Tropical Geometry

Tropical Geometry
Grigory Mikhalkin, Johannes Rau
November 16, 2018

Contents
1 Introduction 5
1.1 Algebraic geometry over what? . . . . . . . . . . . . . . . . . . . 5
1.2 Tropical Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Tropical monomials and integer affine geometry . . . . . . . . 11
1.4 Forgetting the phase leads to tropical numbers . . . . . . . . . 15
1.5 Amoebas of affine algebraic varieties and their limits . . . . . 17
1.6 Patchworking and tropical geometry . . . . . . . . . . . . . . . 22
1.7 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . 28
2 Tropical hypersurfaces in Rn 29
2.1 Polyhedral geometry dictionary I . . . . . . . . . . . . . . . . . . 30
2.2 Tropical Laurent polynomials and hypersurfaces . . . . . . . . 34
2.3 The polyhedral structure of hypersurfaces . . . . . . . . . . . . 40
2.4 The balancing condition . . . . . . . . . . . . . . . . . . . . . . . 47
2.5 Planar Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.6 Floor decompositions . . . . . . . . . . . . . . . . . . . . . . . . . 69
3 Projective space and other tropical toric varieties 76

3.1 Tropical affine space Tn . . . . . . . . . . . . . . . . . . . . . . . 76
3.2 Tropical toric varieties . . . . . . . . . . . . . . . . . . . . . . . . 80
3.3 Tropical projective space . . . . . . . . . . . . . . . . . . . . . . . 84
3.4 Projective hypersurfaces . . . . . . . . . . . . . . . . . . . . . . . 88
3.5 Projective curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
4 Tropical cycles in Rn 100

4.1 Polyhedral geometry dictionary II . . . . . . . . . . . . . . . . . 100
4.2 Tropical cycles and subspaces in Rn . . . . . . . . . . . . . . . . 102
4.3 Stable intersection . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.4 The divisor of a piecewise affine function . . . . . . . . . . . . 116
4.5 Tropical modifications . . . . . . . . . . . . . . . . . . . . . . . . 127
2
Contents
4.6 The projection formula . . . . . . . . . . . . . . . . . . . . . . . . 131

4.7 Examples and further topics . . . . . . . . . . . . . . . . . . . . 136
5 Projective tropical geometry 145

5.1 Tropical cycles in toric varieties . . . . . . . . . . . . . . . . . . 145
5.2 Projective degree . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.3 Modifications in toric varieties . . . . . . . . . . . . . . . . . . . 153
5.4 Configurations of hyperplanes in projective space . . . . . . . 166
5.5 Equivalence of rational curves . . . . . . . . . . . . . . . . . . . 169
5.6 Equivalence of linear spaces . . . . . . . . . . . . . . . . . . . . 170
6 Tropical cycles and the Chow group 172

6.1 Tropical cycles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
6.2 Push-forwards of tropical cycles . . . . . . . . . . . . . . . . . . 173
6.3 Linear equivalence of cycles . . . . . . . . . . . . . . . . . . . . . 175
6.4 Algebraic equivalence of cycles . . . . . . . . . . . . . . . . . . . 176
6.5 Linear systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6.6 Recession Fans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
7 Tropical manifolds 179

7.1 Tropical spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.2 Regularity at Infinity . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.3 Tropical morphisms and structure sheaves . . . . . . . . . . . . 193
7.4 Smooth tropical spaces . . . . . . . . . . . . . . . . . . . . . . . . 196
7.5 Rational functions, Cartier divisors and modifications on ab-
stract spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
8 Tropical curves 210

8.1 Smooth tropical curves . . . . . . . . . . . . . . . . . . . . . . . . 210
8.2 Divisors and linear systems . . . . . . . . . . . . . . . . . . . . . 213
8.3 Line bundles on curves . . . . . . . . . . . . . . . . . . . . . . . . 215
8.4 Riemann-Roch theorem . . . . . . . . . . . . . . . . . . . . . . . 221
9 Tropical jacobians 229

9.1 (Coarse) differential forms, integration, and the Jacobian . . 229
9.2 Abelian varieties, polarization, and theta functions . . . . . . 234
9.3 The Abel-Jacobi theorem . . . . . . . . . . . . . . . . . . . . . . 238
9.4 Riemann’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 250
3
Acknowledgements
We benefited from numerous valuable conversations, in particular with Vic-

tor Batyrev, Benoît Bertrand, Erwan Brugallé, Ana Cannas da Silva, Jim
Carlson, Yakov Eliashberg, Sergey Fomin, Andreas Gathmann, Mark Gross,
Amihay Hanany, Ilia Itenberg, Michael Kapovich, Mikhail Kapranov, Lud-
mil Katzarkov, Yujiro Kawamata, Sean Keel, Viatscheslav Kharlamov, Askold
Khovanskii, Barak Kol, Maxim Kontsevich, Lionel Lang, Conan Leung, An-
drey Losev, Diane Maclagan, Hannah Markwig, Pierre Milman, Yong-Geun
Oh, Mikael Passare, Jean-Jacques Risler, Hans Rullgård, Kristin Shaw, Eu-
genii Shustin, Yan Soibelman, David Speyer, Bernd Sturmfels, Louis Felipe
Tabera, Ilya Tyomkin, Israel Vainsencher, Oleg Viro, Ilia Zharkov and Günter
Ziegler on tropical geometry, especially during early stages of its develop-
ment.
(To be continued.)
We would like to acknowledge support and encouragement from the Clay
Mathematical Institute (USA) at the initial stage of this project. The project
was also supported in part by the European Research Council (TROPGEO
project), the Swiss National Science Foundation (projects no. 125070 and
140666). We would also like to thank the Mathematical Sciences Research
Institute in Berkeley and the Max-Planck-Institut für Mathematik in Bonn
for hospitality at different times during work on this project.
4
1 Introduction
1.1 Algebraic geometry over what?
The two words “Algebraic Geometry” may invoke different images for dif-
ferent people today. For some it might be something very abstract, and even
somewhat formalistic, adapted to work in the most general setting possible.
For many others, including the authors of this book, Algebraic Geometry is
a branch of geometry in the first place — geometry of spaces defined by
polynomial equations.
Note that (unlike many algebraic properties) the resulting geometry de-
pends not only on the type of the defining equations, but also on the choice
of the numbers where we look for solutions. The two most classical choices
are the field R of real numbers and the field C of complex numbers. Both
these fields come naturally enhanced with the so-called “Euclidean topol-
ogy” induced by the metric |x − y| between two points x, y ∈ R (or in C).
Furthermore, real algebraic varieties are differentiable manifolds (perhaps
with singularities) from a topological viewpoint, and complex algebraic va-
rieties are special kind of real algebraic varieties of twice the dimension.
To illustrate the two parallel classical theories let us recall the classical
example of the so-called elliptic curve. Namely, consider a cubic curve in the
complex projective plane CP2 . As long as the defining cubic polynomial is
chosen generically, the resulting curve is topologically a 2-dimensional torus
(see Figure 1.1). This torus is embedded in the complex projective plane
CP2 . As CP2 is 4-dimensional such embedding is beyond our imagination
tools.
Now consider the case of real coefficients. Even if the defining polynomial
is chosen generically, the topological type of the curve is not fixed. But there
are only two possible cases, see Figure 1.2.
As the ambient real projective plane RP2 is indeed 2-dimensional, we can
actually draw how the curve is embedded there. Recall that topologically
5
1 Introduction
Figure 1.1: A complex elliptic curve.
Figure 1.2: Two real elliptic curves.
RP2 can be obtained from a disc D2 by identifying the antipodal points on

its boundary circle, see Figure 1.3.
antipodal points
get identified
Figure 1.3: Real projective plane.
6
1 Introduction
oval
non-
non- trivial
trivial comp.
comp.
Figure 1.4: Elliptic curves in the real projective plane.
Figure 1.4 depicts embeddings of cubic curves in the real projective plane.
Note that there might be different pictures inside D2 before the self-iden-
tification of its boundary, but we get one of the two pictures above in
RP2 = D2 / ∼ for any smooth real curve. An example is given in Figure
1.5.
non-trivial
comp.
oval
Figure 1.5: An elliptic curve in another presentation of RP2 by a disk.
Furthermore, the inclusion R ⊂ C gives us an inclusion RP2 ⊂ CP2 as well

as an inclusion of the real curve into the corresponding complex curve (see
Figure 1.6).
We see that the same equation may yield quite different geometric spaces.
At the same time, real and complex numbers may be the only examples of
fields where algebraic geometry is that much geometric. There is also a
continuation of this series with the quaternions H and octonions O which
leads to very interesting geometry no longer based on fields as we loose
7
1 Introduction
oval non-trivial component
Figure 1.6: Complex elliptic curve with its real locus.
commutativity in the case of H or even associativity in the case of O.

In this book we study geometry based on a predecessor of the entire series
R, C, H, O, called the tropical numbers T.
1.2 Tropical Numbers

We consider the set
T = [−∞, +∞) = R ∪ {−∞}
enhanced with the arithmetic operations
“x + y” = max{x, y}, (1.1)

“x y” = x + y, (1.2)
where we set “(−∞)+ x” = “x +(−∞)” = x and “(−∞)x” = “x(−∞)” =

−∞. These operations are called tropical arithmetic operations. We use
quotation marks to distinguish them from the usual operations on R.
Definition 1.2.1
The set T enhanced with the arithmetic operations (1.1) and (1.2) is called
the set of tropical numbers.
8
1 Introduction
Remark 1.2.2
There are papers where min{x, y} is taken for tropical addition. In such
case one has to modify the set of tropical numbers to include +∞ and ex-
clude −∞. It is hard to say which choice is better. The choice of max may
be more natural from the mathematical viewpoint as we are more used to
taking the logarithm whose base is greater than 1, cf. equation (1.4). Also
when we add two numbers in this way the sum does not get smaller. How-
ever in some considerations in Computer Science and Physics (cf. [IM12])
taking the minimum is more natural. Clearly it does not really matter as
the two choices are isomorphic under x 7→ −x.
Remark 1.2.3
The term “tropical” was borrowed from Computer Science, where it was
reportedly introduced to commemorate contributions of the Brazilian com-
puter scientist Imre Simon. Simon introduced the semiring (N, min, +)
which was later baptized “tropical” following a suggestion of Christian Chof-
furt (see [Sim88; Cho92]). According to Jean-Eric Pin, the suggestion was
(also) made by Dominique Perrin (see [Pin98]). Some years later, in the
first days of what is now known as tropical geometry, the adjective “tropical”
quickly outplayed other suggestions inspired by “non-archimedean amoe-
bas”, “idempotent algebraic geometry” or “logarithmic limit sets”.
The set T is a semigroup with respect to tropical addition. It is commu-

tative, associative and admits the neutral element 0T = −∞. Neverthe-
less we do not get a group as there is no room for subtraction. Indeed, if
“x + y” = 0T , then either x = 0T or y = 0T . Thus the only element ad-
mitting an inverse with respect to tropical addition is the neutral element
0T = −∞.
We may note that tropical addition is idempotent, i.e. we have “x + x” = x
for all x ∈ T. Idempotency makes tropical numbers non-Archimedean. Let
us recall the Archimedes axiom (stated for the case of real numbers).
Axiom 1.2.4 (Archimedes)

