coordinates p i j p_{ij} p ij by the centralizer of T T T in GL 7 \operatorname{GL}_7 GL 7 on ⋀ 2 K 7 \bigwedge^2\mathbb{K}^7 ⋀ 2 K 7 . The root system R R R determines Φ \Phi Φ up to rescaling Φ \Phi Φ and the x i x_i x i . Hence trop ( Q ) \operatorname{trop}(Q) trop ( Q ) , trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) and trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) depend only on R R R and on char ( K ) \operatorname{char}(\mathbb{K}) char ( K ) .
(b) The tropical group G 2 t r o p \mathrm{G}_2^{\mathrm{trop}} G 2 trop in (12) acts on trop ( Q ) \operatorname{trop}(Q) trop ( Q ) , trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) and trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) , induced by the elements of G 2 ( L ) \mathrm{G}_2(\mathbb{L}) G 2 ( L ) that act by monomial matrices. On trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) , the translations are the star trees in Figure 3 , and D 12 \mathrm{D}_{12} D 12 permutes the leaves of G 2 \mathrm{G}_2 G 2 -trees as in § 2 .
We start from the root system R R R of G 2 \mathrm{G}_2 G 2 . Let ε 1 , ε 2 , ε 3 \varepsilon_1,\varepsilon_2,\varepsilon_3 ε 1 , ε 2 , ε 3 be the first three columns in (10) , with ε 1 + ε 2 + ε 3 = 0 \varepsilon_1+\varepsilon_2+\varepsilon_3=0 ε 1 + ε 2 + ε 3 = 0 . The six short roots in Figure 4 are ± ε i \pm\varepsilon_i ± ε i , and the six long roots are ε i − ε j \varepsilon_i-\varepsilon_j ε i − ε j for i ≠ j i\ne j i = j . The roots α = ε 1 \alpha=\varepsilon_1 α = ε 1 and β = ε 3 − ε 1 \beta=\varepsilon_3-\varepsilon_1 β = ε 3 − ε 1 form a basis of R R R and span the lattice X ∗ = Z ε 1 + Z ε 2 X^*=\mathbb{Z}\varepsilon_1+\mathbb{Z}\varepsilon_2 X ∗ = Z ε 1 + Z ε 2 . Since the root lattice of R R R is its weight lattice, X ∗ X^* X ∗ is the character lattice of the maximal torus T T T , see [22 §II.1.6] . Each character χ ∈ X ∗ \chi\in X^* χ ∈ X ∗ is a homomorphism T → K ∗ T\to\mathbb{K}^* T → K ∗ . Since ε 1 , ε 2 \varepsilon_1,\varepsilon_2 ε 1 , ε 2 is a basis of X ∗ X^* X ∗ , the map t ↦ ( ε 1 ( t ) , ε 2 ( t ) , ε 3 ( t ) ) t\mapsto(\varepsilon_1(t),\varepsilon_2(t),\varepsilon_3(t)) t ↦ ( ε 1 ( t ) , ε 2 ( t ) , ε 3 ( t )) identifies T T T with the triples ( t 1 , t 2 , t 3 ) ∈ ( K ∗ ) 3 (t_1,t_2,t_3)\in(\mathbb{K}^*)^3 ( t 1 , t 2 , t 3 ) ∈ ( K ∗ ) 3 with t 1 t 2 t 3 = 1 t_1t_2t_3=1 t 1 t 2 t 3 = 1 . The Weyl group D 12 \mathrm{D}_{12} D 12 acts on X ∗ X^* X ∗ by its reflection representation, hence on the cocharacter lattice X ∗ = Hom ( X ∗ , Z ) X_*=\operatorname{Hom}(X^*,\mathbb{Z}) X ∗ = Hom ( X ∗ , Z ) and on the Lie algebra of T T T , which is the Cartan subalgebra h = X ∗ ⊗ K \mathfrak{h}=X_*\otimes\mathbb{K} h = X ∗ ⊗ K . We now consider tropicalization.
Let L \mathbb{L} L be as in Theorem 8.1, with valuation val \operatorname{val} val . For instance, L \mathbb{L} L can be the field K ( ( z R ) ) \mathbb{K}((z^{\mathbb{R}})) K (( z R )) of generalized Puiseux series; cf. [24 §2.6] . Since T T T is split, the map ν : T ( L ) → X ∗ ⊗ R = \nu:T(\mathbb{L})\to X_*\otimes\mathbb{R}= ν : T ( L ) → X ∗ ⊗ R = Hom ( X ∗ , R ) \operatorname{Hom}(X^{*},\mathbb{R}) Hom ( X ∗ , R ) that sends t t t to the homomorphism χ ↦ − val ( χ ( t ) ) \chi\mapsto-\operatorname{val}(\chi(t)) χ ↦ − val ( χ ( t )) is surjective. This identifies trop ( T ) \operatorname{trop}(T) trop ( T ) with X ∗ ⊗ R X_{*} \otimes\mathbb{R} X ∗ ⊗ R . From these data, Gross et al. [19 ] define the tropical reductive group
G 2 trop = ( X ∗ ⊗ R ) ⋊ D 12 = trop ( T ) ⋊ D 12 . (12) G_{2}^{\operatorname{trop}}=(X_{*}\otimes\mathbb{R})\rtimes D_{12}=\operatorname{trop}(T)\rtimes D_{12}. \tag*{(12)} G 2 trop = ( X ∗ ⊗ R ) ⋊ D 12 = trop ( T ) ⋊ D 12 . (12)
The smallest nontrivial representation of G 2 G_{2} G 2 common to every field K \mathbb{K} K is K 7 \mathbb{K}^{7} K 7 , with basis vector e i e_i e i of weight ε i \varepsilon_i ε i . Here, ε i + 3 = − ε i \varepsilon_{i+3}=-\varepsilon_i ε i + 3 = − ε i and ε 7 = 0 \varepsilon_7=0 ε 7 = 0 , and T T T acts by diag ( t 1 , t 2 , t 3 , t 1 − 1 , t 2 − 1 , t 3 − 1 , 1 ) \operatorname{diag}(t_1,t_2,t_3,t_1^{-1},t_2^{-1},t_3^{-1},1) diag ( t 1 , t 2 , t 3 , t 1 − 1 , t 2 − 1 , t 3 − 1 , 1 ) . Thus x i ( t ⋅ v ) = ε i ( t ) x i ( v ) x_i(t\cdot v)=\varepsilon_i(t)x_i(v) x i ( t ⋅ v ) = ε i ( t ) x i ( v ) , which is the grading on K [ x 1 , … , x 7 ] \mathbb{K}[x_1,\ldots,x_7] K [ x 1 , … , x 7 ] in Figure 4 . The weights of K 7 \mathbb{K}^{7} K 7 , the six short roots and 0 0 0 , have multiplicity one, and their labels are unique up to D 12 D_{12} D 12 . Hence the coordinates of K 7 \mathbb{K}^{7} K 7 are forced. Also, the centralizer of T T T in GL 7 \operatorname{GL}_7 GL 7 is the diagonal torus D D D . In this manner, we arrive at our point of departure in § 1 : the space K 7 \mathbb{K}^{7} K 7 with its standard basis, which we implicitly identified with the imaginary part of the octonions. We did this by introducing a T T T -invariant form Φ ∈ ⋀ 3 K 7 \Phi\in\bigwedge^{3}\mathbb{K}^{7} Φ ∈ ⋀ 3 K 7 . Since R R R has rank two, we want T T T to be a maximal torus of the stabilizer of Φ \Phi Φ , and this leaves no choice.
Proposition 8.2. The T T T -invariant subspace of ⋀ 3 K 7 \bigwedge^{3}\mathbb{K}^{7} ⋀ 3 K 7 is spanned by e 123 e_{123} e 123 , e 456 e_{456} e 456 , e 147 e_{147} e 147 , e 257 e_{257} e 257 , e 367 e_{367} e 367 . Its elements with all five coefficients nonzero have T T T as a maximal torus of their stabilizers and form one orbit K ∗ ⋅ D ⋅ Φ \mathbb{K}^{*}\cdot D\cdot\Phi K ∗ ⋅ D ⋅ Φ . The stabilizer of Φ \Phi Φ in D D D is T × μ 3 T\times\mu_3 T × μ 3 , where μ 3 = { ζ I : ζ 3 = 1 } \mu_3=\{\zeta I:\zeta^3=1\} μ 3 = { ζ I : ζ 3 = 1 } .
Proof. The vectors e i j k e_{ijk} e ij k form a basis of weight vectors for ⋀ 3 K 7 \bigwedge^{3}\mathbb{K}^{7} ⋀ 3 K 7 , with weights ε i + ε j + ε k \varepsilon_i+\varepsilon_j+\varepsilon_k ε i + ε j + ε k . The weight is zero, so that e i j k e_{ijk} e ij k is T T T -invariant exactly for the five listed triples. A torus of the stabilizer that contains T T T commutes with T T T , so it lies in the 7-dimensional torus D D D . The five characters by which D D D scales the coefficients c i j k c_{ijk} c ij k are linearly independent, so the stabilizer in D D D of a T T T -invariant form with k k k nonzero c i j k c_{ijk} c ij k has dimension 7 − k 7-k 7 − k , which is 2 2 2 only for k = 5 k=5 k = 5 . The matrix s = diag ( s 1 , … , s 7 ) ∈ D s=\operatorname{diag}(s_1,\ldots,s_7)\in D s = diag ( s 1 , … , s 7 ) ∈ D multiplies the coefficients c i j k c_{ijk} c ij k by s 1 s 2 s 3 s_1s_2s_3 s 1 s 2 s 3 , s 4 s 5 s 6 s_4s_5s_6 s 4 s 5 s 6 and s i s i + 3 s 7 s_is_{i+3}s_7 s i s i + 3 s 7 , for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 . For r ( c ) = c 147 c 257 c 367 / ( c 123 c 456 ) r(c)=c_{147}c_{257}c_{367}/(c_{123}c_{456}) r ( c ) = c 147 c 257 c 367 / ( c 123 c 456 ) , we have r ( s c ) = r ( c ) s 7 3 r(sc)=r(c)s_7^3 r ( sc ) = r ( c ) s 7 3 . Conversely, D D D maps c c c to every c ′ c' c ′ such that r ( c ′ ) / r ( c ) r(c')/r(c) r ( c ′ ) / r ( c ) is a cube. For any λ ∈ K ∗ \lambda\in\mathbb{K}^{*} λ ∈ K ∗ we have r ( λ c ) = λ r ( c ) r(\lambda c)=\lambda r(c) r ( λ c ) = λ r ( c ) , which proves the orbit statement. If s s s fixes the form Φ \Phi Φ in (1), then s 7 3 = 1 s_7^3=1 s 7 3 = 1 , and ζ = s 7 \zeta=s_7 ζ = s 7 satisfies ζ − 1 s ∈ T \zeta^{-1}s\in T ζ − 1 s ∈ T . Conversely, T T T and μ 3 \mu_3 μ 3 fix Φ \Phi Φ .
Proposition 8.2 holds over every field K \mathbb{K} K . In characteristic 3 3 3 , the group μ 3 \mu_3 μ 3 is trivial, so the stabilizer of Φ \Phi Φ in D D D is T T T .
