combinatorial aspects of tropical geometrypeople.math.gatech.edu/~jyu67/thesis.pdf · first, their...
TRANSCRIPT
Combinatorial Aspects of Tropical Geometry
by
Josephine Thi Mar Lwin Yu
B.S. (University of California, Davis) 2003
A dissertation submitted in partial satisfaction of the
requirements for the degree of
Doctor of Philosophy
in
Mathematics
in the
GRADUATE DIVISION
of the
UNIVERSITY OF CALIFORNIA, BERKELEY
Committee in charge:Professor Bernd Sturmfels, Chair
Professor Lior PachterProfessor Alper Atamturk
Spring 2007
The dissertation of Josephine Thi Mar Lwin Yu is approved:
Chair Date
Date
Date
University of California, Berkeley
Spring 2007
Combinatorial Aspects of Tropical Geometry
Copyright 2007
by
Josephine Thi Mar Lwin Yu
1
Abstract
Combinatorial Aspects of Tropical Geometry
by
Josephine Thi Mar Lwin Yu
Doctor of Philosophy in Mathematics
University of California, Berkeley
Professor Bernd Sturmfels, Chair
Tropical convex geometry and tropical algebraic geometry arise from linear and
polynomial algebra over the tropical semiring (R,min,+). They also appear as images of
convex and algebraic sets over fields with valuations into the real numbers. In this thesis,
the combinatorial aspects of tropical geometry are studied through three topics: tropical
polytopes, tropical linear spaces, and tropical elimination and implicitization.
Tropical polytopes are related to cellular resolutions of monomial ideals in two dif-
ferent ways. First, their natural polyhedral complex structure supports minimal linear free
resolutions of monomial initial ideals of the ideals of 2× 2-minors of matrices of unknowns.
Secondly, for an arbitrary monomial ideal, the tropical convex hull of the exponents of its
minimal generators is the common shadow of a natural family of cellular free resolutions,
generalizing the hull resolution.
Tropical linear spaces are some special types of intersections of tropical hyper-
2
planes. We show that they are also images of certain tropical linear maps. We characterize
the sets of tropical hyperplanes and the parametrizations that define a tropical linear space.
This generalizes the problem of finding minimal tropical bases of matroids. We show that
graphic and cographic matroids have unique minimal tropical bases.
Elimination and implicitization are computational problems of finding the defining
ideals of a projection of an algebraic variety and a variety given by a parametrization,
respectively. We give a combinatorial construction of the tropicalization of these ideals when
the given input polynomials have generic coefficients. When the solution is a principal ideal,
we can recover the Newton polytope of the generating polynomial from the tropical variety,
and in general we can compute its Chow polytope. We also describe an implementation of
this method.
Professor Bernd SturmfelsDissertation Committee Chair
i
To my parents.
ii
Contents
List of Figures iv
List of Tables vi
1 Introduction 1
1.1 Polyhedral geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Commutative algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Tropical mathematics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3.1 Tropical convex geometry . . . . . . . . . . . . . . . . . . . . . . . . 81.3.2 Tropical linear spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 101.3.3 Tropical algebraic geometry . . . . . . . . . . . . . . . . . . . . . . . 11
2 Tropical Polytopes and Cellular Resolutions 13
2.1 Tropical polytopes via cellular resolutions . . . . . . . . . . . . . . . . . . . 142.1.1 An algorithm for computing tropical convex hulls . . . . . . . . . . . 232.1.2 Tropical cyclic polytopes . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 Cellular resolutions via tropical polytopes . . . . . . . . . . . . . . . . . . . 332.2.1 Faces of tropical polytopes . . . . . . . . . . . . . . . . . . . . . . . 40
3 Tropical Linear Spaces 53
3.1 The constant coefficient case . . . . . . . . . . . . . . . . . . . . . . . . . . 543.1.1 Partition matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583.1.2 Uniform matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583.1.3 Fano plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.2 Regular matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613.2.1 Graphic matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613.2.2 Cographic matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . 623.2.3 The matroid R10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.3 Non-constant coefficient case . . . . . . . . . . . . . . . . . . . . . . . . . . 663.4 Parametrizations of tropical linear spaces . . . . . . . . . . . . . . . . . . . 69
iii
4 Implicitization and Elimination 71
4.1 The tropical variety of a generic complete intersection . . . . . . . . . . . . 734.2 Tropical elimination and implicitization . . . . . . . . . . . . . . . . . . . . 76
4.2.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 814.3 From tropical variety to Newton polytope to implicit equation . . . . . . . . 944.4 TrIm: Software for tropical implicitization . . . . . . . . . . . . . . . . . . . 99
4.4.1 Algorithms and implementation . . . . . . . . . . . . . . . . . . . . . 1004.4.2 Computing tropical discriminants . . . . . . . . . . . . . . . . . . . . 106
Bibliography 111
iv
List of Figures
1.1 Tropical hyperplane in TP2 defined by tropical linear form a. . . . . . . . . 10
2.1 Grids representing x21x22x31x32x41x42 and x11x13x31x33x41x42 for r = 4, n =3. These are the labels of v1 and v2 in Figure 2.2. . . . . . . . . . . . . . . . 20
2.2 Tropical convex hull of four points in TP2. . . . . . . . . . . . . . . . . . . . 222.3 (a) Tropical tetrahedron in TP3 with black input vertices. The circles denote
“double vertices”. (b) After wiggling we can see all degenerate edges. . . . 272.4 (a) Tropical hyperplane in TP2 with apex a. (b) The sectors at apex 0 in
TP2. (c) Tropical halfspace (a, {1, 2}) in TP2. . . . . . . . . . . . . . . . . 282.5 (a) Tropical cyclic polytope on four vertices in TP3. (b) A path corresponding
to a generator of the Alexander dual I∗. . . . . . . . . . . . . . . . . . . . . 302.6 (a) Paths in grids corresponding to two one-valent vertices in Cr,n. (b) Hor-
izontal and vertical stripes. . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.7 (a) A diagonal step. (b) Corners indicating that the corresponding monomials
are not minimal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312.8 The staircase diagram of the ideal in Example 2.19 . . . . . . . . . . . . . . 372.9 The three-dimensional part of this is the three cubes in the lower-left corner
of the figure; it also contains a pair of vertical two-dimensional flaps nearvertices C and F, and an extensive top flap connecting vertices A and B tothe other vertices. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.10 Various lifts of the tropical polytope in Figure 2.9. The faces in the back areunshaded. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.11 The J-facet ABED, and two points which demonstrate that it is not anextreme set. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.12 The face lattice of the model (L) and an artist’s rendition of it as a cellcomplex (R). The latter consists of a flat hexagon with a couple of squarespuckering up. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.13 A tropical 3-polytope which is a more complicated version of the model. . . 482.14 A five-vertex tropical polytope in TP3: cube with pendant edge. . . . . . . 502.15 Four lifts of the cube with pendant vertex. Two are bipyramids and two are
square pyramids. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
v
2.16 The face lattice of a cube with pendant edge (L) and its realization as a cellcomplex (R), a tetrahedron with one edge subdivided. . . . . . . . . . . . . 51
2.17 A tropical (2, 4)-hypersimplex, i.e. octahedron. . . . . . . . . . . . . . . . . 52
3.1 Fano plane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 603.2 2-edge-connected graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633.3 4-cycle-complement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1 Tropical plane curves and their Newton polygons . . . . . . . . . . . . . . . 814.2 The Bergman complex of a transversal matroid . . . . . . . . . . . . . . . . 854.3 Tropical construction of a bicubic surface in 3-space . . . . . . . . . . . . . 90
vi
List of Tables
2.1 Monomial matrix M2 in Example 2.8. . . . . . . . . . . . . . . . . . . . . . 232.2 Computation times (in seconds) for I, I∗, and the free resolution for tropical
cyclic polytopes Cr,n. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1 The output of TrIm for the graph in Example 4.18 . . . . . . . . . . . . . . 1044.2 The tropical variety of the implicitization problem in Example 4.18. . . . . 1054.3 The output that describe the Newton polytope of the implicit equation and
the lattice points in it, for Example 4.18 . . . . . . . . . . . . . . . . . . . . 106
vii
Acknowledgments
First and formost, I am tremendously grateful to my advisor Professor Bernd Sturmfels for
providing guidance and encouragement starting even before my first day of graduate school
and for being a constant source of inspiration and optimism ever since. I thank Professors
Lior Pachter and Alper Atamturk for agreeing to be on my qualifying exam and dissertation
committees and providing comments on a draft of this thesis.
I am indebted to great mathematics teachers and professors who nurtured my in-
terest in mathematics over the years — my parents who were my first math teachers, Ms.
Susan Kelley at ALHS in San Francisco, and many inspiring professors at UC Davis. I
am heartened by how much they cared. In particular, I thank late Professor Evelyn Silvia
for teaching me how to write mathematics and Professor Jesus De Loera for much encour-
agement when I was an undergraduate student and continued belief in me as a graduate
student. I owe a great debt of gratitude to my undergraduate advisor Professor Motohico
Mulase, without whom I might not have pursued graduate studies in the first place, for
introducing me to mathematics research and for teaching me at least as much about mathe-
matics as mathematics. I am indebted to to wonderful professors at UC Berkeley for helping
me deepen my mathematical knowledge.
I thank my coauthors Florian Block, Mike Develin, Ingileif Hallgrımsdottir, Peter
Huggins, Michael Joswig, Alex Milowski, Jenia Tevelev, and Debbie Yuster, for sharing
moments of rapture and periods of frustration and for letting me use some of the joint works
in this thesis. I thank my fellow graduate students in the math department, especially my
awesome officemates, for making it a pleasure to go into Evans Hall every day. I also thank
viii
all the people I met at conferences, workshops, and summer schools, for just being there
and talking to me. Like my advisor always says, doing mathematics is a human activity,
and it has been a splendid experience for me.
I would like to acknowledge the National Science Foundation and the Graduate
Opportunity Program at UC Berkeley for funding my graduate education. I also thank the
Institute for Mathematics and Its Applications in Minneapolis for hosting my visit in Spring
2007. The postdocs, visitors, and staff at the IMA have made this semester a highlight of
my graduate studies.
Last but not least, I am immensely grateful to my family and friends for keeping
me sane, providing me invaluable moral support, and believing in me when it is difficult to
do so myself. This thesis would have been meaningless without their love.
1
Chapter 1
Introduction
In this chapter, we introduce the necessary concepts from polyhedral geometry,
and commutative algebra, and we present the basic definitions in tropical geometry.
1.1 Polyhedral geometry
Polyhedra are fundamental objects in geometric combinatorics whose combina-
torial properties and connections to commutative algebra have been intensively studied
[Zie95, MS05]. Since tropical objects are polyhedral complexes, we will first review basic
definitions of polyhedral geometry. This section is based on [Zie95].
A point set A in Rd is convex if for any two points a,b ∈ A, the line segment
{λa + (1 − λ)b : 0 ≤ λ ≤ 1} is also in A. For any set A ⊂ Rd, the convex hull conv(A)
is the smallest convex subset of Rd containing A, or the intersection of all convex subsets
containing A. Equivalently,
conv(A) = {λ1a1 + · · ·+ λkak : ai ∈ A,λi ≥ 0, λ1 + · · ·+ λk = 1}.
2
A polytope is the convex hull of a finite set of points.
A cone in Rd is a non-empty set that is closed under taking finite linear combi-
nations with non-negative coefficients. Every cone contains 0, and the only bounded cone
is the zero cone {0}. For any set A ⊂ Rd, we can define its conical hull or positive hull
cone(A) as the intersection of all cones containing A. Equivalently,
cone(A) = {λ1a1 + · · · + λkak : ai ∈ A,λi ≥ 0}.
We also denote this by R≥0A sometimes. This is the zero cone when A is {0} or empty.
A cone is called polyhedral if it is the conical hull of a finite set of points. All the cones
we encounter in this thesis are polyhedral, hence, we will use the word “cone” to mean
“polyhedral cone” unless otherwise noted.
The Minkowski sum of two sets P,Q ⊂ Rd is the set P+Q = {x+y : x ∈ P,y ∈ Q}.
A polyhedron is a set of the form conv(A) + cone(B) where A and B are finite subsets of
Rd. Note that a polytope is a bounded polyhedron. Familiar polytopes in low dimensions
include polygons, tetrahedra, cubes, and octahedra. The Minkowski sum of two polyhedra
is also a polyhedron.
A halfspace in Rd is a set of the form {x : c1x1 + · · ·+ cdxd ≥ 0}. A fundamental
theorem in polyhedral geometry states that the polyhedra are precisely intersections of
finitely many half spaces. The dimension of a polyhedron P is the dimension its affine hull
{λ1a1 + · · · + λkak : ai ∈ P, λ1 + · · ·λk = 1}. A d-dimensional polyhedron in Rd is called
full-dimensional.
A face of a polyhedron P is the intersection of P with the boundary of a halfspace
containing P . Faces of polyhedra are again polyhedra. The empty set and P itself are
3
always faces of P . The other faces are called proper faces. The smallest and largest proper
faces P are called the vertices and facets of P , respectively. A polytope is the convex hull
of its vertices and, if it is full-dimensional, the intersection of the facet-defining halfspaces.
A k-dimensional polytope is called simple if every vertex is in exactly k facets, and it is
called simplicial if every facet contains exactly k vertices. A k-dimensional polytope with
exactly k + 1 vertices or k + 1 facets is called a k-simplex. A triangle is a 2-simplex, and a
tetrahedron is a 3-simplex. The face lattice of a polytope is the partially ordered set (poset)
of all faces, ordered by inclusion.
A polyhedral complex C in Rd is a finite collection of polyhedra in Rd such that
1. the empty polytope is in C,
2. if a polyhedron is in C, then all of its faces are also in C,
3. the intersection of any two polyhedra in C is a face of both.
The dimension of C is the largest dimension of polyhedra in it. A maximal face of C is a
polyhedron in C that is not contained in any other polyhedra in C. A polyhedral complex
is pure if all of its maximal faces have the same dimension. The ground set C is the union
of polyhedra in it. A polytopal complex is a polyhedral complex in which all polyhedra are
bounded.
A fan F in Rd is a finite collection of non-empty polyhedral cones in Rd such that
1. if a cone is in F , then all of its non-empty faces are also in F ,
2. the intersection of any two cones in F is a face of both.
4
The lineality space of a fan F is the maximal linear space that is contained in all the cones
in F . A fan F is called complete if the union of cones in F is all of Rd. The common
refinement of two complete fans F and G is the fan {C ∩ C ′ : C ∈ F , C ′ ∈ G}.
The (inner) normal fan of a non-empty polytope in Rd is the complete fan consist-
ing of the cones which are all linear functions minimal on a fixed face of the polytope. The
normal fan of a Minkowski sum of two polytopes is the common refinement of the normal
fans of the summands.
A (polytopal) subdivision of a polytope P is a polytopal complex whose ground
set is P . A subdivision is called a triangulation if all the polytopes in the complex are
simplices. Let P be a polytope in Rd with n vertices v1, . . . ,vn. Every vector w ∈ Rn
induces a subdivision of P as follows. Let Q be the polyhedron
conv{(v1, w1), . . . , (vn, wn)}+ cone(0, . . . , 0, 1) ⊂ Rd+1.
The induced subdivision is the polytopal complex consisting of the projections of the
bounded faces of Q onto the first d coordinates. The subdivisions that arise this way
are called regular.
If the polytope P is full-dimensional, the dual F of the regular subdivision is
a polyhedral complex defined as followed. To each bounded facet of Q supported by a
hyperplane c · x ≥ 1, associate the point c ∈ Rd. For each other lower face of Q, the dual
is the convex hull of points corresponding to the facets of Q containing that face. The dual
complex consists of those dual faces.
5
1.2 Commutative algebra
In this section, we will define some basic concepts from commutative algebra.
Much of the material is based on [MS05].
Let R = k[x] = k[x1, . . . , xn] be the ring consisting of polynomials in variables
x1, . . . , xn with coefficients in the field k. An ideal I ⊂ R is a subset that is closed under
addition and multiplication by elements in R.
A monomial in R is a product xa = xa11 · · · x
ann where a = (a1, . . . , an) ∈ Zn is a
vector of nonnegative integers. A monomial is called squarefree if a ∈ {0, 1}n. An ideal
generated by monomials is called a monomial ideal.
Let w ∈ Rn. Let the weight of xi be wi and the weight of a monomial xa =
∏
xai
i ∈ R be the dot product∑
aiwi. For a polynomial f =∑
cixai , we define the initial
form inw(f) to be the sum of terms cixai such that xai has maximal weight. A polynomial
f is called homogeneous with respect to a weight w if f = inw(f). An ideal is homogeneous
with respect to w if it is generated by a set of homogeneous polynomials. The initial ideal
of an ideal I ⊂ R is inw(I) = 〈inw(f) : f ∈ I〉. For most choices of w, the initial ideal is
a monomial ideal. A Grobner basis I with respect to w is a subset {f1, . . . , fr} ⊂ I such
that inw(I) = 〈inw(f1), . . . , inw(fr)〉. If the ideal I is homogeneous with respect to some
positive weight vector w ∈ Rn>0, then we can partition Rn such that all weight vectors w in
the same part give the same initial ideal of I. This partition forms a complete polyhedral
fan called the Grobner fan G of I.
We will now review some notions of monomial ideals. We will denote by R(−a)
the free R-module generated by an element in degree a. An R-module M is Zn-graded
6
if M = ⊕b∈ZnMb, and xaMb ⊆ Ma+b. A monomial ideal is a Zn-graded R-module. A
free R-module of finite rank is a direct sum F ∼= Rr of copies of R, for some nonnegative
integer r. For us, F will be Zn-graded, i.e. F ∼= R(−a1) ⊕ · · · ⊕ R(−ar) for some vectors
a1, . . . ,ar ∈ Zn. A free resolution of an R-module M is a sequence
F· : 0∂−1←−M
∂0←− F0∂1←− F1 ←− · · · ←− Fl−1
∂l←− Fl∂l+1←− 0
of maps where each Fi is a free R-module and ker(∂i−1) = im(∂i) for all 0 ≤ l + 1. Here we
assume that each map ∂i is degree-preserving, i.e. the image of a degree-m element is again
degree-m.
In [BS98], Bayer and Sturmfels defined the hull complex, a complex which yields
a resolution of a monomial ideal I ⊂ R. The construction is as follows: lift each generator
xa ∈ k[x1, . . . , xn] to a vector ta ⊂ Rd−1, where t is some large real number. Form a
polyhedron Pt by adding the positive orthant to the convex hull of these vectors. For t
sufficiently large, it was shown that this polyhedron has constant combinatorial type, and
that the complex of bounded faces of Pt yields a cellular resolution of the monomial ideal.
We give a brief review of the process which leads from a complex to a resolution.
Given any polytopal complex P with the vertices labeled by generators xa, we label
each face by the least common multiple of the generators corresponding to its vertices. We
then form a chain complex XP as follows: each face F of the polytopal complex corresponds
to a generator lying in homological degree equal to its dimension. This chain complex is to
be graded, with the degree of F equal to the label of F (which we denote xF ); each generator
maps to the appropriately homogenized signed sum of the generators corresponding to its
7
facets, i.e.
∂F =∑
±xF
xF ′
F ′, (1.1)
where the sum runs over all facets F ′ of F and the signs are chosen so that ∂2 = 0. The
complex XP is a cellular resolution if it is acyclic, i.e. has homology only in degree 0.
Given any degree b ∈ Zn, the complex X�b is defined to be the subcomplex of this
chain complex given by taking the generators with xF dividing xb. The key result about
cellular resolutions is the following:
Proposition 1.1 ([BS98]). The cellular free (algebraic) complex supported on a labelled
cell complex X is a cellular resolution if and only if X�b is acyclic over the ground field k
for all vectors b with non-negative integer entries. When the complex is acyclic, it is a free
resolution of S/I where I is the ideal generated by the monomial labels of the vertices.
1.3 Tropical mathematics
The tropical semiring T = (R ∪ {∞},⊕,⊙) is the set R of real numbers with two
binary operations called tropical addition ⊕ and tropical multiplication ⊙ defined as
a⊕ b = min(a, b), and a⊙ b = a + b, for all a, b ∈ R.
Then Rn has the structure of a semimodule over the semiring (R,⊕,⊙) with tropical addition
(x1, . . . , xn)⊕ (y1, . . . , yn) = (x1 ⊕ y1, . . . , xn ⊕ yn),
and tropical scalar multiplication
c⊙ (x1, . . . , xn) = (c⊙ x1, . . . , c⊙ xn).
8
Just as ordinary linear and polynomial algebra give rise to polyhedral and algebraic geom-
etry, the tropical linear and polynomial algebra give rise to tropical convex geometry and
tropical algebraic geometry [Stu02, SS04, Mik06, DS04, Jos05].
The term “tropical semiring” was coined by Dominique Perrin and Jean-Eric Pin
in the honor of their Brazilian colleague Imre Simon, a developer of the “min-plus algebra”
[Pin98]. The tropical (min-plus) arithmetics has had successful applications in computer
science and optimization, and recently tropical geometry has found new applications in
commutative algebra and algebraic geometry.
Tropical geometry is a piecewise linear shadow of algebraic geometry. A way in
which tropical objects arise naturally is as images of objects in Kd, a vector space over the
field of generalized Puiseux series, with respect to its non-Archimedean valuation. For a
field k, let K = k((tR)) be the set of formal power series of the form f =∑
a∈R cata where
supp(f) = {a ∈ R : ca 6= 0} is artinian, i.e. every strictly decreasing sequence of elements
in it is finite. Then K forms a field with valuation deg : f 7→ min(supp(f)) onto R. By
[Rib92], K is algebraically closed (resp. real closed) if k is algebraically closed (resp. real
closed). Note that deg(f · g) = deg(f)⊙ deg(g) and deg(f + g) ≥ deg(f)⊕ deg(g).
1.3.1 Tropical convex geometry
Tropical convexity was introduced by Develin and Sturmfels in [DS04] and has
motivated several research projects in this direction [Jos05, BY06, DY07]. A set A ⊂ Rn is
called tropically convex if for all x,y ∈ A and a, b ∈ R also (a⊙x)⊕ (b⊙y) ∈ A. Notice that
we do not put any extra condition on a and b as in usual convexity. The tropical convex
hull tconv(V ) of a set V ⊂ Rn is the minimal tropically convex set containing V in Rn with
9
respect to inclusion. Also,
tconv(V ) = {(a1 ⊙ v1)⊕ · · · ⊕ (ar ⊙ vr) : v1, . . . ,vr ∈ V and a1, . . . , ar ∈ R}.