For any positive real numbers a, b ∈ R, a, b > 0 there exists a natural number
n ∈ N such that
{z. . . + a} > b.
|a + a +
n times
9
1 Introduction
Clearly, the conventional linear order > on T = R ∪ {−∞} makes per-

fect sense tropically. Furthermore, we can express it in terms of tropical
addition: we have a ≥ b if and only if “a + b” = b. However, we have
“ |a + a +
{z. . . + a} ” = a independently of n and thus the Archimedes axiom
n times
does not hold for tropical numbers.
Remark 1.2.5
Note that the Euclidean topology on T is determined by the linear order on
T: it is generated by the sets Ua = {x ∈ T | x < a} and Va = {x ∈ T | a < x}
(as a subbase) for all possible a ∈ T .
The tropical non-zero numbers are T× = R. Of course, they form an
honest group with respect to tropical multiplication as it coincides with
the conventional addition. It is easy to check that the tropical arithmetic
operations satisfy the distribution law
“(x + y)z” = “xz + yz”.
So we see that the only defect of tropical arithmetics is the missing sub-
traction which makes the tropical numbers a so-called semifield instead of
a field. It does not stop us from defining polynomials. Let A ⊂ Nn be a finite
set of integer vectors with non-negative entries. We denote the entries of
j ∈ A by j = ( j1 , . . . , jn ). The function
X
j
“ a j x 11 · · · x njn ” = max{a j + j x} : Tn → T
j∈A
j∈A
of x = (x 1 , . . . , x n ) ∈ Tn is a tropical polynomial P in n variables. Here, a j ∈

T and j x denotes the standard scalar product i ji x i . If we also allow
negative exponents, i.e. A ⊂ Zn , we get a tropical Laurent polynomial. If we
even drop the finiteness condition of A, we obtain a tropical Laurent series.
Tropical polynomials are globally well-defined continuous (with respect
to the Euclidean topology) functions F : Tn → T with F (−∞, . . . , −∞) =
a0 . Laurent polynomials are always defined on Rn ⊂ Tn , but not necessarily
on
∂ Tn = Tn \ Rn = {(x 1 , . . . , x n ) ∈ Tn | x j = −∞ for some j}.
Indeed, if x ∈ Tn is such that x l = −∞, then all monomials “a j x j ” in a Lau-
rent polynomial F have to satisfy jl ≥ 0 whenever a j 6= −∞ as otherwise
the value of such monomial is +∞ ∈ / T.
10
1 Introduction
An infinite tropical Laurent series does not have to be well-defined even

on Rn . However, it is easy to check that the domain of a tropical Laurent
series is convex (though not necessarily open or closed).
1.3 Tropical monomials and integer affine

geometry
Each tropical monomial “a j x j ” = a j + j x with a j 6= −∞ is an affine function
λ : Rn → R, x = (x 1 , . . . , x n ), j = ( j1 , . . . , jn ). We may extend it to a part of
∂ Tn by continuity. Namely, we define a domain D with
Rn ⊆ D ⊆ Tn
as Tn minus all points x = (x 1 , . . . , x n ) with x l = −∞ whenever jl < 0.

Clearly, λ gets naturally extended to a continuous function λ̄ : D → T and
D is the maximal set in Tn where such a continuous extension exists.
The linear part j x of the monomial λ is defined by an integer vector j ∈ Zn
of the dual vector space (Rn )∗ (as we have the classical pairing j x ∈ R, x ∈
Rn ). Thus we may invert our construction and define tropical monomials
(and hereby also polynomials, series, etc.) starting from an arbitrary real
vector space V of dimension n as long as we fix a lattice NV ⊂ V . Here, by a
lattice in a vector space V we mean a discrete free subgroup N ⊂ V of rank
n.
Indeed, once we fixed NV ⊂ V we have a canonical dual lattice NV∗ ⊂ V ∗
which consists of linear functionals V → R taking integer values on NV . A
tropical monomials on V is a function of the form λ(x) + c, λ ∈ NV∗ , c ∈
R. Taking finite tropical sums of tropical monomials we get the notion of
tropical polynomials on V . We call a real vector space V together with a
lattice NV ⊂ V a tropical vector space of dimension n (or just a tropical n-
space, not to be confused with tropical affine space Tn ) and the lattice NV
the tropical lattice.
Definition 1.3.1
An affine map Φ : V → W between tropical vector spaces V, W is called
integer affine if for any x, y ∈ V with x − y ∈ NV we have (Φ(x) − Φ( y)) ∈
NW . The map Φ is called an integer affine transformation of V if V = W , it is
11
1 Introduction
invertible and the inverse function is an integer affine map, too. We define
the differential dΦ : V → W as the conventional differential (identifying
V = Tx V and W = Tφ(x) W ). Note that dΦ takes NV to NW . Clearly, an
integer affine map Φ : V → V is an integer affine transformation if and only
if dΦ|NV is a bijection to NV .
We can generalize these notions to local versions whenever the underly-
ing space comes equipped with a so-called integer affine structure.
Definition 1.3.2
Let M be a smooth n-dimensional manifold. An integer affine structure on
M is given by an open cover Uα and charts φα : Uα → Rn such that for
each pair α, β the overlapping map φβ ◦ φα−1 is locally the restriction of an
integer affine transformation Φβα : Rn → Rn .
Two integer affine structures on M are considered equivalent if the union
of their covers and charts gives an integer affine structure as well. A man-
ifold M together with an equivalence class of integer affine structures is
called an integer affine manifold.
Example 1.3.3
Let A be a real affine space with tangent vector space V . Obviously, by
choosing a lattice NA ⊂ V (i.e. turning V into a tropical vector space) induces
an integral affine structure on A. It is tempting to call such an integral affine
manifold a tropical affine space. However, in this text we reserve this term
for the space Tn (following the algebro-geometric viewpoint here).
Example 1.3.4
Let (V, N ) be a tropical vector space and let Λ ⊂ V be an arbitrary lattice
(unrelated to N ). We declare points x, y ∈ A equivalent if x − y ∈ Λ. Let
T := V /Λ denote the quotient space. Then T is a integer affine manifold
diffeomorphic to (S 1 )n . Indeed, the quotient map V → T can be inverted
locally and the usual atlas given by local inverse maps provides an integer
affine structure on T . We call T a tropical torus.
We will see that the tropical structure on general tropical manifolds can
be thought of as an extension of integer affine structures to polyhedral com-
plexes. Let us preview this (without detailed explanation) by revisiting the
example of a smooth cubic curve in the plane.
12
1 Introduction
Now we do everything tropically borrowing the notions from the main

part of the book. We will see that the tropical projective plane TP2 can be
viewed as a compactification of the tropical 2-space R2 . Combinatorially,
this compactification is a triangle whose interior is identified with the whole
R2 .
Figure 1.7: Tropical projective plane TP2 .
A curve in TP2 might be represented as a picture in R2 which has to be

compactified to get the full picture. Or we can consider already compact
pictures in the triangle whose interior is equipped with an distorted inte-
ger affine structure inherited from a diffeomorphism with R2 . As we shall
see in this book, Figure 1.8 depicts a smooth cubic curve in R2 before the
compactifiction. It is the hypersurfaces in R2 defined by a tropical cubic
polynomial in two variables. In the compactified view, the edges are no
longer straight. Figure 1.9 provides a sketch (where the curvature of the
edges is perhaps still not visible).
We may note that our cubic intersects each side of the triangle three
times. Similarly to the situation over the real numbers, there is more than
one “type” of a smooth tropical curve of given degree, but only finitely many.
To preview the relation between classical and tropical varieties we may
look at Figure 1.10 which depicts the collapse of the complex elliptic curve
from Figure 1.1 to the tropical cubic curve in TP2 . The left-hand side of
Figure 1.10 is topologically a 2-torus S 1 × S 1 minus nine points. These are
the nine points (three from each side of the triangle, cf. Figure 1.9) that are
13
1 Introduction
Figure 1.8: A tropical cubic curve in R2 (before compactification).
Figure 1.9: A tropical cubic curve embedded in TP2 .
attached after the compactification in TP2 . In the next section we look at

the calculus that governs the collapse of complex varieties to tropical ones.
14
1 Introduction
Figure 1.10: Collapse of a complex cubic curve to a tropical cubic curve.
1.4 Forgetting the phase leads to tropical

numbers
To relate complex and tropical numbers consider a complex number
z = ae iα , a, α ∈ R, a ≥ 0.
The exponent α is called the argument, or the phase of z. It gets ignored

when passing to tropical numbers. In order to do this consistently with
addition and multiplication we measure the remaining paramater, the norm
a = |z|, on logarithmic scale. We set
x = log t a = log t |z|,
where the base t of the logarithm is a large real number.

The map z 7→ log t |z| is a surjection C → T which we denote by Log t (of
course, we set log t 0 = −∞). This map “forgets” the phase of z. Further-
more, it rescales the norm. We may use Log t to induce arithmetic operations
on T from C. However, since Log t is not injective the resulting operations
might be multivalued.
Indeed, addition on C induces the following operation on T.
x ¾ t y = Log t (Log−1
t
x + Log−1
t
y) (1.3)
15
1 Introduction
We should stress here that Log−1 t

x+Log−1 t
y denotes the set {z+w : Log t (z) =
x, Log t (w) = y} and we define x ¾ t y to be the image of this set under Log t .
As Log−1t
x is a circle {z ∈ C | |z| = t x }, we get
x ¾ t y = [log t |t x − t y |, log t (t x + t y )] ⊂ T,
i.e. x ¾ t y is an interval in T instead of a specific number.