Proof of Theorem 8.1 (a). We saw that T T T forces x 1 , … , x 7 x_1,\ldots,x_7 x 1 , … , x 7 and has centralizer D D D in GL 7 \operatorname{GL}_7 GL 7 . The 21 weights of D D D on ⋀ 2 K 7 \bigwedge^{2}\mathbb{K}^{7} ⋀ 2 K 7 are distinct, so D D D forces the Plücker coordinates p i j = x i ∧ x j p_{ij}=x_i\wedge x_j p ij = x i ∧ x j , and F ⊂ Q × G \mathcal{F}\subset\mathcal{Q}\times\mathcal{G} F ⊂ Q × G inherits both. Our varieties are built from Φ \Phi Φ alone: Q \mathcal{Q} Q is cut out by the 6 × 6 6\times6 6 × 6 Pfaffians of (3) , and L = P ( E ) \mathcal{L}=\mathbb{P}(E) L = P ( E ) , where E E E is the kernel of the contraction ⋀ 2 K 7 → ( K 7 ) ∨ \bigwedge^{2}\mathbb{K}^{7}\to(\mathbb{K}^{7})^{\vee} ⋀ 2 K 7 → ( K 7 ) ∨ , ω ↦ Φ ( ω , ⋅ ) \omega\mapsto\Phi(\omega,\cdot) ω ↦ Φ ( ω , ⋅ ) , whose coordinates are the linear forms in (4) . Hence Q , G = Gr ( 2 , 7 ) ∩ L \mathcal{Q},\mathcal{G}=\operatorname{Gr}(2,7)\cap\mathcal{L} Q , G = Gr ( 2 , 7 ) ∩ L and F \mathcal{F} F are G 2 G_2 G 2 -stable, and replacing Φ \Phi Φ by λ s ⋅ Φ \lambda s\cdot\Phi λ s ⋅ Φ moves them by s s s . By Proposition 8.2 , any other choice of Φ \Phi Φ and of the scalings of the x i x_i x i is of this form, with λ ∈ K ∗ \lambda\in\mathbb{K}^{*} λ ∈ K ∗ and s ∈ D ( K ) s\in D(\mathbb{K}) s ∈ D ( K ) . So is any other identification of R R R with Figure 4 : it relabels the x i x_i x i by some w ∈ D 12 w\in D_{12} w ∈ D 12 , which permutes the five e i j k e_{ijk} e ij k of weight zero up to sign. Either way, the tropicalizations are translated by the valuation of s s s , which vanishes as val \operatorname{val} val is trivial on K \mathbb{K} K . Finally, only char ( K ) \operatorname{char}(\mathbb{K}) char ( K ) matters: our varieties are defined over the prime field, and initial ideals commute with field extension. ▫-offsetof
Remark 8.3. Fix the basis α = ε 1 \alpha= \varepsilon_1 α = ε 1 and β = ε 3 − ε 1 \beta= \varepsilon_3 - \varepsilon_1 β = ε 3 − ε 1 of R R R . The highest weights of K 7 \mathbb{K}^7 K 7 and ⋀ 2 K 7 \bigwedge^2\mathbb{K}^7 ⋀ 2 K 7 are 2 α + β = ε 5 2\alpha+ \beta= \varepsilon_5 2 α + β = ε 5 and 3 α + 2 β = ε 3 + ε 5 3\alpha+ 2\beta= \varepsilon_3 + \varepsilon_5 3 α + 2 β = ε 3 + ε 5 , with weight lines K e 5 \mathbb{K}e_5 K e 5 and K e 3 ∧ e 5 \mathbb{K}e_3 \wedge e_5 K e 3 ∧ e 5 . Their stabilizers are the parabolic subgroups P 1 P_1 P 1 and P 2 P_2 P 2 of G 2 G_2 G 2 . Hence Q = G 2 / P 1 Q = G_2/P_1 Q = G 2 / P 1 and G = G 2 / P 2 G = G_2/P_2 G = G 2 / P 2 are the orbits of these two lines, and F = G 2 / B F = G_2/B F = G 2 / B is the orbit of the flag ( [ e 5 ] , [ e 3 ∧ e 5 ] ) ([e_5], [e_3 \wedge e_5]) ([ e 5 ] , [ e 3 ∧ e 5 ]) .
Part (b) of Theorem 8.1 rests on a lemma that holds for any split reductive group G G G over K \mathbb{K} K , with maximal torus T T T , character lattice X ∗ X^* X ∗ and Weyl group W = N ( T ) / T W = N(T)/T W = N ( T ) / T . Here N ( T ) N(T) N ( T ) is the normalizer of T T T in G G G . We follow [22 §II.1.1–1.4] . Since G G G is split, every w ∈ W w \in W w ∈ W lifts to some n w ∈ N ( T ) ( K ) n_w \in N(T)(\mathbb{K}) n w ∈ N ( T ) ( K ) , unique up to T ( K ) T(\mathbb{K}) T ( K ) , and every element of N ( T ) ( L ) N(T)(\mathbb{L}) N ( T ) ( L ) is t n w t n_w t n w for unique t ∈ T ( L ) t \in T(\mathbb{L}) t ∈ T ( L ) and w ∈ W w \in W w ∈ W . Conjugation by n w n_w n w is the action of w w w on T T T , which induces actions of W W W on X ∗ X^* X ∗ and on trop ( T ) = Hom ( X ∗ , R ) \operatorname{trop}(T) = \operatorname{Hom}(X^*, \mathbb{R}) trop ( T ) = Hom ( X ∗ , R ) . As for G 2 G_2 G 2 , we have the map ν : T ( L ) → trop ( T ) \nu: T(\mathbb{L}) \to\operatorname{trop}(T) ν : T ( L ) → trop ( T ) . It is surjective, vanishes on T ( K ) T(\mathbb{K}) T ( K ) , and satisfies ν ( n w t n w − 1 ) = w ( ν ( t ) ) \nu(n_wtn_w^{-1}) = w(\nu(t)) ν ( n w t n w − 1 ) = w ( ν ( t )) . We set G trop = trop ( T ) ⋊ W G^{\operatorname{trop}} = \operatorname{trop}(T) \rtimes W G trop = trop ( T ) ⋊ W and trop ( t n w ) = ( ν ( t ) , w ) \operatorname{trop}(t n_w) = (\nu(t), w) trop ( t n w ) = ( ν ( t ) , w ) , which does not depend on the choice of n w n_w n w .
Lemma 8.4. The map trop : N ( T ) ( L ) → G trop \operatorname{trop}: N(T)(\mathbb{L}) \to G^{\operatorname{trop}} trop : N ( T ) ( L ) → G trop is a surjective group homomorphism. Let G → GL m ( K ) G \to\operatorname{GL}_m(\mathbb{K}) G → GL m ( K ) be a representation in which N ( T ) N(T) N ( T ) acts by monomial matrices , so that T T T scales the i i i th coordinate by a character χ i ∈ X ∗ \chi_i \in X^* χ i ∈ X ∗ . Then, G trop G^{\operatorname{trop}} G trop acts on R m / 1 \mathbb{R}^m/\mathbf{1} R m / 1 by tropical monomial matrices : the element ( u , w ) ∈ G trop (u,w) \in G^{\operatorname{trop}} ( u , w ) ∈ G trop permutes the coordinates as n w n_w n w does and then adds ( u ( χ 1 ) , … , u ( χ m ) ) (u(\chi_1),\ldots,u(\chi_m)) ( u ( χ 1 ) , … , u ( χ m )) . With coordinatewise − val -\operatorname{val} − val , this square commutes:
N ( T ) ( L ) × ( L ∗ ) m / L ∗ → ( L ∗ ) m / L ∗ trop × ( − val ) ↓ ↓ − val G trop × ( R m / 1 ) → R m / 1. \begin{CD}
N(T)(\mathbb{L}) \times(\mathbb{L}^*)^m/\mathbb{L}^* @>>> (\mathbb{L}^*)^m/\mathbb{L}^* \\
@V{\operatorname{trop}\times(-\operatorname{val})}VV @VV{-\operatorname{val}}V \\
G^{\operatorname{trop}} \times(\mathbb{R}^m/\mathbf{1}) @>>> \mathbb{R}^m/\mathbf{1}.
\end{CD} N ( T ) ( L ) × ( L ∗ ) m / L ∗ trop × ( − val ) ↓ ⏐ G trop × ( R m / 1 ) ( L ∗ ) m / L ∗ ↓ ⏐ − val R m / 1 . Hence, if a subvariety Y Y Y of P ( K m ) \mathbb{P}(\mathbb{K}^m) P ( K m ) is N ( T ) N(T) N ( T ) -stable, then trop ( Y ) \operatorname{trop}(Y) trop ( Y ) is G trop G^{\operatorname{trop}} G trop -stable.
Proof. The map trop \operatorname{trop} trop is surjective since ν \nu ν is. For w , w ′ ∈ W w,w' \in W w , w ′ ∈ W we write n w n w ′ = s n w w ′ n_wn_{w'} = s n_{ww'} n w n w ′ = s n w w ′ with s ∈ T ( K ) s \in T(\mathbb{K}) s ∈ T ( K ) . Then ( t n w ) ( t ′ n w ′ ) (t n_w)(t'n_{w'}) ( t n w ) ( t ′ n w ′ ) equals t ( n w t ′ n w − 1 ) s n w w ′ t(n_wt'n_w^{-1})s n_{ww'} t ( n w t ′ n w − 1 ) s n w w ′ , and trop \operatorname{trop} trop maps it to
( ν ( t ) + ν ( n w t ′ n w − 1 ) + ν ( s ) , w w ′ ) = ( ν ( t ) + w ( ν ( t ′ ) ) , w w ′ ) = ( ν ( t ) , w ) ( ν ( t ′ ) , w ′ ) . \left(\nu(t) + \nu(n_wt'n_w^{-1}) + \nu(s), ww'\right) = \left(\nu(t) + w(\nu(t')), ww'\right) = (\nu(t),w)(\nu(t'),w'). ( ν ( t ) + ν ( n w t ′ n w − 1 ) + ν ( s ) , w w ′ ) = ( ν ( t ) + w ( ν ( t ′ )) , w w ′ ) = ( ν ( t ) , w ) ( ν ( t ′ ) , w ′ ) .
By assumption, T T T consists of diagonal matrices. The matrix n w n_w n w has entries in K \mathbb{K} K and maps the j j j th basis vector to a multiple of the σ ( j ) \sigma(j) σ ( j ) th, for a permutation σ \sigma σ of [ m ] [m] [ m ] . For n = t n w n = t n_w n = t n w and v ∈ ( L ∗ ) m v \in(\mathbb{L}^*)^m v ∈ ( L ∗ ) m , the coordinate of n ⋅ v n \cdot v n ⋅ v at σ ( j ) \sigma(j) σ ( j ) is χ σ ( j ) ( t ) v j \chi_{\sigma(j)}(t)v_j χ σ ( j ) ( t ) v j times a scalar in K ∗ \mathbb{K}^* K ∗ , so − val -\operatorname{val} − val maps it to u ( χ σ ( j ) ) − val ( v j ) u(\chi_{\sigma(j)}) - \operatorname{val}(v_j) u ( χ σ ( j ) ) − val ( v j ) , where u = ν ( t ) u = \nu(t) u = ν ( t ) . This is the image of − val ( v ) -\operatorname{val}(v) − val ( v ) under ( u , w ) = trop ( n ) (u,w) = \operatorname{trop}(n) ( u , w ) = trop ( n ) . This shows that the square commutes. Its vertical maps are surjective and trop \operatorname{trop} trop is a homomorphism, so our rule is an action.
For the last claim, let g ∈ G trop g \in G^{\operatorname{trop}} g ∈ G trop and d ∈ trop ( Y ) d \in\operatorname{trop}(Y) d ∈ trop ( Y ) . Since val \operatorname{val} val surjects onto R \mathbb{R} R , we have d = − val ( y ) d = -\operatorname{val}(y) d = − val ( y ) for some y ∈ Y ( L ) y \in Y(\mathbb{L}) y ∈ Y ( L ) with nonzero coordinates [29 Theorem 3.2.3] , and g = trop ( n ) g = \operatorname{trop}(n) g = trop ( n ) for some n ∈ N ( T ) ( L ) n \in N(T)(\mathbb{L}) n ∈ N ( T ) ( L ) . Then n ⋅ y ∈ Y ( L ) n \cdot y \in Y(\mathbb{L}) n ⋅ y ∈ Y ( L ) , and g g g maps d d d to − val ( n ⋅ y ) ∈ trop ( Y ) -\operatorname{val}(n \cdot y) \in\operatorname{trop}(Y) − val ( n ⋅ y ) ∈ trop ( Y ) .