A tropical polytope is the tropical convex hull of a finite set of points. Since any tropically
convex set A is closed under tropical scalar multiplication, we identify it with its image
under the projection onto the (n− 1)−dimensional tropical projective space
TPn−1 = Rn/(1, . . . , 1)R.
An important structure of tropical polytopes is the following:
Theorem 1.2 ([DS04]). Tropical polytopes are usual polyhedral complexes dual to regular
subdivisions of products of simplices ∆r−1 ×∆n−1.
Although the definition of tropical polytopes looks like a definition of a “tropical
linear span”, these objects do behave like polytopes. They are bounded in the tropical pro-
jective space. In [DS04], Develin and Sturmfels showed that they satisfy tropical analogues
of the Caratheodory’s theorem and the Farkas Lemma.
Let K = R((tR)) be the field of Puiseux with real coefficients as defined above.
We will see in Chapter 2 that tropical polytopes are precisely the images of polytopes in
(K>0)n under the non-archimedean valuation.
In Chapter 2, we will prove two results that relate tropical polytopes and cellular
resolutions, one using the polyhedral complex structure of tropical polytopes, and another
using the combinatorial structure of their lifts to the field of Puiseux series. This chapter
is based on [BY06] and [DY07].
10
1.3.2 Tropical linear spaces
For a tropical linear form (a1⊙ x1)⊕ · · · ⊕ (an⊙ xn) where ai ∈ R∪ {−∞}, define
its tropical hyperplane to be the set of (x1, . . . , xn) ∈ Rn such that the minimum value in
the tropical linear form is attained at least twice. See figure 1.1. Recall that the tropical
addition is taking minimum.
(1, 0, 0) = (0,−1,−1)
(0, 1, 0)
(0, 0, 1)
−a
Figure 1.1: Tropical hyperplane in TP2 defined by tropical linear form a.
In an ordinary vector space, linear spaces can be defined either as intersections
of hyperplanes or linear spans of some vectors. However, in the tropical world, not every
intersection of tropical hyperplanes or tropical linear span behaves well. In fact, arbitrary
tropical linear spans give tropical polytopes as seen before. However, for some special sets
of hyperplanes, called tropical bases, and for some special sets of vectors called cocircuits,
the intersections and tropical linear spans respectively give tropical linear spaces. This is
the topic of study in Chapter 3. We will explore the tropical bases of some families of
tropical linear spaces. We will also show that a tropical linear space is the “tropical linear
span” or the tropical convex hull of a special set of vectors at infinity.
The constant coefficient tropical linear spaces are also called Bergman fans of
11
matroids [AK06] and they play a crucial role in the theory of tropical discriminants [DFS].
1.3.3 Tropical algebraic geometry
While working with tropical varieties, we will adopt the convention in [SS04,
STY06] and define the initial form of a polynomial to be the sum of its minimal degree
terms instead of the more commonly used maximal degree. Otherwise, we will be working
in the max-tropical semiring. All the theory, of course, is the same.
Let I ⊂ C[x] be an ideal in the polynomial ring over complex numbers. The tropical
variety T (I) of the ideal I is defined to be the set of w ∈ Rn such that the initial ideal inw(I)
does not contain a monomial [SS04, BJS+]. If I is homogeneous, T (I) is naturally a subfan of
(the negative of) the Grobner fan of I. Otherwise, one can put a polyhedral fan structure on
T (I), for example, the Grobner fan structure of a homogenization of I. If I is prime, T (I) is
pure, has dimension the same as the Krull dimension of I [BG84, Stu02], and is connected
in codiimension 1 [BJS+]. They also satisfy tropical analogues of Bezout and Bernstein
theorems [Stu02]. Since they are polyhedral objects, the invariants of the original varieties
such as the dimesion and the degree can be found using polyhedral computations. Tropical
varieties are also known under the names logrithmic limit sets [Ber71], non-archimedean
amoebas [EKL], and Bieri–Groves sets [BG84].
For a polynomial f =∑
caxa ∈ C[x], the Newton polytope P of f is the convex
hull of points a ∈ Rn such that ca 6= 0. The tropical hypersurface T (f) of f is the union
of codim-1 cones in the inner normal fan of the Newton polytope P . The tropical variety
T (I) is the intersection of tropical hypersurfaces T (f) for all f ∈ I. The Newton polytope
of a product of two polynomials f and g is the Minkowski sum of the Newton polytopes of
12
f and g. Recall that the normal fan of a Minkowski sum is the common refinement of the
summands, so T (fg) = T (f) ∪ T (g). If I is the principal ideal generated by a polynomial
f , then T (I) = T (f).
Every maximal cone Γ of the fan T (I) naturally comes with a multiplicity, which is
a positive integer. The multiplicity MΓ of a d-dimensional cone Γ is the sum of multiplicities
of all monomial-free minimal associate primes of the initial ideal inv(I) in C[x1, . . . , xn]
where v ∈ Rn is any point in the relative interior of the cone Γ. The multiplicities MΓ
are an important piece of data which must also be determined when computing a tropical
variety T (I). In the hypersurface case, a maximal cone Γ of T (f) correspond to an edge
E of the Newton polytope of f , and the multiplicity of Γ is the lattice length of E, i.e. one
less than the number of integer points in E. It is with these multiplicities that a tropical
variety satisfies the balancing condition which ensures that tropical intersection numbers
are independent of choices; see [Mik06, §3] and [Kat, §6.4].
In Chapter 4, we will present a tropical approach to the implicitization problem,
which is the problem of finding the defining ideal of a parametrized variety. We give a
combinatorial construction of the corresponding tropical variety, from which one can obtain
information about the original variety, such as the Chow polytope. In the important special
case when the parametrized variety is a hypersurface, this gives a description of the Newton
polytope of the implicit equation in terms of the Newton polytopes of the polynomials that
define the parametrization, hence answering a decade-old problem posted in [SY94]. We will
also describe a computer implementation of this method. This chapter is based on [STY06].
13
Chapter 2
Tropical Polytopes and Cellular
Resolutions
Recall that a set A ⊂ Rn is called tropically convex if for all x, y ∈ A and a, b ∈ R
we have (a⊙x)⊕(b⊙y) ∈ A, where the tropical addition and scalar multiplications are taken
coordinatewise. The tropical convex hull tconv(V ) of a set V ⊂ Rn is the inclusion-wise
minimal, tropically convex set containing V in Rn. The tropical convex hulls of finite sets
of points are called tropical polytopes. In [DS], Develin and Sturmfels showed that tropical
convex hulls have a natural polyhedral cell structure which is polar dual to subdivisions of
products of simplices.
In Section 2.1, we will exploit this cellular structure to realize the tropical polytopes
as minimal cellular resolutions of some particular monomial ideals. We will use this result
to give a practical algorithm for computing tropical polytopes in Section 2.1.1 and describe
an implementation using Macaulay 2 [GS02].
14
Tropical polytopes arise naturally as images of polytopes under a non-Archimedean
valuation. This new point of view gives rise to a construction of a family of cellular res-
olutions of arbitrary monomial ideals, which generalizes the hull resolution by Bayer and
Sturmfels [BS]. This is the topic of Section 2.2
Although Develin and Sturmfels gave a polyhedral cell structure of tropical poly-
topes in [DS04], this description is unsatisfactory in many ways as a definition of the face
poset of tropical polytopes. For instance, the tropical convex hull of three points can have
as many as six vertices. A more natural approach is to define faces of tropical polytopes as
the “shadows” of faces of ordinary polytopes over the Puiseux series field. A new definition
will be proposed in Section 2.2.1. This chapter is based on joint works with Florian Block
[BY06] and Mike Develin [DY07].
2.1 Tropical polytopes via cellular resolutions
Let V = {v1, . . . ,vr} ⊂ TPn−1, vi = (vi1, . . . , vin), and P = tconv(V ). We first
describe a natural polyhedral complex structure of the tropical polytope P. Define an
unbounded polyhedron as follows:
PV = {(y, z) ∈ Rr+n/(1, . . . , 1,−1, . . . ,−1)R : yi + zj ≤ vij for all i ∈ [r], j ∈ [n]}.
By [DS04], there is a piecewise linear isomorphism between the complex of bounded faces
of PV and the tropical polytope P = tconv(V ) given by the projection (y, z) 7→ z. The
boundary complex ∂PV of PV is polar to the regular polyhedral subdivision of the product
of simplices ∆r−1 × ∆n−1 induced by the weights vij. We denote this regular subdivision
by (∂PV )∗. More precisely, a subset of vertices (ei, e′j) of ∆r−1 ×∆n−1 forms a cell of the
15
subdivision (∂PV )∗ if and only if the equations yi+zj = vij indexed by these vertices specify
a face of the polyhedron PV .
Let S = R[x11, . . . , xrn] be the polynomial ring with indeterminates xij for i ∈
[r] = {1, . . . , r} and j ∈ [n] = {1, . . . , n}. Let the weight of xij be vij .
Let A denote the (r+n)×rn integer matrix whose column vectors are the vertices
(ei, e′j) of ∆r−1 ×∆n−1, where i ∈ [r], j ∈ [n]. This defines a homomorphism Zrn → Zr+n
by eij 7→ (ei, e′j). Let L denote its kernel. The lattice ideal
J = 〈xa − xb : a,b ∈ Nrn with a− b ∈ L〉.
of ∆r−1 × ∆n−1 is generated by the 2 × 2 minors of the r × n matrix of indeterminates
[xij ]. Let I = inV (J) be the initial ideal of J with respect to weight V . Since the ideal
J is homogeneous with respect to any grading assigning the same weight to the variables
in each row, the initial ideal I is independent of the representatives of the points vi in the
tropical projective space TPn−1. In other words, if c · (1, 1, . . . , 1) is added to any vi, the
initial ideal I remains the same.
In the rest of this section, we will assume that the points v1, . . . ,vr are in generic
position, i.e., they satisfy the following equivalent conditions.
Proposition 2.1. The following are equivalent.
(1) The initial ideal I is a monomial ideal.
(2) The regular subdivision (∂PV )∗ of ∆r−1 ×∆n−1 induced by the weights vij is a trian-
gulation.
(3) The polyhedron PV is simple.
16
(4) For any k distinct points in V , their projections onto a k-dimensional coordinate
subspace do not lie in a tropical hyperplane, for any 2 ≤ k ≤ n.
(5) No k × k submatrix of the r × n matrix [vij ] is tropically singular, i.e., has vanishing
tropical determinant (e.g. see [DS04]), for any 2 ≤ k ≤ n.
Proof. (2) ⇐⇒ (3) follows directly from the polarity between the regular subdivisions of
∆r−1 ×∆n−1 and ∂PV .
(2) ⇐⇒ (5) is proven in [DS04, Proposition 24].
(4) ⇐⇒ (5) is proven in [RGST05, Lemma 5.1].
(1) ⇐⇒ (2): Statement (1) is equivalent to V being in the interior of a full dimensional
cone in the Grobner fan of the lattice ideal J . Statement (2) means that V is in the interior
of a full dimenional cone in the secondary fan N (Σ(A)) which is the normal fan of the
secondary polytope of A (for details see [Stu96]). By [Stu96, Proposition 8.15(a)], these
two fans coincide if A is unimodular, i.e., all invertible rank(A) × rank(A) submatrices
have the same determinant up to sign. We will check criterion (iv) of [Sch86, Theorem 19.3]
for total unimodularity. Fix a collection of rows of A. Split it according to containment in
the upper r× rn submatrix of the (r +n)× rn matrix A. Then the sum of the rows in each
part is a 0-1 vector. This implies that all submatrices of A have determinants 0 or ±1, so
A is unimodular.
It also follows from the unimodularity that all monomial initial ideals of J are
square free [Stu96, Corollary 8.9]. The square free Alexander dual I∗ of a square free
monomial ideal I = 〈xa1, . . . ,xak
〉 is
I∗ = ma1∩ · · · ∩ma
k
,
17
where each ai is a 0-1 vector and ma = 〈xj : aj = 1〉. See [MS05, Stu96] for details. The
following is the main result of this section.
Theorem 2.2. For a sufficiently generic set of points V in TPn−1, the tropical convex hull
P = tconv(V ) supports a minimal linear free resolution of the ideal I∗, as a cellular complex.
Moreover, the cellular structure of the minimal free resolution is unique.
Since the set of 2×2 minors of a matrix is fixed under transposition of the matrix,
we immediately see the duality between tropical convex hulls of r points in TPn−1 and n
points in TPr−1, as shown in [DS04].
Let ∆V (J) be the initial complex of J , i.e. a simplicial complex whose faces are
supports of monomials that are not in I = inV (J). In other words, the Stanley-Reisner
ideal of ∆V (J) is I = inV (J). We can identify a square free monomial m ∈ S with the
set of indeterminates xij dividing m. The vertices of ∆V (J) are xij , and the minimal
generators of I are the minimal non-faces of ∆V (J). Moreover, the minimal generators of
the Alexander dual I∗ are the complements of the maximal cells of ∆V (J). The following
lemma establishes a connection between the ideal J and the tropical convex hull.
Lemma 2.3. The cell complexes ∆V (J) and (∂PV )∗ are isomorphic. In particular, there is
a bijection between maximal cells of ∆V (J) and those of (∂PV )∗ induced by xij ←→ (ei, e′j).
Proof. The result follows from applying [MS05, Theorem 7.33] or [Stu96, Theorem 8.3] to
the product of simplices ∆r−1 ×∆n−1.
We will label the vertices of PV by the minimal generators of I∗ so that PV gives
a cellular resolution of I∗. First, we have a general lemma about simple polyhedra and
cellular resolutions.
18
Lemma 2.4. Let P be a simple polyhedron (possibly unbounded) with facets F1, . . . , Fm. La-
bel each face G of P by xaG =∏
Fi+G xi ∈ R[x1, . . . , xm]. Then the complex of bounded faces
of P supports a minimal linear free resolution of the square free monomial ideal generated
by the vertex labels.
Proof. Let X be the cellular free complex supported on the bounded faces of P . To see that
the complex is a free resolution, by Proposition 1.1, it suffices to check that each subcomplex
X�b is acyclic for all b ∈ Nm. Since the monomial labels are squarefree, it suffices to check
for b ∈ {0, 1}m. Notice that a face G is in X�b if and only if xaG =∏
Fi+G xi divides xb if
and only if G ⊂⋂
i:bi=0 Fi. Since⋂
i:bi=0 Fi is contractible, X�b is acyclic over any ground
field. Hence X is a free resolution. Moreover, every face of P has a unique label since a
face is uniquely determined by the set of facets containing it. So the resolution is minimal.
Since the polytope P is simple, any codimension-k face is contained in k facets, and its
label has degree m− k, so the resolution is linear.
We will apply this to PV to prove Theorem 2.2.
Proof of Theorem 2.2. Since V is generic, PV is simple. Hence, by Lemma 2.4, the tropical
convex hull P, which is isomorphic to the complex of bounded faces of PV , supports a
minimal linear free resolution of the ideal generated by the monomial labels of its vertices.
We will show that the labels from Lemma 2.4 coincide with the minimal generators of I∗.
The facets Fij of PV are defined by equations yi + zj = vij . Let xij be the
19
indeterminate corresponding to Fij . For a square free monomial m,
m is a vertex label of PV
⇐⇒⋂
{Fij : xij does not divide m} is a vertex of PV
⇐⇒ {(ei, ej) : xij does not divide m} is a maximal cell of (∂PV )∗
⇐⇒ {xij : xij does not divide m} is a maximal cell of ∆V (J)
⇐⇒ m is a minimal generator of I∗.
The third equivalence follows from Lemma 2.3.
By construction, the monomial labels are unique, so all the multi-graded Betti
numbers are at most one. The uniqueness of the cellular structure follows from the following
lemma.
Lemma 2.5. The monomial matrix Mi is unique up to rescaling the columns by non-zero
scalars in k if and only if βi,a ≤ 1 for all a and the monomials in Ai = {xa : βi,a 6= 0} do
not divide each other.
Proof.
⇐: Let a ∈ Nn be such that βi,a = 1 and no other monomial in Ai divides xa. Then
the a-graded piece of the ith syzygy module, which is the image of Mi in Fi−1, is a 1-
dimensional k-vector space. There is a unique linear transformation, up to rescaling, be-
tween two 1-dimensional vector spaces, so the column in Mi corresponding to S(−a) is
unique up to rescaling.
⇒: Suppose there are a,b ∈ Ai such that xa divides xb. Then the b-graded piece of the
ith syzygy module is at least 2-dimensional. In the monomial matrix Mi, we can add the
a-column to the b-column without changing its image. By the minimality of the resolution,
20
these columns are linearly independent, so this operation is not rescaling, i.e., we get a
different monomial matrix.
For the case βi,a > 1, take a = b in the previous argument.
However, the multi-graded Betti numbers already determine the tropical polytope
because in this case a face F contains a face G if and only if the monomial label of F is
divisible by the monomial label of G. Moreover, the vertex labels (the minimal generators
of I∗) determine all the other monomial labels by Lemma 2.12.
The dimension dim(U) of any subset U of TPn−1 is the affine dimension of its
projection {u ∈ Rn−1 : (0, u) ∈ U} onto the last n− 1 coordinates.
Corollary 2.6. For any face F ⊂ P, dim(F ) = deg(xaF )− (n− 1)(r − 1).
Proof. Since the simple polyhedron PV has dimension r + n − 1 and rn inequalities, by
construction the degree of a vertex label is rn − (r + n − 1) = (r − 1)(n − 1). Then the
assertion follows from the fact that the resolution is linear.
Figure 2.1: Grids representing x21x22x31x32x41x42 and x11x13x31x33x41x42 for r = 4, n = 3.These are the labels of v1 and v2 in Figure 2.2.
The monomial labels have a geometric meaning. To have a more intuitive notation,
we will represent each squarefree monomial m ∈ S with an r×n grid shaded at position (i, j)
if xij does not divide m. Hence the support of xaF is left unshaded in the grid (see Figure
2.1). Let Cj = cone{ei : i 6= j} be the closed cone which is the usual conical (positive) hull
21
of all but one standard unit vector. Suppose z = (z1, . . . , zn) ∈ P is in the relative interior
of a cell with label xaz , and it is the image of the point (y, z) ∈ PV . Then
xij 6 | xaz ⇐⇒
yi + zj = vij
yi + zk ≤ vik ∀k
⇐⇒ vij − zj ≤ vik − zk ∀k
⇐⇒ vi − z ∈ Cj ⇐⇒ vi ∈ z + Cj.
So the box (i, j) is shaded if and only if the input vertex vi lies in the sector z + Cj . See
Figure 2.4(b).
This monomial labeling is essentially the same as the labeling by types introduced
in [DS04]. Specifically, for any point z in the relative interior of a cell F in P with type(z) =
(S1, . . . , Sn), we have i ∈ Sj if and only if xij does not divide xaF . The following result
follows from [DS04, Lemma 10].
Lemma 2.7. Given the monomial label xaz of a vertex z, its coordinates can be computed
by solving the linear system
{zl − zk = vil − vik : i ∈ [r], k, l ∈ [n], xik and xil do not divide xaz}.
Example 2.8 (Four Points in TP2). Suppose we are given the following points in TP2
(r = 4, n = 3):
v1 = (0, 3, 6),v2 = (0, 5, 2),v3 = (0, 0, 1),v4 = (0, 4,−1).
They determine the tropical polytope in Figure 2.2.
The points give the weight vector V = (0, 3, 6, 0, 5, 2, 0, 0, 1, 0, 4,−1) in the poly-
nomial ring
S = R[x11, x12, x13, x21, x22, x23, x31, x32, x33, x41, x42, x43].
22
v = (0,0,1)
v = (0,5,2)2
1v = (0,3,6)
f1
2
f3
e
v = (0,4,−1)4
3
f
Figure 2.2: Tropical convex hull of four points in TP2.
The initial ideal I and its Alexander dual are
I = inV
⟨
2× 2 minors of
x11 x12 x13
x21 x22 x23
x31 x32 x33
x41 x42 x43
⟩
= 〈x33x41, x23x41, x23x31, x12x31, x31x11x42, x11x42, x31x41, x31x31, x13x21,
x33x42, x22x41, x13x32, x22x31, x11x22, x23x42, x22x33, x13x42, x13x22, x12x21x33〉,
I∗ = 〈x12x13x22x23x33x42, x12x13x22x23x41x42, x13x22x23x31x33x42,
x12x13x22x31x41x42, x13x22x31x33x41x42, x13x21x22x31x41x42,
x11x13x22x23x31x33, x21x22x31x32x41x42, x11x13x31x33x41x42,
x11x13x23x31x33x41〉.
Note that I is not generated in degree 2. Compare the minimal generators of I∗ with the
23
grids in Figure 2.2. The minimal free resolution of I∗ is of the form
0←− I∗M0←− S10 M1←− S12 M2
←− S3 ←− 0.
The tropical convex hull consists of 10 zero-dimensional faces (vertices), 12 one-dimensional
faces (edges), and 3 two-dimensional faces.
Table 2.1 shows the monomial matrix M2 in the monomial matrix notation of
[MS05, Section 1.4]. The rows correspond to the edges of the tropical polytope, and the three
columns, whose labels are omitted here, correspond to the faces f1, f2, and f3, respectively.
x12x13x22x23x31x33x42
x12x13x22x23x33x41x42
x12x13x22x23x31x41x42
x12x13x22x31x33x41x42
x11x13x22x23x31x33x42
x12x13x21x22x31x41x42
x13x22x23x31x33x41x42
x13x21x22x31x32x41x42
x11x13x22x31x33x41x42
x13x21x22x31x33x41x42
x11x13x22x23x31x33x41
x11x13x23x31x33x41x42
2
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
6
4
1 0 0
−1 0 0
−1 0 0
−1 1 0
0 0 −1
0 −1 0
1 0 −1
0 0 0
0 0 1
0 1 0
0 0 −1
0 0 1
3
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
7
5
Table 2.1: Monomial matrix M2 in Example 2.8.
2.1.1 An algorithm for computing tropical convex hulls
We will now discuss an algorithm and implementation for computing tropical poly-
topes. This algorithm has immediate applications to tropical linear spaces that will be
studied in Chapter 3.
24
Since the cellular structure of the minimal free resolution is unique, we get the
following algorithm for computing the tropical convex hull of a finite set of points in tropical
projective space.
Algorithm 2.9.
Input A list of points v1, . . . ,vr ∈ TPn−1 in generic position.
Output The tropical convex hull of the input points.
Algorithm
1. Set J = 〈2× 2 minors of the r × n matrix [xij]〉.
2. Compute I = inV (J).
3. Compute the Alexander dual I∗ of I.
4. Find a minimal free resolution of I∗.
5. Output the desired data about the tropical polytope.
Let 0 ← I∗ ← F0 ← · · · ← Fm be the free resolution computed by the algorithm,
and let Mi : Fi → Fi−1 denote the monomial matrices defining the boundary maps. Since
the free resolution is linear, the row labels of the matrix Mi are in one-to-one correspondence
with the faces of dimension i− 1, its column labels with the faces of dimension i. An entry
in Mi is nonzero if and only if its row label divides its column label, which happens if and
only if the face corresponding to its column contains the face corrresponding to its row.