To get a unique number we define
“x + t y” = max{x ¾ t y} = log t (t x + t y ). (1.4)
This expression has a well-defined limit when t → +∞. We get
“x + y” = lim “x + t y” = lim log t (t x + t y ) = max{x, y},

t→+∞ t→+∞
thus recovering tropical addition as a certain limit of addition of complex

numbers with the help of rescaling by log t once we forgot the phase.
Remark 1.4.1
Note also that if x 6= y, then we have
“x + y” = lim min{x ¾ t y} = lim log t |t x − t y | = max{x, y}.

t→+∞ t→+∞
We see that in this case “x + y” is the single limit of the multi-valued oper-
ation x ¾ t y and therefore independent of the chosen phases for the preim-
ages of x and y. If x = y then min{x ¾ t y} = −∞, so it is independent of
t. Hence, in the realm of multivalued operations, the limit of ¾ t for t → ∞
is the multivalued tropical addition
¨
max(a, b) if a 6= b,
a+b=
[∞, a] if a = b.
The operations ¾ t and + are examples of hypergroup additions, see [Vir10]

for the relevant treatment in the context of tropical calculus (more explic-
itly, the definitions of ¾ t and + are given in sections 5.4 and 5.3). Viro
suggests an another useful viewpoint on tropical calculus using the multi-
valued addition + instead of the more conventional single-valued version
given by max.
16
1 Introduction
Multiplication in C induces a well-defined single-valued operation

“x y” = Log t (Log−1
t
x · Log−1
t
y) = log t (t x t y ) = x + y,
as the norm of the product of two complex numbers is independent of their
phases.
1.5 Amoebas of affine algebraic varieties

and their limits
The map Log t : C → T can be applied coordinatwise to generalize to the
case of several variables.
Log t : Cn → Tn ,
(z1 , . . . , zn ) 7→ (log t |z1 |, . . . , log t |zn |)
Clearly, Log t ((C× )n ) = Rn . Images of algebraic subvarieties V ⊆ (C× )n under
Log t are called amoebas. They were introduced in [GKZ08, Chapter 6]. The
most well-known example is the amoeba
A t = Log t ({(z, w) ∈ (C× )2 | z + w + 1 = 0}) (1.5)
depicted in Figure 1.11.
It is the closed set in R2 bordered by three arcs
t x + t y = 1, t y + 1 = tx, 1 + tx = t y.
We add to the picture the three “asymptotics of the tentacles”: the nega-
tive part of the x-axis, the negative part of the y-axis and the diagonal ray
{(x, x)|x ≥ 0}. The union of these three rays, with the origin as vertex, is
denoted by Γ . The tripod Γ separates the amoeba A t into three equal parts.
To see that these parts are equal we note that the whole picture — without
2
the boundary points — is symmetric with respect to the linear action on R
0 −1
by the symmetric group S3 generated by 1 −1 and this action interchanges
the three parts of the amoeba. Also, the action is volume-preserving as our
S3 is a subgroup of SL2 (Z). Thus the area Z t (which clearly depends on the
parameter t > 1) of each part is the same. One can show that
π2
Zt = (1.6)
6(log t)2
17
1 Introduction
Figure 1.11: The amoeba A t and three rays inside.
(in the following, the logarithm without index always refers to the natural
base logarithm log := loge ) as we see in the next remark.
Remark 1.5.1
It was observed by Passare (see [Pas08]) that two different ways of comput-
ing the area Ze produces yet another proof of Euler’s formula ζ(2) = π2 /6.
log
Indeed, as log t = log(t) coordinatewise it suffices to establish (1.6) for t = e.
Let us compute the area of the part of Ae in the negative quadrant. It is
given by
Z0 Z0 Z0
dxd y = − log(1 − e x )d x
−∞ log(1−e x ) −∞
Z 0 X
∞ ∞ Z 0 ∞
e nx X e nx X 1
= dx = dx = = ζ(2).
−∞ i=1
n i=1 −∞
n i=1
n2
On the other hand, one can show (using the fact that {z + w + 1 = 0} ∈ C2
is a holomorphic submanifold) that the area of Ae is half of the area of the
corresponding coamoeba, which is the image of {z + w + 1 = 0} ∈ (C× )2
under the argument map
Arg : (C× )2 → (R/2πZ)2 , (a1 e iα1 , a2 e iα2 ) 7→ (α1 , α2 ).
Using elementary triangle geometry (for all 0 < β1 , β2 , β1 + β2 < π there
exists a triangle with vertices 0, 1 ∈ C and interior angles β1 and β2 ) one can
18
1 Introduction
show that the coamoeba consists of two congruent triangles in (R/2πZ)2

(neglecting the boundary), each of area π2 /2. Hence area(Ae ) = π2 /2.
But as ζ(2) represents a third of Ae , it follows ζ(2) = π2 /6.
If we consider a line {az + bw + c = 0} with a, b, c non-zero, then the
amoeba Log t {az + bw + c = 0} can be obtained from A t by the translation
x 7→ x + Log t c − Log t a,
y 7→ y + Log t c − Log t b,
as {az + bw + c = 0} can be obtained from {z + w + 1 = 0} by the rescaling
c c
z 7→ z, w 7→ w.
a b
Note that if V ⊆ Cn is fixed the only effect of varying t is the scaling of
the target Rn with the coefficient log1 t , i.e.
1
Log t (V ) = Log(V ).
log t
In particular, the limit of Log t {az + bw + c = 0} when t → +∞ does not
depend on the coefficients (as long as they are non-zeroe) and is equal to
Γ , the union of the three rays inside A t , see Figure 1.11.
The situation changes if we vary V simultaneously with varying the base
t, i.e. if we consider a family of complex varieties Vt ⊆ (C× )n with a real
parameter t.
For example, let us take a family of lines Vt = {a(t)z + b(t)w + c(t) = 0}
in (C× )2 , where a(t) = αt A + o(t A), b(t) = β t B + o(t B ), c(t) = γt C + o(t C ),
α, β, γ ∈ C× , A, B, C ∈ R are functions for large positive values of t with
highest order terms as described. Then the limit L of Log t (Vt ) for t → +∞
(which we may consider in the topology induced by the Hausdorff metric
on neighbourhoods of compact sets in Rn ) depends only on A, B, C and is
equal to the translation of Γ by
x 7→ x + C − A,
y 7→ y + C − B.
Thus the numbers A, B, C define asymptotics of the amoebas Log t (Vt ). Within
the paradigm of tropical geometry we regard the limits L as geometric ob-
jects of their own, the so called tropical lines, and A, B, C ∈ R ⊂ T as tropical
19
1 Introduction
Figure 1.12: Tropical lines intersect in one point. They can be used to con-
nect any pair of points in T2 .
coefficients defining this line. Indeed, we will later define the notion of a
tropical hypersurface of a tropical polynomial, and in the case of the linear
tropical polynomial “Ax +B y +C” this hypersurface will be equal to L. Note
that the tropical lines we can get in this way only differ by translations in
R2 .
Remark 1.5.2
Let us consider a special case when b(t) = −1, a(t) = t A, c(t) = t C . Then
the curve Vt is a graph of the function w(z) = t Az + t C while its amoeba can
be written as
Log t (Vt ) = { y = “Ax” ¾ t C},
i.e. it can be thought of as the graph of the multivalued addition (1.3) of
“Ax” = A + x ∈ T and C ∈ R. The statement continues to hold in the limit
case t → ∞ if we use multivalued tropical addition +, i.e.
lim Log t (Vt ) = { y = “Ax” + C},

t→∞
We easily get some familiar properties of lines for the new piecewise lin-
ear objects. Two generic lines intersect in a single point. And for two generic
points in R2 there is a unique line connecting them. We get more tropical
lines in R2 if we allow the coefficients A, B, C to be −∞ = 0T as well: con-
ventional horizontal lines, vertical lines and diagonal lines parallel to the
vector (1, 1), see Figure 2.21.
Remark 1.5.3
For an affine variety V ⊂ Cn we call Log t (V ) ⊂ Tn its affine amoeba. Note
20
1 Introduction
Figure 1.13: Amoeba of a parabola.
that if V ⊂ Cn is irreducible and is not contained in a coordinate hyperplane

z j = 0, j = 1, . . . , n then V is the topological closure of V ∩ (C× )n in Cn .
Furthermore, the amoeba Log t (V ) is the topological closure of Log t (V ∩
(C× )n ) in Tn . For example, the affine amoeba A¯t ⊂ T2 of the line {z+w+1 =
0} ⊂ C2 can be obtained from A t (see (1.5)) by adding two points
A¯t = A t ∪ {(−∞, 0)} ∪ {(0, −∞)}
at the far left and far lower apex of A t at Figure 1.11.

As a second example type of amoebas, let us consider the graph G t =
{(z, w) ∈ C2 | w = f t (z)} of a polynomial
d
X
f t (z) = a j (t)z j
j=0
in z whose coefficients a j (t) are C-valued functions in the positive real vari-
able t (defined for sufficiently large values of t) such that there exist α j ∈ C×
and A j ∈ R
a j (t) − α j t A j
lim = 0,
t→+∞ tAj
in other words, a j (t) = α j t A j + o(t A j ), t → +∞.
We may note that the amoeba Log t (G t ) is contained in the graph of the
multivalued polynomial
F t = A0 ¾ t “A1 x” ¾ t · · · ¾ t “Ad x d ” (1.7)
21
1 Introduction
Figure 1.14: A tropical parabola as the limit of amoebas when t → +∞.
if a j (t) = α j t A j . The case of d = 2, i.e. when G t is a parabola is depicted on

Figure 1.13.
Let us compare F t against the corresponding tropical polynomial
d
X
F (x) = “ A j x j ”.
j=0
By Remark 1.4.1 the set F t (x) ⊂ T has a one-point limit {F (x)} when t →
+∞ if x is such that the collection {“Ad x d ”} has a unique maximum. If
this maximum is not unique then this limit is [−∞, F (x)]. Again, using
multivalued tropical addition this can be reformulated by saying that the
limit of Log t (G t ) is equal to the graph of F∞ = A0 + “A1 x” + · · · + “Ad x d ”.
Clearly the same holds if we slightly perturb the coefficients a j (t) from
α j t A j to α j t A j + o(t A j ). The limit of Log t (G t ) is shown on Figure 1.14.
1.6 Patchworking and tropical geometry

We saw that if we change the coefficients a, b, c in az + bw + c = 0, the
amoeba only gets translated. If we choose a, b, c to be real, then the bound-
ary of the amoeba A t will be the image of the real locus of the line L. In
formulas, ∂ A t = Log t (RL) where RL = {(z, w) ∈ (R∗ )2 |az + bw + c = 0} =
L ∩ (R∗ )2 .
If we assume a, b, c 6= 0, then the three arcs in the boundary of A t cor-
respond to the three components of RL ∩ (R∗ )2 which in turn correspond
22
1 Introduction
Figure 1.15: Log t maps the real line z + w + 1 = 0 to the three boundary
arcs of the amoeba.
to three out of four quadrants of R2 . Which arc corresponds to which

quadrant is determined by signs of the coefficients a, b, c. Furthermore,
even if a, b, c are functions of a real positive parameter t → +∞ of the
form a(t) = αt −A + o(t −A) and the function is real we may still speak of
the sign of a(t) as the sign of the leading coefficient α ∈ R. Consider
Γ = lim t→+∞ Log t (L t ) where L t is given by a(t)z + b(t)w+c(t) = 0 with the
real coefficients a(t), b(t), c(t). We saw already that Γ is a graph in R2 with
one vertex. Now let us focus on RL again. First, we split (R∗ )2 into its four
quadrants R2>0 × {+, −}2 and compute the limits of RL for each quadrant
separately (see Figure 1.16). As described before, these components corre-
spond to the three boundary arcs of A t , so the limit is easy to compute: For
each quadrant, we just get a part of the tropical line consisting of two rays.
In our example, the signs of a(t), b(t), c(t) are all positive, i.e. α, β, γ > 0.
We can summarize these pictures by drawing the whole limit as before,
but now labeling the edges of Γ with the signs of the real quadrants whose
part of RL converge to the edge (see Figure 1.17).
Let us now construct RP2 (as a topological space) by gluing together four
copies of TP2 along the sides at infinity as indicated in Figure 1.18. The
four boundary sides of the picture are also glued together by identifying
antipodal points. The inside of each triangle is homeomorphic to R2 .
Now we can just redraw the quadrant pictures from above in this repre-
sentation of RP2 , see Figure 1.19. Equivalently, we could first draw a copy
of Γ in each of the quadrants and then throw away those edges which are
not labelled with the corresponding sign in the previous picture.
23
1 Introduction
(−, +) R2>0 (−, +) R2