Proof of Theorem 8.1 (b). Let N = N ( T ) N = N(T) N = N ( T ) be the normalizer of T T T in G 2 G_2 G 2 . By [3 (2.11) and (3.4)] , the Weyl group of T T T is N / T ≅ D 12 N/T \cong D_{12} N / T ≅ D 12 . The elements of N N N permute the weight lines of T T T , so they are monomial matrices. Conversely, a monomial matrix in G 2 G_2 G 2 normalizes G 2 ∩ D G_2 \cap D G 2 ∩ D . This group is T T T : by Proposition 8.2 , the diagonal matrices fixing Φ \Phi Φ form T × μ 3 T \times\mu_3 T × μ 3 , and det ( ζ I ) = ζ 7 = ζ \det(\zeta I)=\zeta^{7}=\zeta det ( ζ I ) = ζ 7 = ζ . Hence N ( L ) N(\mathcal{L}) N ( L ) consists of the elements of G 2 ( L ) G_{2}(\mathbb{L}) G 2 ( L ) that act by monomial matrices. Let ρ = ( 1 5 3 4 2 6 ) \rho=(1\,5\,3\,4\,2\,6) ρ = ( 1 5 3 4 2 6 ) and τ = ( 2 3 ) ( 5 6 ) \tau=(2\,3)(5\,6) τ = ( 2 3 ) ( 5 6 ) , with signed permutation matrices
n ρ : e i ↦ e ρ ( i ) for i ≤ 6 , e 7 ↦ − e 7 , n τ : e i ↦ − e τ ( i ) for i ≤ 6 , e 7 ↦ e 7 . n_{\rho}: e_{i}\mapsto e_{\rho(i)}\text{ for }i\leq6,\quad e_{7}\mapsto-e_{7},\qquad n_{\tau}: e_{i}\mapsto-e_{\tau(i)}\text{ for }i\leq6,\quad e_{7}\mapsto e_{7}. n ρ : e i ↦ e ρ ( i ) for i ≤ 6 , e 7 ↦ − e 7 , n τ : e i ↦ − e τ ( i ) for i ≤ 6 , e 7 ↦ e 7 .
Both fix Φ \Phi Φ and have determinant 1 1 1 , so they lie in N N N . Sending a monomial matrix to its permutation of [ 7 ] [7] [ 7 ] is a homomorphism N → S 7 N\to S_{7} N → S 7 . Its kernel contains T T T , so its image has at most ∣ N / T ∣ = 12 |N/T|=12 ∣ N / T ∣ = 12 elements, and it contains ρ \rho ρ and τ \tau τ . Hence it identifies N / T N/T N / T with the group D 12 ⊂ S 7 D_{12}\subset S_{7} D 12 ⊂ S 7 of § 2 . Products n w n_{w} n w of n ρ n_{\rho} n ρ and n τ n_{\tau} n τ lift all w ∈ D 12 w\in D_{12} w ∈ D 12 . They map x i x_{i} x i to ± x w ( i ) \pm x_{w(i)} ± x w ( i ) and p i j p_{ij} p ij to ± p w ( i ) w ( j ) \pm p_{w(i)w(j)} ± p w ( i ) w ( j ) . The signs matter if char ( K ) ≠ 2 \operatorname{char}(\mathbb{K})\neq2 char ( K ) = 2 : the permutation matrix of τ \tau τ maps p 23 + p 47 p_{23}+p_{47} p 23 + p 47 to p 47 − p 23 p_{47}-p_{23} p 47 − p 23 , so it does not preserve the linear space L \mathcal{L} L in (4) .
We now apply Lemma 8.4 to Q ⊂ P ( K 7 ) \mathcal{Q}\subset\mathbb{P}(\mathbb{K}^{7}) Q ⊂ P ( K 7 ) and G ⊂ P ( ⋀ 2 K 7 ) \mathcal{G}\subset\mathbb{P}(\bigwedge^{2}\mathbb{K}^{7}) G ⊂ P ( ⋀ 2 K 7 ) , with G 2 t r o p = G 2 t r o p G_{2}^{\mathrm{trop}}=G_{2}^{\mathrm{trop}} G 2 trop = G 2 trop as in (12) . Its proof applies verbatim to F \mathcal{F} F in their product, with one square for each factor. The group N N N acts on both spaces by monomial matrices, and T T T scales x i x_{i} x i and p i j p_{ij} p ij by ε i \varepsilon_{i} ε i and ε i + ε j \varepsilon_{i}+\varepsilon_{j} ε i + ε j . The three varieties are G 2 G_{2} G 2 -stable by the proof of (a). So G 2 t r o p G_{2}^{\mathrm{trop}} G 2 trop acts on trop ( Q ) \operatorname{trop}(\mathcal{Q}) trop ( Q ) , trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) and trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) , induced by N ( L ) N(\mathcal{L}) N ( L ) .
Explicitly, we identify u ∈ trop ( T ) u\in\operatorname{trop}(T) u ∈ trop ( T ) with the vector ( u ( ε i ) ) i = ( a , − a , 0 ) ∈ R 7 (u(\varepsilon_{i}))_{i}=(a,-a,0)\in\mathbb{R}^{7} ( u ( ε i ) ) i = ( a , − a , 0 ) ∈ R 7 , where a 1 + a 2 + a 3 = 0 a_{1}+a_{2}+a_{3}=0 a 1 + a 2 + a 3 = 0 . Then ( u , w ) (u,w) ( u , w ) acts on trop ( Q ) \operatorname{trop}(\mathcal{Q}) trop ( Q ) by d ↦ w ( d ) + u d\mapsto w(d)+u d ↦ w ( d ) + u , and on trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) by d ↦ w ( d ) + ( u i + u j ) i j d\mapsto w(d)+(u_{i}+u_{j})_{ij} d ↦ w ( d ) + ( u i + u j ) ij , where w ( d ) w ( i ) w ( j ) = d i j w(d)_{w(i)w(j)}=d_{ij} w ( d ) w ( i ) w ( j ) = d ij . On trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) , both happen at once. The translation adds the star tree with edge lengths u u u . These u u u span the lineality space, by Lemma 2.2 : if d d d and − d -d − d satisfy the four-point condition (8) , then d d d is a star tree, so the lineality of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) lies in that of the Dressian. The permutation relabels the leaves by w w w , as D 12 D_{12} D 12 does in § 2 . The same holds for the two Dressians: N N N preserves Gr ( 2 , 7 ) \operatorname{Gr}(2,7) Gr ( 2 , 7 ) and L \mathcal{L} L , and it permutes the 78 tropical polynomials defining the flag Dressian. Finally, the n w n_{w} n w have entries 0 , ± 1 0,\pm1 0 , ± 1 , so all of this holds over every field, including char ( K ) = 2 \operatorname{char}(\mathbb{K})=2 char ( K ) = 2 of § 4 .
We now discuss the zero weight plane E 0 E_{0} E 0 of the representation E E E , where L = P ( E ) \mathcal{L}=\mathbb{P}(E) L = P ( E ) ; cf. proof of Theorem 8.1 (a). The linear forms in (4) have coefficients ± 1 \pm1 ± 1 and disjoint variables, so E E E has dimension 14 over every field. By Figure 4 , the weights ε i + ε j \varepsilon_{i}+\varepsilon_{j} ε i + ε j of the p i j p_{ij} p ij are the long roots once, for the hexagon edges, the short roots twice, such as ε 1 \varepsilon_{1} ε 1 for { p 17 , p 56 } \{p_{17},p_{56}\} { p 17 , p 56 } , and zero three times, for { p 14 , p 25 , p 36 } \{p_{14},p_{25},p_{36}\} { p 14 , p 25 , p 36 } . Removing the weights of ( K 7 ) ∗ (\mathbb{K}^{7})^{*} ( K 7 ) ∗ , each root is a weight of E E E once and zero twice. So the root system chooses all coordinates of nonzero weight: T T T alone forces the 18 coordinates p i j p_{ij} p ij whose weight is a root, where the two of a short root agree up to sign. It is blind only on E 0 E_{0} E 0 , where D D D supplies p 14 , p 25 , p 36 p_{14},p_{25},p_{36} p 14 , p 25 , p 36 , with sum zero. Thus p i , i + 3 ↦ ε i p_{i,i+3}\mapsto\varepsilon_{i} p i , i + 3 ↦ ε i identifies E 0 ∨ E_{0}^{\vee} E 0 ∨ with V ∗ ⊗ K = K ∨ V^{*}\otimes\mathbb{K}=\mathbb{K}^{\vee} V ∗ ⊗ K = K ∨ , D 12 D_{12} D 12 -equivariantly. Each line in X Q ∗ = X Q ∗ ⊗ Q X_{Q}^{*}=X_{Q}^{*}\otimes\mathbb{Q} X Q ∗ = X Q ∗ ⊗ Q thus gives a coordinate on E 0 E_{0} E 0 . Its primitive vector, up to sign. Let A \mathcal{A} A be a set of lines whose coordinates span E 0 ∨ E_{0}^{\vee} E 0 ∨ . With one Plücker coordinate per root, it defines a tropical variety trop A ( G ) \operatorname{trop}_{\mathcal{A}}(\mathcal{G}) trop A ( G ) . One is tempted to choose a basis, for instance A = { Q ε 1 , Q ε 2 } \mathcal{A}=\{\mathbb{Q}\varepsilon_{1},\mathbb{Q}\varepsilon_{2}\} A = { Q ε 1 , Q ε 2 } . This forgets d 36 d_{36} d 36 , and the projection trop ( G ) → trop A ( G ) \operatorname{trop}(\mathcal{G})\to\operatorname{trop}_{\mathcal{A}}(\mathcal{G}) trop ( G ) → trop A ( G ) of [29 Corollary 3.2.13] keeps 227 of the 248 maximal cells in Theorem 3.1 and contracts the others. Worse, the Weyl group action is lost since no basis of E 0 ∨ E_{0}^{\vee} E 0 ∨ is D 12 D_{12} D 12 -stable. This destroys the beauty of the forest and the orchard.
The D 12 D_{12} D 12 -orbits of lines in X Q ∗ X_{Q}^{*} X Q ∗ have three or six elements. The smallest D 12 D_{12} D 12 -stable choices are the short tripod A s h o r t = { Q ε 1 , Q ε 2 , Q ε 3 } \mathcal{A}_{\mathrm{short}}=\{\mathbb{Q}\varepsilon_{1},\mathbb{Q}\varepsilon_{2},\mathbb{Q}\varepsilon_{3}\} A short = { Q ε 1 , Q ε 2 , Q ε 3 } , the long tripod A l o n g = { Q ( ε 2 − ε 3 ) , Q ( ε 3 − ε 1 ) , Q ( ε 1 − ε 2 ) } \mathcal{A}_{\mathrm{long}}=\{\mathbb{Q}(\varepsilon_{2}-\varepsilon_{3}),\mathbb{Q}(\varepsilon_{3}-\varepsilon_{1}),\mathbb{Q}(\varepsilon_{1}-\varepsilon_{2})\} A long = { Q ( ε 2 − ε 3 ) , Q ( ε 3 − ε 1 ) , Q ( ε 1 − ε 2 )} , and their union. The short tripod gives the coordinates p 14 , p 25 , p 36 p_{14},p_{25},p_{36} p 14 , p 25 , p 36 forced by D D D , so trop A s h o r t ( G ) \operatorname{trop}_{\mathcal{A}_{\mathrm{short}}}(\mathcal{G}) trop A short ( G ) is trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) without the duplicate short root coordinates. The long tripod replaces them by p 14 − p 25 p_{14}-p_{25} p 14 − p 25 , p 14 − p 36 p_{14}-p_{36} p 14 − p 36 , p 25 − p 36 p_{25}-p_{36} p 25 − p 36 , and trop A l o n g ( G ) \operatorname{trop}_{A_{\mathrm{long}}}(\mathcal{G}) trop A long ( G ) has f f f -vector ( 42 , 211 , 275 ) (42,211,275) ( 42 , 211 , 275 ) instead of ( 51 , 241 , 248 ) (51,241,248) ( 51 , 241 , 248 ) . The union projects onto both: it adds − val ( p i , i + 3 − p j , j + 3 ) -\operatorname{val}(p_{i,i+3}-p_{j,j+3}) − val ( p i , i + 3 − p j , j + 3 ) , which is max ( d i , i + 3 , d j , j + 3 ) \max(d_{i,i+3},d_{j,j+3}) max ( d i , i + 3 , d j , j + 3 ) unless d i , i + 3 = d j , j + 3 d_{i,i+3}=d_{j,j+3} d i , i + 3 = d j , j + 3 . Neither the long tripod nor the union gives a space of trees, because these differences are not Plücker coordinates.