Therefore the number of i-dimensional faces with k facets in the tropical convex hull is
equal to the number of columns of Mi having k nonzero entries. A face F is maximal if
25
and only if either it has dimension n − 1 or the row in Mdim(F )+1 labeled by xaF contains
zeroes only.
We can also compute the f -matrix [fij ] (0 ≤ i ≤ n − 1, 1 ≤ j) where fij is the
number of faces having dimension i and j vertices. We already know the f -vector∑
j fij
which is the sum of columns in the f -matrix. The following result in [DS04] was obtained
by counting regular triangulations of ∆r−1 ×∆n−1.
Proposition 2.10 ([DS04, Corollary 25]). All tropical convex hulls of r generic points
in TPn−1 have the same f -vector. The number of faces of dimension i is equal to the
multinomial coefficient
(
r + n− i− 2
r − i− 1, n− i− 1, i
)
=(r + n− i− 2)!
(r − i− 1)! · (n− i− 1)! · i!.
Example 2.11 (Four Points in TP2 - continuation of Example 2.8). The eighth row of
M1 corresponds to the vertex v1 (see Figure 2.2): It contains only one nonzero entry, so
the vertex v1 is contained in only one edge. The first column of M2 represents the two-
dimensional face f1 having five facets. This is reflected in five nonzero entries in that column.
The face f2 with three facets correspond to the second column which has 3 nonzero entries,
and the last column of M2 to f3. The eighth row of M2 is a zero row. This shows that the
one-dimensional face e is not contained in any other face, so it is maximal. The f−matrix
of this polytope is
10 0 0 0 0 0 · · ·
0 12 0 0 0 0 · · ·
0 0 1 0 2 0 · · ·
.
26
An alternative to free resolutions
The multi-graded Betti numbers already determine the tropical polytope because
in this case a face F contains a face G if and only if the monomial label of F is divisible
by the monomial label of G. Moreover, the vertex labels (the minimal generators of I∗)
determine all the other monomial labels by the following result, which follows from [DS04,
Corollary 14] and Corollary 2.6.
Lemma 2.12. Let F be a face of P with grid aF and let b be a grid arising from aF by
unshading one box such that no row or column is completely unshaded. Then there is a face
G ⊃ F with label aG = b and of one dimenstion higher.
Conversely, every face can be obtained this way starting from the vertices. So,
instead of algebraically computing the free resolution, we can build the fact poset com-
binatorially if we know the vertex labels, i.e., the minimal generators of I∗. We have an
implementation of this method using Macaulay 2 [GS02], Maple, and JavaView.
Non-generic input vertices
When the input vertices V are not in generic position, the initial ideal I is not
monomial. In that case, we can replace the weights V with any refinement which makes I
a monomial ideal and proceed as before to build the face poset. We can then compute the
coordinates of the vertices using Lemma 2.7 and identify vertices with the same coordinates.
Example 2.13. We can compute the following tropical tetrahedron in TP3 :
tconv{−e1,−e2,−e3,−e4}
= tconv{(0, 1, 1, 1), (0,−1, 0, 0), (0, 0,−1, 0), (0, 0, 0,−1)}.
27
These input vertices are not in generic position, so we end up with “double vertices” which
have the same coordinates (see Figure 2.3(a)). The degenerate edges can be seen when we
wiggle the input vertices to (0, 1, 1, 1), (0,−1,−0.1, 0), (0, 0,−1,−0.1) and (0,−0.1, 0,−1).
The “wiggled tropical tetrahedron” is shown in Figure 2.3(b). It is an example of a generic
tropical polytope in TP3 which is also a usual polytope.
����
����
����
����
��������������������������������
��������������������������������
��������������������������������
��������������������������������
��������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������
���������������
���������������
��������������������������������
��������������������������������
���������������
���������������
��������������������������������
����������������������������������������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������
��������������������������������
����������������������������������������������������������������������������������������������������������������������������
�������������������������������������������������������������������������������������������� �
��������������
���������������
��������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������
��������������������������������
��������������������������������
����
����
����
����
(a) (b)
Figure 2.3: (a) Tropical tetrahedron in TP3 with black input vertices. The circles denote“double vertices”. (b) After wiggling we can see all degenerate edges.
Tropical halfspaces
Tropical halfspaces introduced in [Jos05] give us an exterior description of tropical
polytopes. We can extend our algorithm to find such a description.
The tropical hyperplane at the apex a ∈ TPn−1 is the set which is the union
of boundaries of the sectors a + Ci (see Figure 2.4). For a ∈ TPn−1, ∅ 6= A ( [n], the
set a +⋃
i∈A Ci is a closed tropical halfspace (see Figure 2.4(c)). Tropical halfspaces are
tropically convex, and a tropical polytope P is the intersection of the inclusionwise minimal
halfspaces containing it [Jos05]. The apex of such a minimal halfspace must be a vertex
28
−b2
C3
C1C0 a
(a) (b) (c)
Figure 2.4: (a) Tropical hyperplane in TP2 with apex a. (b) The sectors at apex 0 in TP2.(c) Tropical halfspace (a, {1, 2}) in TP2.
of P on the boundary [Jos05, Lemma 3.6]. Recall that the box (i, j) in the grid label of a
vertex v is shaded if and only if vi ∈ v + Cj . Hence P is the intersection of the halfspaces
v +⋃
i∈A Ci such that v is a vertex of P and A is a minimal subset of columns in the
corresponding grid of v such that the shaded boxes in those column cover all the rows. This
description is redundant in general. We may be able to refine this result as follows.
Conjecture 2.14. In the generic case, a minimal halfspace with respect to P has the form
v +⋃
i∈A Ci where v is a vertex of P, and, in the grid label of v, the shaded boxes in the
columns in A form a partition of [r].
For example, in Figure 2.2 (Example 2.8), no halfspace at v2 is minimal because
no set of columns form a partitions of the rows. However, the sectors 2 and 3 at the point
(0, 3, 2) form a minimal halfspace, and shaded boxes in columns 2 and 3 of the grid form
a partitions of the rows. The converse of the conjecture above is not true, i.e., there are
non-minimal halfspaces of the form described.
Experiments with computation time
We experimented with computing tropical cyclic polytopes Cr,n (which will be
defined in the next section) with r input vertices in n − 1 (projective) dimensions. We
29
used Macaulay 2 [GS02] on a Sun Blade 150 (UltraSPARC-IIe 550MHz) computer with
512MB memory. The computation became infeasible when rn > 80 or so, although r =
30, n = 3 worked. The main problem was the insufficient amount of memory. Some sample
computation times for tropical cyclic polytopes are given in Table 2.2. We see from the
data that computing the Alexander dual can be a problem. This can be made faster using
the software Frobby by Bjarke Roune [Rou07].
n r Initial ideal Alexander dual Free resolution
3 30 74 433 24 21 64 944 236 10 15 221 278 10 70 4169 1106
Table 2.2: Computation times (in seconds) for I, I∗, and the free resolution for tropicalcyclic polytopes Cr,n.
A computation problem we didn’t address is the following:
Question 2.15. How does one obtain a vertex description of a tropical polytope from its
hyperplane description?
2.1.2 Tropical cyclic polytopes
As an application to our method, we will compute and analyze the tropical cyclic
polytopes. Define tropical cyclic polytopes as Cr,n = tconv{v1, . . . ,vr} ⊂ TPn−1, where
vij = (i− 1)(j − 1) for i ∈ [r], j ∈ [n]. Since (i − 1)⊙(j−1) = (i − 1)(j − 1), this is tropical
exponentiation. The Cr,n are generic because the minimum in any k×k minor of the matrix
[vij ] is attained uniquely by the antidiagonal. An example of a tropical cyclic polytope is
shown in Figure 2.5(a).
We will show that the set of 2 × 2 minors of [xij] forms a Grobner basis with
30
respect to V , using the Buchberger Algorithm. If the leading terms of two 2× 2 minors do
not contain a common indeterminate, then both terms in the corresponding S-polynomial
are divisible by one of the original 2×2 minors. Suppose they have the form xijxkl−xilxkj,
and xijxpq − xiqxpj. Then the S-polynomial is xilxkjxpq − xiqxpjxkl. The weights of the
monomials are two of the terms in the 3 × 3 minor of [vij] corresponding to rows i, k, p
and columns j, l, q. The unique minimum of this minor is attained by the antidiagonal.
Therefore the leading term of the S-polynomial cannot be the antidiagonal, so it is divisible
by the leading term of a 2 × 2 minor. Hence the initial ideal I is the diagonal initial ideal
generated by the binomials which are on the diagonals of the 2×2 minors. This correspond
to the staircase triangulation of ∆r−1 ×∆n−1.
(0,0,0,0)
(0,1,2,3)
(0,1,2,3)
(0,2,4,6)
(0,2,4,6)
(0,3,6,9)
���������
���������
������������
�����������������
����������
���������������
���������
���������
���������
���������
���������
���������
���������
���������
.....
x x
x x x
11 12
x 2r2221
n1 n2 nr
... 1r
...
xxx
... ... .
...
(a) (b)
Figure 2.5: (a) Tropical cyclic polytope on four vertices in TP3. (b) A path correspondingto a generator of the Alexander dual I∗.
Consider a path in an r×n grid representing indeterminates xij , which goes from
the lower left corner to the upper right corner, only moving either right or up at each step
as in Figure 2.5(b). Such paths are precisely the maximal sets, with respect to inclusion,
that do not contain any set {xij , xkl}, i < k, j < l. Hence their complements correspond
31
to the minimal generators of the Alexander dual I∗, which are the monomial labels of the
vertices of Cr,n.
��������
��������
����������������������������������������������������������������������������������������
��������
��������
��������
��������
��������
��������
��������
��������
����������������������������������������
����������
������.
������������������������
��������
...����������������
��������
..��������
(a) (b)
Figure 2.6: (a) Paths in grids corresponding to two one-valent vertices in Cr,n. (b) Hori-zontal and vertical stripes.
The labels of the faces of the tropical cyclic polytope Cr,n are obtained by unshad-
ing the boxes on the paths so that the remaining shaded set still intersects every row and
every column. For example, there are two 1−valent vertices with grids corresponding to
the paths in Figure 2.6(a). The two edges containing these vertices are the only maximal
1−faces, whose labels are obtained by unshading the lower right corner and the upper left
corner, respectively.
We can identify a vertex of Cr,n with the Young diagram above (or below) the
corresponding path in the r × n grid. Then the 1−skeleton of Cr,n is the Hasse diagram of
the Young lattice of the Young diagram fitting in an (r − 1)× (n− 1) grid.
���������
������������
������
���������
���������
���������
���������
������������������
���������
���������
������������������
���������
���������
���������
(a) (b)
Figure 2.7: (a) A diagonal step. (b) Corners indicating that the corresponding monomialsare not minimal.
The shaded part in the label of a k−dimensional face contains the diagonal steps as
32
in Figure 2.7(a) exactly k times because every time we shade in such a corner, the dimension
goes down by one. By straightforward counting, we get that
#k − faces in Cr,n =
(
r + n− k − 2
r − k − 1, n− k − 1, k
)
as seen in Proposition 2.10. That is, out of the r + n− k − 2 steps we take from the lower
left corner to the upper right corner, we take r − k − 1 steps up, n− k − 1 steps right, and
k steps diagonally.
Proposition 2.16. The ordinary generating function for the number Mr,n,k of maximal
k−faces of the tropical cyclic polytope Cr,n equals
∑
r≥1,n≥1,z≥0
Mr,n,kxrynzk =
(
xy
1− y+
yx2
1− x
)/(
1− z
(
xy
1− y−
yx2
1− x
))
.
Proof. A face is maximal if and only if the set of shaded boxes in the r × n grid does not
contain any corners as in Figure 2.7(b). Then Mr,n,k is equal to the number of (k+1)−tuples
of either horizontal or vertical stripes of boxes, as in Figure 2.6(b), such that the sum of
the widths equals n and the sum of heights equals r. The proposition follows from basic
properties of ordinary generating functions.
Moreover, every k−dimensional face contains precisely 2k vertices because every
diagonal step as in Figure 2.7(a) gives 2 ways of shading in the corners, and there are exactly
k such diagonal steps. From this it is easy to see that every k−dimensional face has the
combinatorial structure of a k−dimensional hypercube. Therefore, the f−matrix of Cn,r is
very simple: fk,2k =(
r+n−k−2r−k−1,n−k−1,k
)
, and all other entries are 0.
33
2.2 Cellular resolutions via tropical polytopes
In the previous section, we saw that for a generic tropical polytope, we can con-
struct a monomial ideal whose minimal free resolution is given by the tropical polytope. Not
all monomial ideals arise this way. More specifically, the monomial ideals constructed this
way are precisely the initial monomial ideals of the lattice ideal of a product of simplices.
In this section, we will explore a complementary direction. For an arbitrary monomial
ideal, we will construct a family of cellular resolutions which have a tropical polytope as a
common “shadow”.
Two important operations on monomials are compatible with tropical operations
on the exponent. For monomials xa and xb, their least common multiple is x−(−a⊕−b), and
their product is xa⊙b.
As mentioned in the introduction, tropical discrete geometry arises naturally is as
the image of ordinary discrete geometry over the Puiseux series field with real exponents
K := R((tR)) under the degree map deg : K → R sending an element to its leading (lowest)
exponent. It is naturally an ordered field, where a < b if the leading coefficient of b − a
is positive; its positive elements K>0 comprise the set of all elements with positive leading
coefficient. The field K is real-closed, with t as an infinitesimal. As such, the usual theory
of discrete geometry applies in Kd, and in particular we can define the convex hull of a
point set as the set of all convex combinations of the points. We can also extend the degree
map to deg : Kd → Rd where the degree map is applied coordinatewise. Then the tropical
polytopes are simply the images of objects up above:
Proposition 2.17. Suppose P = tconv(v1, . . . , vk) ⊂ TPd−1, where each vi has first coordi-
34
nate zero. Define lifts vi ∈ deg−1(vi) ∈ Kd>0, and define P = conv({vi}). Then P = deg(P ).
Proof. First, we show that P ⊂ deg(conv({vi})). Suppose that we have a point x =
⊕
ci ⊙ vi ∈ P . Since P ⊂ TPd−1, we can add a constant to each ci such that the largest
ci, without loss of generality c1, is equal to 0. It is easy to lift the ci’s to ci ∈ K with
deg(ci) = ci such that ci > 0 and∑
ci = 1: lift every ci < 0 to tci , and lift ci = 0 to
(1−∑
cj<0 tcj )/|{j | cj = 0}|.
But then we claim that deg(∑
civi) = x. This follows immediately: since there is
no cancellation of the leading terms as all leading coefficients of the ci’s and vi’s are positive,
for each coordinate j we have
deg
(
∑
i
cijvij
)
= max(deg(cijvij)) = max((ci)j + (vi)j) = xj
as desired.
For the reverse direction, using the same logic, it is easy to see that deg(∑
αivi) =
⊕
deg(αi)⊙ vi, again due to the lack of cancellation of leading terms, which shows that the
set deg(conv({v1, . . . , vk})) is a subset of P , completing the proof.
We call the polytope P a lift of P . Note that we can take any lifts of the vertices of
the polytope, as long as the lifted points are in the positive orthant Kd>0; the lift operation
always preserves convex hulls. As an aside, the stipulation that each vi has first coordinate
zero is merely for simplicity; we could ignore this and just look at the facial structure of
the cone generated by the lifts of the vertices. Giving each vi first coordinate zero amounts
to slicing this cone with the hyperplane x1 = 1.
Theorem 2.18. Let P be the tropical polytope given by the convex hull of the points (0,−a),
where a ranges over the exponent vectors of the minimal generators of a monomial ideal
35
I in k[x1, . . . , xd−1]. Take any lift P of P and add the positive orthant in the last d − 1
coordinates, {0}× (K>0)d−1, to P to obtain a polyhedron P
+. Then the complex of bounded
faces of P+
yields a cellular resolution of I.
Proof. We need only to check that for each b = (b1, . . . , bd−1), the bounded faces of P+
with labels dividing b form an acyclic complex. Let the coordinates of the space Kd that
P+
lives in be given by z0, z1, . . . , zd−1, and consider the halfspace H in Kd given by
t−1/2 − tb1z1 − · · · − tbd−1zd−1 ≥ 0. Then a given vertex v ∈ P+
with deg(v) = (0,−a) is in
the halfspace H if and only if xa divides xb. Thus the bounded faces in the halfspace H
are precisely those in the subcomplex X�b. Applying a projective transformation sending
this hyperplane to infinity yields X�b as the complex of bounded faces of some polytope,
which is acyclic. Therefore, each X�b is acyclic, and thus by Proposition 1.1 this complex
yields a cellular resolution of I.
These cellular resolutions can be smaller than the hull resolution. For instance, it
was shown in [Dev04] that no face of the hull complex can have more than (d− 1)! vertices;
this is not true for these cellular resolutions. For example, we can have arbitrarily many
generators of an ideal all lying on the boundary of the same tropical halfspace, and we can
find a lift where these all lay on a single facet, as in the example below.
Example 2.19. Consider the ideal
I = 〈x2z3, xyz3, y2z3, x2y3, xy3z, y3z2, x3y2, x3yz, x3z2, x6, y6, z6〉.
The staircase diagram of I, which is drawn using Gfan [Jen], is depicted in Figure 2.8. The
36
tropical convex hull P of the following points
(0,−2, 0,−3), (0,−1,−1,−3), (0, 0,−2,−3),
(0,−2,−3, 0), (0,−1,−3,−1), (0, 0,−3,−2),
(0,−3,−2, 0), (0,−3,−1,−1), (0,−3, 0,−2),
(0,−6, 0, 0), (0, 0,−6, 0), (0, 0, 0,−6)
corresponding to the twelve generators of I is the negative of the complement of the staircase
diagram inside the 6×6×6 box in R3≥0 including the origin. According to [Dev04], no face of
the hull complex can have more than six vertices. However, notice that the first nine vertices
of P lie on the boundary of the same valid tropical halfspace H with apex (0,−3,−3,−3),
i.e. the first nine vertices lie in the tropical hyperplane defined by the tropical linear form
x0 ⊕ (3⊙ x1)⊕ (3⊙ x2)⊕ (3⊙ x3), and they are all on the boundary between sector 0 and
sectors 1, 2, 3. We can lift the halfspace H to a halfspace H in K4>0, say,
H = {(X0,X1,X2,X3) : −(1 + t + t3)X0 + t3X1 + t3X2 + t3X3 ≥ 0}.
Any lift of the last three points to K4>0 lie in the interior of H, and the following lifts of the
first nine points lie on the boundary of H ∩K4>0:
(1, t−2, 1, t−3), (1, t−1, t−1, t−3 + t−2 − 2t−1 + 1), (1, 1, t−2, t−3),
(1, t−2, t−3, 1), (1, t−1, t−3 + t−2 − 2t−1 + 1, t−1), (1, 1, t−3, t−2),
(1, t−3, t−2, 1), (1, t−3 + t−2 − 2t−1 + 1, t−1, t−1), (1, t−3, 0, t−2),
so these nine points lie on a facet of the lift P+, and the cellular resolution given by this
lift is smaller than the hull resolution.
These lifts include the hull complex as a special case, where each point a is simply
lifted to ta. Like the hull complex, each of them includes the Scarf complex as a subcomplex;
37
Figure 2.8: The staircase diagram of the ideal in Example 2.19
this complex is a simplicial complex where a set of vertices forms a face if the corresponding
set of generators has a unique least common multiple among all LCM’s of sets of generators.
Proposition 2.20. Suppose that a set S of the generators of I has a unique least common
multiple. Then S forms a (simplicial) bounded face in every P+.
Proof. The proof is by induction. We claim that for every v ∈ S, S \ {v} has a unique least
common multiple. Suppose it did not; let T be another set with the same least common
multiple. Then T ∪ {v} has the same least common multiple as S \ {v} ∪ v, so since S has
a unique least common multiple, we must have T ∪ {v} = S. The only other set for which
this is true is T = S, but this contradicts the statement that S has a unique LCM.
Therefore, by induction on the dimension, every proper subset of S forms a face.
Let the least common multiple of S be xb, and as before consider the linear functional
given by f(x) =∑
tbixi. No generator not in S divides xb (otherwise adding it to S gives
the same LCM), and so the hyperplane f(x) = t−1/2 separates the vertices in S from the
vertices outside of S. Again applying a projective transformation to map this hyperplane to
infinity, the induced subcomplex of bounded faces on S must be acyclic. Since it contains
38
every proper subset of S as a face, it must also contain S itself. So S is a bounded face of
P+
as desired.
The following result relates the genericity of tropical polytope to generic mono-
mial ideals defined in [MS05, Chapter 6]. It follows from [MS05, Proposition 6.26] and
Propositions 2.26 and 2.20.
Proposition 2.21. If the negatives of exponent vectors of the minimal generators of a
monomial ideal I are in general position tropically, then I is a generic monomial ideal.
In this case, all the lifts have the same combinatorial type, and the cellular reso-
lution is given by the Scarf complex.
Bayer and Sturmfels referred to the hull complex as a canonical free resolution of
a monomial ideal. Here, however, we see that the hull complex is just one of a family of
resolutions arising from different lifts of the corresponding tropical polytope. For instance, in
Example 2.31 (Figure 2.10), we come up with different free resolutions of the corresponding
monomial ideal. Note that we can always obtain a simplicial resolution by taking a generic
lift. It would be interesting to answer the following questions:
Question 2.22. What does the family of lifts of a given tropical polytope look like? Are
there generalized hull resolutions we can only get by lifting to nonconvergent power series? Is
there a reasonable algorithm for picking a small resolution from among these hull complexes?
Example 2.23. Consider the ideal I = 〈x2z, x2y, xy2z5, xy3z4, xy4z3, xy5z2〉 be an ideal
in the polynomial ring R = k[x, y, z] over any field k. The tropical polytope of interest in
this case is the tropical convex hull of {A,B,C,D,E, F} = {(0,−2, 0,−1), (0,−2,−1, 0),
39
A
A A
B
B
B
C
D
D
E
D
E
F
E
F
Figure 2.9: The three-dimensional part of this is the three cubes in the lower-left corner ofthe figure; it also contains a pair of vertical two-dimensional flaps near vertices C and F,and an extensive top flap connecting vertices A and B to the other vertices.