(+, +) (+, +)
Log t , t → ∞
(+, −)
(+, −)
(−, −) (−, −)
Figure 1.16: The logarithmic limit for each quadrant separately.
(−, −),
(−, +), (+, −)
(−, +)
(−, −), (+, −)
Figure 1.17: A tropical line whose edges are labelled with signs referring to
the 4 quadrants.
As indicated, we denote the resulting set in RP2 by RΓ . Note that RΓ is

closed in RP2 and that topologically it is an embedded circle S 1 = RP1 ⊂
RP2 . It can be considered as the limit of RL t under a certain reparameter-
ization (the so-called phase-tropical limit). Furthermore, RL t is isotopic to
RΓ .1 A similar construction works not only for lines but for smooth alge-
braic curves in RP2 . It was introduced by Viro in 1979 and is now known
as Viro patchworking (cf. [Vir79]; for a list of references see [Vir06]). It
is the most powerful construction tool currently known in Real Algebraic
Geometry. One of the major breakthroughs obtained with the help of this
technique was the construction by Itenberg of a counterexample to the so-
1
In our case of a line, this is true for any t (as long as the functions a(t), b(t), c(t) are
defined at t and do not vanish simultaneously). In more general cases we get similar
isotopies for large values of t (whenever RΓ is smooth as a tropical variety).
24
1 Introduction
R2
Figure 1.18: RP2 obtained from gluing four copies of TP2 (antipodal points
are also identified).
RΓ
Figure 1.19: A real tropical line.
called Ragsdale conjecture standing open since 1906 (see [Rag06] for the
conjecture and [IV96] for the counterexample). Let us review the relevant
background of this conjecture.
There are only two homology types of circles embedded in RP2 . The line
RL is an example of the non-trivial class (as the complement RP2 \ RL is
still connected). A circle which bounds a disc in RP2 is zero-homologous
and is called an oval. If we consider a smooth real algebraic curve RC ⊂
RP2 , then its connected components are embedded circles. Note that non-
trivially embedded circles must intersect by topological reasons and that
their intersection points must be singular points of RC. Thus if the degree
of RC is even then all its components are ovals; if it is odd, then all but one
component are ovals.
25
1 Introduction
Figure 1.20: A tropical curve of degree 10.
Each oval separates its complement into two connected components, the
interior (homeomorphic to a disc) and the exterior (homeomorphic to a
Möbius band). An oval is called even if it sits in the interior of an even
number of other ovals (and odd otherwise). The Ragsdale conjecture stated
that the number of even ovals, denoted by p, of a smooth real curve of even
degree 2k is bounded by
3k(k − 1)
p≤ + 1.
2
It was noted by Viro [Vir80] that this inequality comes as a special case of
the more general conjecture
b1 (RX ) ≤ h1,1 (X ), (1.8)
26
1 Introduction
Figure 1.21: The quasitropical limit of the counterexample to the Ragsdale

conjecture from [Ite93].
where X is a smooth complex algebraic surface defined over R and RX is its

real locus (as usual, b1 stands for the first Betti number and h1,1 stands for
the (1, 1)-Hodge number). Furthermore, the inequality (1.8) implies the
bound
3k(k − 1)
n≤ + 1.
2
for the number n of odd ovals that is weaker than the historical Ragsdale
conjecture [Rag06] by 1.
In one of the first spectacular applications of the patchworking technique
in [Vir80] Viro disproved the original Ragsdale conjecture for the number
of odd ovals by that very one, leaving the more general conjecture (1.8) still
plausible. Then the final counterexample was given by Itenberg [Ite93] in
27
1 Introduction
yet another striking application of patchworking.

Figure 1.20 shows a very particular tropical curve of degree 10. As in
the example of the line, this curve can be obtained as the limit of a family
of real algebraic curves, given by a family of equations. After a choice of
signs for the (leading terms of) the coefficients of these equations, we can
again draw the quasitropical limit in the four quadrants (see Figure 1.21).
The result corresponds to a smooth algebraic curve of degree 10 in RP2 with
p = 32 even ovals (don’t forget the big one which crosses the line at infinity
several times), which exceeds the Ragsdale bound 3·5·4 2 + 1 = 31.
Tropical geometry can be viewed as a further development and gener-
alization of patchworking. The first volume of this book takes an intrinsic
point of view — Tropical Geometry per se. In some applications (e.g. some
simple cases in Gromov-Witten theory) in can completely replace Complex
Geometry. In the first volume we will look only at the most elementary
instances of such applications (rational curves in toric varieties). In the
second volume we plan to take a more detailed look at amoebas and at the
so-called phase-tropical varieties linking tropical and complex geometries.
1.7 Concluding Remarks
28
2 Tropical hypersurfaces in Rn
We start our presentation of tropical geometry with tropical hypersurfaces
in Rn . Recall from the introduction that R can be regarded as the set of
tropical non-zero numbers T× . Hence, using the terminology from algebraic
geometry, Rn = (T× )n is the tropical algebraic torus. Subvarieties of the
algebraic torus are usually called very affine varieties in algebraic geometry.
In this sense, we are dealing with tropical very affine geometry in this (and
the next) chapter.
As many examples in the introduction showed, tropical algebraic geom-
etry can typically be translated into notions from ordinary affine geometry
(unfortunately in slight conflict with the meaning of “affine” in the previous
paragraph) after unwrapping the definition of tropical arithmetics. Table
2.1 shows an incomplete dictionary between tropical algebro-geometric and
ordinary affine language and the present chapters focuses on exploring this
interplay in the case of hypersurfaces. It is therefore appropriate to start by
recalling some basic notions from affine and polyhedral geometry.
Throughout the following, any reference to topology refers to the Eu-
clidean topology of Rn .
algebraic geometry affine geometry

algebraic torus (T× )n vectorspace Rn
(T× )n -torsor affine space A
monomials “x j ” integer linear forms j x
monomial maps Φ : (T× )n P
→ (T× )m integer linear maps A : Rn → Rm
polynomials f = j a j x j PL convex functions f = maxi {a j + j x}
hypersurfaces V ( f ) n − 1-dim’l polyhedral subspaces X ⊂ Rn
Table 2.1: A dictionary between tropical algebraic and affine geometry of

Rn
29
2.1 Polyhedral geometry dictionary I

In this first section, we collect the very basic notions from polyhedral ge-
ometry that will be used throughout the rest of the text. In this way, we fix
our terminology while avoiding unwanted interruptions in the subsequent
sections. For more details and proofs of the theory needed here, we refer
the reader to the (far more comprehensive) expositions in [GKZ08; Zie95;
BG09].
2.1.1 Tropical vector spaces

Let V be a real vector space. A subgroup N ⊂ V is called lattice in V if it
is free of rank n and discrete (equivalently, free of rank n and spans V ). A
vector space V together with lattice N is called a tropical vector space. We
use this terminology for lack of better alternatives. In particular, note that
does not carry a T-module structure, as one might expect in analogy with
real and complex vector spaces.
If not specified otherwise, we always regard Rn as tropical vector space
with lattice Zn . A subspace W ⊂ V is called a rational subspace if dim(W ) =
rank(W ∩ Λ). It follows that W ∩ Λ resp. Λ/(W ∩ N ) form lattices for W
resp. V /W and turn these spaces into tropical vector spaces. The dual vector
space V ∗ is a tropical vector space with dual lattice N ∗ = Hom(N , Z) (where
we identify λ ∈ N ∗ with the function V → R given by linear extension). The
functions λ ∈ N ∗ are called integer linear forms or integer linear functions
on V .
A function κ : V → R is called integer affine (or a tropical monomial) if
it is the sum of an integer linear form and a real constant. In other words,
they have the form
κ(x) = a + j x
for some j ∈ N ∗ , a ∈ R.
A linear map Φ : V → W between two tropical vector spaces V = N ⊗ R
and W = M ⊗ R is called an integer linear map if it is induced by an linear
map N → W , i.e. if it sends lattice vectors to lattice vectors. A map Ψ :
V → W is called integer affine if it is of the form Ψ(x) = Φ(x) + w with α
integer linear and w ∈ W . The map Ψ is called an integer affine isomorphism
or tropical isomorphism if there exists an integer affine inverse map. An
30
Figure 2.1: An unbounded and a bounded polyhedron
integer affine map Ψ : Rn → Rm can be written as x 7→ Ax + b, where

A ∈ Mat(m × n, Z) is a matrix with integer entries and b ∈ Rm . The map Ψ
is an isomorphism if and only if A is a square matrix with determinant ±1,
i.e. if A ∈ GL(n, Z).
2.1.2 Polyhedra
For each integer affine function κ on V , we define the rational halfspace of
κ to be
Hκ := {x ∈ V : κ(x) ≥ 0} = {x ∈ V : j x ≥ −a}.
A subset σ ⊆ V is called a rational polyhedron of X if it is the intersection
of finitely many rational halfspaces. Let us emphasize that the adjective
“rational” refers to the fact that the bounding inequalities have linear parts
in N ∗ (or, equivalently, N ∗ ⊗ Q). As we will not work with other halfspaces
or polyhedra, and also to avoid conflicts with other usage of the attribute
“rational”, we mostly drop it in the following. Figure 2.1 depicts an un-
bounded and a compact polyhedron.
The Minkowski sum of two polyhedra
σ + σ0 = {x + y : x ∈ σ, y ∈ σ0 }
is a polyhedron again. A face τ of a polyhedron σ is given as σ∩H−κ , where