The representation E E E is a familiar one. If char ( K ) ≠ 3 \operatorname{char}(K)\ne3 char ( K ) = 3 , then E E E is the adjoint representation g 2 \mathfrak{g}_2 g 2 and E 0 E_0 E 0 is the Cartan subalgebra h \mathfrak{h} h . In characteristic 0, this is [14 §22.3] . In general, the highest root θ \theta θ is the highest weight of E E E , and a vector of weight θ \theta θ generates a subrepresentation that maps onto the simple module of highest weight θ \theta θ [22 §II.2.13] . This module has dimension 14 [20 Part I, Corollary 2.7a] , so it is isomorphic to both E E E and g 2 \mathfrak{g}_2 g 2 . In characteristic 3, the short roots generate a 7-dimensional ideal of g 2 \mathfrak{g}_2 g 2 [20 Part I, Table 1] , so the simple module has dimension at most 7, and we leave open whether E ≅ g 2 E\cong\mathfrak{g}_2 E ≅ g 2 . Our results do not depend on this identification. Characteristic 3 also affects the long tripod: it does not span E 0 E_0 E 0 , since the long roots span a sublattice of index 3. In characteristic 2, every primitive vector of X ∗ X^* X ∗ is congruent to some ε i \varepsilon_i ε i modulo 2 X ∗ 2X^* 2 X ∗ , so every A A A yields coordinates among p 14 , p 25 , p 36 p_{14},p_{25},p_{36} p 14 , p 25 , p 36 . Hence the discrepancy between trop 2 ( G ) \operatorname{trop}_2(\mathcal{G}) trop 2 ( G ) and trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) in Theorem 4.1 is an intrinsic feature of the group G 2 \mathrm{G}_2 G 2 . Characteristic-dependence did not come from a choice.
Hexagons
Section 8 explained how the group G 2 \mathrm{G}_2 G 2 determines our coordinates. In hindsight, this is a story about buildings [38 , 39 ] . All relevant buildings have already appeared. Linear subspaces of P 6 \mathbb{P}^6 P 6 over K K K form the spherical building of G L 7 ( K ) \mathrm{GL}_7(K) GL 7 ( K ) , of type A 6 A_6 A 6 . Its apartments correspond to the maximal split tori of G L 7 \mathrm{GL}_7 GL 7 , and the diagonal torus D D D gives the apartment spanned by the seven coordinate points. Inside P 6 \mathbb{P}^6 P 6 , the points in Q \mathcal{Q} Q and the lines in G \mathcal{G} G , with incidences given by F \mathcal{F} F , form the spherical building H ( K ) \mathcal{H}(K) H ( K ) of G 2 ( K ) \mathrm{G}_2(K) G 2 ( K ) . Its apartments correspond to the maximal split tori of G 2 \mathrm{G}_2 G 2 . The hexagon (6) is the apartment of T T T : its points are the weight lines K e 1 , … , K e 6 Ke_1,\ldots,Ke_6 K e 1 , … , K e 6 of T T T in Q \mathcal{Q} Q , and its lines are coordinate lines, so it lies in the apartment of D D D . Hence the coordinates forced in Theorem 8.1 (a) amount to the choice of an apartment.
Over the field L L L , the valuation adds the affine buildings of P G L 7 ( L ) \mathrm{PGL}_7(L) PGL 7 ( L ) and of G 2 ( L ) \mathrm{G}_2(L) G 2 ( L ) [6 ] . Their standard apartments are the tropical torus R 7 / 1 \mathbb{R}^7/\mathbf{1} R 7 / 1 [25 ] and trop ( T ) = X ∗ ⊗ R \operatorname{trop}(T)=X_*\otimes\mathbb{R} trop ( T ) = X ∗ ⊗ R [15 ] , which is the lineality space of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . We leave the interplay between tropicalized spherical buildings and affine buildings, for G 2 \mathrm{G}_2 G 2 and for other split reductive groups, to future work. Here we single out the spherical buildings H ( K ) \mathcal{H}(K) H ( K ) and H ( L ) \mathcal{H}(L) H ( L ) . We tropicalize their apartments.
The incidence graph of H ( K ) \mathcal{H}(K) H ( K ) is the bipartite graph whose nodes are the points in Q \mathcal{Q} Q and the lines in G \mathcal{G} G , and whose edges are given by the flags in F \mathcal{F} F . It has diameter 6 and girth 12: it is the split Cayley generalized hexagon over the field K K K ; cf. [3 §5] and [39 Definition 2.4.9] . The apartments of H ( K ) \mathcal{H}(K) H ( K ) are the 12-cycles in this graph, which we call hexagons ; an example is (6). Their Zariski closure is an irreducible subvariety H \mathcal{H} H of dimension 12 in F 6 ⊂ Q 6 × G 6 \mathcal{F}^6\subset\mathcal{Q}^6\times\mathcal{G}^6 F 6 ⊂ Q 6 × G 6 .
The prime ideal of H \mathcal{H} H lives in a polynomial ring with 6 × 7 + 6 × 21 = 168 6\times7+6\times21=168 6 × 7 + 6 × 21 = 168 variables. We were not able to compute its generators. Among them are the ideal generators for Q \mathcal{Q} Q and G \mathcal{G} G , together with 12 incidence relations given by F \mathcal{F} F as displayed in Figure 14 . More precisely, all twelve pairs ( x ( i ) , p ( i − 1 ) ) (x^{(i)},p^{(i-1)}) ( x ( i ) , p ( i − 1 ) ) and ( x ( i ) , p ( i + 1 ) ) (x^{(i)},p^{(i+1)}) ( x ( i ) , p ( i + 1 ) ) sit in F \mathcal{F} F , where labels are taken modulo 6.
Figure 14. A hexagon for G 2 G_2 G 2 consists of six points in Q \mathcal{Q} Q and six lines in G \mathcal{G} G .
A hexagon is non-degenerate when all its points and lines are distinct. The group G 2 \mathrm{G}_2 G 2 acts transitively on Q \mathcal{Q} Q , G \mathcal{G} G , F \mathcal{F} F , and on the non-degenerate part of H \mathcal{H} H [39 Theorem 5.2.9] . In particular, every non-degenerate hexagon is obtained as g ⋅ ( 6 ) g \cdot(6) g ⋅ ( 6 ) for some g ∈ G 2 g \in G_2 g ∈ G 2 , with pointwise stabilizers g T g − 1 gTg^{-1} g T g − 1 [3 (5.8), (5.9) and §5] . By Theorem 8.1 (a), the torus g T g − 1 gTg^{-1} g T g − 1 forces the coordinates moved by g ∈ G 2 ( L ) g \in G_2(L) g ∈ G 2 ( L ) : in these coordinates a G 2 G_2 G 2 -stable variety Y Y Y tropicalizes as g − 1 Y = Y g^{-1}Y=Y g − 1 Y = Y does. Hence, the tropical varieties of § 2 –§ 7 do not depend on the apartment: every non-degenerate hexagon gives the same trop ( Q ) \operatorname{trop}(\mathcal{Q}) trop ( Q ) , trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) and trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) .
Consider the following projection, mapping a hexagon to its pair of opposite flags [39 ] :
H ⟶ F 2 , ( ( x ( 1 ) , x ( 2 ) , … , x ( 6 ) ) , ( p ( 12 ) , p ( 23 ) , … , p ( 61 ) ) ) ⟼ ( ( x ( 1 ) , p ( 12 ) ) , ( x ( 4 ) , p ( 45 ) ) ) . \mathcal{H} \longrightarrow\mathcal{F}^{2},\quad((x^{(1)},x^{(2)},\ldots,x^{(6)}),(p^{(12)},p^{(23)},\ldots,p^{(61)})) \longmapsto((x^{(1)},p^{(12)}),(x^{(4)},p^{(45)})). H ⟶ F 2 , (( x ( 1 ) , x ( 2 ) , … , x ( 6 ) ) , ( p ( 12 ) , p ( 23 ) , … , p ( 61 ) )) ⟼ (( x ( 1 ) , p ( 12 ) ) , ( x ( 4 ) , p ( 45 ) )) .
In the case of the split Cayley hexagon, by Lemma 9.2 , this map is birational with inverse
ψ : F 2 ⇢ H . (13) \psi:\mathcal{F}^{2}\dashrightarrow\mathcal{H}. \tag*{(13)} ψ : F 2 ⇢ H . (13)
This is a parametric representation of the hexagon variety H \mathcal{H} H , to be explained further below. We here tropicalize the map ψ \psi ψ . Our main result is the following correspondence theorem for hexagons of G 2 G_2 G 2 . For the sake of brevity, we assume char ( K ) = 0 \operatorname{char}(\mathbb{K})=0 char ( K ) = 0 for the rest of this section.
Theorem 9.1. Fix fields K ⊂ L \mathbb{K}\subset\mathbb{L} K ⊂ L of characteristic zero, where L \mathbb{L} L has a valuation as before. We give a piecewise-linear map trop ( ψ ) \operatorname{trop}(\psi) trop ( ψ ) from the dominant region of trop ( F ) 2 \operatorname{trop}(\mathcal{F})^{2} trop ( F ) 2 to trop ( H ) \operatorname{trop}(\mathcal{H}) trop ( H ) which yields − val ∘ ψ -\operatorname{val}\circ\psi − val ∘ ψ on points of F ( L ) 2 \mathcal{F}(\mathbb{L})^{2} F ( L ) 2 that are generic in their fiber under the valuation.
We now explain the statement. A maximal cone in trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) is dominant if its images in both trop ( Q ) \operatorname{trop}(\mathcal{Q}) trop ( Q ) and trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) are full-dimensional. Among the 2714 maximal cones, precisely 1311 are dominant. Their union is the dominant region D ⊂ trop ( F ) \mathcal{D}\subset\operatorname{trop}(\mathcal{F}) D ⊂ trop ( F ) , and D 2 \mathcal{D}^{2} D 2 is the dominant region of trop ( F ) 2 \operatorname{trop}(\mathcal{F})^{2} trop ( F ) 2 . Each maximal cone of D 2 \mathcal{D}^{2} D 2 is divided into smaller cones on which trop ( ψ ) \operatorname{trop}(\psi) trop ( ψ ) is linear. For v v v in the relative interior of such a cone, − val ( ψ ( z ) ) = trop ( ψ ) ( v ) -\operatorname{val}(\psi(z))=\operatorname{trop}(\psi)(v) − val ( ψ ( z )) = trop ( ψ ) ( v ) holds for generic points z ∈ F 2 z\in\mathcal{F}^{2} z ∈ F 2 in the fiber trop − 1 ( v ) \operatorname{trop}^{-1}(v) trop − 1 ( v ) . We focus on D 2 \mathcal{D}^{2} D 2 because of Lemma 9.2 .
The image of trop ( ψ ) \operatorname{trop}(\psi) trop ( ψ ) is a polyhedral subset of trop ( H ) \operatorname{trop}(\mathcal{H}) trop ( H ) . Its points are tropicalized hexagons , like that in Figure 15 . These have points in trop ( Q ) \operatorname{trop}(\mathcal{Q}) trop ( Q ) , lines in trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) , and incidence given by trop ( F ) \operatorname{trop}(\mathcal{F}) trop ( F ) . Thus, they are cycles of six G 2 G_2 G 2 -trees, with six apples, such that two consecutive trees share an apple whose distance to the leaves is the same in both trees.
Figure 15. A tropicalized hexagon consists of six apples in trop ( Q ) \operatorname{trop}(Q) trop ( Q ) and six trees in trop ( G ) \operatorname{trop}(G) trop ( G ) .
To prepare for the proof of Theorem 9.1, we now describe the parametrization ψ \psi ψ in the classical case [13 ] where domain and target are algebraic varieties. We use the polar form
⟨ x , y ⟩ = q ( x + y ) − q ( x ) − q ( y ) = x 1 y 4 + x 2 y 5 + x 3 y 6 + x 4 y 1 + x 5 y 2 + x 6 y 3 − 2 x 7 y 7 . (14) \langle x,y\rangle=q(x+y)-q(x)-q(y)=x_{1}y_{4}+x_{2}y_{5}+x_{3}y_{6}+x_{4}y_{1}+x_{5}y_{2}+x_{6}y_{3}-2x_{7}y_{7}. \tag*{(14)} ⟨ x , y ⟩ = q ( x + y ) − q ( x ) − q ( y ) = x 1 y 4 + x 2 y 5 + x 3 y 6 + x 4 y 1 + x 5 y 2 + x 6 y 3 − 2 x 7 y 7 . (14)
With this, we compute the projection of a point onto a line at maximal distance [39 §1.5] .