(0,−1,−2,−5), (0,−1,−3,−4), (0,−1,−4,−3), (0,−1,−5,−2)}. It is depicted in Figure
2.9. Figure 2.10 shows several lifts of this polytope, with different combinatorial types (face
lattices). The shaded faces are those that stay bounded after adding the positive orthant
to the lift P . The first two lifts give free resolutions of the form
0←− I ←− R6 ←− R8 ←− R3 ←− 0
and they are both minimal, although they have different cell structures. The last lift gives
a resolution of the form
0←− I ←− R6 ←− R9 ←− R4 ←− 0
which is not minimal.
A natural question then is:
40
C
E
F
A B
C
D E
F
A
B
C
D E
F
A
B
D
Figure 2.10: Various lifts of the tropical polytope in Figure 2.9. The faces in the back areunshaded.
Question 2.24. Do all cellular resolutions of a monomial ideal occur in this way?
It was shown by Velasco that there exist monomial ideals whose minimal free resolution is
not cellular [Vel]. An affirmative answer to this question may give us a new approach to
this result.
2.2.1 Faces of tropical polytopes
We will now discuss the notion of faces of tropical polytopes. Ultimately we hope
to be able to read off from these the family of cellular resolutions defined above. Since the
facial structure of tropical polytopes is not the main focus of this chapter, the reader is
referred to [DY07] for more discussion.
The idea is that the faces of tropical polytopes should come from the faces of the
lifts to the Puiseux series field. In a sense, each one of these lifts yields a candidate for
the face lattice of the tropical polytope P . The problem is that when the points are not in
(tropically) general position, the lifts can have different combinatorial structures as seen in
Example 2.23.
Let A = [aij] be a d×d matrix whose columns are considered as d points in TPd−1.
The tropical determinant of A is defined by the formula ⊕σ∈Sd(⊙ai,σi
) where Sd denotes
41
the group of permutations of d elements. We say that A is tropically non-singular if the
maximum in its tropical determinant is attained uniquely. In this case, the tropical sign of
A is defined (as in [Jos05]) to be the sign of the permutation that attains the maximum.
Otherwise, the tropical sign is defined to be 0. Let A be a d × d matrix with entries in
K>0 whose degree is A. If the tropical sign of A is not zero, then the sign of the unique
permutation that attains the maximum in the tropical determinant is also the sign of the
leading term of the determinant of A. This observation leads to the following.
Lemma 2.25. For a tropical polytope P with at least d vertices in TPd−1, the oriented
matroid structure of any lift P must refine the partial oriented matroid structure of P given
by the tropical signs on each subset of the d vertices.
On the other hand, there may be a point configuration whose oriented matroid
refines the partial oriented matriod of vertices of P but cannot be obtained as a lift. The
oriented matroid of the tropical polytope in Example 2.23 is the same as that of a square
pyramid with two points at the cone point, but this point configuration cannot be attained
as a lift since distinct points must be lifted to distinct points.
Proposition 2.26. If a tropical polytope P with at least d vertices in TPd−1 is in general
position, then all lifts P are simplicial and have the same oriented matroid structure.
Proof. Let V be a matrix whose columns are vertices of P . The assumption that P is in
general position (in the sense of [BY06, Proposition 4]) implies that all maximal d × d-
submatrices of V are tropically non-singular. By the previous lemma, the tropical signs of
these submatrices determine the chirotope of all the lifts P . Moreover, since these signs are
all non-zero, the lifts are simplicial.
42
It is still possible for lifts of a non-generic tropical polytope to be simplicial and
have the same face lattice. See the tropical (2,4)-hypersimplex in Example 2.35.
For a generic (or simplicial) tropical polytope, every such lift to the Puiseux series
is simplicial and has the same oriented matroid as in Proposition 2.26, hence the same
face lattice. In this case the faces of tropical polytopes should be images of faces of the
lifts. In fact, Joswig’s definition of faces in [Jos05] coincides with these images for generic
tropical polytopes. Unfortunately, when the points are not in generic position, Joswig’s
definition of face lattice has undesirable properties. We propose a different generalization
of this definition to non-generic tropical polytopes as follows:
Definition 2.27. A k-face of a tropical polytope is a minimal k-dimensional subset which
is the union of images of k-faces of every lift under the degree map v.
In other words, the face poset of a tropical polytope is the “common coarsening” of face
posets of all the lifts. The images of faces of lifts under the degree map were called fatoms
(face atoms). The following proposition provides the link between lifted hyperplanes and
tropical hyperplanes.
Proposition 2.28. Let H = {X ∈ Kd : f1X1 + · · ·+fdXd = 0} ⊂ Kd be a hyperplane, and
define H+ := {X ∈ Kd : f(X) ≥ 0}. The image of H under the degree map is a tropical
hyperplane with apex (−deg(f1), . . . ,−deg(fd)), the image of H+ ∩ (K>0)d consists of the
union of the closed sectors indexed by {i : fi > 0}, and the image of H ∩ (K>0)d is the
boundary of this tropical halfspace. (If some fi is equal to 0, then the apex of this tropical
hyperplane has i-th coordinate equal to infinity.)
Proof. Let z = deg(X). If f(X) = 0, then the leading term of f(X) has to cancel. The
43
leading exponent of fiXi is deg(fi) + zi, so the maximum of these d expressions has to
be achieved at least twice. This means that z is in the indicated tropical hyperplane.
Conversely, if z is in the tropical hyperplane, then f(tz) has ties in the leading terms, and
it is trivial to adjust leading coefficients and fill in subleading terms to find a lift of z which
lies on H.
Suppose now that X ∈ H+ ∩ (K>0)d. If z is not in any closed sector indexed by
fi > 0, then each fi for which the maximum of all deg(fi) + zi is achieved has negative
leading coefficient. Since these are the terms which contribute to the leading term of f(X),
f(X) must be negative, a contradiction. Conversely, if z is in a closed sector indexed by,
without loss of generality, f1 > 0, then deg(fi) + zi is maximized by i = 1. Therefore
f1X1 contributes to the leading term of f(X). Finding a lift with sufficiently large leading
coefficient of X1 then ensures that f(X) > 0. This proves the second assertion.
The third statement can be proven the same way as the first, and by noting that
the leading term of f(X) can cancel for X ∈ (K>0)d if and only if deg(X) lies in both the
sectors {i : fi > 0} and sectors {j : fj < 0}.
Armed with this, we can prove a crucial step in this discussion, namely that the
boundary of the lifted polytope indeed maps to the boundary of the tropical polytope.
Proposition 2.29. Let P be a lift of the tropical polytope P ⊂ TPd−1. Then the image of
the boundary of P under the degree map is precisely the boundary of P .
Proof. Every point in the boundary of P lies in some facet. Since the degree map preserves
convex hulls, it must map to a point in the convex hull of the vertices of this facet. However,
by Proposition 2.28, these vertices all lie in a tropical hyperplane which bounds P (namely
44
the image of the hyperplane defining the facet of the lifted polytope), and, since tropical
hyperplanes are convex, the image of our point must also lie on the boundary of P .
For the converse, suppose we have some point v in the boundary of P . This point
is the image of some point in P , since lifts preserve convex hulls; however, a priori, this
point need not lie on the boundary. Consider a tropical hyperplane H with apex v. Not
every open sector of this hyperplane contains a point in P , since that would imply that v
is in the interior of P . Therefore, we can partition the sectors of this hyperplane into a
pair (A,B) such that the A-sectors contain P , with B nonempty. Suppose without loss of
generality that A = {1, . . . , k} and B = {k + 1, . . . , d}. Define a linear functional f on the
lift space via
f = c1t−v1x1 + · · · + ckt
−vkxk − t−vk+1xk+1 − · · · − tvdxd, (2.1)
where the ci’s are positive real constants. Since P lies in the union of the A-sectors of H,
every vertex w of P has wi − vi maximized for some i ∈ [k]. Therefore, the leading term of
the expression given by f(w) contains some positive summand from among the first k terms
of f . So, if we make the ci’s large enough, we can ensure that f(w) > 0 for all vertices w
of P . Fix a set of such ci’s.
But we can easily lift the point v to a point v with f(v = 0); for instance, simply
taking vi = 1ci
tvi for i ∈ [k] and vi = kd−k tvi otherwise. Hence v has a lift which lies
outside of P . So, we have constructed a lift of v which lies inside of P , and a lift which
lies outside of it. The line segment between these two lifts, each of which has positive
leading coefficients, consists entirely of other lifts of v, and must intersect the boundary of
P somewhere, completing the proof.
45
The degree map is well-behaved; for instance, it also does not increase the dimen-
sion of a face.
Proposition 2.30. If F is a k-face of P , then the image of F under the degree map is at
most k-dimensional.
Proof. Suppose we have a k-face F of P . Triangulate this with its original vertex set,
dividing it into a number of k-simplices which are the convex hulls of k + 1 lifts of vertices
of P . The image of each such simplex under the degree map, by Proposition 2.17, is the
convex hull of k+1 points in TPd−1, and is therefore at most k-dimensional (see e.g. [DS04].)
Therefore, the image of F under the degree map is the finite union of k-dimensional things,
and is therefore k-dimensional.
Example 2.31 (The model). Let P be the tropical convex hull of
(A,B,C,D,E, F ) = (0201, 0210, 0125, 0134, 0143, 0152) ⊂ TP3,
as seen in Example 2.23.
This is a tropical 3-polytope lying in three-space, shown in Figure 2.9. Its nat-
ural polyhedral decomposition dual to a subdivision of ∆5 × ∆3 consists of three three-
dimensional cells, all unit cubes, and a number of two-dimensional flaps. It has 29 vertices,
clearly too many for the tropical convex hull of 6 points.
The facets defined in [Jos05], which we will call J-facets, are ABDC, ABED, ABFE,
ABFC, and CDEF. They exhibits a number of the problems. For example, the facet ABED
is not an extreme set, i.e. we can find two points not on the facet but whose convex hull
intersects the facet (Figure 2.11). The fundamental problem here is that in some lifts,
46
������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������������
A
B
C
E
F
D
x
y
z
Figure 2.11: The J-facet ABED, and two points which demonstrate that it is not an extremeset.
ABED itself is not a facet of the lift. In those lifts, we can take two points not in the lift
of ABED for which the tropical line segment connecting them pierces the convex hull of
ABED.
Moreover, the facets ABDC, ABED, ABFE, and ABFC intersect badly. To be
precise, their intersections are two-dimensional; the convex hulls of ABDC and ABED
intersect in the convex hull of ABD, a two-dimensional object. Furthermore, according to
Joswig’s definition, A and B are not actually vertices of this polytope. If we intersect the
facet vertex sets setwise, we never get the singletons A and B, but merely the atom AB. So
according to Joswig’s definition, this object is a pyramid over a square, with vertices AB
(apex), C, D, E, and F. But this misses some aspects of the tropical polytope, as A and B
are both vertices in the sense that if we remove either, the convex hull changes. There are
also examples where the J-face lattice does not have correct height, as in Example 2.32.
By our new Definition 2.27, the tropical polytope P has three facets; the upper
47
C F
A B
D E
A B C D E F
CDEFABCFunderbelly
edges (L to R): AB, BF, AC, CD, DE, EF, CF
Figure 2.12: The face lattice of the model (L) and an artist’s rendition of it as a cell complex(R). The latter consists of a flat hexagon with a couple of squares puckering up.
shell comprising the convex hull of {A,B,C, F}, the lower face comprising the convex hull of
{C,D,E,F}, and the “underbelly” given by the union of the convex hulls of {A,B,C,D},
{A,B,D,E}, and {A,B,E, F}. The problematic pairs of faces which intersect improperly,
such as (ABCD,ABDE), have all been combined into the underbelly. This underbelly is a
facet because in every lift, it is the union of facets of the lifted polytope, and it is minimal
with this property. The face poset consists of three 2-faces, seven 1-faces (AB, AC, CD,
DE, EF, CF, BF), and six 0-faces (the vertices); see Figure 2.12.
Example 2.32. Let P be the tropical convex hull of
(A,B,C,D,E, F,G,H, I) = (0301, 0310, 0224, 0233, 0242, 0158, 0167, 0176, 0185) ⊂ TP3.
This polytope (Figure 2.13) is merely a three-tiered version of the two-tiered pre-
vious example. The J-facets are ABFI, ABDC, ABED, CDEGF, CDEHG, and CDEIH. By
Joswig’s definition, taking setwise intersections of these J-facets, we have a chain of faces
given by D ⊂ CDE ⊂ CDEG ⊂ CDEGF . This chain is too long to live in the boundary
of a 3-polytope; the face poset does not have the correct height.
If P is in general position, by Proposition 2.26, the face lattice of P is the same as
48
Figure 2.13: A tropical 3-polytope which is a more complicated version of the model.
that of any lift. In this case, the face lattice is determined by combinatorially intersecting
the vertex sets of facets, and this new definition is the same as that of Joswig’s. In essence,
this definition gets rid of many of the problems with Joswig’s definition in the non-generic
case via the following method: whenever two J-facets fundamentally intersect improperly,
they are declared to be part of the same facet of P . Note also that it is clear that the 0-faces
of P are always precisely the vertices.
These and other examples lead us to formulate the following conjectures about our
new definition, each of which represents an improvement on Joswig’s.
Conjecture 2.33.
1. The k-faces of tropical polytopes are extreme sets, i.e. there should not be two points in
the polytope not in a face such that the tropical line segment connecting them intersects
the face.
2. The (topological) boundary of a k-face is a union of (k − 1)-faces. Thus, the faces fit
49
together to form a cell complex.
3. The homology of this cell complex is that of a sphere.
4. The intersection of two faces is well-defined (i.e. does not depend on the lift), and is
itself a contractible union of faces. Note that in some cases, a (for instance) 3-face
in a lift which maps to a two-dimensional object under the degree map may be part of
the lifts of two 3-faces of P . In this case, though, its contribution to the intersection
will still be two-dimensional.
5. k-faces of tropical polytopes are always contractible.
6. The faces of a tropical polytope do not depend on the provided vertex set. In other
words, if we consider P as the convex hull of a different set of points (and form lifts
by lifting this different set of points), the k-faces of P will be the same for all k.
We now present two more examples of tropical polytopes.
Example 2.34. Let P be the convex hull of {A,B,C,D,E} = (0101, 0011, 0002, 0001, 0110) ⊂
TP3, as depicted in Figure 2.14.
Discussion: This is a cube with a pendant edge. We previously encountered the
convex hull of the vertices ABC; this tropical polytope consists of the union of three facets
of a cube surrounding a vertex (0112 here). Adding the point D realizes the unit cube as
a tropical polytope; its facets are simply the convex hull of ABC, which consists of three
facets of the cube, and the convex hulls of ABD, ACD, and BCD, each of which is another
facet of the cube. This is a perfectly normal tropical tetrahedron (albeit not one in general
position.) All lifts of this polytope are simplices.
50
x
y
z
B
C
D
E
A
Figure 2.14: A five-vertex tropical polytope in TP3: cube with pendant edge.
C
D
A
BE
C
D
A
BE
C
D
EA
B
C
E
D
A
B
Figure 2.15: Four lifts of the cube with pendant vertex. Two are bipyramids and two aresquare pyramids.
51
DE
A
B
C
A B C D E
ACD BCD AEBC AEBD
edges (L to R): AC, BC, AD, BD, CD, AE, BE
Figure 2.16: The face lattice of a cube with pendant edge (L) and its realization as a cellcomplex (R), a tetrahedron with one edge subdivided.
Adding the pendant vertex, however, produces a variety of possible lifts. There
exist lifts where the lifted point E lies in various places with respect to the tetrahedron
ABCD; for instance, it could be coplanar with 0, e1, and e2 (as these four points lie on
a tropical hyperplane), or with e1, e2, and e3. Even if it is in general position, it can be
beyond different faces of the tetrahedron. See Figure 2.15 for some possibilities.
In all, the face lattice of this tropical 3-polytope is depicted in Figure 2.16. This
is a tetrahedron formed by ABCD, with the edge AB subdivided by point E. It has four
facets: two triangles ACD and BCD, and two squares ABCE and ABDE. The two squares
intersect in the union of the edges AE and BE. Its f-vector is (5, 7, 4). The fact that point
E subdivides edge AB is interesting; intuitively, this makes sense, since the point through
which point E is connected to the rest of the polytope lies on edge AB. If we had placed
point E on point AB, the diagram would be the same, except that point E would not appear
as a vertex at all.
Example 2.35. Let P be the convex hull of {A,B,C,D,E, F} = (0011, 0101, 0110, 1001, 1010, 1100) ⊂
52
x
y
z
Figure 2.17: A tropical (2, 4)-hypersimplex, i.e. octahedron.
TP3, as depicted in Figure 2.17.
Discussion: This polytope is an instance of a tropical hypersimplex, to be precise
the (2, 4)-hypersimplex, whose vertex set consists of all 4-tuples with two 1’s and two 0’s.
These (n, d)-hypersimplices always lie on a hyperplane, to be precise one with apex 0.
However, their hull lifts are bona fide hypersimplices, and in particular are full-dimensional;
their face lattice will accordingly be d-dimensional. Indeed, it turns out that every lift of
this object is a (2, 4)-hypersimplex, which is an octahedron, so its face lattice is simply that
of an octahedron. In general, the face lattice of the tropical hypersimplex is identical to
that of the corresponding ordinary hypersimplex.
Interestingly, four of the facet-defining tropical hyperplanes are the same (the one
with vertex 0); this hyperplane cuts out different facets by virtue of taking different sectors
to comprise the relevant halfspace.
53
Chapter 3
Tropical Linear Spaces
Tropical linear spaces are tropical analogues of usual linear spaces. The tropical
hyperplane of a tropical linear form (c1 ⊙ x1) ⊕ · · · ⊕ (cn ⊙ xn), c ∈ TPn−1, is the set of
points x ∈ TPn−1 such that the minimum in the linear form is attained at least twice. Let
us recall the setup from [Spe]. Let d ≤ n be positive integers. A point p ∈ TP([n]d )−1 is
called a tropical Plucker vector if for every d− 2-subset S of [n] and four distinct elements
i, j, k, l ∈ [n]\S, the minimum in (pSij ⊙ pSkl)⊕ (pSik ⊙ pSjl)⊕ (pSil ⊙ pSjk) is attained at
least twice. Given a tropical Plucker vector p ∈ TP([n]d )−1, for each d + 1-subset I ⊂ [n] we
can define a tropical linear form ⊕i∈I
(
pI\{i} ⊙ xi
)
called a circuit. The tropical linear space
corresponding to p is the intersection of the tropical hyperplanes of these circuits. A tropical
basis of a tropical linear space is a set of defining linear forms for the space. Tropical bases
are not unique, and need not be minimal in any sense. Much of this chapter is concerned
with finding minimal tropical bases.
A tropical linear space whose defining tropical linear forms have coefficients all 0
54
or ∞ is called constant coefficient. Its associated tropical hyperplanes are determined by
the supports of the tropical linear forms, that is, the entries with non-∞ coefficients. As
a result, the conditions for being a tropical basis depend only on those supports. In this
case, we can deal with the matroid whose dependent sets are supports of these tropical
linear forms. In Sections 3.1 and 3.2, we deal exclusively with the constant coefficient case,
which amounts to finding tropical bases of matroids. We describe minimal tropical bases for
several classes of matroids. Our main findings, appearing in Section 3.2, are that graphic
matroids, cographic matorids, and the matroid R10 have unique minimal tropical bases. We
hope to extend these findings to all regular matroids.
In Section 3.3, our main result concerns the tropical rank of a matrix whose rows
form a tropical basis. Furthermore, we conjecture a criterion for being a tropical basis in the
non-constant case. Finally, in Section 3.4, we show that there is a natural parametrization
of tropical linear spaces in terms of cocircuits. This chapter is based on joint work with
Debbie Yuster [YY].
3.1 The constant coefficient case
As discussed in the Introduction, constant coefficient tropical linear spaces can
be described in terms of their associated matroids. We will deal with the matroid whose
circuits are supports of circuits of a tropical linear space.
Let M be a matroid on the ground set {1, 2, . . . , n} and C be the collection of
its circuits. For a circuit C ∈ C, let T (C) be the set of points x ∈ TPn−1 such that the
minimum value in {xi : i ∈ C} is attained at least twice. The set T (C) :=⋂
C∈C T (C) is a
55
polyhedral fan called the tropical variety or the Bergman fan of M . Given a subset B ⊂ C,
define T (B) :=⋂
C∈B T (C). The set B is called a tropical basis of M if T (B) = T (C). We
wish to study the following problem.
Problem 3.1. Identify a minimal tropical basis for any matroid.
It was shown in [AK06] that the intersection of the tropical variety of a matroid
M and a sphere centered at the origin is a geometric realization of the order complex of
the lattice of flats of M . In a matroid, an element e is called a loop if {e} is a circuit, and
two elements e1, e2 are said to be parallel if {e1, e2} is a circuit. Since removing the loops
and replacing each parallel class with a single element in a matroid does not change the
lattice of flats, we may assume that our matroids are simple, i.e. contain no loops or parallel
elements. Since the circuits of the direct sum (or 1-sum) of two matroids is the union of
circuits of the summands, the following is clear.
Lemma 3.2. If matroids M , M1, and M2 are such that M = M1 ⊕M2, then the tropical
bases of M are precisely unions of tropical bases of M1 and M2.
Since every matroid is the direct sum of its connected components, we can restrict
attention to connected matroids.
Each circuit in the tropical basis “excludes” certain points from being in the trop-
ical variety, namely those values which induce a unique minimum on the terms of that
circuit. We can cut down a tropical basis of a matroid as long as the smaller circuit set
excludes the same points as the larger one. In comparing the excluded points, it suffices to
consider the 0/1 points:
56
Lemma 3.3. For any B ⊂ C,
[T (B) ∩ {0, 1}n = T (C) ∩ {0, 1}n] =⇒ [T (B) = T (C)].
In other words, to test equality of T (B) and T (C), it is sufficient to check that they agree
on 0/1 points.
Proof. Suppose we are given a subset B ⊂ C of circuits of a matroid M , and we would like
to know if B is a tropical basis of M . Since each circuit of a matroid excludes points, T (B)
contains, or is equal to, T (C). Thus, we need only worry about “extra” points, i.e. points
which are in T (B) but not in T (C). We show that if T (B) contains an extra point x, then
T (B) contains an extra 0/1 point. Thus checking that there are no extra 0/1 points shows
equality of T (B) and T (C).
Suppose T (B) contains an extra point x. Then there exists a circuit C ∈ C\B
which excludes x, and therefore the set {xi : i ∈ C} has a unique minimum m. From x,
construct the 0/1 vector v as follows:
vi =
1 if xi > m
0 if xi ≤ m
We see that v is excluded by C, so v /∈ T (C). However, for any circuit C such that
{xi : i ∈ C} attains its minimum at least twice, {vi : i ∈ C} also attains its minimum at
least twice, so v ∈ T (B).