κ is an integer affine-linear function such that P ⊂ Hκ . If τ = σ ∩ H−κ 6= ;,
Hκ is called a supporting halfspace of σ. The (relative) boundary ∂ σ of σ is
the union of all proper faces. The complement σ := σ \ ∂ σ is called the
(relative) interior of σ. We denote by L(σ) the real subspace of V spanned
31
by σ and called the linear span of σ. More precisely, L(σ) is spanned by all
differences x − y, x, y ∈ σ. We define the dimension of σ by dim(σ) :=
dim(L(σ)). We set LZ (σ) = L(σ) ∩ Zn . If σ is is rational, then LZ (σ) is a
lattice in L(σ).
Let σ ⊂ V be a polyhedron. The recession cone rc(σ) of σ is the set of
direction vectors of all rays contained in σ,
rc(σ) :={v ∈ V : ∃ x ∈ σ such that x + µv ∈ σ ∀ µ ≥ 0}

={v ∈ V : x + µv ∈ σ ∀ x ∈ σ, µ ≥ 0}.
Equivalently, if σ is given as the intersection of the affine halfspaces ji x ≥

−ai , i = 1, . . . , k, then rc(σ) is the cone obtained as the intersection of the
linear halfspaces ji x ≥ 0, i = 1, . . . , k (see [BG09, Proposition 1.23]).
Any polyhedral cone σ ⊂ V is equal to
σ = R≥0 v1 + . . . R≥0 vk
for suitable vectors v1 , . . . , vk ∈ V , called generators of σ. Any bounded

polyhedron σ ⊂ V is the convex hull Conv(A) of a finite set of points A ⊂
V , |A| < ∞. Finally, an arbitrary polyhedron σ ⊂ V can be written as a
Minkowski sum σ = τ + ρ, where τ is bounded and ρ is a cone. Here, τ is
not unique, but ρ = rc(σ). See [BG09, Sections 1.B and 1.C].
2.1.3 Polyhedral complexes and fans

A collection P = {σ1 , . . . , σm } of polyhedra is called a polyhedral complex if
for each σi all faces are also contained in P and if each intersection σi ∩σ j
produces a face of both σi and σ j (if nonempty). The S elements σi of P
are called the cells of P . The support of P is |P | := i σi . If |P | is equal
to a polyhedron P (e.g. Rn ) we call P a polyhedral subdivision of σ (or
Rn , respectively). Given two polyhedral complexes X and Y , we call X a
refinement of Y if each cell of X is contained in a cell of Y . In this case,
the cells of X contained in a fixed cell σ ∈ Y form a polyhedral subdivision
of σ.
We say that P is of pure dimension n if all maximal polyhedra in P have
dimension n. We will mostly deal with pure-dimensional complexes in the
following. The cells of dimension n, 1 and 0 are called facets, edges and
32
vertices, respectively. An edge containing a single vertex is called a ray. The

k-skeleton of a polyhedral complex is the set of polyhedra
P (k) := {P ∈ P : dim(P) ≤ k}.
It forms a polyhedral complex again.

A polyhedron σ is called a cone if for each x ∈ σ the whole ray R≥0 x is
contained in σ. Equivalently, it is the finite intersection of rational halfs-
paces given by integer linear (not affine) functions. A polyhedral complex
of cones is called a polyhedral fan. A fan is called pointed if it contains {0}.
In this case, all of its cones are pointed, which means have {0} as a face.
Any polyhedron σ in V gives rise to a fan in dual space V ∗ which is called
its normal fan and constructed as follows. For each face τ of σ, let Cτ be
the cone in V ∗ consisting of those linear forms which are bounded on σ
from below and whose minimum on σ is attained on τ,
Cτ := {λ ∈ V ∨ : λ(x) ≥ λ( y) for all x ∈ σ, y ∈ τ}.
The collection of cones Cτ for all faces τ of σ forms a polyhedral fan which
is the normal fan of σ, denoted by N (P) (see [BG09, Proposition 1.67]).
Let X be a polyhedral complex in V and fix a cell τ ∈ X . For any cell
σ ∈ X containing τ, we define σ/τ to be the polyhedral cone in V / L(τ)
generated by all differences x − y, x ∈ σ, y ∈ τ. The collection of σ/τ for
all σ ⊇ τ forms a pointed fan in V / L(τ) denoted by StarX (τ) and called
the star of X at τ.
Exercise 2.1.1
Let σ and σ0 be polyhedra such that σ ∩ σ0 6= ;. Show that rc(σ ∩ σ0 ) =
rc(σ) ∩ rc(σ0 ).
Exercise 2.1.2
Prove that StarX (τ) = {σ/τ : τ ⊂ σ} is a pointed fan.
Exercise 2.1.3
Let X and Y be two polyhedral complexes in Rn . Show that
X ∩ Y := {σ ∩ σ0 : σ ∈ X , σ0 ∈ Y }
is a polyhedral complex and |X ∩ Y | = |X | ∩ |Y |.
33
Exercise 2.1.4
Let X and Y be two polyhedral complexes in Rn and Rm , respectively. Show
that
X × Y := {σ × σ0 : σ ∈ X , σ0 ∈ Y }
is a polyhedral complex in Rn+m and |X × Y | = |X | × |Y |.
2.2 Tropical Laurent polynomials and

hypersurfaces
The natural class of functions on the tropical algebraic torus Rn = (T× )n are
tropical Laurent polynomials (in contrast to tropical polynomials, which
are the those Laurent polynomials that extend to Tn ). In this section, we
develop the basic properties of these functions.
Recall that a tropical monomial κ on Rn is a function of the form
j
κ(x 1 , . . . , x n ) = “ax 11 . . . x njn ”,
where a ∈ R, j1 , . . . , jn ∈ Z and x 1 , . . . , x n are the coordinates of Rn . In

standard arithmetic operations this can be expressed as
κ(x 1 , . . . , x n ) = a + j1 x 1 + · · · + jn x n .
As before we will use multi-index notation j = ( j1 , . . . , jn ) ∈ Zn resp. x =

(x 1 , . . . , x n ) ∈ Rn and write
κ(x) = “ax j ” = a + j x,
where j x is the standard scalar product.

A tropical Laurent polynomial f : Rn → T is a finite sum of tropical mono-
mials, i.e., a function of the form
X
f (x) = “ a j x j ” = max{a j + j x},
j∈A
j∈A
with A ⊆ Zn finite and a j ∈ R for all j ∈ A. By definition, a Laurent poly-

nomial is a convex, piecewise integer affine function. We will presently
give an example (see example 2.2.2) showing that the coefficients a j are in
34
general not uniquely determined by the function f . In this book, we will

adopt the convention that a Laurent polynomial is a function whereas an
assignment of coefficients Zn → T, j 7→ a j (with a j = −∞ for almost all
j) is called a representation of f . We will say a few more words about the
ambiguity later in this chapter.
Example 2.2.1
If a tropical Laurent polynomial f can be represented by only one term,
then it is just an integer affine function, as discussed before. If f consists
of two terms, say f = “ai x i + a j x j ”, then it divides Rn into two halfspaces
along the hyperplane ai + i x = a j + j x. In one half, f equals ai + i x, in the
other f is a j + j x. In particular, f is locally integer affine except for points
in the hyperplane, where f is non-differentiable and strictly convex.
Example 2.2.2
We take three polynomials f1 = “0 + 1x + x 2 ”, f2 = “0 + x + x 2 ” and f3 =
“0 + (−1)x + x 2 ” in one variable x. First, let us point out some possibly
confusing facts. Note that neither the constant term 0 not the coefficient
1 of 1x can be omitted here, because 0 6= 0T = −∞ and 1 6= 1T = 0. In
fact, x 2 = “0x 2 ”, but “1x” 6= x = “0x”. After this piece of warning, let us
just draw the graphs of the three functions (see Figure 2.2). The important
max{0, 1 + x, 2x} max{0, x, 2x} max{0, (−1) + x, 2x}
Figure 2.2: Three tropical polynomials
observation is that f2 and f3 are actually the same functions coming from
different representations as a sum of monomials. For f3 , the linear term is
never maximal. For f2 , the linear term is maximal at 0, but only together
with the other two terms. In fact, the “shortest” representation of both
functions is given by f2 = f3 = “0+ x 2 ”. Apart from that, note that again the
functions are non-differentiable and strictly convex only at a finite number
35
of breaking points. Away from these points, only one monomial attains the
maximum and hence the function is locally affine linear.
Example 2.2.3
The expression
f (x, y) = “(−1)x 2 + (−1) y 2 + 1x y + x + y + 0”
defines a tropical polynomial of degree 2 in two variables. Its graph is

displayed in Figure 2.3.
Figure 2.3: The graph of max{(−1) + 2x, (−1) + 2 y, 1 + x + y, x, y, 0}
Classical algebraic geometry is the study of zero-sets of classical polyno-

mials. Consequently, the object of study in this chapter should be zero-sets
of tropical polynomials. However, note that the naive definition {x ∈ Rn :
f (x) = 0T = ∞} produces the empty set whenever f 6≡ −∞ (i.e., a j 6= ∞
for at least one j ∈ A) and therefore we obviously need some alternative
definition. We will give this definition now and afterwards provide some
first, intrinsically tropical evidence that our definition is indeed the “right”
one. The extrinsic, but probably more convincing motivation for our notion
of tropical hypersurfaces is the fact that they describe limits of amoebas of
hypersurfaces as exemplified to in the introduction. We will not pursue this
line of thought here.
Definition 2.2.4
Let f be a tropical Laurent polynomial in n variables. Then we define the
36
hypersurface V ( f ) ⊆ Rn to be the set of points in Rn where f is not differ-

entiable.
In classical arithmetics, the zero element of a group is distinguished by

the property of being idempotent, i.e. 0 + 0 = 0. In tropical arithmetics,
this is true for any number as “x + x” = max{x, x} = x. However, we
might consider this as a hint that a tropical sum should be called “zero” if
the maximum is attained by at least two terms. This leads to an alterna-
tive definition of the tropical hypersurface of a tropical polynomial. The
following proposition shows that both definitions coincide.
Proposition P 2.2.5 j
Let f = “ j∈A a j x ” be a representation of tropical polynomial. Then the
hypersurface V ( f ) is equal to the set of points x ∈ Rn where the maximum
f (x) is attained by at least two monomials, i.e.
V ( f ) = {x ∈ Rn : ∃i 6= j ∈ A such that f (x) = ai x i = a j x j }.
Proof. If the maximum is attained by only one monomial, then f is locally

affine and thus differentiable. If two monomials attain the maximum, then
f is strictly convex at x and therefore cannot be differentiable.
Example 2.2.6
For the three (in fact, two) polynomials from example 2.2.2, the hypersur-
faces are just finite sets of points (see Figure 2.4).
V ( f1 ) = {−1, 1} V ( f2 ) = V ( f3 ) = {0}
Figure 2.4: The “zeros” of tropical polynomials
The polynomial f from example 2.2.3 produces a connected hypersurface