Lemma 9.2. Let p ( 12 ) ∈ G p^{(12)}\in G p ( 12 ) ∈ G and x ( 4 ) ∈ Q x^{(4)}\in Q x ( 4 ) ∈ Q be such that the line p ( 12 ) p^{(12)} p ( 12 ) is not contained in the hyperplane ( x ( 4 ) ) ⊥ (x^{(4)})^{\perp} ( x ( 4 ) ) ⊥ . There exist unique points x ( 2 ) , x ( 3 ) ∈ Q x^{(2)},x^{(3)}\in Q x ( 2 ) , x ( 3 ) ∈ Q and lines p ( 23 ) , p ( 34 ) ∈ G p^{(23)},p^{(34)}\in G p ( 23 ) , p ( 34 ) ∈ G such that
p ( 12 ) ∋ x ( 2 ) ∈ p ( 23 ) ∋ x ( 3 ) ∈ p ( 34 ) ∋ x ( 4 ) . p^{(12)}\ni x^{(2)}\in p^{(23)}\ni x^{(3)}\in p^{(34)}\ni x^{(4)}. p ( 12 ) ∋ x ( 2 ) ∈ p ( 23 ) ∋ x ( 3 ) ∈ p ( 34 ) ∋ x ( 4 ) . The entries of x ( 2 ) x^{(2)} x ( 2 ) , x ( 3 ) x^{(3)} x ( 3 ) , and hence those of p ( 23 ) p^{(23)} p ( 23 ) , p ( 34 ) p^{(34)} p ( 34 ) , depend polynomially on p ( 12 ) p^{(12)} p ( 12 ) , x ( 4 ) x^{(4)} x ( 4 ) .
Proof. By [39 Remark 2.4.5] , two points have distance 6 if and only if they are not perpendicular under (14) . Since p ( 12 ) ⊈ ( x ( 4 ) ) ⊥ p^{(12)}\not\subseteq(x^{(4)})^{\perp} p ( 12 ) ⊆ ( x ( 4 ) ) ⊥ , there exists a point on p ( 12 ) p^{(12)} p ( 12 ) that is opposite to x ( 4 ) x^{(4)} x ( 4 ) . This proves that the distance of p ( 12 ) p^{(12)} p ( 12 ) and x ( 4 ) x^{(4)} x ( 4 ) is exactly 5. By [39 Lemma 1.3.5] , they have a unique shortest path. Let x ( 2 ) , x ( 3 ) ∈ Q x^{(2)},x^{(3)}\in Q x ( 2 ) , x ( 3 ) ∈ Q and p ( 23 ) , p ( 34 ) ∈ G p^{(23)},p^{(34)}\in G p ( 23 ) , p ( 34 ) ∈ G be the vertices of this path.
We compute formulas for x ( 2 ) x^{(2)} x ( 2 ) , x ( 3 ) x^{(3)} x ( 3 ) , and hence for p ( 23 ) p^{(23)} p ( 23 ) , p ( 34 ) p^{(34)} p ( 34 ) , in terms of p ( 12 ) = ( p i j ) ∈ P 20 p^{(12)}=(p_{ij})\in P^{20} p ( 12 ) = ( p ij ) ∈ P 20 and x ( 4 ) = ( x 1 : ⋯ : x 7 ) ∈ P 6 x^{(4)}=(x_{1}:\cdots:x_{7})\in P^{6} x ( 4 ) = ( x 1 : ⋯ : x 7 ) ∈ P 6 . The intersection p ( 12 ) ∩ ( x ( 4 ) ) ⊥ p^{(12)}\cap(x^{(4)})^{\perp} p ( 12 ) ∩ ( x ( 4 ) ) ⊥ is the point x ( 2 ) x^{(2)} x ( 2 ) with coordinates
( x ( 2 ) ) i = x 4 p 1 i + x 5 p 2 i + x 6 p 3 i + x 1 p 4 i + x 2 p 5 i + x 3 p 6 i − 2 x 7 p 7 i for i = 1 , 2 , … , 7. (x^{(2)})_{i}=x_{4}p_{1i}+x_{5}p_{2i}+x_{6}p_{3i}+x_{1}p_{4i}+x_{2}p_{5i}+x_{3}p_{6i}-2x_{7}p_{7i}\qquad\text{for }i=1,2,\ldots,7. ( x ( 2 ) ) i = x 4 p 1 i + x 5 p 2 i + x 6 p 3 i + x 1 p 4 i + x 2 p 5 i + x 3 p 6 i − 2 x 7 p 7 i for i = 1 , 2 , … , 7.
The cross product with respect to our trilinear form Φ \Phi Φ in (1) is defined as
x × y : = Φ ( x , y , ⋅ ) ∈ ( K 7 ) ∨ ≅ K 7 . x\times y:=\Phi(x,y,\mathord{\cdot})\in(\mathbb{K}^{7})^{\vee}\cong\mathbb{K}^{7}. x × y := Φ ( x , y , ⋅ ) ∈ ( K 7 ) ∨ ≅ K 7 .
The last identification is via ⟨ ⋅ , ⋅ ⟩ \langle\mathord{\cdot},\mathord{\cdot}\rangle ⟨ ⋅ , ⋅ ⟩ . Since G 2 G_{2} G 2 preserves Φ \Phi Φ and ⟨ ⋅ , ⋅ ⟩ \langle\mathord{\cdot},\mathord{\cdot}\rangle ⟨ ⋅ , ⋅ ⟩ , the cross product is G 2 G_{2} G 2 -stable. We set x ( 3 ) = x ( 2 ) × x ( 4 ) x^{(3)}=x^{(2)}\times x^{(4)} x ( 3 ) = x ( 2 ) × x ( 4 ) . The coordinates of x ( 3 ) x^{(3)} x ( 3 ) are polynomials of bidegree ( 2 , 1 ) (2,1) ( 2 , 1 ) in ( x , p ) (x,p) ( x , p ) . For example, x 1 ( 3 ) = p 12 x 5 x 7 + p 13 x 6 x 7 + ⋯ + p 67 x 1 x 3 − 2 p 67 x 5 x 7 x_{1}^{(3)}=p_{12}x_{5}x_{7}+p_{13}x_{6}x_{7}+\cdots+p_{67}x_{1}x_{3}-2p_{67}x_{5}x_{7} x 1 ( 3 ) = p 12 x 5 x 7 + p 13 x 6 x 7 + ⋯ + p 67 x 1 x 3 − 2 p 67 x 5 x 7 has 24 terms.
The line spanned by two points x , y ∈ Q x,y\in Q x , y ∈ Q lies in G G G if and only if x × y = 0 x\times y=0 x × y = 0 . One checks that x ( 2 ) × x ( 3 ) = x ( 3 ) × x ( 4 ) = 0 x^{(2)}\times x^{(3)}=x^{(3)}\times x^{(4)}=0 x ( 2 ) × x ( 3 ) = x ( 3 ) × x ( 4 ) = 0 , that x ( 3 ) ∈ Q x^{(3)}\in Q x ( 3 ) ∈ Q , and these points are distinct. Hence the lines p ( 23 ) = x ( 2 ) ∧ x ( 3 ) p^{(23)}=x^{(2)}\wedge x^{(3)} p ( 23 ) = x ( 2 ) ∧ x ( 3 ) and p ( 34 ) = x ( 3 ) ∧ x ( 4 ) p^{(34)}=x^{(3)}\wedge x^{(4)} p ( 34 ) = x ( 3 ) ∧ x ( 4 ) are in G G G . Their Plücker coordinates are polynomials of bidegree ( 3 , 2 ) (3,2) ( 3 , 2 ) and ( 3 , 1 ) (3,1) ( 3 , 1 ) respectively. Our formulas for coordinates of x ( 2 ) x^{(2)} x ( 2 ) , x ( 3 ) x^{(3)} x ( 3 ) , p ( 23 ) p^{(23)} p ( 23 ) , p ( 34 ) p^{(34)} p ( 34 ) are not unique. They can be changed by adding polynomials that vanish on F 2 \mathbb{F}^{2} F 2 .
Lemma 9.2 implies the birationality of the map ψ \psi ψ in (13) . To see this, consider two flags ( x ( 1 ) , p ( 12 ) ) , ( x ( 4 ) , p ( 45 ) ) ∈ F (x^{(1)},p^{(12)}),(x^{(4)},p^{(45)}) \in\mathcal{F} ( x ( 1 ) , p ( 12 ) ) , ( x ( 4 ) , p ( 45 ) ) ∈ F whose points x ( 1 ) , x ( 4 ) x^{(1)},x^{(4)} x ( 1 ) , x ( 4 ) are opposite, and apply Lemma 9.2 to the mixed pairs ( x ( 1 ) , p ( 45 ) ) , ( x ( 4 ) , p ( 12 ) ) (x^{(1)},p^{(45)}),(x^{(4)},p^{(12)}) ( x ( 1 ) , p ( 45 ) ) , ( x ( 4 ) , p ( 12 ) ) . Our formulas now yield the unique hexagon completion.
Proof of Theorem 9.1. We first split up ψ \psi ψ into the two functions corresponding to completing the half-hexagons from mixed pairs, as in Lemma 9.2 . Let I Q × G I_{\mathcal{Q}\times\mathcal{G}} I Q × G be the ideal of Q × G \mathcal{Q}\times\mathcal{G} Q × G , in the polynomial ring with 28 variables x k , p i j x_k,p_{ij} x k , p ij . For v ∈ trop ( Q ) × trop ( G ) v \in\operatorname{trop}(\mathcal{Q})\times\operatorname{trop}(\mathcal{G}) v ∈ trop ( Q ) × trop ( G ) , the initial ideal in v ( I Q × G ) \operatorname{in}_v(I_{\mathcal{Q}\times\mathcal{G}}) in v ( I Q × G ) contains no monomials. Let f f f be any of the coordinates of the functions constructed in the proof of Lemma 9.2 . For any f ~ ∈ f + I Q × G \widetilde{f} \in f+I_{\mathcal{Q}\times\mathcal{G}} f ∈ f + I Q × G and any lift z ∈ Q × G z \in\mathcal{Q}\times\mathcal{G} z ∈ Q × G with − val ( z ) = v -\operatorname{val}(z)=v − val ( z ) = v we have trop ( f ~ ) ( v ) ≥ − val ( f ( z ) ) \operatorname{trop}(\widetilde{f})(v)\geq-\operatorname{val}(f(z)) trop ( f ) ( v ) ≥ − val ( f ( z )) . We then replace f f f by another representative f ~ ∈ f + I Q × G \widetilde{f}\in f+I_{\mathcal{Q}\times\mathcal{G}} f ∈ f + I Q × G such that in v ( f ~ ) \operatorname{in}_v(\widetilde{f}) in v ( f ) does not vanish on any component of the initial variety V ( in v ( I Q × G ) ) ∩ ( K ∗ ) 28 V(\operatorname{in}_v(I_{\mathcal{Q}\times\mathcal{G}}))\cap(\mathbb{K}^{*})^{28} V ( in v ( I Q × G )) ∩ ( K ∗ ) 28 . We call such f ~ \widetilde{f} f tropically faithful at v v v . Consider a lift z ∈ Q × G z\in\mathcal{Q}\times\mathcal{G} z ∈ Q × G and a representative α ∈ ( K ∗ ) 28 \alpha\in(\mathbb{K}^{*})^{28} α ∈ ( K ∗ ) 28 of its residue class such that in v ( f ~ ) ( α ) ≠ 0 \operatorname{in}_v(\widetilde{f})(\alpha)\ne0 in v ( f ) ( α ) = 0 . After rescaling the coordinates of z z z to valuation 0, the equality trop ( f ~ ) ( v ) = − val ( f ( z ) ) \operatorname{trop}(\widetilde{f})(v)=-\operatorname{val}(f(z)) trop ( f ) ( v ) = − val ( f ( z )) holds. The lift z z z is generic in the sense of the claim if it also avoids the forbidden locus { p ⊂ x ⊥ } \{p\subset x^\perp\} { p ⊂ x ⊥ } of Q × G \mathcal{Q}\times\mathcal{G} Q × G from Lemma 9.2 .