So it is possible to remove some circuits from a tropical basis of a matroid, if the
same 0/1 vectors are excluded in the smaller set. It was shown in [AK06] that a 0/1-vector
is in the tropical variety if and only if its support, the set of coordinates with non-zero
57
values, is a flat of the matroid. Hence the excluded points are non-flats. We conclude that
a collection of circuits is a tropical basis if and only if it excludes all 0/1 non-flats.
If two circuits C1, C2 have a unique element in their intersection, pasting them
means taking their symmetric difference C1△C2 = (C1\C2) ∪ (C2\C1).
Lemma 3.4. If a collection S of circuits of a matroid has the property that every other
circuit of the matroid can be obtained by successively pasting circuits in S, then S is a
tropical basis.
Proof. Suppose we have a weight assignment x ∈ Rn such that all circuits in S have their
minimum value attained at least twice. Consider the circuit C obtained from pasting A
and B together along some element e, where A,B ∈ S. We claim that C too attains its
minimum at least twice.
The minima may be attained in one of three ways:
Case 1) A attains its minimum on a and a′, neither of which is e, and B attains its
minimum on b and b′, neither of which is e. In this case, C attains its minimum twice, on
either a and a′, or b and b′ (or all four), depending on which element weights are minimal.
Case 2) A attains its minimum on elements a and e and B attains its minimum on
edges b and b′, neither of which is e. Then b and b′ have lower weight than e, so C attains
its minimum on b and b′.
Case 3) A attains its minimum on elements a and e, and B attains its minimum on
edges b and e. Then weight(a) = weight(e) = weight(b), and thus C attains its minimum
on a and b.
58
3.1.1 Partition matroids
Given a partition Π of [n], the partition matroid MΠ is the matroid whose circuits
are pairs of elements in the same block of the partition [Oxl92]. This is the case when the
defining ideal of the linear space is also binomial.
A minimal tropical basis consists of enough pairs in each block to form a “spanning
tree” on that block. This ensures uniform weighting within each block, and thus forces the
minimum to be attained twice on each circuit.
3.1.2 Uniform matroids
The uniform matroid Ud,n is the matroid arising from a generic set of n points
in Rd. The circuits of Ud,n are the (d + 1)-subsets of [n]. Any tropical basis of Ud,n must
contain at least 1d+1
(
nd
)
elements [BJS+, Theorem 2.10], and the bound is not tight.
Lemma 3.5. An inclusion minimal tropical basis B for a uniform matroid U is given by
B = {C : i ∈ C},
where i is any fixed element of the matroid.
Proof. We must show that given a point x ∈ Rn that is excluded by some circuit of U , we
can find a circuit in B which excludes x. As shown above, we can restrict our attention to
0/1 points. Without loss of generality let us fix i = 1, so that the tropical basis B consists
of all circuits of U containing the element 1. Consider a 0/1 point x ∈ Rn that is excluded
by some circuit Cx. The minimum of {xj : j ∈ Cx} is attained uniquely. Consequently,
there is exactly one element k of Cx such that xk = 0.
59
If Cx contains the element 1, then it is in B and we’re done. Otherwise, consider
the following two cases:
Case 1 (x1 = 1): Let Cx be the circuit obtained from Cx by replacing any element
of Cx other than k by the element 1.
Case 2 (x1 = 0): Let Cx be the circuit obtained from Cx by replacing k with 1.
In both cases, Cx is a circuit of U , since every (d + 1)-subset of [n] is a circuit.
It contains the element 1, so it is in B. Finally, Cx excludes x, since the minimum of
{xj : j ∈ Cx} is attained uniquely. Thus x is excluded by the tropical basis B.
To see that B is inclusion minimal, consider a circuit C∗ ∈ B. The point x, where
xi = 1 for all i ∈ C∗\1 and xi = 0 otherwise, would be excluded by B, but not by B\{C∗}.
Thus it is necessary to include C∗ in this tropical basis.
Note that the tropical bases given for uniform matroids are not unique. For ex-
ample, consider U2,4. It has four circuits, namely 123, 124, 134, and 234. Any collection
of three out of the four circuits forms an inclusion- and cardinality- minimal tropical basis.
Yet, there is no single circuit that must be in a tropical basis of U2,4. Additionally, the
tropical bases given above for uniform matroids are not in general cardinality-minimal, de-
spite being inclusion-minimal. For example, the smallest tropical bases of U2,5 contain five
elements (123, 124, 125, 134, and 345 for example), however there are inclusion-minimal
tropical bases containing six elements, namely the six circuits which contain the element ‘1’.
60
1
2 3
4 7
5
6
Figure 3.1: Fano plane
3.1.3 Fano plane
The Fano plane gives rise to a matroid whose ground set consists of its 7 vertices.
The circuits are the “lines” in the point configuration (see Figure 3.1) and any 4-subset not
containing a line.
The unique minimal tropical basis for the Fano plane consists of the seven 3-
element circuits, one arising from each line (124, 137, 156, 235, 267, 346, & 457).
The lines are necessary: If one of the lines, L, were to be excluded, we could assign
weight 1 to two of the line’s points, weight 0 to the remaining point, and weight 0 to all
points off the line. Thus every circuit except L would have its minimum attained twice,
while L would have a unique minimum.
The lines are sufficient: Each 4-element circuit (1236, 1467, 2456, 3567, 1257,
1345, and 2347) can be obtained by pasting two 3-element circuits together along a shared
element, and deleting that element. For example, we can think of the circuit 1236 as being
pasted from the circuits 124 and 346. By Lemma 3.4, the 3-element circuits are sufficient.
Problem 3.6. Find a minimal tropical basis for transversal matroids [STY06, Section 3.3].
61
Problem 3.7. What happens to tropical bases under deletion and contraction?
3.2 Regular matroids
A regular matroid is one that is representable over every field. In this section we
characterize minimal tropical bases of certain important classes of regular matroids, namely
graphic matroids, cographic matroids, and the matroid R10. Seymour showed that every
regular matroid can be constructed by piecing together matroids of these types [Sey80].
3.2.1 Graphic matroids
A graphic matroid is formed from a graph G. The edges of G form the ground set,
and the circuits of the matroid are the minimal edge collections corresponding to cycles of
G, where a cycle is a closed walk all of whose vertices have degree two in the cycle. If G
contains n edges, we can think of a point in Rn as giving edge weight assignments for the
edges of G. A point in Rn is in the tropical variety of the graphic matroid arising from G
if and only if each cycle in G attains its minimum edge weight at least twice. We restrict
our attention to graphic matroids arising from graphs with no loops and no parallel edges.
Theorem 3.8. The unique minimal tropical basis for a graphic matroid is the collection of
its induced cycles.
An induced cycle is an induced subgraph that is itself a cycle.
Proof. Induced cycles are sufficient: Suppose we have a weight vector which achieves its
minimum twice on each induced cycle. Consider a circuit arising from a non-induced cycle
C. Then the induced subgraph on the vertices contained in C contains a chord. Divide
62
C into cycles C1 and C2 along this chord. Continuing in this manner, we can decompose
C into a collection of induced cycles pasted together. Thus by Lemma 3.4, induced cycles
form a tropical basis.
Induced cycles are necessary: Suppose we have an induced cycle C which is not in
our tropical basis. Then it is possible to construct an edge weighting for which the minimum
is attained at least twice on every cycle except C. Assign all but one edge of C weight 1.
Assign weight 0 to the remaining edge of C and to all other edges of the graph. Every
cycle incident with C has two or more edges that are not contained in C, since otherwise
C would contain a chord and thus not be an induced cycle. Thus all cycles except C have
their minimum attained at least twice, while C has a unique minimum. Thus it is necessary
to have C in every tropical basis.
3.2.2 Cographic matroids
A cographic matroid is formed from a connected graph G, having the edges of G as
its ground set. Circuits of cographic matroids are the inclusion-minimal edge cuts, i.e., sets
of edges such that removing them makes G disconnected. In order to insure every circuit
contains at least 3 elements, we will restrict our attention to cographic matroids arising
from 3-edge-connected graphs. A graph is called k-edge-connected if it remains connected
after removing any k − 1 edges. A bridge of a connected graph is an edge whose removal
disconnects the graph. A connected graph is 2-edge-connected if and only if it does not
contain a bridge.
Theorem 3.9. The unique minimal tropical basis of a cographic matroid M on a graph G
63
V
0
0
1 1 1 0
UA B
Figure 3.2: 2-edge-connected graph
consists of the edge cuts that split G into two 2-edge-connected subgraphs.
Proof. For an edge cut C that splits G into connected subgraphs U and V , let the index of
C be the sum of the number of bridges in U and V . We will show by induction on the index
that the circuits of positive index can be obtained by successively pasting index 0 circuits,
as in Lemma 3.4.
Suppose an edge cut C splits G into connected subgraphs U and V . Let e be a
bridge in U that splits it into subgraphs A and B. Let brA, brB , and brV be the number of
bridges in A, B, and V respectively. Then the index of C is equal to 1 + brA + brB + brV .
Let C ′ be the edge cut that splits A and B ∪ V . Here B ∪ V denotes the induced subgraph
of G on the vertices in B and V . In B ∪ V , there are at least two edges between B and V
because otherwise the edge e and the unique edge would disconnect G, contradicting the
3-edge-connectedness of G. Hence a bridge in B ∪ V must be either a bridge in B or a
bridge in V . Therefore, the index of the edge cut C ′ is at most brA + brB + brV . Similarly,
the index of the edge cut C ′′ that splits B and A∪ V is at most brA + brB + brV . The edge
64
cut C is obtained by pasting the lower index circuits C ′ and C ′′ along e. By induction on
the index, we see that C is obtained by successively pasting index 0 circuits. This proves
(by Lemma 3.4) that the index 0 circuits form a tropical basis.
Now we show that all such edge cuts are necessary. Suppose an edge cut C ⊂
edges(G) splits the graph G into subgraphs U and V , each of them 2-edge-connected.
Consider edge weights as follows: each edge contained in U and V gets weight 0, one edge
going between U and V gets weight 0, and the other edges between U and V get weight 1.
Consider any other edge cut C ′ 6= C. It must cause either U or V to become disconnected.
Suppose it cuts U into subgraphs A and B (see Figure 3.2). Since U is 2-edge-connected,
there are at least 2 edges going between A and B, and those edges are in C ′. Hence the
minimal weight 0 is attained at least twice on C ′, but not on C. Thus C must be in our
tropical basis.
3.2.3 The matroid R10
The matroid R10 is a regular matroid whose elements are given by the edges of
the complete graph on five vertices, K5. Its circuits are given by the fifteen 4-cycles of K5,
along with the complement of each 4-cycle. Thus R10 has 30 circuits.
Proposition 3.10. The unique minimal tropical basis of R10 consists of the fifteen 4-cycles.
Proof. The 4-cycles are necessary: For any 4-cycle C = {a1, a2, a3, a4}, we can construct a
0/1-point x which is excluded by C, but not any other circuit of R10. Set xa1 = xa2 = xa3 =
1, and assign 0 to all other entries of x. Then x attains its minimum value on C uniquely,
while on every other circuit its minimum is attained at least twice. This is because no other
65
4-cycle can share 3 elements with C, and any 4-cycle-complement will contain at least three
elements disjoint from C, corresponding to ‘zero’ entries of x.
In order to show the 4-cycles are sufficient, we show that any 0/1-point excluded
by a 4-cycle-complement will also be excluded by a 4-cycle. Consider a point x which is
excluded by a 4-cycle-complement C = {a1, a2, a3, a4, a5, a6}. Since x attains its minimum
uniquely on the terms of C, without loss of generality suppose xa1 = 0 and xa2 = xa3 =
xa4 = xa5 = xa6 = 1. Note that no 4-cycle-complement contains a 4-cycle. In order for a
4-cycle to exclude x, it must intersect C in exactly three elements other than a1. All 4-cycle-
complements look the same up to permuting the vertices. Figure 3.3 shows one such picture.
One can easily see that for any 5-element subset of the 6 edges, there exists a 4-cycle meeting
the subset in exactly 3 edges. Therefore every point excluded by a 4-cycle-complement is
also excluded by a 4-cycle.
Figure 3.3: 4-cycle-complement
Problem 3.11. Do all simple regular matroids have unique minimal tropical bases?
The Fano plane is not regular but has a unique minimal tropical basis.
66
3.3 Non-constant coefficient case
We now discuss the non-constant coefficient case. A matroid polytope of a rank
d matroid on the ground set [n] is the convex hull of the characteristic vectors in Rn of
the bases of the matroid. The hypersimplex ∆d,n is the convex hull of 0/1 vectors in Rn
with coordinate sum d. A matroidal subdivision of ∆d,n is a regular subdivision in which
every face is a matroid polytope. It was shown in [Spe] that a point p ∈ TP([n]d )−1 is a
tropical Plucker vector if and only if it induces a matroidal subdivision of ∆d,n. Moreover,
the corresponding tropical linear space is the polyhedral subcomplex of the dual of the
subdivision, consisting of the cells dual to matroid polytopes of loop-free matroids, i.e.
matroids without any one-element circuits.
A coordinate of p being ∞ is equivalent to deleting the corresponding vertex from
the hypersimplex ∆d,n. The case when all coordinates of p are either 0 or ∞ is the constant
coefficient case discussed in Sections 2 and 3. In this case, the subdivision contains a single
matroid polytope, and the circuits of the corresponding matroid are precisely the supports
of the circuits defined in the Introduction. The tropical linear space of a matroid is also
called the Bergman fan.
Since the hypersimplex ∆d,n lies on the hyperplane of points whose coordinates
sum to d, the duals of the subdivisions, hence the tropical linear spaces, have a lineality
space containing R(1, . . . , 1), as seen above. Since each cell in a matroidal subdivision is a
matroid polytope, the star of any face in the tropical linear space is isomorphic to a constant
coefficient tropical linear space.
Let K be the field of Puiseux series∑
a∈I caxa where I is a locally finite subset of
67
R with a least element and ca ∈ C. Let the degree map deg : K → T = R∪{∞} be the map
that sends an element in K\{0} to its leading (lowest) exponent of x and sends 0 to ∞. It
induces a map from a projective space over K to a tropical projective space. Let L be a d-
dimensional vector subspace in Kn, and let P ∈ P([n]
d )−1
K be its Plucker vector, i.e. the vector
of maximal minors of a matrix whose row space is L. Then deg(P ) is a tropical Plucker
vector. If a tropical Plucker vector arises in this way, then the corresponding tropical
linear space coincides with the image deg(L) ⊂ Tn and is called realizable. If there is a
representative Plucker vector containing only complex numbers, then we get the realizable
constant coefficient case. For a linear form f = a1X1 + · · · + anXn with ai ∈ K, let its
tropicalization be the tropical linear form (deg(a1)⊙x1)⊕· · ·⊕ (deg(an)⊙xn). The circuits
defined in the Introduction are precisely the tropicalizations of linear forms with minimal
support in the orthogonal complement of L. We will refer to a set of linear forms over Kn
as a tropical basis if their tropicalizations define the tropical linear space.
The tropical determinant of an r × r square matrix A = [aij ] is defined to be
⊕
σ∈Sr
(a1σ1 ⊙ · · · ⊙ arσr), where Sr is the symmetric group of order r!.
The tropical rank of a matrix is the largest integer r such that there exists a tropically non-
singular submatrix of size r, i.e. the minimum in the tropical determinant is not unique.
In the special case of realizable tropical linear spaces all of whose tropical Plucker
coordinates are non-∞, we establish a necessary condition for a set of linear forms to form
a tropical basis, which generalizes Theorem 5.3 in [RGST05].
Theorem 3.12. Let L be a n− k dimensional linear subspace in Kn all of whose Plucker
coordinates are non-zero. Let M ∈ Km×n,m ≥ k, be a matrix whose rows are non-zero
68
elements in the orthogonal complement of L. If the rows of M form a tropical basis for L,
then any k columns of deg(M) have tropical rank k.
Proof. Suppose that there is a m× k submatrix A of deg(M) with tropical rank less than
k. Then by [DSS05, Theorem 5.5] the Kapranov rank of A is less than k, which means that
there is an m×k matrix A′ over K with rank less than k such that A = deg(A′). Let v ∈ Kk
be a non-zero vector in the kernel of A′. Then deg(v) is in the tropical prevariety in Tk
defined by the m rows of A. We can augment deg(v) to a vector in the tropical prevariety
of the rows of deg(M) by putting ∞ in the other n − k coordinates. The support of this
vector has size at most k, however, the points in the tropical linear space have support size
at least k + 1 because of the hypothesis that all the tropical Plucker coordinates are finite.
Hence the prevariety is not equal to the tropical linear space, so the rows of M do not form
a tropical basis.
Conjecture 3.13. The converse of Theorem 3.12 holds.
Example 3.14. The proposition and the conjecture do not apply when some of the tropical
Plucker coordinates are ∞. Consider the 2-dimensional linear subspace of K4 which is the
kernel of the matrix
M =
1 0 1 1
0 1 1 1
1 −1 0 0
, whose degree is
0 ∞ 0 0
∞ 0 0 0
0 0 ∞ ∞
.
The circuits are precisely these three rows, hence the 0/1 points in the tropical variety are
(0, 0, 0, 1), (0, 0, 1, 0), (1, 1, 0, 0). The last two columns of deg(M) have tropical rank only 1,
but the rows of M form a tropical basis. Notice that the corresponding tropical Plucker
69
coordinate is ∞. The first two rows of the matrix deg(M) have the same tropical rank as
the whole matrix in every subset of columns, but they do not form a tropical basis. Hence
we cannot determine whether a set of linear forms is a tropical basis just from the tropical
ranks of columns.
In the cases when Theorem 3.12 and its converse are applicable, we would get an
algorithm for checking if a given matrix is a tropical basis.
3.4 Parametrizations of tropical linear spaces
So far we have been looking at tropical linear spaces as intersections of tropical
hyperplanes. In this section, we will look at them as images of tropical linear maps.
Let A be an n × d matrix over K whose image (column space) is L. Let deg(A)
be the matrix whose entries are the degrees of entries in A. This matrix defines a tropical
linear map deg(A) : Td → Tn, v 7→ deg(A) ⊙ v, where the tropical matrix multiplication ⊙
is defined by replacing sums with minima and products with sums in the evaluation of the
ordinary matrix product.
For any such A, we have
deg(L) = deg(im(A)) ⊇ im(deg(A)) (3.1)
The containment holds because the columns of deg(A) are in the tropical linear space
deg(L), and so are their tropical linear combinations, since tropical linear spaces are closed
under taking tropical linear combinations. Similar expressions hold in much more generality
for tropical varieties, as shown in [PS04, Theorem 2]. We are interested in knowing when
the equality deg(L) = im(deg(A)) is attained.
70
A cocircuit of the linear space L is an element in L whose support is minimal
with respect to inclusion. They are circuits of the orthogonal complement L⊥. Two cocir-
cuits with the same support must be constant multiples of each other, since otherwise one
coordinate can be cancelled to get a vector with a smaller support.
Lemma 3.15. Every nonzero v ∈ L can be written as v = v1 + · · ·+ vd for some cocircuits
vi ∈ L such that deg(v) = deg(v1)⊕ · · · ⊕ deg(vd).
Proof. If v is a cocircuit, then we are done.
Suppose not. Let u ∈ L be a cocircuit with supp(u) ⊂ supp(v). Then for c ∈ K
with large enough degree, we have deg(cu) ≥ deg(v) coordinatewise. Pick such a c so that
cu and v coincide in at least one coordinate, i.e. supp(v−cu) ( supp(v). Let v1 = cu. Since
deg(v1) ≥ deg(v), we have deg(v) = deg(v1)⊕ deg(v − v1).
If v−v1 is not a cocircuit, then we can repeat the same argument on v−v1, which
has a strictly smaller support. We will end up with cocircuits with the desired property.
Theorem 3.16. The equation deg(L) = im(deg(A)) holds if and only if every cocircuit in
L is represented in A.
Proof. The “if” direction follows immediately from previous Lemma. For the “only if”
direction, suppose there is a cocircuit c ∈ L whose support is not represented in A. Then
deg(c) ∈ deg(L)\im(deg(A)) since any vector in im(deg(A)) with finite coordinates in
supp(c) also has finite coordinates outside supp(c).
This theorem for the constant coefficient case appeared in [DSS05, Proposition 7.5].
71
Chapter 4
Implicitization and Elimination
Implicitization and elimination are fundamental operations in computational al-
gebraic geometry. Elimination is the problem of finding the defining ideal of the image
of an algebraic variety under a projection onto a coordinate subspace. Implicitization is
the problem of transforming a given parametric representation of an algebraic variety into
its implicit representation as the zero set of polynomials. Most algorithms for elimina-
tion and implicitization are based on multivariate resultants or Grobner bases, but current
implementations of these algorithms tend to be too slow to handle large instances.
The tropical approach to these problems is based on the following idea. Rather
than computing the variety V (I) of an ideal I by algebraic means, we shall compute the
tropical variety T (I) ⊂ Rn by combinatorial means. Here we provide an alternative to get
some information about the unknown ideals when Grobner bases and resultants fail.
Our setup for elimination problems is the following. Given a closed subvariety
X ⊂ (C∗)n and a morphism of tori α : (C∗)n → (C∗)k, our objective is to compute the
72
Zariski closure of the image of X under α. This is a closed subvariety of (C∗)k which denote
by α(X). In concrete instances the morphism α is represented by a k × n integer matrix
A = (aij), and the subvariety X is given by its ideal IX in the Laurent polynomial ring
C[x±] = C[x±11 , . . . , x±1
n ]. The desired elimination ideal Iα(X) is the inverse image of the
given ideal IX under the C-algebra homomorphism
α∗ : C[u±] 7→ C[u±11 , . . . , u±1
k ] → C[x±11 , . . . , x±1
n ], ui 7→n∏
j=1
xaij
j .
Let f1, . . . , fn ∈ C[t±11 , . . . , t±1
d ] be Laurent polynomials with Newton polytopes
P1, . . . , Pn ⊂ Rd and with supports Ai ⊂ Pi. We assume that each fi is generic relative to
its support, which means that the coefficient vector of fi lies in a Zariski open subset of CAi
where Ai ⊂ Pi is the support of fi. The aim of implicitization is to compute the prime ideal
I of all polynomials g ∈ C[x1, . . . , xn] which satisfies g(f1, . . . , fn) ≡ 0 in C[t±11 , . . . , t±1
d ]. We
here seek to read off as much information as possible about the variety V (I) from P1, . . . , Pn.
Implicitization is a special case of elimination where X is the graph of the parametrization
and α is the projection onto the image coordinates. The desired prime ideal I is the defining
ideal Iα(X) of the image α(X).
Of particular interest is the case when n = d + 1 and I = 〈 g 〉 is a principal ideal.