V ( f ) consisting of 4 vertices, 3 bounded edges and 6 unbounded rays (see
Figure 2.5).
37
Figure 2.5: The hypersurface of “(−1)x 2 + (−1) y 2 + 1x y + x + y + 0”
Remark 2.2.7
As mentioned in the introduction, there is an alternative definition of the
tropical sum of two numbers given by multivalued addition. This was sug-
gested by Viro (cf. [Vir10]) and is given, for each two numbers a, b ∈ T =
R ∪ {−∞}, by ¨
max(a, b) if a 6= b,
a+b= (2.1)
[−∞, a] if a = b.
Note that when a = b, the sum a + b is not just a single element but the
set [−∞, a] := {−∞} ∪ {x ∈ R : x ≤ a}. The set T equipped with the
multivalued addition + and the uni-valued multiplication “ · ” = + forms a
structure which is called a hyperfield (in our case, the tropical hyperfield).
One reason to consider such generalizations of ordinary algebra is that they
allow to describe “zero-sets” in a natural way. Of course, when considering
a multi-valued sum, “being zero” should be replaced by “containing zero”.
Note that a sum a1 + · · · + an in the tropical hyperfield contains −∞ if and
only if we find i 6= j such that ai = a j ≥ ak , k = 1, . . . , n, i.e., if the maximum
of the summands occurs at least twice. Hence, if we replace a tropical
polynomial
f (x) = max(a j1 + j1 x, . . . , a jn + jn x)
by the hyperfield version
f + (x) = (a j1 + j1 x) + · · · + (a jn + jn x),
we find
V ( f ) = {x ∈ Rn : −∞ ∈ f + (x)} =: V ( f + ).
38
In summary, sacrificing uni-valued addition leads to a more natural defini-

tion of tropical zero-sets.
Exercise 2.2.8
Let f : R → R be a univariate tropical polynomial of degree d and with non-
vanishing constant term. Show that T is algebraically closed in the following
sense. There exist c ∈ R and α1 , . . . , αn ∈ R (unique up to reordering) such
that n
Y
f (x) = “c (x + αi )”
i=1
for all x ∈ R (as functions, not as polynomial representations). Show that

V ( f ) = {α1 , . . . , αn }.
Exercise 2.2.9
Let f , g : Rn → R be tropical Laurent polynomials. Show that V (“ f g”) =
V ( f ) ∪ V (g).
Exercise 2.2.10
Given a set H, we denote by Pot(H) the set of subsets of R. A hyperfield is a
set H together with binary operations + : H ×H → Pot(H) and · : H ×H → H
and two elements 0H , 1H ∈ H such that for all a, b, c ∈ H
a + (b + c) = (a + b) + c, 0H + a = {a},
∃! − a ∈ H s.t. 0H ∈ a + (−a), a + b = b + a,
a ∈ b + c ⇒ b ∈ a + −c, 0H · a = 0H ,
(H \ {0H }, ·) is a commutative group, a · (b + c) = (a + b) · (a + c).
Note that some of the axiom (e.g. associatitvity) require to consider the
canonical extension of + : H × H → Pot(H) to Pot(H) × Pot(H) → Pot(H).
(a) With + defined as in Equation 2.1, show that (T, +, +, −∞, 0) forms a
hyperfield.
(b) With ¾ t defined as in Equation 1.3, show that (T, ¾ t , +, −∞, 0) forms
a hyperfield for all t > 0.
(c) Show that lim t→∞ ¾ t = + (cf. Remark 1.4.1).
39
2.3 The polyhedral structure of

hypersurfaces
Of course, tropical hypersurfaces are not just sets but carry much more
structure. In this section, we describe their structure as polyhedral com-
plexes. For each monomial parameterized by j ∈ A, set
σ j := {x ∈ Rn : f (x) = “a j x j ”}
to be the locus of points where the chosen monomial is maximal. The sets
σ j subdivide Rn into the domains of linearity of f . If σ j is a neighbourhood
of a point x, then f is obviously differentiable at x, with differential d f x = j
(regarding j as a covector). Note that σ j is a rational polyhedron in Rn as it
is the intersection of the halfspaces a j + j x ≥ ai + i x for all i ∈ A. Moreover,
the intersection of two polyhedra σi and σ j is either empty or a common
face (given by intersecting with the plane ai + i x = a j + j x). It follows that
the collection of polyhedra
S ( f ) := {faces τ of σ j for some j ∈ A}
forms a polyhedral subdivision of Rn . While the definition of σ j may depend
on the representation of f (see Example 2.3.1), we prove in Lemma 2.3.4
that S ( f ) only depends on the function f .
Example 2.3.1
Our polynomials from the previous examples give subdivisions as indicated
in Figure 2.6. Note that in the second and third case the subdivisions S ( f2 )
and S ( f3 ) are identical, but the definition of σ1 depends on the chosen
representation. For the polynomial in two variables, the subdivision S ( f )
is shown in Figure 2.7.
σ1 = ;
σ0 σ1 σ2 σ0 σ1 σ2 σ0 σ2
Figure 2.6: The subdivision induced by polynomials
40
σ(0,2)
σ(1,1)
σ(0,1)
σ(0,0) σ(2,0)
σ(1,0)
Figure 2.7: Another subdivision
Obviously, a point x ∈ Rn lies in n − 1-cell of S ( f ) if and only if at

least two monomials attain the maximum f (x) at x. We get the following
corollary.
Corollary 2.3.2
The hypersurface V ( f ) is equal to the n − 1-skeleton of S ( f ). In particu-
lar, V ( f ) is canonically equipped with the structure of a rational polyhedral
complex of pure dimension n − 1.
It remains to show that S ( f ) does not depend on the representation of

f . On the way, we define the reduced representation of f given by throwing
away all monomials “a j x j ” for which σ j is empty or not full-dimensional.
Definition
P2.3.3
Let f = “ j∈A a j x j ” be a tropical Laurent polynomial. The set
Ared := { j ∈ A : dim(σ j ) = n} = {d f x : x ∈ Rn \ V ( f )}
is called the reduced support of f . The truncated polynomial

X
f red = “ aj x j”
j∈Ared
is called the reduced representation of f .
Lemma 2.3.4
Let f be a tropical Laurent polynomial. On the level of functions on Rn , we
have f ≡ f red . Moreover, two tropical Laurent polynomials describe the same
41
function if and only if their reduced representations agree (as abstract poly-
nomials). In particular, the subdivision S ( f ) is independent of the represen-
tation of f .
Proof. As S ( f ) forms a subdivision of Rn , every point x ∈ Rn is contained

in some σ j with dim(σ j ) = n, so
f (x) = “a j x j ” = f red (x).
Moreover, if two polynomials describe the same function, they have the
same reduced support (by the second description of Ared ). The value f (x)
for any x ∈ Rn with d f x = i determines the coefficient of x i uniquely. Thus
the second claim follows. Finally, since obviously S ( f ) = S ( f red ) (by the
first description of Ared ), it follows that S ( f ) is completely determined by
the underlying function.
We have seen that each tropical Laurent polynomial subdivides Rn into

its domains of linearity S ( f ). Next, we want to describe the the dual sub-
division to S ( f ) which is a certain subdivision of the Newton polytope of
f . The Newton polytope NP( f ) of f = “ j a j x ” is given by
P j
NP( f ) = Conv{ j ∈ Zn : a j 6= −∞},
i.e. the convex hull of all appearing exponents. Note that NP( f ) naturally
lives in the dual space of Rn .
Definition
P2.3.5
Let f = “ j∈A a j x j ” be a tropical Laurent polynomial. We set
e := {( j, −a j ) ∈ Zn × R : j ∈ A and a j 6= −∞}
A
and Pe = Conv(A e). The projection of the lower faces of Pe (i.e. those which
are also faces of Pe + ρ with half-ray ρ := {0} × R≥0 ) produces a subdivision
SD( f ) of NP( f ), which we call the dual subdivision of f .
Example 2.3.6
Our polynomials in one variable from example 2.2.2 all have the same New-
ton polytope Conv{0, 2}. Only for the first polynomial, SD( f1 ) is non-trivial,
42
e and P
A e
SD( f i )
Figure 2.8: The dual subdivision
i.e. the segment is divided into two unit segments with vertex {1} (see Fig-
ure 2.8).
For the polynomial f of example 2.2.3, the dual subdivision is again more
interesting. In Figure 2.9, only the lower faces of P
e are drawn solidly. We
Figure 2.9: The dual subdivision of “(−1)x 2 + (−1) y 2 + 1x y + x + y + 0”
obtain a subdivision of the triangle of size 2 into 4 triangles of size 1. More-

over, the subdivision consists of 3 internal edges and 6 edges in the bound-
ary of the big triangle.
In our examples, we can easily observe an inclusion-reversing duality
between the cells of S ( f ) and SD( f ). Let us formulate this systematically.
For each cell σ ∈ S ( f ) we define
Aσ := { j ∈ A : σ ⊆ σ j }.
In other words, Aσ is the set of monomials which are maximal on σ. We

find \
σ= σj. (2.2)
j∈Aσ
43
Analogously, for a point x ∈ Rn we set A x := { j ∈ A : x ∈ σ j }. We have

x ∈ σ if and only if A x = Aσ .
Theorem 2.3.7
The subdivisions S ( f ) of Rn and SD( f ) of NP( f ) are dual in the following
sense. There is an inclusion-reversing bijection of cells given by
S ( f ) → SD( f )
(2.3)
σ 7→ Dσ := Conv(Aσ )
such that dim σ + dim Dσ = n and L(σ)⊥ = L(Dσ ). Moreover, SD( f ) is inde-
pendent of the representation of f and the set of vertices of SD( f ) is equal to
the reduced support of f .
Proof. First let us show that Dσ is indeed a cell of SD( f ). To do so, we pick a
point x in the relative interior of σ. Let us consider the linear form (x, −1)
on Pe ⊂ Rn+1 . For each vertex ( j, −a j ) of P
e we find (x, −1)( j, −a j ) = a j + j x.
Thus the face of P e on which (x, −1) is maximal is exactly the convex hull
of the points Aσ with
e
eσ := {( j, −a j ) : j ∈ Aσ }.
A
The projection of this face to Rn is Dσ , and therefore Dσ is indeed a face of
SD( f ). Next we see that all lower faces of P e are obtained in this way for
a suitable (x, −1). It follows that equation (2.3) indeed describes a well-
defined bijection whose inverse map is given by
\
D 7→ σj.
j∈D∩A
The previous arguments also provide another way to construct S ( f ). Let