We performed an exhaustive computation that subdivides each of the 6 × 251 = 1506 6\times251=1506 6 × 251 = 1506 maximal Gröbner cones of trop ( Q ) × trop ( G ) \operatorname{trop}(\mathcal{Q})\times\operatorname{trop}(\mathcal{G}) trop ( Q ) × trop ( G ) . The subdivision is constructed such that each subcone τ \tau τ admits a tropically faithful representative f τ ∈ f + I Q × G f_\tau\in f+I_{\mathcal{Q}\times\mathcal{G}} f τ ∈ f + I Q × G on its relative interior.
We now pass from half-hexagons to hexagons. Let a tropical flag ( a ( 1 ) , d ( 12 ) ) (a^{(1)},d^{(12)}) ( a ( 1 ) , d ( 12 ) ) be contained in the relative interior of a maximal Gröbner cone σ \sigma σ of D \mathcal{D} D . Then any component of V ( in ( a ( 1 ) , d ( 12 ) ) I F ) V(\operatorname{in}_{(a^{(1)},d^{(12)})}I_{\mathcal{F}}) V ( in ( a ( 1 ) , d ( 12 ) ) I F ) in the torus is an orbit of the dim ( σ ) \dim(\sigma) dim ( σ ) -dimensional subtorus T σ ⊂ ( K ∗ ) 28 T_\sigma\subset(\mathbb{K}^{*})^{28} T σ ⊂ ( K ∗ ) 28 with trop ( T σ ) = span ( σ ) \operatorname{trop}(T_\sigma)=\operatorname{span}(\sigma) trop ( T σ ) = span ( σ ) . Its images in V ( in a ( 1 ) I Q ) V(\operatorname{in}_{a^{(1)}}I_{\mathcal{Q}}) V ( in a ( 1 ) I Q ) and V ( in d ( 12 ) I G ) V(\operatorname{in}_{d^{(12)}}I_{\mathcal{G}}) V ( in d ( 12 ) I G ) are the orbits of the projections of T σ T_\sigma T σ , of full dimension by assumption, so these projections are dominant.
Thus, for a pair of general flags in the dominant region D \mathcal{D} D , a generic lift of the flags will be a generic lift of the mixed pairs ( a ( 1 ) , d ( 45 ) ) (a^{(1)},d^{(45)}) ( a ( 1 ) , d ( 45 ) ) and ( a ( 4 ) , d ( 12 ) ) (a^{(4)},d^{(12)}) ( a ( 4 ) , d ( 12 ) ) . Hence, we can additionally require incidences x ( 1 ) ∈ p ( 12 ) x^{(1)}\in p^{(12)} x ( 1 ) ∈ p ( 12 ) and x ( 4 ) ∈ p ( 45 ) x^{(4)}\in p^{(45)} x ( 4 ) ∈ p ( 45 ) without changing valuations of the computed half-hexagons. In particular, on a dense open subset of D 2 ⊂ ( trop ( Q ) × trop ( G ) ) 2 \mathcal{D}^{2}\subset(\operatorname{trop}(\mathcal{Q})\times\operatorname{trop}(\mathcal{G}))^{2} D 2 ⊂ ( trop ( Q ) × trop ( G ) ) 2 , the valuation of a generically lifted hexagon is computed by trop ( f τ ) \operatorname{trop}(f_\tau) trop ( f τ ) applied to the mixed pair. These tropicalizations specify the piecewise linear map trop ( ψ ) \operatorname{trop}(\psi) trop ( ψ ) promised in Theorem 9.1. □
Theorem 9.1 holds more generally. The restriction to the dominant region D 2 \mathcal{D}^{2} D 2 of trop ( F ) 2 \operatorname{trop}(\mathcal{F})^{2} trop ( F ) 2 should not be needed. We leave the details to a future project on tropicalizing morphisms.
Example 9.3. The hexagon in Figure 15 was constructed by applying our piecewise-linear map trop ( ψ ) \operatorname{trop}(\psi) trop ( ψ ) to the following pair of tropical flags in D \mathcal{D} D . As always for tree metrics, we use the max-convention, and write a ( i ) = − val ( x ( i ) ) ∈ trop ( Q ) a^{(i)}=-\operatorname{val}(x^{(i)})\in\operatorname{trop}(\mathcal{Q}) a ( i ) = − val ( x ( i ) ) ∈ trop ( Q ) and d ( i j ) = − val ( p ( i j ) ) ∈ trop ( G ) d^{(ij)}=-\operatorname{val}(p^{(ij)})\in\operatorname{trop}(\mathcal{G}) d ( ij ) = − val ( p ( ij ) ) ∈ trop ( G ) :
a ( 1 ) = ( 11 , 12 , 4 , 15 , 0 , 14 , 13 ) , d ( 12 ) = ( 9 , 9 , 12 , 7 , 11 , 10 , 10 , 5 , 8 , 0 , 7 , 13 , 0 , 12 , 11 , 11 , 7 , 10 , 10 , 9 , 9 ) a ( 4 ) = ( 9 , 12 , 5 , 15 , 12 , 11 , 0 ) , d ( 45 ) = ( 11 , 12 , 14 , 11 , 10 , 9 , 15 , 5 , 0 , 9 , 12 , 18 , 15 , 14 , 5 , 5 , 12 , 15 , 9 , 12 , 11 ) . \begin{aligned}
a^{(1)}&=(11,12,4,15,0,14,13), & d^{(12)}&=(9,9,12,7,11,10,10,5,8,0,7,13,0,12,11,11,7,10,10,9,9)\\
a^{(4)}&=(9,12,5,15,12,11,0), & d^{(45)}&=(11,12,14,11,10,9,15,5,0,9,12,18,15,14,5,5,12,15,9,12,11).
\end{aligned} a ( 1 ) a ( 4 ) = ( 11 , 12 , 4 , 15 , 0 , 14 , 13 ) , = ( 9 , 12 , 5 , 15 , 12 , 11 , 0 ) , d ( 12 ) d ( 45 ) = ( 9 , 9 , 12 , 7 , 11 , 10 , 10 , 5 , 8 , 0 , 7 , 13 , 0 , 12 , 11 , 11 , 7 , 10 , 10 , 9 , 9 ) = ( 11 , 12 , 14 , 11 , 10 , 9 , 15 , 5 , 0 , 9 , 12 , 18 , 15 , 14 , 5 , 5 , 12 , 15 , 9 , 12 , 11 ) .
The apples and trees in their tropical hexagon are computed by tropicalizing representatives f ~ \widetilde{f} f of coordinate functions f f f of ψ \psi ψ that are faithful in the sense of the proof above. We find
a ( 2 ) = ( 0 , 3 , 3 , 6 , 1 , 5 , 4 ) , d ( 23 ) = ( 3 , 3 , 6 , 1 , 5 , 4 , 6 , 9 , 4 , 8 , 7 , 7 , 0 , 6 , 5 , 5 , 7 , 6 , 4 , 3 , 3 ) a ( 3 ) = ( 2 , 5 , 3 , 4 , 1 , 0 , 0 ) , d ( 34 ) = ( 6 , 7 , 12 , 9 , 8 , 4 , 10 , 15 , 12 , 11 , 7 , 13 , 10 , 9 , 0 , 0 , 7 , 10 , 4 , 7 , 6 ) a ( 5 ) = ( 0 , 3 , 3 , 6 , 3 , 2 , 0 ) , d ( 56 ) = ( 2 , 3 , 6 , 3 , 2 , 0 , 6 , 9 , 6 , 5 , 3 , 7 , 4 , 3 , 3 , 3 , 6 , 0 , 3 , 2 ) a ( 6 ) = ( 2 , 5 , 3 , 4 , 1 , 0 , 2 ) , d ( 61 ) = ( 11 , 9 , 12 , 7 , 11 , 10 , 10 , 15 , 8 , 14 , 13 , 13 , 0 , 12 , 11 , 11 , 13 , 10 , 10 , 9 , 11 ) . \begin{aligned}
a^{(2)}&=(0,3,3,6,1,5,4), & d^{(23)}&=(3,3,6,1,5,4,6,9,4,8,7,7,0,6,5,5,7,6,4,3,3)\\
a^{(3)}&=(2,5,3,4,1,0,0), & d^{(34)}&=(6,7,12,9,8,4,10,15,12,11,7,13,10,9,0,0,7,10,4,7,6)\\
a^{(5)}&=(0,3,3,6,3,2,0), & d^{(56)}&=(2,3,6,3,2,0,6,9,6,5,3,7,4,3,3,3,6,0,3,2)\\
a^{(6)}&=(2,5,3,4,1,0,2), & d^{(61)}&=(11,9,12,7,11,10,10,15,8,14,13,13,0,12,11,11,13,10,10,9,11).
\end{aligned} a ( 2 ) a ( 3 ) a ( 5 ) a ( 6 ) = ( 0 , 3 , 3 , 6 , 1 , 5 , 4 ) , = ( 2 , 5 , 3 , 4 , 1 , 0 , 0 ) , = ( 0 , 3 , 3 , 6 , 3 , 2 , 0 ) , = ( 2 , 5 , 3 , 4 , 1 , 0 , 2 ) , d ( 23 ) d ( 34 ) d ( 56 ) d ( 61 ) = ( 3 , 3 , 6 , 1 , 5 , 4 , 6 , 9 , 4 , 8 , 7 , 7 , 0 , 6 , 5 , 5 , 7 , 6 , 4 , 3 , 3 ) = ( 6 , 7 , 12 , 9 , 8 , 4 , 10 , 15 , 12 , 11 , 7 , 13 , 10 , 9 , 0 , 0 , 7 , 10 , 4 , 7 , 6 ) = ( 2 , 3 , 6 , 3 , 2 , 0 , 6 , 9 , 6 , 5 , 3 , 7 , 4 , 3 , 3 , 3 , 6 , 0 , 3 , 2 ) = ( 11 , 9 , 12 , 7 , 11 , 10 , 10 , 15 , 8 , 14 , 13 , 13 , 0 , 12 , 11 , 11 , 13 , 10 , 10 , 9 , 11 ) .
A visualization of the six trees d ( i j ) d^{(ij)} d ( ij ) and their six apples a ( k ) a^{(k)} a ( k ) is shown in Figure 15 . We observe that all G2 _2 2 -trees in this hexagon lie in maximal cones of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) . Their orbit types are:
d ( 12 ) : 1 , d ( 23 ) : 24 , d ( 34 ) : 10 , d ( 45 ) : 15 , d ( 56 ) : 12 , d ( 61 ) : 23. d^{(12)}:1,\quad d^{(23)}:24,\quad d^{(34)}:10,\quad d^{(45)}:15,\quad d^{(56)}:12,\quad d^{(61)}:23. d ( 12 ) : 1 , d ( 23 ) : 24 , d ( 34 ) : 10 , d ( 45 ) : 15 , d ( 56 ) : 12 , d ( 61 ) : 23.
Thus, the hexagon appears in Figure 8 –Figure 9 , after relabeling the taxa via the Weyl group D12 _{12} 12 .
Curves of Genus Ten
In this section we discuss an application of the G2 _2 2 -Grassmannian G \mathcal{G} G in algebraic geometry, namely the construction of canonical curves of genus 10 in Mukai’s theory [32 ] . To describe this, we write Λ c \Lambda_c Λ c for a general linear subspace of codimension c c c in the ambient projective space P 20 \mathbb{P}^{20} P 20 . By Bertini’s Theorem, the intersection G ∩ Λ c \mathcal{G}\cap\Lambda_c G ∩ Λ c is a smooth variety of dimension 5 − c 5-c 5 − c . The intersection, which is irreducible if c ≤ 4 c\leq4 c ≤ 4 , lives in the projective space L ∩ Λ c ≃ P 13 − c \mathcal{L}\cap\Lambda_c\simeq\mathbb{P}^{13-c} L ∩ Λ c ≃ P 13 − c . This construction, which is due to Mukai, leads to the following cascade of interesting varieties:
(a) If c = 2 c=2 c = 2 , then G ∩ Λ c \mathcal{G}\cap\Lambda_c G ∩ Λ c is a Fano threefold of genus 10 in P 11 \mathbb{P}^{11} P 11 .
(b) If c = 3 c=3 c = 3 , then G ∩ Λ c \mathcal{G}\cap\Lambda_c G ∩ Λ c is a K3 surface of genus 10 in P 10 \mathbb{P}^{10} P 10 .