Here the problem is to predict the Newton polytope Q of the hypersurface V (I) = {g = 0}
from P1, . . . , Pn. This problem was posed in [SY94] and has remained open for over a
decade. It reappeared in recent work of Emiris and Kotsireas [EK05]. We here present a
general solution to the problem of [SY94], namely, a construction of the Newton polytope
Q in terms of the input polytopes Pi. When the variety V (I) is not a hypersurface but has
codimension greater than one, the role of Q is played by the Chow polytope [KSZ92], and
73
our construction generalizes to that case.
The motivation behind both [EK05] and [SY94] is that a priori knowledge of the
Newton polytope Q would greatly facilitate the subsequent computation of recovering the
coefficients of g(x) from the coeffcients of f1(t), . . . , fn(t). This is a problem of linear algebra,
and it will be discussed in Subsection 4.4.1. Part of this chapter is based on a joint work
with Bernd Sturmfels and Jenia Tevelev [STY06].
4.1 The tropical variety of a generic complete intersection
Let IX = 〈f1, . . . , fd〉 ⊂ C[x1, . . . , xn] be a generic complete intersection, i.e. the va-
riety V (IX) has dimension n−d and the coefficients of f1, . . . , fd are generic. Let P1, . . . , Pd
be the Newton polytopes of f1, . . . , fd, and F1, . . . , Fd be their normal fans. Let F be
the common refinement of F1, . . . ,Fd in Rn. It is the normal fan of the Minkowski sum
P = P1+· · ·+Pd. For w ∈ Rn, let facew(Pi) denotes the face of the polytope Pi at which the
linear functional u 7→ u · w attains its minimum. The collection of facew(Pi), i = 1, . . . , d,
depends only on which cone of F contains w in its relative interior.
Let Γ be an n − d-dimensional cone in F and w ∈ Γ be a point in its relative
interior. Then Γ is contained in a cone of dimension at least n − d in each normal fan
Fi, so the polytopes facew(Pi) have dimensions at most d and lie in a d-dimensional affine
subspace orthogonal to Γ. Hence it makes sense to talk about their mixed volume.
The following theorem allows us to compute the tropical variety T (IX) by doing
only polyhedral computations.
Theorem 4.1. The tropical variety T (IX) is the union of n− d-dimensional cones Γ in F
74
such that the mixed volume MV (faceΓ(P1), . . . , faceΓ(Pd)) is positive. The multiplicity of Γ
is equal to this mixed volume.
Consider the system of equations
inw(f1)(t) = · · · = inw(fd)(t) = 0. (4.1)
The Newton polytopes of these Laurent polynomials are facew(P1), . . . , facew(Pd). Let s
denote the dimension of the polytope
facew(P ) = facew(P1) + · · ·+ facew(Pd). (4.2)
Notice that s ≤ d because all the facew(Pi) lie in parallel d-dimensional affine spaces. The
equations (4.1) can be rewritten as a system of Laurent polynomials in s unknowns. We
wish to determine whether or not the system (4.1) has a solution t in the algebraic torus
(C∗)n. Since the coefficients of the fi are assumed to be generic, a necessary condition for
(4.1) to be solvable is that s = d. Otherwise, we have d polynomials with generic coefficients
in fewer than d variables, which have no solution. By Bernstein’s Theorem [Ber75, Kho99],
the number of solutions to the system (4.1) equals
MixedVolume(
facew(P1), . . . , facew(Pd))
, (4.3)
where the mixed volume is normalized with respect to the lattice parallel to the affine span
of facew(P ). In degenerate cases this mixed volume may be zero.
The following lemma is an important step in the proof of Theorem 4.1.
Proposition 4.2. For any w ∈ Rn the following are equivalent:
1. The vector w lies in T (IX).
75
2. The ideal 〈 inw(fi) : i = 1, . . . , d 〉 contains no monomial.
3. The initial system (4.1) has at least one solution t ∈ (C∗)n.
4. The mixed volume (4.3) is positive.
5. For all subsets K ⊆ {1, . . . , d}, the polytope facew(∑
i∈K Pi) has dimension at least |K|.
Proof. Since monomials are units in the Laurent polynomial ring, Hilbert’s Nullstellensatz
shows that (2) and (3) are equivalent. The equivalence of (3) and (4) follows from Bernstein’s
Theorem [Ber75]. The equivalence of (4) and (5) appears in [Ewa96, Theorem 4.13].
Now (1) holds if and only if there is no monomial in inw(IX) . This condition
implies (2) but, a priori, it may be stronger. To see that they are equivalent, we will use
the hypothesis that the coefficients of the fi are generic and show that (3) implies (1).
Suppose that (3) holds, and let t0 ∈ (C∗)d be a solution of the system of equations (4.1).
Using the polyhedral homotopy of [HS95], we can construct a solution t1 of the equations
f1(t) = · · · = fd(t) = 0 which degenerates to t0 under the one-parameter torus given by w.
The deformation t0 → t1 ensures that t0 is a point in the variety of the initial ideal inw(IX),
which therefore contains no monomial.
Proof of Theorem 4.1. Since IX is a complete intersection, by the Unmixedness Theorem
[Eis95, Corollary 18.14], all minimal primes of IX have codimension d, and every associated
prime is a minimal prime. Since the tropical variety T (IX) is pure of dimension n− d, the
Proposition 4.2 above implies the first statement of the theorem. Recall that the multiplicity
of a maximal cone Γ in T (IX) is the sum of multiplicities of all monomial-free minimal
associated primes of inw(IX), where w is any vector in the relative interior of Γ. This is the
76
number of solutions, counted with multiplicities, of the initial system (4.1) after a monomial
change of coordinate so that the ideal becomes zero-dimensional. By Proposition 4.2 agin,
this multiplicity equals the mixed volume (4.3).
4.2 Tropical elimination and implicitization
In the tropical setting, the problem of computing the image α(X) of the given
variety X is replaced by the problem of computing the tropical variety of the image, along
with multiplicities for the maximal cones.
The tropicalization of the image of X under the morphism α is the tropical variety
obtained by taking the image of the fan T (IX) under the linear map A:
T (Iα(X)) = A · T (IX). (4.4)
This formula follows from [Tev, Remark 3.2]. The main problem we are facing is to compute
the correct multiplicities on this fan.
In [ST], Sturmfels and Tevelev show the following formula for the push-forward
of multiplicities. We will assume that the morphism α is generically finite of degree δ. We
also put a fine enough fan structure on T (Iα(X)) such that the image A(Γ) of any cone Γ
in T (IX) is a union of cones in T (Iα(X)). Then the multiplicity of a maximal cone Π in
T (Iα(X)) is
mΠ =1
δ
∑
Γ:A(Γ)⊇Π
mΓ · index(Γ,Π),
where mΓ is the multiplicity of Γ in T (IX), and index(Γ,Π) is the index of the sublattice
generated by A(Γ ∩ Zk) inside the sublattice generated by Π ∩ Zk.
77
We now turn our attention to implicitization, which will be our main application.
Our implicitization problem is specified by a collection of n Laurent polynomials
fi(t) =∑
a∈Ai
ci,a · ta11 · · · t
ad
d (i = 1, 2, . . . , n). (4.5)
Here each Ai is a finite subset of Zd, and the ci,a are generic complex numbers. Our ultimate
aim is to compute the ideal I ⊂ C[x1, . . . , xn] of algebraic relations among f1(t), . . . , fn(t),
or at least, some information about its variety V (I).
To phrase the implicitization problem as an elimination problem, we will consider
the graph G ⊂ (C∗)d × Cn of the morphism f : (C∗)d → Cn defined by f : (t1, . . . , td) 7→
(f1(t), . . . , fn(t)), i.e. the graph G is the closure of {(t, x) : x = f(t)} in (C∗)d ×Cn. Hence
the defining ideal G is
IG = 〈xi − fi(t1, . . . , td) : i = 1, . . . , n〉. (4.6)
Our parametrized variety is the closure of the image of f , which is the same as the projection
of the graph G on the last n coordinates.
Let Ψ : Rd → Rn be the tropicalization of the map f = (f1, . . . , fn), i.e., let
Ψi(w) = min{w ·v : v ∈ Pi} be the support function of the Newton polytope Pi = conv(Ai).
The image of Ψ is contained in the tropical variety T (I), by [PS04, Theorem 2], but this
containment is usually strict. In other words, the image of the tropicalization of f is usually
a proper subset of the tropicalization of the image of f . The point of the following result is
to characterize the difference T (I)\image(Ψ).
Let e1, . . . , en be the standard basis of Rn. For J ⊆ {1, . . . , n}, we abbreviate the
orthant R≥0{ej : j ∈ J} by RJ≥0 and the Minkowski sum
∑
j∈J Pj by PJ .
78
Theorem 4.3. Let f1, . . . , fn be Laurent polynomials in d variables which are generic rel-
ative to their support, and let I ⊂ C[x1, . . . , xn] be the prime ideal of all algebraic relations
among f1, . . . , fn. Then the following subsets of Rn coincide:
1. the tropical variety T (I),
2. the union of all sets Ψ(T (〈fj : j ∈ J〉)) + RJ≥0, where J ⊆ {1, 2, . . . , n},
3. the union of the cones Ψ(w) + RJ≥0, where w ∈ Rd and J runs over all subsets of
{1, 2, . . . , n} such that 〈 inw(fj) : j ∈ J 〉 contains no monomial,
4. the union of the cones Ψ(w) + RJ≥0 such that, for all subsets K ⊆ J , the linear form
w attains its minimum over PK at a face of dimension at least |K|.
Characterization (4) gives us a combinatorial algorithm for computing the tropical
variety T (I) directly from the given Newton polytopes P1, . . . , Pn.
It is instructive to note that the contribution of the empty set J = ∅ in Theorem
4.3 (2) is precisely the image of the tropicalization Ψ of the given map f :
Ψ(T (〈∅〉)) + R∅≥0 = Ψ(T ({0})) = Ψ(Rd) = image(Ψ). (4.7)
Thus it is the contributions made by non-empty subsets J which make up the difference
between the tropicalization of the image and the image of the tropicalization.
Proof of Theorem 4.3. The Theorem 4.1 gives us a combinatorial description of the tropical
variety T (IG) of the graph G defined by ideal IG in (4.6). The Newton polytopes of the
generators in (4.6) are Pi = conv(Pi ∪ (0, ei)) ⊂ Rd+n for i = 1, . . . , n, i.e. each Pi is
a pyramid over Pi with apex (0, ei). Here we view Pi as lying in the first d-coordinate
subspace in Rd+n, and e1, . . . , en are standard unit vectors in the last n coordinates.
79
We now describe the rays of the normal fan F of the Minkowski sum P = P1 +
· · ·+ Pn. The vector (0, ei) ∈ Rd+n is a ray in F if
face(0,ei)(P ) = P1 + · · ·+ Pi + · · ·+ Pn
is a facet. Suppose a vector (w, z) ∈ Rd+n with w 6= 0 is a ray in F . Then we must have
z = Ψ(w). Otherwise, face(w,z)(P ) is strictly contained in face(w,Ψ(w))(P ). The tropical
variety T (IG) is d-dimensional, and the d-dimensional cones in F have the form {(w,Ψ(w)) :
w ∈ C} + (0, RJ≥0) ⊂ Rd+n where C is a d − |J | dimensional cone in the normal fan of
P1 + · · · + Pn. The image of this cone under the projection onto the last n coordinates
is Ψ(C) + RJ≥0. Note that face(C,J)(Pi) is equal to conv(faceC(Pi) ∪ (0, ei)) if i /∈ J and
conv(faceC(Pi)) if i ∈ J . Now we have to determine which pairs (C, J) actually give a
d-dimensional cone in the tropical variety T (IG). Using the equivalence of (1) and (5) in
Proposition 4.2, these are the pairs (C, J) where
dim
(
∑
i∈K
face(C,J)(Pi)
)
≥ |K| for all K ⊂ {1, . . . , d}. (4.8)
We saw above that
dim
(
∑
i∈K
face(C,J)(Pi)
)
= |K\J |+ dim
(
∑
i∈K
faceC(Pi)
)
.
Hence the condition (4.8) is equivalent to
dim
(
∑
i∈K
faceC(Pi)
)
≥ |K ∩ J | for all K ⊂ {1, . . . , d}. (4.9)
However, in condition (4.9) it suffices to check for K ⊂ J . This proves the equivalence of
(1) and (4). The equivalence of (2),(3), and (4) follows from Proposition 4.2.
80
We now consider the prime ideal I of algebraic relations among the generic Laurent
polynomials f1(t), . . . , fn(t). Generators for the ideal I are (still) unknown, but in Theorem
4.3 we can compute its tropical variety T (I) in combinatorial terms. In what follows we
similarly compute the multiplicity MΓ for every maximal cone Γ in a fan structure on T (I).
The image of the cones in T (IX) under A do not generally form a fan, and they
need to be subdivided to give a fan structure on the tropical variety T (Iα(X)). We choose a
fine enough fan structure on T (Iα(X)) so that the image of any cone in T (IX) is the union
of some cones of T (Iα(X)). We say that a pair (C, J) covers a d-dimensional cone Γ in
T (Iα(X)) if C is d-dimensional and contains Γ. There may be more than one pair (C, J)
which covers a fixed cone Γ. Suppose that (C, J) covers Γ. Then the sublattice generated by
Ψ(C ∩Zd) + ZJ has rank d in Zn. We define index(C, J) to be the index of that sublattice
in the maximal rank d sublattice of Zn that contains it.
Let Fj = facew(Pj) and FJ = facew(∑
j∈J Pj) =∑
j∈J Fj for some (hence any)
vector w in the relative interior of C. Then the |J |-dimensional mixed volume
MixedVolume(Fj : j ∈ J) (4.10)
is the same as the one in equation (4.3). This mixed volume is normalized with respect
to the affine lattice spanned by the |J |-dimensional polytope FJ . Multiplying (4.10) by
index(C, J) we obtain the scaled mixed volume
MixedVolume(Fj : j ∈ J) · index(C, J) (4.11)
The following theorem follows from the push forward of multiplicities formula seen above.
Theorem 4.4. The multiplicity MΓ of a maximal cone Γ in the fan structure on T (I) is
the sum of all scaled mixed volumes (4.11) where the pair (C, J) covers Γ.
81
4.2.1 Examples
We now present some families of examples which illustrate Theorems 4.3 and 4.4.
Let us begin by examining the simplest case: a parametrized curve in the plane.
Example 4.5 (Plane curves). Let d = 1 and n = 2. We wish to compute the equation
g(x1, x2) of the plane algebraic curve parametrized by two Laurent polynomials x1 = f1(t)
and x2 = f2(t). The Newton polytopes of f1 and f2 are segments on the line R1,
P1 = [ a , b ] and P2 = [ c , d ]. (4.12)
Here a ≤ b and c ≤ d are integers.
Figure 4.1: Tropical plane curves and their Newton polygons
Now let’s see how we get the tropical variety and the multiplicities. For J = {1, 2}
we just get the empty set since 〈f1, f2〉 is the unit ideal in C[t, t−1]. For J = {1} we get
the ray spanned by (1, 0) since T (〈f1〉) = {0}. Its multiplicity is the the volume of the first
Newton polytope, which is equal to b− a. Similarly for J = {2} we get the ray spanned by
82
(0, 1) with multiplicity d − c. Finally, for J = ∅ we get the image of the tropicalization Ψ
which consists of the two rays spanned by (a, c) and by (−b,−d). See Figure 4.1.
The Newton polygon Q of the implicit equation g(x1, x2) is a (possibly degenerate)
quadrangle in R2. There are four cases:
• if a ≥ 0 and c ≥ 0 then Q = conv{
(0, b), (0, a), (c, 0), (d, 0)}
,
• if b ≤ 0 and d ≤ 0 then Q = conv{
(0,−a), (0,−b), (−d, 0), (−c, 0)}
,
• if a ≤ 0, d ≥ 0 and bc ≥ ad then Q = conv{
(0, b − a), (0, 0), (d − c, 0), (d,−a)}
,
• if b ≥ 0, c ≤ 0 and bc ≤ ad then Q = conv{
(0, b− a), (0, 0), (d − c, 0), (−c, b)}
.
The possible quadrangles Q and their normal fans are illustrated in Figure 4.1. We note
that the case a = 48, b = 63, c = d = 32 appears in [EK05, Example 3.3].
Monomials
Suppose that each fi is a monomial, so each Pi is just a point, say Pi = {ai} ⊂ Zd.
Then I is the toric ideal generated by all binomials xu − xv where u− v is in the kernel of
the matrix A = [a1, a2, . . . , an]. The tropical variety T (I) equals the row space of A. The
tropical morphism Ψ : Rd → Rn is the linear map w 7→ wA given by left multiplication
with A. For any nonempty subset J of {1, 2, . . . , n}, the ideal 〈fj : j ∈ J〉 is the unit ideal,
which contributes nothing to the union in (2) of Theorem 4.3. The ideal corresponding to
J = ∅ is 〈∅〉 = {0}, and T (I) = rowspace(A) is indeed the image of T (〈∅〉) = Rd under
the map Ψ.
83
The Unmixed Case
Suppose that all Newton polytopes are dilations of a common polytope, say Pi =
ciP for some positive integers ci and a full-dimensional polytope P ⊂ Rd. We distinguish
two cases.
First suppose that P contains the origin. Then the image of Ψ is the halfline
C− = R≤0(c1, . . . , cn). In fact, the set Ψ(T (〈fj : j ∈ J〉)) in Theorem 4.3 (2) equals C−
if |J | < d, it equals {0} if |J | = d, and it is empty if |J | > d. Therefore, the tropical
variety T (I) is the union of the cones RJ≥0 for J ⊂ {1, . . . , n} with |J | = d and the cones
C− + RJ≥0 for J ⊂ {1, . . . , n} with |J | = d− 1.
Next suppose that P does not contain the origin. Then the image of Ψ is the line
C = R(c1, . . . , cn), and we conclude that the tropical variety T (I) is the union of the cones
RJ≥0 for |J | = d, and the cones C + RJ
≥0 for |J | = d− 1.
In the case n = d + 1, the tropical variety T (I) just constructed has codimension
one in Rn. It is the union of all codimension one cones in the normal fan of the Newton poly-
tope Q of the implicit equation. The polytope Q was constructed with a different technique
in [SY94, Theorem 9], and the previous two paragraphs are consistent with that result.
In the first case when 0 ∈ P , the polytope Q is a simplex, which is the convex
hull of 0,m1e1,m2e2, . . . ,md+1ed+1. Each mi is equal the length of the edge from 0 to
miei, hence the multiplicity of its normal cone spanned by {ej : j 6= i}. This is the mixed
volume of the corresponding d polytopes {Pj : j 6= i} in Rd. The total degree of the implicit
equation g is the maximum of the total degrees of the extreme monomials, which correspond
to vertices of the Newton polytope Q, so the degree of g is the maximum of m1, . . . ,md+1.
84
Example 4.6 (Full support). An interesting special case is when the input polynomials
f1, . . . , fd+1 have full support of degrees a1, . . . , ad+1 respectively. Then the Newton poly-
topes are ai∆d wehre ∆d is the standard simplex conv{0, e1, . . . , ed}. In that case, the mixed
volume of {Pj : j 6= i} is mi =∏
j 6=i aj by the Bernstein and Bezout theorems. As shown
above, the Newton polytope of the implicit equation g is conv(0,m1e1, . . . ,md+1ed+1).
Linear maps
Suppose each support Ai is a subset of {e1, . . . , ed}, the unit vectors in Rd. Thus
fi is a generic linear form in the set of unknowns { tj : ej ∈ Ai}. The ideal I is generated
by an (n − d)-dimensional space of linear forms in C[x1, . . . , xn]. We introduce the subset
Bj = {i : ej ∈ Ai} for j = 1, 2, . . . , d. The rank d matroid corresponding to the linear space
V (I) ⊂ Cn is the transversal matroid [Bru87] of the set family {B1, . . . , Bd}. The tropical
variety T (I) is the Bergman complex of that transversal matroid, as shown in [AK06].
Thus, in the linear case, Theorem 4.3 offers an interesting new representation of
the Bergman fans of transversal matroids. It can be described as follows. The sum of
simplices P =∑n
i=1 conv(Ai) is a generalized permutohedron [Pos], which means that the
normal fan of P is a coarsening of the Sn-arrangement. Let C be any cone in that fan. Each
initial form inC(fj) is supported on a subset faceC(Aj) of Aj. Then with C we associate
the simplicial complex ∆C on {1, 2, . . . , n} whose simplices are the subsets J such that these
initial forms have a solution in (C∗)d. This is a combinatorial condition on the set family
{ faceC(Aj) : j ∈ J }, which is essentially Postnikov’s Dragon Marriage Condition [Pos, §5].
To construct the Bergman fan, we map the normal fan of P from Rd into Rn using the map
Ψ, and to each image cone Ψ(C) we attach the family of orthants RJ≥0 indexed by ∆C .
85
Example 4.7. Let d = 3, n = 6 and consider six linear forms
f1 = c11 · t1 + c12 · t2
f2 = c21 · t1
f3 = c31 · t1 + c33 · t3
f4 = c42 · t2
f5 = c53 · t3
f6 = c62 · t2 + c63 · t3,
where the cij are general complex numbers. The tropical linear space T (I) is three-
dimensional in R6, but each maximal cone contains the line spanned by (1, 1, 1, 1, 1, 1),
so we can represent T (I) by a graph. This graph is the Bergman complex of a rank 3
matroid on {1, 2, 3, 4, 5, 6}, namely, the transversal matroid of
(B1, B2, B3) =(
{1, 2, 3}, {1, 4, 6}, {3, 5, 6})
.
6
����
����
��������
��������
����
����
����
����
456
4
1
32
235
5
124
����
Figure 4.2: The Bergman complex of a transversal matroid
The drawing of T (I) in Figure 4.2 is the same as [DFS, Figure 2], but we now
86
derive it from Theorem 4.3 (2). The polytope P = P1 + · · ·+P6 is a planar hexagon, and Ψ
maps the six two-dimensional cones in its normal fan to the six edges in the outer hexagon in
Figure 4.2. Each of the edges {124, 1}, {2, 6}, {235, 3}, {5, 1}, {456, 6}, and {4, 3} is between
the image of a ray in the normal fan of P and a ray generated by e1, e3, or e6. For instance,
the edge {124, 1} corresponds to the ray generated by Ψ ((1, 1, 0)) = (1, 1, 0, 1, 0, 0) plus the
ray R≥0 e1 in R6. None of the simplicial complexes ∆C contains 2, 4, or 5 because f2, f4,
and f5 are monomials. However, e2, e4, and e5 are images of some rays in the normal fan
of P under the tropical map Ψ. The triangle in the middle of Figure 4.2 represents the
simplicial complex associated with the 0-dimensional face {0} of the normal fan:
∆{0} ={
{1, 3}, {1, 6}, {3, 6}}
.