N (P e ) be the normal fan of Pe , then S ( f ) is the subdivision obtained from
intersecting N ( Pe ) with the plane Rn × {−1} ∼ = Rn (see Figure 2.10). Using
this description the orthogonality and dimension statements follow directly
from the corresponding statements for the dual cells of a polyhedron and
its normal fan.
Finally, the equality (x, −1)( j, −a j ) = a j + j x also implies that the ver-
tices of Pe detected by linear forms of type (x, −1) (which correspond to the
vertices of SD( f )) are in bijection to Ared . Hence the lower faces of P e and
hence SD( f ) only depend on the function f and the last claims follow.
44
N (P
e)
P
e
R × {−1}
Figure 2.10: Dual subdivisions via normal fans
Example 2.3.8
Let A ∈ Zn be a finite set and let f = “ j∈A x j ” be the Laurent polynomial
P
with only trivial coefficients a j = 0, j ∈ A. Then SD( f ) is just the trivial

subdivision of NP( f ) = Conv(A) (i.e. SD( f ) contains NP( f ) and all its faces)
and S ( f ) is just the normal fan of NP( f ).
Remark 2.3.9
Let S be a subdivision of a polyhedron P ⊆ Rn . S is called a regular sub-
division if it can be obtained by projecting down the lower faces of some
polyhedron P e ⊆ Rn+1 . Equivalently, S is regular if there exists a convex
function on P which is affine on each cell of S (to get the polyhedron in
Rn+1 , we take the convex hull of the graph of the function; the other way
around, the union of lower faces of the polyhedron describes the graph of
a suitable function). Note that both S ( f ) and SD( f ) are regular subdi-
visions. Such subdivisions are sometimes also called convex or coherent.
They play a prominent role in the context of the Viro’s patchworking tech-
nique (see [Vir06] for references) as well as the study of discriminants and
resultants (see [GKZ08]).
Note that not all subdivisions are regular. An example is given in Figure
2.11. Assume there is convex function g inducing this subdivision. We can
assume that g is constant zero on the inner square. Then if we fix one
g(v1 ) on an outer vertex, on the next vertex v2 in clockwise direction, we
need g(v2 ) ≥ g(v1 ), due to the diagonal edge subdividing the corresponding
trapezoid. Going around the square completely gives a contradiction.
Another way to describe the duality of S ( f ) and SD( f ) can be formu-
lated in terms of Legendre transforms. Let g : S → R be a function on an
arbitrary set S ⊆ Rn . The Legendre transform g ? is a function on covectors
45
of Rn given by
g ? (w) = sup{wx − g(x)}.
x∈S
?
It is easy to check that g takes finite values on a convex set of Rn (possibly
empty) and that g ? is convex. Moreover, if g is a convex function in the
beginning,
P than (g ? )? = g. By definition, a tropical Laurent polynomial f =
“ j∈A a j x j ” is equal to the Legendre transform of its inverted coefficient
map A → R, j 7→ −a j .
Lemma 2.3.10
Let f be a Laurent polynomial and let g be the convex function on NP( f )
whose graph is the lower hull of
e = Conv({( j, −a j ) : j ∈ A}).
P
Then f ? = g and equivalently f = g ? .
Proof. We set g 0 : A → R, j 7→ −a j . As f = g 0? , it suffices to show g 0? = g ? .

For a fixed x ∈ Rn , we can compute g 0? (x) resp. g ? (x) as the maximum
value of (x, −1) on {( j, −a j ) : j ∈ A} resp. P e . This maximum is always
attained on at least one vertex of P e and the vertices of P e are contained in
{( j, −a j ) : j ∈ A}. Hence the claim follows.
Exercise 2.3.11
Let f be a tropical Laurent polynomial. Show that the Newton polytope of
f can be described as
NP( f ) = {w ∈ Rn : wx − f (x) is bounded from above}.
Figure 2.11: A non-regular subdivision
46
Exercise 2.3.12
Let f , g : Rn → R be tropical Laurent polynomials. Show that NP(“ f g”) =
NP( f ) + NP(g).
Exercise 2.3.13
Let f be a tropical Laurent polynomial. In this exercise, we want to prove
N (NP( f )) = {rc(σ) : σ ∈ S ( f )}. Recall that rc(σ) denotes the recession
cone of σ.
(a) Let σ ⊂ Rn × R be a (closed) cone. For k = 0, 1 set σk := σ ∩ (Rn ∩

{−k}) ⊂ Rn . Show that rc(σ1 ) = σ0 .
(b) For D ∈ SD( f ), let F ⊂ NP( f ) be the unique face of NP( f ) such that
D ⊂ F . Let σ D and σ F be the dual cell/cone of D and F in S ( f )
and N (NP( f )), respectively. Show that rc(σ D ) = rc(σ F ).
(c) Conclude that N (NP( f )) = {rc(σ) : σ ∈ S ( f )}.
2.4 The balancing condition

So far, we have seen that to each tropical Laurent polynomial f we can
associate a subdivision S ( f ) of Rn and a dual subdivision SD( f ) of NP( f )
which is induced by the Legendre transform of f . We described the hy-
persurface V ( f ) as a set — the points where f is not differentiable — and
as polyhedral complex of pure dimension n − 1 — the n − 1-skeleton of
S ( f ). We now add yet another layer to our description, namely weights of
points. We will later say that smooth points of V ( f ) are those with weight 1.
Moreover, these weights are necessary to formulate the most fundamental
structure property of tropical objects, the balancing condition.
For each polytope P, we define the volume Vol(P) to be the volume of
σ measured in the affine space spanned by P and normalized such that
Vol(Sn ) = 1, where Sn denotes the standard simplex in Rn given by x 1 +
. . . x n ≤ 1 and x i ≥ 0, i = 1, . . . , n. In other words, the volume of a full-
dimensional P is the usual volume of P in Rn divided by n!. Let P be a
lattice polytope, i.e. all vertices are integer points. We call P a minimal
simplex if Vol(P) = 1. This is the case if and only if σ can be mapped to
the standard simplex Sn by an integer affine transformation. For any lattice
47
polytope P we have Vol(P) ∈ N. This follows from the fact that we can
always triangulate P into lattice simplices (not necessarily minimal ones).
The volume of a simplex with vertices v0 , . . . , vn is given by the determi-
nant of the vectors v1 − v0 , . . . , vn − v0 , which is obviously integer for lattice
simplices. For one-dimensional lattice polytopes E (i.e., edges) we find
Vol(E) = #{E ∩ Zn } − 1. For two-dimensional lattice polytopes P, Pick’s
theorem states Vol(P) = 2i + b − 1, where i := #{σ ∩ Zn } is the number of
interior lattice points and b := #{∂ σ ∩ Zn } is the number of lattice points
in the boundary.
Definition 2.4.1
Let f be a tropical Laurent polynomial. We turn its hypersurface V ( f ) into a
weighted polyhedral complex as follows. The weight of each cell σ of V ( f )
is the volume of the corresponding cell Dσ in the dual subdivision SD( f ),
ω(σ) := Vol(Dσ ) ∈ N.
For each point x ∈ V ( f ) in the relative interior of σ we define the weight
ω(x) ∈ N (or multiplicity of x) to be ω(σ).
A point resp. cell of V ( f ) is called smooth or singular point/cell of V ( f ) if
its weight is equal to 1 or greater than 1, respectively. If V ( f ) does not have
any singular points, it is called a smooth hypersurface. This is equivalent to
the condition that all cells of SD( f ) are minimal simplices, in which case
we call SD( f ) a unimodular subdivision. Of course, it suffices to check this
condition for the maximal cells of SD( f ), as faces of minimal simplices are
minimal simplices again.
Remark 2.4.2
Let σ be a facet of V ( f ). The dual cell Dσ is an edge whose endpoints,
say i and j, are the exponents of the two monomials in the reduced rep-
resentation of f which attain the maximum at F . Thus the weight ω(F )
can be considered as a measure of the change of slope when crossing from
the linearity domain Pi to P j through F . This is precisely what happens for
polynomials in one variable. In general, one should think of the change of
slope relative to F .
Example 2.4.3
Let us revisit our running examples. The polynomials from example 2.2.2
48
SD( f1 ) V ( f1 )
0 1 2 mult 1 1
SD( f2 ) = SD( f3 ) V ( f2 ) = V ( f3 )
0 2 mult 2
Figure 2.12: Multiplicities of tropical zeros
Figure 2.13: V (“(−1)x 2 + (−1) y 2 + 1x y + x + y + 0”) is smooth.
describe hypersurfaces of either two points with multiplicity 1 or one single

point with multiplicity 2. Correspondingly, the dual subdivision divides the
interval [0, 2] in either two segments of volume 1 or one segment of volume
2 (see Figure 2.12). In the case of the planar conic from example 2.2.3, the
dual subdivision only consists of four minimal triangles of volume 1 and
therefore the conic is smooth (see Figure 2.13).
Remark 2.4.4
For a smooth hypersurface V ( f ), the reduced support of f is NP( f ) ∩ Zn .
This follows from the fact that a minimal simplex does not contain integer
points other than its vertices and thus each point in NP( f ) ∩ Zn must be a
49
Figure 2.14: A tropical 2-dimensional hyperplane
vertex of SD( f ).
Example 2.4.5
The tropical hyperplane in Rn is the hypersurface V ( f ) with f = “x 1 + · · · +
x n + 0”. It is obviously smooth, as NP( f ) is the standard simplex itself.
Figure 2.14 depicts the 2-dimensional hyperplane in R3 .
A hypersurface V ( f ), equipped with multiplicities as above, satisfies the
so-called balancing condition — a fundamental property of tropical vari-
eties.
Definition 2.4.6
Let X be a pure-dimensional polyhedral complex in Rn . A weight function
on X is given by a weight ω(σ) ∈ Z for each facet σ of X . In this case, we
say X is a weighted polyhedral complex. We define its support |X | as the
union of facets of non-zero weight.
Let X be a positively weighted polyhedral complex in Rn . We say X
is balanced or satisfies the balancing condition if for every cell τ ∈ X of
codimension one the following equation holds:
X
ω(σ)vσ/τ ∈ L(τ)
σ facet
τ⊂σ
Here vσ/τ ∈ Zn denotes a primitive generator of σ modulo τ, i.e. an integer

vector that points from τ to the direction of σ and satisfies
LZ (σ) = LZ (τ) + Zvσ/τ .
50
Figure 2.15: The balancing condition
The balancing condition requires that the weighted sum of primitive vec-
tors around a codimension one cell τ is parallel to τ. An equivalent formu-
lation is as follows. Consider the ray σ/τ ∈ StarX (τ) let uσ/τ ∈ Zn / LZ (τ)
denote its unique primitive generator. By definition, vσ/τ + L(τ) = uσ/τ and
hence the balancing condition can be equivalently expressed as
X
ω(σ)uσ/τ = 0 ∈ Rn / L(τ).
σ facet
τ⊂σ
Allowing non-zero weights will be convenient later, but for all practical
purposes (cf. definition of support) we may always throw away all facets of
weight 0 and hence assume that all weights are non-zero. Obviously, the
balancing condition is not affected by this operation.
Theorem 2.4.7
Any hypersurface V ( f ) of a Laurent polynomial f = “ j∈A a j x j ” forms a
P
balanced polyhedral complex.
Proof. Let us first consider the two-dimensional case, i.e. assume f is a