(c) If c = 4 c=4 c = 4 , then G ∩ Λ c \mathcal{G}\cap\Lambda_c G ∩ Λ c is a canonical curve of genus 10 in P 9 \mathbb{P}^{9} P 9 .
(d) If c = 5 c=5 c = 5 , then G ∩ Λ c \mathcal{G}\cap\Lambda_c G ∩ Λ c is a self-dual configuration of 18 points in P 8 \mathbb{P}^{8} P 8 .
Analogous constructions for g = 7 , 8 , 9 g=7,8,9 g = 7 , 8 , 9 yield birational models for the moduli space M g \mathcal{M}_g M g . For further context see [26 , 27 ] . Our curves in (c) form a divisor K \mathcal{K} K in M 10 \mathcal{M}_{10} M 10 , i.e. they obey one constraint in the 27-dimensional moduli space of all curves. The canonical divisor of the curve is the configuration in (d). The varieties (a)–(d) are arithmetically Gorenstein of degree 18. Their common Betti table records the degrees of the higher syzygies. It equals
[ 28 105 162 84 ⋅ ⋅ ⋅ 84 162 105 28 ] (15) \begin{bmatrix}
28 & 105 & 162 & 84 \\
\cdot& \cdot& \cdot& 84 & 162 & 105 & 28 \tag*{(15)}
\end{bmatrix} [ 28 ⋅ 105 ⋅ 162 ⋅ 84 84 162 105 28 ] (15)
This means that the ideal is minimally generated by 28 quadrics, these quadrics satisfy 105 independent linear syzygies, and those in turn satisfy 162 independent second syzygies, etc.
Here we examine these objects through the lens of tropical geometry. The tropical moduli space trop ( M 10 ) \operatorname{trop}(\mathcal{M}_{10}) trop ( M 10 ) is a (stacky) polyhedral fan of dimension 27; see [8 ] . It has 2317172 maximal cones, one for each trivalent graph of genus 10, with 18 vertices and 27 edges. Lower-dimensional cones of trop ( M 10 ) \operatorname{trop}(\mathcal{M}_{10}) trop ( M 10 ) are indexed by non-trivalent graphs, some with vertex labels that remember contracted loops. All our graphs are undirected, connected, and they can have loops and multiple edges. Only 41301 of the trivalent graphs of genus 10 are simple.
We aim to parametrize the tropical divisor trop ( K ) \operatorname{trop}(\mathcal{K}) trop ( K ) in trop ( M 10 ) \operatorname{trop}(\mathcal{M}_{10}) trop ( M 10 ) by intersecting our Grassmannian trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) with tropical linear spaces L 4 \mathcal{L}_4 L 4 of codimension 4. These are points in trop ( Gr ( 17 , 21 ) ) \operatorname{trop}(\operatorname{Gr}(17,21)) trop ( Gr ( 17 , 21 )) . This poses a challenge: it is infeasible to compute this tropical Grassmannian, and we know of no way to properly sample from it. Note that such an L 4 \mathcal{L}_4 L 4 can be very large. According to Speyer’s formula [35 ] , it can have up to 735471 maximal cells. We used the following heuristics to sample from Gr ( 21 − c , 21 ) \operatorname{Gr}(21-c,21) Gr ( 21 − c , 21 ) . Here, the min-convention was used.
Method 1: Shift the uniform Bergman fan. The Bergman fan B c \mathcal{B}_c B c consists of all ( 21 c + 1 ) \binom{21}{c+1} ( c + 1 21 ) cones of codimension c c c that are spanned by unit vectors in R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 . We pick a random point q q q in R 21 / 1 \mathbb{R}^{21}/1 R 21 /1 , and, as in [29 Corollary 3.6.16] , we take our tropical linear space to be L c = q + B c \mathcal{L}_c=q+\mathcal{B}_c L c = q + B c .
Method 2: Intersect successively with hyperplanes. We pick hyperplanes H i = q i + B S i \mathcal{H}_i=q_i+\mathcal{B}_{S_i} H i = q i + B S i , where S i S_i S i is a subset of the 21 coordinates. These define L c = H 1 ∩ ⋯ ∩ H c \mathcal{L}_c=\mathcal{H}_1\cap\cdots\cap\mathcal{H}_c L c = H 1 ∩ ⋯ ∩ H c . We compute
( ( ( trop ( G ) ∩ H 1 ) ∩ H 2 ) ∩ H 3 ) ∩ ⋯ to get to trop ( G ) ∩ L c . (((\operatorname{trop}(\mathcal{G})\cap\mathcal{H}_1)\cap\mathcal{H}_2)\cap\mathcal{H}_3)\cap\cdots\quad\text{to get to}\quad\operatorname{trop}(\mathcal{G})\cap\mathcal{L}_c. ((( trop ( G ) ∩ H 1 ) ∩ H 2 ) ∩ H 3 ) ∩ ⋯ to get to trop ( G ) ∩ L c .
Using random sampling, we obtained the following result, illustrated in Figure 16 and Figure 17 .
Proposition 10.1. Methods 1 and 2 can both yield curves in maximal cones of trop ( M 10 ) \operatorname{trop}(\mathcal{M}_{10}) trop ( M 10 ) . We obtain a sample of the 2317172 trivalent metric graphs with 18 vertices and 27 edges.
Proof. We present input data for each of the two methods that leads to a trivalent graph. For Method 1, we select the following shift vector to define L c = q + B c \mathcal{L}_c = q + \mathcal{B}_c L c = q + B c for c = 1 , 2 , 3 , 4 , 5 c = 1, 2, 3, 4, 5 c = 1 , 2 , 3 , 4 , 5 :
q = − 204958 e 12 − 1849487 e 13 − 1813412 e 14 + 1389486 e 15 − 611356 e 16 − 155086 e 17 − 128988 e 23 − 1698926 e 24 + 1863394 e 25 − 602010 e 26 − 2239681 e 27 − 1904628 e 34 + 124320 e 35 − 1025736 e 36 − 87111 e 37 − 1360439 e 45 + 705715 e 46 − 619752 e 47 + 2004970 e 56 + 1344973 e 57 + 232336 e 67 . \begin{aligned}
q ={}& -204958e_{12} - 1849487e_{13} - 1813412e_{14} + 1389486e_{15} - 611356e_{16} - 155086e_{17} \\
&- 128988e_{23} - 1698926e_{24} + 1863394e_{25} - 602010e_{26} - 2239681e_{27} - 1904628e_{34} \\
&+ 124320e_{35} - 1025736e_{36} - 87111e_{37} - 1360439e_{45} + 705715e_{46} - 619752e_{47} \\
&+ 2004970e_{56} + 1344973e_{57} + 232336e_{67}.
\end{aligned} q = − 204958 e 12 − 1849487 e 13 − 1813412 e 14 + 1389486 e 15 − 611356 e 16 − 155086 e 17 − 128988 e 23 − 1698926 e 24 + 1863394 e 25 − 602010 e 26 − 2239681 e 27 − 1904628 e 34 + 124320 e 35 − 1025736 e 36 − 87111 e 37 − 1360439 e 45 + 705715 e 46 − 619752 e 47 + 2004970 e 56 + 1344973 e 57 + 232336 e 67 .
For Method 2, our four hyperplanes are defined by the following shift vectors q 1 , q 2 , q 3 , q 4 q_1, q_2, q_3, q_4 q 1 , q 2 , q 3 , q 4 :
q 1 = − 217936 e 13 − 84861 e 37 + 34008 e 46 , q_1 = -217936e_{13} - 84861e_{37} + 34008e_{46}, q 1 = − 217936 e 13 − 84861 e 37 + 34008 e 46 ,
q 2 = 1013667 e 14 + 132815 e 15 + 744071 e 25 − 794071 e 26 + 721948 e 37 − 60165 e 46 + 360635 e 47 + 705771 e 67 , \begin{aligned}
q_2 ={}& 1013667e_{14} + 132815e_{15} + 744071e_{25} - 794071e_{26} + 721948e_{37} \\
&- 60165e_{46} + 360635e_{47} + 705771e_{67},
\end{aligned} q 2 = 1013667 e 14 + 132815 e 15 + 744071 e 25 − 794071 e 26 + 721948 e 37 − 60165 e 46 + 360635 e 47 + 705771 e 67 ,
q 3 = − 806183 e 15 − 732664 e 26 − 107361 e 27 − 721205 e 34 − 846796 e 35 − 183135 e 36 + 542064 e 46 + 278271 e 47 + 78144 e 67 , \begin{aligned}
q_3 ={}& -806183e_{15} - 732664e_{26} - 107361e_{27} - 721205e_{34} - 846796e_{35} \\
&- 183135e_{36} + 542064e_{46} + 278271e_{47} + 78144e_{67},
\end{aligned} q 3 = − 806183 e 15 − 732664 e 26 − 107361 e 27 − 721205 e 34 − 846796 e 35 − 183135 e 36 + 542064 e 46 + 278271 e 47 + 78144 e 67 ,
q 4 = 722953 e 12 + 1004831 e 14 − 757760 e 15 − 217944 e 16 + 733597 e 17 + 589026 e 23 + 213921 e 24 + 650364 e 26 − 890415 e 34 − 891447 e 35 − 636797 e 36 + 914965 e 37 + 315068 e 46 − 862815 e 56 + 692897 e 57 . \begin{aligned}
q_4 ={}& 722953e_{12} + 1004831e_{14} - 757760e_{15} - 217944e_{16} + 733597e_{17} \\
&+ 589026e_{23} + 213921e_{24} + 650364e_{26} - 890415e_{34} - 891447e_{35} \\
&- 636797e_{36} + 914965e_{37} + 315068e_{46} - 862815e_{56} + 692897e_{57}.
\end{aligned} q 4 = 722953 e 12 + 1004831 e 14 − 757760 e 15 − 217944 e 16 + 733597 e 17 + 589026 e 23 + 213921 e 24 + 650364 e 26 − 890415 e 34 − 891447 e 35 − 636797 e 36 + 914965 e 37 + 315068 e 46 − 862815 e 56 + 692897 e 57 .
The divisor ( d ) (d) ( d ) is found by intersecting with one additional tropical hyperplane q 5 q_5 q 5 . For each sample, we compute the graph given by trop ( G ) ∩ L 4 \operatorname{trop}(G) \cap\mathcal{L}_4 trop ( G ) ∩ L 4 , and we check whether it is trivalent.
A picture says more than a thousand words. Our two genus 10 curves are shown in Figure 16 and Figure 17 . The left panel in each figure is the actual tropical curve. This curve has degree 18 in dimension 15, so it has up to 15 × 18 = 270 15 \times18 = 270 15 × 18 = 270 unbounded rays. These form trees that are attached to the Berkovich skeleton . This is our metric graph, with 2-valent vertices removed, shown here with dark edges. Miraculously, that graph is trivalent, with 27 edges.
The 18 red points form the canonical divisor ( d ) (d) ( d ) . With Method 1, they are the 18 vertices. With Method 2, the 18 self-dual points sit on edges of the curve. We see this in Figure 17 .
We now explain the panels on the right. The picture shows the Berkovich skeleton and explains how it arises from intersecting trop ( G ) \operatorname{trop}(G) trop ( G ) with the linear space L 4 \mathcal{L}_4 L 4 . First consider Figure 16 , where the graph is planar. Only five of the 248 maximal cones of trop ( G ) \operatorname{trop}(G) trop ( G ) meet the skeleton. These are the cones of types 11, 12, 7, 11, 12, whose G 2 G_2 G 2 -trees are displayed. The five shaded regions indicate the intersections of these five cones with the shifted Bergman fan L 4 \mathcal{L}_4 L 4 . For instance, the blue region shows that the cone of type 7 intersects L 4 \mathcal{L}_4 L 4 in two pieces on the skeleton: one edge and one tree with five edges. The two blue vertices are the contribution of the G 2 G_2 G 2 -tree 7 to the canonical divisor, which equals trop ( G ) ∩ L 5 \operatorname{trop}(G) \cap\mathcal{L}_5 trop ( G ) ∩ L 5 . Broken lines between two colored regions are 3-valent vertices of the curve that become 2-valent in the Berkovich skeleton. They are intersections of L 4 \mathcal{L}_4 L 4 with codimension-1 cones in trop ( G ) \operatorname{trop}(G) trop ( G ) .