This shows how tropical implicitization works for the linear forms f1, . . . , f6.
Binomials
Consider the case where the map f is given by n binomials
fi = ci1 · tai + ci2 · t
bi (i = 1, 2, . . . , n).
Each Newton polytope Pi = conv(ai, bi) is a line segment, and their Minkowski sum P =
P1 + · · · + Pn is a zonotope (i.e. a projection of the n-cube). The normal fan of P is the
hyperplane arrangement H ={
{u · aj = u · bj}}
j=1,...,n. The map Ψ : Rd → Rn is the
tropical morphism associated with the arrangement H, that is,
Ψ(u) =(
min(u · a1, u · b1),min(u · a2, u · b2), . . . ,min(u · an, u · bn))
.
This map was recently studied by Ardila [Ard] for certain graphic arrangements H.
87
Theorem 4.3 shows that the tropical variety T (I) is the union of the cones Ψ(C)+
RJ≥0, where C is any cone of the hyperplane arrangement H, and the hyperplanes indexed
by J ⊂ {1, . . . , n} contain the cone C and are linearly independent.
Example 4.8. Let d = 2 and consider the following n = 3 binomials in t1 and t2:
f1 = (t1 − α) · tu11 tu2
2 , f2 = (t2 − β) · tv11 tv2
2 , f3 = (t1 − γt2) · tw11 tw2
2 ,
where α, β, γ are general complex numbers, and u1, u2, v1, v2, w1, w2 are integers. Then H
is the arrangement of three lines in the plane R2 which consists of the two coordinate axes
and the main diagonal. The images of the six one-dimensional cones of the arrangement H
under the map Ψ are spanned by the vectors
Ψ(e1) = (u1, v1, w1),
Ψ(e1 + e2) = (u1 + u2, v1 + v2, w1 + w2 + 1),
Ψ(e2) = (u2, v2, w2),
Ψ(−e1) = (−u1 − 1,−v1,−w1 − 1),
Ψ(−e1 − e2) = (−u1 − u2 − 1,−v1 − v2 − 1, w1 − w2 − 1),
Ψ(−e2) = (−u2,−v2 − 1,−w2 − 1).
The image of Ψ consists of the six two-dimensional cones in R3 which are spanned by
(cyclically) consecutive vectors in this list. Next, the tropical surface T (I) contains the six
two-dimensional cones which are spanned by the following pairs in R3:
{Ψ(e1), e2} , {Ψ(e1 + e2), e3} , {Ψ(e2), e1} ,
{Ψ(−e1), e2} , {Ψ(−e1 − e2), e3} , {Ψ(−e2), e1}.
88
Finally, the zero-dimensional cone C = {0} contributes the three two-dimensional cones
which are spanned by the pairs of standard basis vectors
{e1, e2} , {e1, e3} , {e2, e3}.
These pairs determine a non-planar graph with 9 vertices and 15 edges. It is isomorphic to
the one depicted in Figure 4.2. Thus T (I) is a two-dimensional fan in R3 whose intersection
with the unit sphere is an immersion of that graph. The resulting embedded graph, which
is dual to the Newton polytope of the implicit equation, depends on the numerical values
of the exponents u1, u2, v1, v2, w1, w2. This example shows how our method can be applied
to a family of implicitization problems where the exponents of the fi are not fixed integers
but are unknowns.
Surfaces
Suppose that d = 2, so our given input is a list of lattice polygons P1, . . . , Pn in
R2. The tropical variety T (I) is a two-dimensional fan in Rn which we represent by an
embedded graph in the (n − 1)-sphere Sn−1. This graph is constructed as follows. We fix
the node ei for each polygon Pi which is not just a point, and we fix the node Ψ(w) for each
vector w that is an inner normal to an edge of the polygon P = P1 + · · ·+Pn and such that
Ψ(w) 6= 0. We identify such a vector Ψ(w) with the point Ψ(w)/||Ψ(w)|| on Sn−1.
We now connect pairs of nodes by great circles in Sn−1 according to the follow-
ing rules: a pair {Ψ(w),Ψ(w′)} gets connected if their edges on P are adjacent, a pair
{Ψ(w), ei} gets connected if w is the inner normal to an edge of Pi, and a pair {ei, ej}
gets connected if Pi + Pj is two-dimensional. Some pairs of great circles intersect and thus
89
create new nodes in Sn−1. This intersection need not be transversal, i.e., two great circles
may intersect in a smaller great circle. The result of this construction is the tropical surface
T (I), represented by a graph in Sn−1.
The edges of this graph are weighted with multiplicities as follows. The weight of
a pair {Ψ(w),Ψ(w′)} is the greatest common divisor of the 2×2-minors of the n×2-matrix
(Ψ(w),Ψ(w′)), divided by the determinant of the 2×2-matrix (w,w′). The weight of a pair
{Ψ(w), ei} is the normalized length of the edge facew(Pi) of Pi times the greatest common
divisor of all coordinates of Ψ(w) but the ith. The weight of a pair {ei, ej} is the mixed area
MixedVolume(Pi, Pj) = area(Pi + Pj)− area(Pi)− area(Pj).
Now, when forming the embedded graph T (I) ⊂ Sn−1, each edge Γ may be covered by more
than one of the great circles created by these pairs, and we take MΓ to be the sum of their
weights. With these multiplicities, the graph T (I) is balanced.
Example 4.9. Consider the family of surfaces in 3-space which is given by
A1 = {(1, 0), (3, 1), (2, 2)}, A2 = {(−1, 0), (0,−1), (0, 0)}, A3 = {(2, 1), (0, 2), (1, 3)}.
To construct the graph T (I) on the sphere S2, we first draw the three nodes e1, e2, e3
and the nine nodes Ψ(w) which are the images under Ψ of the inner normals of the 9-gon
P1 + P2 + P3. These nine directions are given by the columns of the matrix
e2 e3 e1 e2 e3 e1 e2 e3 e1
1 1 −1 −3 −7 −4 −2 0 2
−1 −2 −2 0 0 0 0 −1 −2
2 4 0 −2 −5 −4 −3 −2 −2
90
We connect these twelve nodes with great circles as described above, namely, by forming
the 9-cycle of these columns, by connecting them to the ei as indicated, and by forming the
triangle e1, e2, e3. This creates an embedded graph in S2 which has 14 vertices, 27 edges
and 15 regions. The Newton polytope Q of the implicit equation is dual to this graph, so
it has 15 vertices, 27 edges and 14 facets. The cones {e1, e2}, {e1, e3}, and {e2, e3} have
multiplicities 2, 6, and 3 respectively. These are the corresponding mixed volumes of the
input Newton polytopes. All the other cones have multiplicity 1. Using the methods to
be described in Sections 4.3 and 4.4, we construct Q metrically and find that the largest
coordinate sum of a vertex is 14 and that there are 155 lattice points in it. Hence the
implicit equation g(x1, x2, x3) is a polynomial of degree 14 having at most 155 terms.
Here is a concrete numerical example taken from the computer algebra literature.
3
������������
������������
������
������
���������
���������
������������
������������
������
������
���������
���������
e
ee
3
1 2
5
6
7
4 8 2
����
����
��������
���������������
���������������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������������������������
��������������������������������������������������������������������������������������������������������������
e
48
57
2
ee
6
1 2
Figure 4.3: Tropical construction of a bicubic surface in 3-space
Example 4.10. Let d = 2 and consider the well-known problem of implicitizing bicubic
91
surfaces. The following specific surface was discussed in [EK05, Example 3.4]:
f1(s, t) = 3t3 − 6t2 + 3t + s3 − 3s2 + 6s− 1,
f2(s, t) = 3s3 − 6s2 + 3s + t3 + 3t,
f3(s, t) = −3s3t3 + 15s2t3 − 15st3 − 3s3t2 − 18s2t2 + 27st2
−3t2 + 6s3t + 9s2t− 18st + 3t− 3s2 + 3s.
The vertices of the Newton polygons are indicated in boldface font. The polygon P1 is a
triangle, P2 is a quadrangle, and P3 is a 7-gon. Their Minkowski sum P1 + P2 + P3 is an
octagon. The eight primitive inner normal vectors w = (w1, w2) of this octagon are listed
in the first two rows of the following matrix:
e1e2e3 e2e3 e1e2e3 e3 e3 e1e2 e3 e2
w1 1 1 0 −1 −1 −1 0 1
w2 0 1 1 1 0 −1 −1 −1
Ψ(w)1 0 0 0 −3 −3 −3 −3 −3
Ψ(w)2 0 1 0 −3 −3 −3 −3 −3
Ψ(w)3 0 1 0 −2 −3 −6 −3 −2
The last three rows contain the coordinates of their images Ψ(w) under the tropical mor-
phism Ψ. We now construct the graph as in 4.2.1. The first and third inner normal vector
is mapped to zero under Ψ, so they do not contribute to the tropical surface T (I). The
remaining six columns 2, 4, 5, 6, 7, 8 and the unit vectors e1, e2, e3 form the following 15 great
92
circles on the sphere S2, as shown in Figure 4.3:
pair index mixed volume on edge
{e1, e2} 1 9 {e1, e2}
{e1, e3} 1 18 {e1, e3}
{e2, e3} 1 17 {e2, e3}
{2, e2} 1 1 {e2, e3}
{2, e3} 1 1 {e2, e3}
{6, e1} 3 3 {6, e1}
{6, e2} 3 3 {6, e2}
{4, e3} 3 1 {6, e3}
{8, e3} 3 1 {6, e3}
{5, e3} 3 2 {6, e3}
{7, e3} 3 2 {6, e3}
{4, 5} 3 1 {6, e3}
{5, 6} 9 1 {6, e3}
{6, 7} 9 1 {6, e3}
{7, 8} 3 1 {6, e3}
There is considerable overlap among these great circles. For instance, the points 4 and 8
coincide, as do the points 5 and 7, and they all lie on the great circle between e3 and 6.
Likewise, the point 2 lies on the great circle between e2 and e3. Thus the embedded graph
T (I) ⊂ S2 is the complete graph K4, and, by adding up the contributions of each index
93
times mixed volume, we get the corresponding weights:
edge Γ {e1, e2} {e1, e3} {e2, e3} {6, e1} {6, e2} {6, e3}
weight MΓ 9 18 18 9 9 18
(4.13)
From this table we can now determine the Newton polytope Q of the implicit
equation g(x1, x2, x3). The general method for this is explained in the next section. In this
example we find that the proposed Newton polytope is the tetrahedron
Q ={
(u1, u2, u3) ∈ R3 : u1, u2, u3 ≥ 0 and u1 + u2 + 2 · u3 ≤ 18}
.
The number of lattice points in Q equals 715. What we can conclude at this point is that
Q would be the Newton polytope of the implicit equation g if f1, f2, f3 were replaced by
polynomials with the same support but with generic coefficients.
However, the coefficients of the specific polynomials f1, f2, f3 we took from [EK05,
Example 3.4] are not generic. For instance, the equations f2(s, t) = f3(s, t) = 0 have only 15
solutions (s, t) ∈ (C∗)2, which is less than the number 17 predicted by the mixed volume.
Yet, it turns out that Q is the correct Newton polytope and all 715 possible monomials
appear with non-zero coefficients in the implicit equation:
g(x1, x2, x3) = 387420489x181 + 387420489x18
2 − 18014398509481984x93 + · · · · · ·
−12777985432959891776936639829x22 + · · · − 3707912273492242256259566313.
We included the x22-term because it has the coefficient of largest absolute value.
94
4.3 From tropical variety to Newton polytope to implicit
equation
In the previous sections we constructed the tropical variety T (I) along with the
multiplicity MΓ for each maximal cone Γ. Assuming I = 〈g〉 to be a principal ideal, we now
show how this information reveals the Newton polytope Q of g. The polynomial g is then
recovered from Q and the fi’s using numerical linear algebra. If V (I) is not a hypersurface
then the role of Q will be played by the Chow polytope.
Generic hypersurfaces
The hypersurface case, when n = d + 1, has received the most attention in
the computer algebra literature [CGKW01, EK05, SY94]. Suppose we are given d + 1
Laurent polynomials f0, f1, . . . , fd whose coefficients are generic relative to their supports
A0, A1, . . . , Ad ⊂ Zd. Their Newton polytopes are denoted by P0, P1, . . . , Pd. Note that we
shifted indices by one. The prime ideal of algebraic relations among the fi is denoted, as
before, by I ⊂ C[x0, x1, . . . , xd]. There is a simple combinatorial criterion for when the
ideal I is principal, i.e., when the variety V (I) parametrized by (f0, f1, . . . , fd) is actually
a hypersurface in Cd+1.
Proposition 4.11. The prime ideal I is principal if and only if there exist points ai ∈ Ai
such that the d× (d + 1)-matrix (a0, a1, . . . , ad) has maximal rank d.
Proof. This can be seen by writing down the d × (d + 1)-Jacobian matrix(
∂fi/∂tj)
, and
using the fact that the coefficients of the fi are generic.
95
We now assume that the condition in Proposition 4.11 is satisfied, and we let
g(x0, x1, . . . , xd) denote the unique (up to scaling) generator of the ideal I. Let Q ⊂ Rd+1
be the Newton polytope of the irreducible polynomial g. Then the tropical hypersurface
T (I) = T (g) is the union of the codimension one cones in the normal fan of Q, and the
multiplicity of a maximal cone Γ in any fan structure on T (I) equals to lattice length of
the edge of Q whose normal cone contains Γ.
Suppose we constructed the tropical hypersurface T (I) using Theorem 4.3, and
we computed the multiplicities MΓ of each cone Γ in T (I) using Theorem 4.4. These data,
combined with the requirement that Q lies in the non-negative orthant and intersects each
of the d + 1 coordinate hyperplanes, determine the polytope Q uniquely. Hence we can
construct the desired Newton polytope Q ⊂ Rd+1 combinatorially from the given Newton
polytopes P0, P1, . . . , Pd ⊂ Rd.
The previous paragraph almost solves the problem stated in [SY94]. The only
shortcoming is that the description of the set T (I) given in Theorem 4.3 does not come
with a nice fan structure as we will see in Examples 4.9 and 4.10. In our view, the following
theorem provides a better solution, as it does not require the knowledge of any fan structure.
All that is needed are the integers Mγ = MΓ > 0 where γ is a smooth point on T (I) and
Γ is any sufficiently small relatively open cone in T (I) which contains γ. These numbers
Mγ are computed using the formula in Theorem 4.4, simply by replacing “(C, J) covers Γ”
with “(C, J) covers γ”.
The following theorem gives a formula for each coordinate of each vertex of the
Newton polytope Q. It does not rely on any particular fan structure on T (I).
96
Theorem 4.12. For a generic vector v ∈ Rd+1, the ith coordinate of the vertex facev(Q)
equals the number of intersection points, each counted with its intersection multiplicity, of
the tropical hypersurface T (I) with the halfline v + R≥0 ei.
This result and its generalization to larger codimension, to be stated in Theorem
4.15 below, is due to Dickenstein, Feichtner, and Sturmfels [DFS]. We consider the con-
junction of Theorems 4.3, 4.4 and 4.12 to be a satisfactory solution to the problem stated
in [SY94].
What remains is for us to explain the meaning of the term “intersection multiplic-
ity” in the statement of Theorem 4.12. Since v is generic, each intersection of the tropical
hypersurface T (I) with the halfline v + R≥0ei occurs in a smooth point γ of T (I). The
intersection multiplicity at γ is the product of the multiplicity Mγ and the index of the
finite index sublattice Zei + (RΓ ∩ Zd+1) of Zd+1. In [DFS, §2], these two factors were
called the intrinsic and extrinsic multiplicities of the intersection.
Non-generic hypersurfaces
The following proposition ensures that we can apply our tropical method even if
the given polynomials fi(t) do not have generic coefficients. If there is a unique irreducible
relation among the given fi, then our linear algebra method will find that relation, up to a
polynomial multiple.
Proposition 4.13. Let f0, . . . , fd ∈ C[t±11 , . . . , t±1
d ] be any Laurent polynomials whose ideal
of algebraic relations is principal, say I = 〈g〉, and Pi ⊂ Rd the Newton polytope of fi. Then
the polytope Q, which is constructed combinatorially from P0, . . . , Pd as in Subsection 4.3,
97
contains a translate of the Newton polytope of g.
Proof. We introduce a family of Laurent polynomials f ǫ0(t), f
ǫ1(t), . . . , f
ǫd(t) with coefficients
in C(ǫ) which are generic for their support and which satisfy limǫ→0fǫi (t) = fi(t) for all
i. There exists an irreducible polynomial gǫ(x0, x1, . . . , xd) with coefficients in C(ǫ) which
vanishes if we replace each xi by f ǫ(t). Removing common factors of ǫ from the terms of
gǫ, we may assume that
g(x) := limǫ→0 gǫ(x)
is not the zero polynomial. The Newton polytope of g is contained in Q. Now, we take the
limit for ǫ→ 0 in the identify
gǫ(
f ǫ0(t), f
ǫ1(t), . . . , f
ǫd(t)
)
≡ 0 in C(ǫ)[t±11 , . . . , t±1
d ].
This implies that g(x) is an algebraic relation among f0(t), f1(t), . . . , fd(t), so the irreducible
polynomial g(x) is a factor of g(x). This means that the Newton polytope of g is a Minkowski
summand of the Newton polytope of g. This implies that a translate of the Newton polytope
of g lies in Q.
Example 4.14. Let d = 2 and consider the three homogeneous quadrics:
f0(s, t) = s2 + st− 2t2
f1(s, t) = s2 − 2st + t2
f2(s, t) = −2s2 + st + t2
Here, g = x0 + x1 + x2, so its Newton polytope is the triangle ∆ = conv{e0, e1, e2}. The
Newton polytope of the implicit equation for a parametric surface given by three general
98
quadrics is twice that triangle: Q = 2 ·∆. This shows that ∆ is not a subpolytope of Q but
a (non-unique) translate of ∆ is a subpolytope of Q.
The approach suggested by Proposition 4.13 is to not worry at all whether the
given polynomials have generic coefficients or special coefficients. This will not be relevant
until the very end, when the numerical linear algebra detects that the solution space to the
linear system of equations for the coefficients of g(x) is larger than expected. This should
work well when the fi are not too far from the generic case. However, for the kind of special
morphisms f which typically arise in algebraic geometry and its applications (for instance,
in statistics), this approach is likely to fail in practice.
Lower-dimensional varieties
We now consider the case when the parametric variety V (I) ⊂ Cn and its tropi-
calization T (I) ⊂ Rn do not have codimension one. For technical reasons, we here assume
that the fi are homogeneous polynomials of the same degree, so that I is a homogeneous
prime ideal in C[x1, . . . , xn]. Here the role of the Newton polytope Q is played by the Chow
polytope Chow(I). This is the lattice polytope in Rn which is defined as follows. For a
generic vector v ∈ Rn, the ith coordinate of the vertex facev(Chow(I)) of Chow(I) is the
sum of µP (inv(I)) where P runs over all monomial primes that contain the variable xi and
µP (inv(I)) denotes the multiplicity of the monomial ideal inv(I) along P . The Chow poly-
tope was introduced and studied in [KSZ92]. If V (I) has codimension one, then Chow(I)
is precisely the Newton polytope of the irreducible generator g of I.
The following result from [DFS, Theorem 2.2] shows that in the generic case we
99
can construct Chow(I) from the tropical variety T (I) and its multiplicities Mγ in the same
way as constructing the Newton polytope of the implicit equation of a hypersurface. Let
I be any prime ideal of height c in C[x1, . . . , xn] and v a generic vector in Rn, so that the
initial ideal inv(I) is generated by monomials.
Theorem 4.15 (DFS). A prime ideal P = 〈xi1 , . . . , xic〉 is a minimal prime of inv(I) if
and only if the tropical variety T (I) meets the cone v + R≥0{ei1 , . . . , eic}. The number of
intersections, each counted with its intersection multiplicity, coincides with the multiplicity
µP (inv(I)) of the monomial ideal inv(I) along P .
Geometrically, the vertices of Chow(I) correspond to the toric degenerations of the
algebraic cycle underlying I. All points in Chow(I) have the same coordinate sum, namely,
the degree of the projective variety of I. Thus we get a combinatorial rule for computing
the degree of the image of any projective morphism. Moreover, if I is homogeneous with
respect to some multigrading, then Theorem 4.15 gives us a formula also for the multidegree
of I. The multidegree is a multivariate polynomial which refines the classical notion of
degree, and which has received much attention recently in algebraic combinatorics through
the work of Knutson and Miller [KM05].
4.4 TrIm: Software for tropical implicitization
We present software called TrIm for tropical implicitization. It also computes
the tropical variety of an elimination ideal of a generic complete intersection. A special
case of interest to us is when the elimination ideal is principal. In this case, according to
Khovanskii and Esterov [KE], our computation is equivalent to computing a mixed fiber
100
polytope [McM04].
4.4.1 Algorithms and implementation
We will compute the tropical variety of the implicit ideal by projecting the tropical
variety of the graph. So our first task is to compute the tropical variety of a generic complete
intersection ideal generated by n polynomials with generic coefficients in d + n variables.
The correctness of the following algorithm is the content of Theorem 4.1 above.
Algorithm 4.16 (Tropical variety of a generic complete intersection). Input A set of n
Laurent polynomials in d + n variables.
Output The tropical variety, with multiplicities, of the corresponding generic complete in-
tersection.
Do 1. List the exponent vectors of the input polynomials.
2. Compute the Minkowski sum of the Newton polytopes P1, . . . , Pn, and output the
inner normal vectors of the facets.
3. For each d dimensional cone C in the normal fan,
• Let w be a vector in the interior of C.
• Compute the n-dimensional mixed volume MV (facew(P1), . . . , facew(Pn)).
• If the mixed volume is positive, output the generators of the cone and the
multiplicity, which is the mixed volume times the index of the lattice gen-
erated by points in facew(P1), . . . , facew(Pd) (shifted to the origin) in its
saturation.
101
In our implementation of TrIm, the Minkowski sum is computed using the iB4e
library [Hug06], enumerating the d-dimensional cones is done by polymake [GJ00], the
mixed volumes are computed using the mixed volume library [EC95], and the integer
linear algebra for lattice indices is done using NTL [Sho]. What oversees all the computations
is a perl script. In this case, the output format is consistent with that of the current version
of Gfan [Jen] and is intended to interface with the polyhedral software Polymake [GJ00]
when it supports polyheral complexes and fans in the future.
When n = d + 1, the image variety is hypersurface and the tropical variety is
the normal fan of Newton polytope P of this hypersurface. In this case, TrIm recovers the
polytope P from the tropical variety. The maximal cones and their multiplicities in the
tropical variety gives us the directions and lengths of edges in the polytope. We put them
together using the following algorithm from Theorem 4.12.
Given a generic vector w ∈ Rk such that facew(P ) is a vertex v, the ith coordinate
vi is the number of intersections, counted with multiplicities, of the ray w + R>0ei with the
tropical variety. Here, the multiplicity of the intersection with a cone Γ is the multiplicity
of Γ times the absolute value of the ith coordinate of the primitive normal vector to the
cone Γ.
Intuitively, what we are doing here is the following. We wish to determine the
coordinates of the extreme vertex v = facew(P ) of a polytope P in a given direction w.
Our polytope is placed so that it lies in the positive orthant and touches all the coordinate
hyperplanes. To compute the ith coordinate of the vertex v, we can walk from v toward
102
the ith hyperplane along the edges of P , while keeping track of the edge lengths in the
ith direction. A systematic way to do the walk is to follow the edges whose inner normal
cone intersect the ray w + R>0ei. Recall that the multiplicity of a codim-1 normal cone is
the lattice length of the corresponding edge. Using this subroutine for computing extreme
vertices, we can build the whole polytope using the following algorithm.
Algorithm 4.17 (Beneath/Beyond). Input A codim-1 tropical variety with multiplicities
on the maximal cones.
Output The vertices and facets of the Newton polytope P .
Initialization For some generic vectors w ∈ Rk, compute the vertices facew(P ) and take
their convex hull P ′.
Iteration For each inner facet normal vector w of P ′, compute a vertex v ∈ facew(P ). If
v · w < minx∈P ′ x · w then set P ′ ←conv(P ′ ∪ {v}).
Termination There are no new vertices in any facet normal directions of P ′.
Output Vertices and facets of P = P ′.
The intialization step can be performed systematically so that P ′ is initialized to
be a simplex of dimension dim(P ) by testing O(dim(P )2) choices of w. Our algorithm is
essentially the classical Beneath/Beyond algorithm for computing convex hulls, except that
points are given to us implicitly via a linear programming (LP) oracle rather than a list.
The software iB4e [Hug06] that we used here is designed so that any LP oracle can be
plugged in and then the Beneath/Beyond algorithm is performed with the specified oracle.
103
Example 4.18. Now let us look at how TrIm handles the polynomials in Example 4.10.
The input is the following:
[s,t]
[3*t^3-6*t^2+3*t+s^3-3*s^2+6*s-1,
3*s^3-6*s^2+3*s+t^3+3*t,
-3*s^3*t^3+15*s^2*t^3-15*s*t^3-3*s^3*t^2-18*s^2*t^2+27*s*t^2-3*t^2
+6*s^3*t+9*s^2*t-18*s*t+3*t-3*s^2+3*s]
The tropical variety of the graph of this polynomial map is shown in Table 4.1, in
Gfan / polymake format. The projection onto the last three coordinates is shown in table
4.2. The tropical variety is represented as a union of cones with multiplicities. Notice that
we do not have a nice fan structure for this image. This set of cones is slightly different from
that in Example 4.10. However, the tropical variety as a set and multiplicities at smooth
points are the same. Compare the multiplicites here with those in display formula 4.13.
Table 4.3 shows the Newton polytope of the implicit equation and lattice points in it.
In the hypersurface case, the exponents of monomials with non-zero coefficients in
the implicit equation g all lie in the Newton polytope Q. After computing the vertices of
Q, enumerating the integer points in Q gives us a list of all possible monomials in g with
indeterminate coefficients. Finally, we apply numerical or integer linear algebra to compute
the indeterminate coefficients of the polynomial g.
This is done as follows. Recall that we are given a parametrization (f0, . . . , fd)
of the hypersurface V (I). We can thus pick any point τ ∈ (C∗)d and substitute its image
(x0, . . . , xd) = (f0(τ), . . . , fd(τ)) ∈ V (I) into the equation g(x0, . . . , xd) = 0. This gives us
one linear equation for the coefficients of g. We now pick a different point τ ′ in (C∗)d to get
a second equation, and so on. In this manner, we can generate a system of linear equations
104
AMBIENT_DIM
5
DIM
2
LINEALITY_DIM
0
RAYS
1 -1 -3 -3 -2
1 0 0 0 0
0 0 0 1 0
-1 -1 -3 -3 -6
-1 0 -3 -3 -3
0 0 0 0 1
1 1 0 1 0
0 1 0 0 0
0 0 1 0 0
0 -1 -3 -3 -3
-1 1 -3 -3 -2
MAXIMAL_CONES
0 1
0 5
0 9
1 2
1 5
1 6
1 8
2 3
2 5
2 6
2 7
2 8
3 4
3 8
3 9
4 5
4 10
5 7
5 8
5 9
5 10
6 7
7 8
7 10
MULTIPLICITIES
1
1
1
2
2
1
3
3
18
1
2
9
1
3
1
2
1
2
18
2
1
1
3
1
N_RAYS
11
N_CONES
24
N_INCIDENCES
3
5
6
4
3
8
3
5
5
3
3
Table 4.1: The output of TrIm for the graph in Example 4.18
105
IMAGE_RAYS
-3 -3 -2
0 0 0
0 1 0
-3 -3 -6
-3 -3 -3
0 0 1
0 1 0
0 0 0
1 0 0
-3 -3 -3
-3 -3 -2
IMAGE_CONES
0 1
0 5
0 9
1 2
1 5
1 6
1 8
2 3
2 5
2 6
2 7
2 8
3 4
3 8
3 9
4 5
4 10
5 7
5 8
5 9
5 10
6 7
7 8
7 10
IMAGE_MULTIPLICITIES
0
3
3
0
0
0
0
9
18
0
0
9
9
9
9
6
3
0
18
6
3
0
0
0
Table 4.2: The tropical variety of the implicitization problem in Example 4.18.
106
VERTICES
1 0 18 0
1 18 0 0
1 0 0 9
1 0 0 0
FACETS
0 0 162 0
0 162 0 0
0 0 0 324
2916 -162 -162 -324
VERTICES_IN_FACETS
{1 2 3}
{0 2 3}
{0 1 3}
{0 1 2}
LATTICE_POINTS
1 0 0 0
1 1 0 0
1 2 0 0
1 3 0 0
....
1 1 0 8
1 2 0 8
1 0 1 8
1 1 1 8
1 0 2 8
1 0 0 9
N_LATTICE_POINTS
715
Table 4.3: The output that describe the Newton polytope of the implicit equation and thelattice points in it, for Example 4.18
whose solution space is one-dimemensional, and is spanned by the vector of coefficients of
the desired implicit equation g. We believe that implicitization is a fruitful direction of
further study in numerical or integer linear algebra.
All our Linux software along with user manual and details about the implementa-
tion can be downloaded from my website, which is currently:
http://math.berkeley.edu/∼jyu
4.4.2 Computing tropical discriminants
Given a d × n matrix A, considered as d points in Zn, the A-discriminant is the
dual variety of the toric variety associated to A. It is an object of great interest in algebraic
geometry and combinatorics [GKZ94]. In [DFS], Dickenstein, Feichtner, and Sturmfels
107
gave a construction for the tropicalization of the A-discriminant. Their main theorem is
the following.
Theorem 4.19. [DFS, Theorem 1.1] The tropical discriminant of A is the Minkowski sum
of T (kernel(A)) and the row space of A.
Hence the key to computing the tropical discriminant is the tropical linear space of
kernel of A. The point configuration A is called non-defective if the (tropical) A-discriminant
is a (tropical) hypersurface. In this case, the defining polynomial of the hypersurface is
called the A-discriminant. We can use the same method as above to construct its Newton
polytope from the tropical discriminant.
Example 4.20. We will compute the mixed discriminant of the two quadrangles in [MO,
Figure 3]. The point configuration is the following
A =
1 1 1 1 0 0 0 0
0 0 0 0 1 1 1 1
1 2 0 0 2 0 1 1
0 0 1 2 1 1 2 0
.
The Bergman fan of kernel(A) has the following rays with a one dimensional
lineality space spanned by (1, 1, 1, 1, 1, 1, 1, 1) and 72 four-dimensional maximal cones. This
is just one possible fan structure of the tropical linear space. The number of rays and cones
will vary depending on the method used to compute them.
RAYS
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
108
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
-1 -1 -1 -1 -1 -1 -1 0
1 1 1 1 0 0 0 0
1 1 0 0 1 1 0 0
1 0 1 0 1 0 1 0
0 -1 0 -1 -1 0 -1 0
0 1 0 1 1 0 1 0
-1 0 -1 0 -1 0 -1 0
-1 -1 0 0 -1 -1 0 0
-1 -1 -1 -1 0 0 0 0
The A discriminant is a polynomial in 8 variables with 222 monomials. Its Newton
polytope has the following 28 vertices and 12 facets.
VERTICES
1 0 3 0 7 2 0 0 6
1 0 6 4 0 0 4 4 0
1 0 4 4 2 4 4 0 0
1 0 2 8 0 4 0 2 2
1 8 0 0 2 2 2 4 0
1 0 6 0 4 2 6 0 0
1 6 0 0 4 5 3 0 0
1 4 0 6 0 5 1 2 0
1 4 2 4 0 0 0 6 2
1 4 0 0 6 4 0 0 4
1 0 2 6 2 5 1 0 2
1 3 0 3 4 5 0 0 3
1 6 0 4 0 2 0 5 1
1 7 0 3 0 2 1 5 0
1 4 4 0 2 0 4 4 0
1 0 4 0 6 0 0 2 6
1 0 7 0 3 0 6 2 0
1 0 4 6 0 0 0 5 3
1 4 2 0 4 0 0 4 4
1 6 0 0 4 2 0 3 3
1 6 2 0 2 0 2 5 1
1 3 4 3 0 0 3 5 0
1 5 2 3 0 0 1 6 1
1 3 0 7 0 5 0 2 1
1 4 0 4 2 6 2 0 0
1 3 0 5 2 6 1 0 1
109
1 0 2 4 4 4 0 0 4
1 0 4 6 0 3 3 2 0
Example 4.21 (Hyperdeterminant of the 3-cube). Let A be the vertices of the 3-cube, i.e.
A =
1 1 1 1 1 1 1 1
0 0 0 0 1 1 1 1
0 0 1 1 0 0 1 1
0 1 0 1 0 1 0 1
.
The A-discriminant is
D222 = 4c000c011c101c110 + 4c001c010c100c111
+c2000c
2111 + c2
001c2110 + c2
010c2101 + c2
011c2100
−2c000c001c110c111 − 2c000c010c101c111 − 2c000c011c100c111
−2c001c010c101c110 − 2c001c011c110c100 − 2c010c011c101c100,
which is a polynomial of degree 4 in 8 variables cv where v is a vertex of the 3-cube.
The Bergman fan of the 3-cube contains 80 four-dimensional cones on the following
rays and the lineality space (1, 1, 1, 1, 1, 1, 1, 1).
RAYS
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
-1 -1 -1 -1 -1 -1 -1 0
1 1 1 1 0 0 0 0
1 1 0 0 1 1 0 0
110
0 0 -1 -1 -1 -1 0 0
1 0 1 0 1 0 1 0
0 -1 0 -1 -1 0 -1 0
0 -1 -1 0 0 -1 -1 0
0 1 1 0 0 1 1 0
0 1 0 1 1 0 1 0
-1 0 -1 0 -1 0 -1 0
0 0 1 1 1 1 0 0
-1 -1 0 0 -1 -1 0 0
-1 -1 -1 -1 0 0 0 0
The A-discriminant contains 12 monomials, and the following 6 vertices.
VERTICES
1 0 0 0 2 2 0 0 0
1 0 0 2 0 0 2 0 0
1 2 0 0 0 0 0 0 2
1 0 2 0 0 0 0 2 0
1 0 1 1 0 1 0 0 1
1 1 0 0 1 0 1 1 0
As we increase the dimension, the number of vertices of the Newton polytope is
much less than the number lattice points in it, hence the number monomials in the polyno-
mial. For instance, the A-discriminant of the 4-cube has 2,894,276 monomials, although its
11-dimensional Newton polytope has only 25,448 vertices [HSYY07]. The tropical approach
allows us to compute the Newton polytope of the A-discriminant without first computing
the entire polynomial.
111
Bibliography
[AK06] Federico Ardila and Caroline J. Klivans. The Bergman complex of a matroid
and phylogenetic trees. J. Combin. Theory Ser. B, 96(1):38–49, 2006.
[Ard] Federico Ardila. A tropical morphism related to the hyperplane arrangement
of the complete bipartite graph, arXiv:math.CO/0404287.
[Ber71] George M. Bergman. The logarithmic limit-set of an algebraic variety. Trans.
Amer. Math. Soc., 157:459–469, 1971.
[Ber75] D. N. Bernstein. The number of roots of a system of equations. Funkcional.
Anal. i Prilozen., 9(3):1–4, 1975.
[BG84] Robert Bieri and J. R. J. Groves. The geometry of the set of characters induced
by valuations. J. Reine Angew. Math., 347:168–195, 1984.
[BJS+] Tristram Bogart, Anders Jensen, David Speyer, Bernd Sturmfels, and Rekha
Thomas. Computing Tropical Varieties, arXiv:math.AG/0507563.
[Bru87] Richard A. Brualdi. Introduction to matching theory. In Combinatorial geome-
112
tries, volume 29 of Encyclopedia Math. Appl., pages 53–71. Cambridge Univ.
Press, Cambridge, 1987.
[BS98] Dave Bayer and Bernd Sturmfels. Cellular resolutions of monomial modules.
J. Reine Angew. Math., 502:123–140, 1998.
[BY06] Florian Block and Josephine Yu. Tropical convexity via cellular resolutions.
Journal of Algebraic Combinatorics, 24(1):103–114, August 2006.
[CGKW01] Robert M. Corless, Mark W. Giesbrecht, Ilias S. Kotsireas, and Stephen M.
Watt. Numerical implicitization of parametric hypersurfaces with linear alge-
bra. In Artificial intelligence and symbolic computation (Madrid, 2000), volume
1930 of Lecture Notes in Comput. Sci., pages 174–183. Springer, Berlin, 2001.
[Dev04] Mike Develin. A complexity bound on faces of the hull complex. Discrete
Comput. Geom., 32(4):471–479, 2004.
[DFS] Alicia Dickenstein, Eva Maria Feichtner, and Bernd Sturmfels. Tropical Dis-
criminants, arXiv:math.AG/0510126.
[DS04] Mike Develin and Bernd Sturmfels. Tropical convexity. Doc. Math., 9:1–27
(electronic), 2004.
[DSS05] Mike Develin, Francisco Santos, and Bernd Sturmfels. On the rank of a tropical
matrix. In Combinatorial and computational geometry, volume 52 of Math. Sci.
Res. Inst. Publ., pages 213–242. Cambridge Univ. Press, Cambridge, 2005.
113
[DY07] Mike Develin and Josephine Yu. Tropical polytopes and cellular resolutions.
Experimental Mathematics, 2007, arXiv:math.CO/0605494.
[EC95] Ioannis Z. Emiris and John F. Canny. Efficient incremental algorithms for the
sparse resultant and the mixed volume. J. Symbolic Comput., 20(2):117–149,
1995.
[Eis95] David Eisenbud. Commutative algebra, volume 150 of Graduate Texts in Math-
ematics. Springer-Verlag, New York, 1995. With a view toward algebraic
geometry.
[EK05] Ioannis Z. Emiris and Ilias S. Kotsireas. Implicitization exploiting sparseness.
In Geometric and algorithmic aspects of computer-aided design and manufac-
turing, volume 67 of DIMACS Ser. Discrete Math. Theoret. Comput. Sci., pages
281–297. Amer. Math. Soc., Providence, RI, 2005.
[EKL] Manfred Einsiedler, Mikhail Kapranov, and Douglas Lind. Non-archimedean
amoebas and tropical varieties. AIM 2004 -31, arXiv:math.AG/0408311.
[Ewa96] Gunter Ewald. Combinatorial convexity and algebraic geometry, volume 168 of
Graduate Texts in Mathematics. Springer-Verlag, New York, 1996.
[GJ00] Ewgenij Gawrilow and Michael Joswig. polymake: a framework for analyzing
convex polytopes. In Polytopes—combinatorics and computation (Oberwolfach,
1997), volume 29 of DMV Sem., pages 43–73. Birkhauser, Basel, 2000.
[GKZ94] I. M. Gel′fand, M. M. Kapranov, and A. V. Zelevinsky. Discriminants, resul-
114
tants, and multidimensional determinants. Mathematics: Theory & Applica-
tions. Birkhauser Boston Inc., Boston, MA, 1994.
[GS02] D. R. Grayson and M.E. Stillman. Macaulay 2, a software system for research
in algebraic geometry. Available at http://www.math.uiuc.edu/Macaulay2/,
2002.
[HS95] Birkett Huber and Bernd Sturmfels. A polyhedral method for solving sparse
polynomial systems. Math. Comp., 64(212):1541–1555, 1995.
[HSYY07] Peter Huggins, Bernd Sturmfels, Josephine Yu, and Debbie Yuster. The Hyper-
determinant and Triangulations of the 4-Cube. Mathematics of Computation,
2007, arXiv:math.CO/0602149.
[Hug06] Peter Huggins. ib4e: A software framework for parametrizing specialized lp
problems. In Mathematical Software – ICMS 2006, volume 4151 of Lecture
Notes in Computer Science, pages 245–247. Springer, 2006.
[Jen] Anders N. Jensen. Gfan, a software system for Grobner fans. Available at
http://home.imf.au.dk/ajensen/software/gfan/gfan.html.
[Jos05] Michael Joswig. Tropical halfspaces. In Combinatorial and computational ge-
ometry, volume 52 of Math. Sci. Res. Inst. Publ., pages 409–431. Cambridge
Univ. Press, Cambridge, 2005.
[Kat] Eric Katz. The tropical degree of cones in the secondary fan,
arXiv:math.AG/0604290.
115
[KE] Askold Khovanskii and Alexander Esterov. Elimination theory and Newton
polyhedra, arXiv:math.AG/0611107.
[Kho99] A. Khovanskii. Newton polyhedra, a new formula for mixed volume, product
of roots of a system of equations. In The Arnoldfest (Toronto, ON, 1997), vol-
ume 24 of Fields Inst. Commun., pages 325–364. Amer. Math. Soc., Providence,
RI, 1999.
[KM05] Allen Knutson and Ezra Miller. Grobner geometry of Schubert polynomials.
Ann. of Math. (2), 161(3):1245–1318, 2005.
[KSZ92] M. M. Kapranov, B. Sturmfels, and A. V. Zelevinsky. Chow polytopes and
general resultants. Duke Math. J., 67(1):189–218, 1992.
[McM04] Peter McMullen. Mixed fibre polytopes. Discrete Comput. Geom., 32(4):521–
532, 2004.
[Mik06] Grigory Mikhalkin. Tropical geometry and its applications. In Proceedings of
the International Congress of Mathematicians, Vol. II (Madrid, 2006), pages
827–852, Zurich, Switzerland, 2006. European Mathematical Society, arXiv:
math.AG/0601041.
[MO] Grigory Mikhalkin and Andrei Okounkov. Geometry of planar log-fronts,
arXiv:math.AG/0608018.
[MS05] Ezra Miller and Bernd Sturmfels. Combinatorial commutative algebra, volume
227 of Graduate Texts in Mathematics. Springer-Verlag, New York, 2005.
116
[Oxl92] James G. Oxley. Matroid theory. Oxford Science Publications. The Clarendon
Press Oxford University Press, New York, 1992.
[Pin98] Jean-Eric Pin. Tropical semirings. In Idempotency (Bristol, 1994), volume 11
of Publ. Newton Inst., pages 50–69. Cambridge Univ. Press, Cambridge, 1998.
[Pos] Alexander Postnikov. Permutohedra, associahedra, and beyond,
arXiv:math.CO/0507163.
[PS04] Lior Pachter and Bernd Sturmfels. Tropical geometry of statistical models.
Proc. Natl. Acad. Sci. USA, 101(46):16132–16137 (electronic), 2004.
[RGST05] Jurgen Richter-Gebert, Bernd Sturmfels, and Thorsten Theobald. First steps
in tropical geometry. In Idempotent mathematics and mathematical physics,
volume 377 of Contemp. Math., pages 289–317. Amer. Math. Soc., Providence,
RI, 2005.
[Rib92] Paulo Ribenboim. Fields: algebraically closed and others. Manuscripta Math.,
75(2):115–150, 1992.
[Rou07] Bjarke Hammersholt Roune. Frobby version 0.1 - software for computations
with monomial ideals. Available at http://broune.com/, 2007.
[Sch86] Alexander Schrijver. Theory of linear and integer programming. Wiley-
Interscience Series in Discrete Mathematics. John Wiley & Sons Ltd., Chich-
ester, 1986. , A Wiley-Interscience Publication.
117
[Sey80] P. D. Seymour. Decomposition of regular matroids. J. Combin. Theory Ser. B,
28(3):305–359, 1980.
[Sho] Victor Shoup. NTL: A library for doing number theory. Available at http:
//shoup.net/ntl/.
[Spe] David Speyer. Tropical linear spaces, arXiv: math.CO/0410455.
[SS04] David Speyer and Bernd Sturmfels. The tropical Grassmannian. Adv. Geom.,
4(3):389–411, 2004.
[ST] Bernd Sturmfels and Jenia Tevelev. Elimination theory for tropical varieties,
arXiv:0704.3471v1.
[Stu96] Bernd Sturmfels. Grobner bases and convex polytopes, volume 8 of University
Lecture Series. American Mathematical Society, Providence, RI, 1996.
[Stu02] Bernd Sturmfels. Solving systems of polynomial equations, volume 97 of CBMS
Regional Conference Series in Mathematics. Published for the Conference
Board of the Mathematical Sciences, Washington, DC, 2002.
[STY06] Bernd Sturmfels, Jenia Tevelev, and Josephine Yu. The newton poly-
tope of the implicit equation. Moscow Mathematical Journal, 2006, arXiv:
math.CO/0607368.
[SY94] Bernd Sturmfels and Jie Tai Yu. Minimal polynomials and sparse resultants. In
Zero-dimensional schemes (Ravello, 1992), pages 317–324. de Gruyter, Berlin,
1994.
118
[Tev] Eugene Tevelev. Compactifications of subvarieties of tori,
arXiv:math.AG/0412329.
[Vel] Mauricio Velasco. Minimal free resolutions that are not supported by a CW-
complex.
[YY] Josephine Yu and Debbie S. Yuster. Representing tropical linear spaces by
circuits, arXiv:math.CO/0611579.
[Zie95] Gunter M. Ziegler. Lectures on polytopes, volume 152 of Graduate Texts in
Mathematics. Springer-Verlag, New York, 1995.