The trivalent genus 10 graph in Figure 17 is not planar. Here, the skeleton meets many more maximal cones of trop ( G ) \operatorname{trop}(G) trop ( G ) . The numbers 17, 12, 16, … refer to their types of G 2 G_2 G 2 -trees in Figure 8 and Figure 9 . For both approaches, the intersection trop ( G ) ∩ L 3 \operatorname{trop}(G) \cap\mathcal{L}_3 trop ( G ) ∩ L 3 is the tropicalization of the K3-surface in (b). With Method 1, its Berkovich skeleton collapses onto a polyhedral 2-sphere, its essential skeleton , and hence onto the boundary of a 3-polytope by Steinitz’s Theorem. In Figure 16 , our curve is the edge graph of that simple polytope. We conclude that that it would be worthwhile to develop tropical Mukai theory for curves of genus 7, 8, 9, 10.
Authors’ addresses:
Michael Joswig joswig@math.tu-berlin.de
Technische Universität Berlin, Chair of Discrete Mathematics / Geometry, Berlin & MPI Mathematics in the Sciences, Leipzig, Germany
Kevin Kuehn kuehn@math.tu-berlin.de
Technische Universität Berlin, Chair of Discrete Mathematics / Geometry, Germany
Bernd Sturmfels bernd@mis.mpg.de
MPI Mathematics in the Sciences, Leipzig, Germany
Mathieu Vallée vallee@mis.mpg.de
MPI Mathematics in the Sciences, Leipzig, Germany
Acknowledgments: MJ and KK received funding by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy – “The Berlin Mathematics Research Center MATH+” (EXC-2046/1, EXC-2046/2, project ID 390685689), “Symbolic Tools in Mathematics and their Application” (TRR 195, project ID 286237555), and the “Mathematical Research Data Initiative (MaRDI)” (project ID 460135501). BS and MV were supported by the European Research Council (ERC) through the synergy grant UNIVERSE+, 101118787. Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.
Data
Our data are available in the Zenodo repository 10.5281/zenodo.24104003 ; this repository comprises the 222 files of Table 4 and its folder data, the intermediate results of our computations, from which these files are copied and from which our code checks the statements of this article. They were computed with Oscar [33 , 12 ] , polymake [16 ] and Gfan [23 ] . The files in the folders 00_ideals to 05_curves are written by Oscar 1.8.2 in its format .mrdi, and its function load reads them.
Table 4. Directory layout of the Zenodo repository.
Conventions. The suffixes qq and gf2 name the field Q \mathbb{Q} Q or F 2 \mathbb{F}_2 F 2 . By Theorems 4.1 , 4.2 and 6.3 , the fans and polytopes over Q \mathbb{Q} Q work for every characteristic other than 2. The coordinates of P 6 \mathbb{P}^6 P 6 are x 1 , … , x 7 x_1,\ldots,x_7 x 1 , … , x 7 , those of P 20 \mathbb{P}^{20} P 20 are p 12 , … , p 67 p_{12},\ldots,p_{67} p 12 , … , p 67 in lexicographic order, and those of P 6 × P 20 \mathbb{P}^6 \times\mathbb{P}^{20} P 6 × P 20 are x 1 , … , x 7 , p 12 , … , p 67 x_1,\ldots,x_7,p_{12},\ldots,p_{67} x 1 , … , x 7 , p 12 , … , p 67 . The fans lie in R 21 \mathbb{R}^{21} R 21 and R 28 \mathbb{R}^{28} R 28 . All tropical objects in the files use the min-convention of Oscar. A point w w w in a file is − w -w − w in the max-convention of this paper, so the G 2 G_2 G 2 -tree of a ray r r r is − r -r − r , up to lineality, and the coordinates x k x_k x k give
the d k 8 d_{k8} d k 8 of the flag Dressian. The fans are PolyhedralFans with the coarsest fan structure of § 2 –§ 4 . Their rays are numbered as in rays_modulo_lineality, and their maximal cones as in maximal_cones.
Orbit types. Our computation numbers the rays and maximal cones of each fan, and their orbits under D 12 D_{12} D 12 . The four tables *_orbits.csv translate these orbits into the types of this paper, one orbit per line. The members of the orbits in grassmannian_orbits.csv and flag_orbits.csv are numbered as in the files of the same names in 02_tropical_varieties, those in the other two tables as in 01_dressians. An orbit over Q \mathbb{Q} Q and an orbit over F 2 \mathbb{F}_2 F 2 share a line when they consist of the same rays or cones. The columns are:
kind: ray or cell.
orbit_type: the type α 1 , … , ϵ \alpha_1,\ldots,\epsilon α 1 , … , ϵ of a ray, as in Figure 7 , or the type 1 , … , 35 1,\ldots,35 1 , … , 35 of a maximal cone, as in Figure 8 and Figure 9 , written alpha-1, ..., epsilon and 1, ..., 35. A ray or maximal cone of a flag fan has the type of the ray or maximal cone of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) below it, and a ray over the lineality space has type star. The tetrahedra of the G 2 G_2 G 2 -Dressian, and the cones of the flag Dressian over them, have type epsilon; see Figure 6 .
apple, for the flag fans: the edges of the G 2 G_2 G 2 -tree of the cell type that carry the apple, as in Figure 11 , separated by |. Here pendant k is the pendant edge of the leaf k k k , and inner a-b is the inner edge between the inner nodes a , b ∈ { 8 , … , 12 } a,b\in\{8,\ldots,12\} a , b ∈ { 8 , … , 12 } .
size: the size of the orbit.
orbit, qq_orbit, gf2_orbit: for a cell, the number of its orbit, which is the number NNN in 03_no_polytopes for a toric cell type; for a ray, the smallest index of a ray in the orbit. The entry is empty if the orbit does not occur over that field.
members, qq_members, gf2_members: the indices of the rays or maximal cones of the orbit.
For example, the line ray,alpha-2,3,8,1,8;33;34,1;19;22 of grassmannian_orbits.csv says that the rays of type α 2 \alpha_2 α 2 form an orbit of size 3. This orbit consists of the rays 8, 33, 34 of grassmannian_qq. In characteristic p = 2 p=2 p = 2 , it consists of the rays 1, 19, 22 of grassmannian_gf2.
Newton–Okounkov polytopes. File 03_no_polytopes/grassmannian_qq/orbit_NNN is the polytope P C P_C P C of § 6 for a maximal cone C C C in the orbit NNN of 02_tropical_varieties/grassmannian_qq, and likewise for the other three folders. There is one file for each orbit of toric degenerations in Theorem 6.3 . The polytopes are lattice polytopes of dimension 5 in R 5 \mathbb{R}^5 R 5 for G \mathcal{G} G , and of dimension 6 in R 6 \mathbb{R}^6 R 6 for F \mathcal{F} F , where P C P_C P C is the Minkowski sum of the polytopes of the coordinates x 1 , … , x 7 x_1,\ldots,x_7 x 1 , … , x 7 and of the coordinates p 12 , … , p 67 p_{12},\ldots,p_{67} p 12 , … , p 67 .
Hexagon and curves. The file tropicalized_hexagon is a named tuple with the fields points, the six vectors val ( x ( 1 ) ) , … , val ( x ( 6 ) ) \operatorname{val}(x^{(1)}),\ldots,\operatorname{val}(x^{(6)}) val ( x ( 1 ) ) , … , val ( x ( 6 ) ) in Q 7 \mathbb{Q}^7 Q 7 , and lines, the six vectors val ( p ( 12 ) ) , val ( p ( 23 ) ) , … , val ( p ( 61 ) ) \operatorname{val}(p^{(12)}),\operatorname{val}(p^{(23)}),\ldots,\operatorname{val}(p^{(61)}) val ( p ( 12 ) ) , val ( p ( 23 ) ) , … , val ( p ( 61 ) ) in Q 21 \mathbb{Q}^{21} Q 21 , each shifted to the minimum 0. For such a vector v v v , the vector a ( i ) a^{(i)} a ( i ) or d ( i j ) d^{(ij)} d ( ij ) of Example 9.3 is max ( v ) 1 − v \max(v)\mathbf{1}-v max ( v ) 1 − v . The files genus_10_method_1 and genus_10_method_2 hold the tropical curves of genus 10 obtained with Methods 1 and 2 of § 10 , as one-dimensional PolyhedralComplexes in R 21 \mathbb{R}^{21} R 21 with one point over each point of R 21 / 1 \mathbb{R}^{21}/\mathbf{1} R 21 / 1 . They have 449 and 447 maximal cells, of which 239 and 210 are unbounded, and their bounded cells form graphs of genus 10 on 201 and 228 vertices. The first file is a named tuple with the field complex. The second also has the field multiplicities, the multiplicities 1, 2, 4 or 5 of the maximal cells in the order of maximal_polyhedra. The file genus_10_method_1_pieces is the vector of the 454 one-dimensional polyhedra computed for Method 1 from its shift q q q ; they make up the 449 maximal cells of genus_10_method_1, five of which consist of two collinear pieces.
Computation
Public repository. Our code is available at github.com/dmg-lab/TropicalGeometryG2 .
Software and hardware. We used Oscar 1.8.2 [12 , 33 ] in Julia 1.13, with polymake [16 ] inside Oscar for polyhedral geometry, Gfan 0.8 [23 ] for tropical prevarieties, and, for the curves of § 10 , the LP solver HiGHS 1.15 [21 ] as a floating-point filter whose answers are checked over Q \mathbb{Q} Q , with SoPlex [17 ] (via Oscar) as an exact alternative. The fans of § 2 –§ 4 were computed on a laptop with an Intel Core i7-12700H processor (14 cores) and 16 GB of memory. The curves of § 10 were searched on a cluster whose nodes have two Intel Xeon Platinum 8268 processors (48 cores at 2.9 GHz) and 384 GB of memory.
Timings. Table 5 lists the wall-clock times to compute each fan from its ideal, before coarsening. The Dressians are the prevarieties of the generators; they do not depend on the field. Oscar 1.8.2 tropicalizes prime ideals with the trivial valuation only over Q \mathbb{Q} Q ; over F 2 \mathbb{F}_2 F 2 , we use the same method, which calls the same function of Singular, and it returns the fan of Theorem 4.1 . Our census computes it differently: it certifies each orbit of cells of the prevariety of the generators by a Gröbner basis, and adds 6 polynomials of these Gröbner bases to the generators whenever some cells fail; after two such rounds, the 54 polynomials form a tropical basis. For the flag Dressian, Oscar’s tropical_prevariety ran on the cluster on a node with two Intel Xeon Gold 6226R processors (32 cores at 2.9 GHz) and 754 GB of memory, after eliminating the six coordinates that the binomial linear forms in (4) identify with others.
Table 5. Time to compute the tropical fans of § 2 –§ 4 ; loading Oscar takes 14 s more.
Curves of genus 10. Table 6 summarizes the random searches behind Proposition 10.1 . Method 1 intersects trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) with L 4 = q + B 4 \mathcal{L}_4=q+\mathcal{B}_4 L 4 = q + B 4 exactly, cell by cell. The part of q q q orthogonal to the 15-dimensional linear span of trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) is fixed, taken from an earlier sample, and the part in this span is random. With uniformly random shifts q q q instead, 19200 samples gave 791 curves of genus 10, only one of them with a trivalent core. Method 2 cuts trop ( G ) \operatorname{trop}(\mathcal{G}) trop ( G ) with four tropical hyperplanes in turn, shifted by q 1 , … , q 4 q_1,\ldots,q_4 q 1 , … , q 4 with supports of random sizes. It decides which cells meet in floating point, and recomputes the curves of genus 10 exactly, in 7 to 33 minutes each. In both methods, the random vectors have integer entries in [ − 2 20 , 2 20 ] [-2^{20},2^{20}] [ − 2 20 , 2 20 ] , and sample k k k draws them with the random number generator Xoshiro(k) of Julia. In Method 1, such a vector is projected onto the span, and the shift q q q in the proof of Proposition 10.1 is twice that of sample 2722, so that it is integral. A curve has 18 distinct points in Table 6 when its intersection with L 5 \mathcal{L}_5 L 5 consists of 18 points of multiplicity 1.
Table 6. The random searches on the cluster for the tropical curves in Proposition 10.1 .