polynomial in two variables and hence V ( f ) is a piecewise linear graph in
the plane. We have to check the balancing condition around each vertex
ν of V ( f ). Let Dν be the 2-cell in the dual subdivision. By duality we see
that locally around ν the subdivision S ( f ) looks like the normal fan of Dν .
In particular, for each edge σ containing ν the primitive generator vσ/ν is
51
orthogonal to the dual edge Dσ in Dν , and ω(σ) is by definition just the

integer length of Dσ . Thus, when we concatenate all the vectors ω(σ)vσ/ν
σV
2
V
2
2 2
Figure 2.16: The balancing condition in two dimensions
in say clockwise order, the prescribed chain of edges is equal to a right angle
rotation of the boundary of Dν . The fact that this boundary closes up is thus
equivalent to the fact X
ω(σ)vσ/ν = 0.
σ edge
ν⊂σ
This finishes the proof for R2 , and essentially the same argument can be
applied in the general case. Let τ be a codimension one face of V ( f ) and
let Dτ be the dual 2-cell of SD( f ). Then L(Dτ ) is canonically the dual space
of Rn / L(τ) and for each facet σ containing τ, the primitive generator uσ/τ ∈
Rn / L(τ) is orthogonal to the corresponding edge Dσ ∈ SD( f ). Thus, once
again the “closedness” of Dτ implies the balancing condition.
Remark 2.4.8
Note that the balancing condition only involves the multiplicity of the facets
of V ( f ) or, in other words, of generic points of V ( f ). Since we defined mul-
tiplicities for all cells of S ( f ), one might ask whether S ( f )(k) is balanced
for all k. Indeed, we will prove this in Theorem 4.7.6.
Remark 2.4.9
Let us one more time revisit our question of defining the “zero set” of a
tropical polynomial. Now that we know that the balancing condition is a
fundamental property of tropical hypersurfaces, we might be bothered by
52
the fact that the graph of a tropical Laurent polynomial
Γ = {(x, y) ∈ Rn × R : y = f (x)}
does not satisfy the balancing condition (of course, Γ carries a natural poly-
hedral structure by lifting S ( f ), and we use multiplicity 1 for all facets
here). Note that in classical algebraic geometry, the graph of a polynomial
Figure 2.17: The graph of max{(−1) + 2x, (−1) + 2 y, 1 + x + y, x, y, 0}
is a hypersurface again given by an algebraic equation, namely y = f (x).

Tropically, however, the graph of a polynomial is obviously a polyhedral
complex, but the balancing condition fails exactly at the points over V ( f )
where the function is strictly convex. We can repair this by using the mul-
tivalued tropical addition + from remark 2.2.7 again. Let us consider the
multivalued function f + and its “graph”
Γ + = {(x, y) ∈ Rn × T : y ∈ f + (x)}.
We find that Γ + contains the ordinary graph Γ , but additionally, over each
point x ∈ V ( f ), contains the half-infinite interval {(x, y) : y ≤ f (x)}.
Moreover, Γ + is canonically balanced. To that end, we equip each facet of
Γ + which is not already contained in Γ with the multiplicity of its projection
to Rn which is a facet of V ( f ). It is easy to see that Γ + is the only balanced
completion of Γ if we only allow to add facets in direction (0, −1) ∈ Rn × R.
This is why Γ + is sometimes called the completed graph of f . Note that, in
contrast to Γ , the completed version Γ + intersects the “zero-section” Rn ×
{−∞} and this intersection gives back V ( f ) — yet another reason for our
53
Figure 2.18: The graph of ((−1) + 2x) + ((−1) + 2 y) + (1 + x + y) + x + y + 0
definition of the zeros of a tropical polynomial. We will meet the completed

graph again when introducing the concept of tropical modifcations later.
We close this section by proving an inverse statement to theorem 2.4.7.

This emphasizes the importance of the balancing condition by showing it
is not only a necessary but also sufficient condition for characterizing trop-
ical hypersurfaces (at least if we extend the notion to quotients of tropical
polynomials).
Theorem 2.4.10
Let X be a balanced polyhedral complex in Rn which is of dimension n − 1
and all of whose weights are positive. Then there exist tropical polynomials f
such that X is a weighted refinement of V ( f ). Moreover, f is unique up to
adding an integer affine function.
Here, X is a weighted refinement of V ( f ) if |X | = V ( f ), for each σ ∈

X there exists σ0 ∈ S ( f )(n−1) with σ ⊂ σ0 , and if σ a facet, ωX (σ) =
ωV ( f ) (σ0 ).
Proof. Let C1 , . . . , Cl be the connected components of Rn \ |X |. P Our plan

is to find a monomial f i for each component Ci such that f = “ f i ” and
f | Ci = f i .
Let us consider the graph Γ with vertices v1 , . . . , vl corresponding to the
connected components and edges corresponding to facets of X . An edge
connects vi and v j if the associated cell separates Ci from C j . Note that Γ
54
is connected since X (n−2) is of codimension 2 and hence Rn \ |X (n−2) | is

connected.
We pick one vertex, say v1 , and set f1 ≡ 0. Assume that v2 is connected
to v1 by an edge and let σ ∈ X be the corresponding facet. Then we define
f2 = f1 + ω(σ)hσ ,
where hσ is the integer affine function such that hσ |σ = 0 and dhσ (vC2 /σ ) =
1 (where vC2 /σ is a primitive generator of C2 modulo σ). Using this proce-
dure recursively, we may define integer affine functions f i for all i = 1, . . . , l.
We need to check that the definition of f i does not depend on the chosen
path from v1 to vi . Equivalently, we show that for any loop in Γ adding
up the various functions ω(σ)hσ produces 0. Let τ ∈ X be a n − 2-cell.
Going around τ in a small loop gives rise to a simple loop lτ in Γ . Note that
Rn \ |X (n−3) | is simply connected since X (n−3) is of codimension 3 in Rn . It
follows that it suffices to check the addition to zero for the loops lτ .
Let σ1 , . . . , σh be the facets adjacent to τ, ordered according to lτ , and
let hσ j be the integer affine functions defined above. Then the condition
X
ω(σ j )hσ j = 0
is just the dual version (and thus equivalent to) the balancing condition of
X at R. We have seen this in more details in the proof of theorem 2.4.7.
So far we proved that the above procedure produces a well-defined mono-
mial f i for each connected component Ci . Let f : Rn → R be the unique
continuous function with f |Ci := f i for all i. It is clear from the construc-
tion of the f i that f is well-defined and convex on each line segment in
Rn \ X (n−2) . By continuity, f is convex on all of Rn and therefore equal to
Pl
twice its Legendre dual “ i=1 f i ”. Hence f is a tropical polynomial. It is
now easy to check that X is a weighted refinement of V ( f ).
To verify uniqueness, let g be another tropical polynomial satisfying the
hypothesis. Then g is also integer affine on each component of Rn \ |X | =
Rn \ V ( f ). Moreover, since the weights of X and V (g) agree, g needs to
satisfy the same transition formula than f when traversing a facet σ ∈ X .
It follows that g = f + g1 , where g1 is the integer affine functions which
describes g on C1 .
55
Exercise 2.4.11
Let f1 , f2 , f3 : R2 → R be the tropical polynomials given by
f1 = “(−1) + x + 1 y + x 2 + x y + (−1) y 2 ”,
f2 = “(−2) + (−2)x 3 + (−2) y 3 + x + y + x 2 + y 2 + x 2 y + x y 2 + 1x y”,
f3 = “0 + (−1)x + (−3)x 2 + y + 1x y + x 2 y + (−2) y 2 + x y 2 + x 2 y 2 ”.
In order to compute the associated tropical curves V ( f i ), proceed as follows.
(a) Compute the Newton polytopes NP( f i ) and the dual subdivisions SD( f i ).
(b) Compute (some of) the vertices of V ( f i ). (Each triangle in SD( f i )

singles out three terms of f i . The corresponding vertex is the point
where these three terms attain the maximum simultaneously.)
(c) Draw the curves V ( f i ) ⊂ R2 by adding the edges.
Exercise 2.4.12
Let f : R → R be a univariate polynomial and α ∈ R. Let ordα ( f ) be the
maximal number k such that (x − α)k | f (cf. Exercise 2.2.8). Show that
ωV ( f ) (α) = ordα ( f ).
Exercise 2.4.13
Let X be a weighted polyhedral complex of pure dimension k. For any
cell τ the fan StarX (τ) carries induced weights given by ω(σ/τ) = ω(σ).
Show that the following statements are equivalent.
(a) The polyhedral complex X is balanced.
(b) The stars StarX (τ) are balanced for any τ ∈ X .
(c) The stars StarX (τ) are balanced for any k − 1-cell τ.
2.5 Planar Curves

This section is devoted to the study planar curves as a particular example of
tropical hypersurfaces. While they are easily visualizable, they still exhibit
56
f (x, y) = y
L
f (x, y) = x
f (x, y) = 0
Figure 2.19: A tropical planar line
interesting combinatorial, but also geometric properties with have counter-

parts in higher dimensions.
A tropical planar curve is the hypersurface C = V ( f ) of a tropical Laurent
polynomial X
f =“ a j x j1 y j2 ”
j∈Z2
in two variables x, y. Let ∆d , d ∈ N denote the convex hull of the points

(0, 0), (d, 0), (0, d). We call ∆d the d-fold standard simplex. Mostly, we will
consider polynomials whose Newton polytope is NP( f ) = ∆d for a suitable
d. In this case we call V ( f ) a curve of degree d. We may extend this def-
inition to arbitrary polynomials by saying V ( f ) is of degree d if d is the
smallest natural number such that (a shift of) ∆d contains NP( f ). If NP( f )
is not equal to this minimal (shift of) ∆d , we call V ( f ) degenerated.
Of course, we start with planar lines L = V ( f ), where f is a polynomial
of degree 1. The general form of such a polynomial is
f (x, y) = “ax + b y + c”.
Let us start with a = b = c = 0 (see Figure 2.19). Then L = V ( f ) con-

sists of three rays R≥ (1, 1), R≥ (−1, 0) and R≥ (0, −1). The point (0, 0) is
the single vertex of L. Each ray corresponds to two of the three terms of f
being maximal, while (0, 0) is the single point where all three terms attain
the maximum. What happens if we change a, b, c? There is still a unique
point where all three monomials are maximal, (c − a, c − b). Therefore
57
L a,b,c = V (“ax + b y + c”) is just the translation of L to this point. If we de-

crease one coefficient at a time, we move L in one of the directions (−1, −1),
(1, 0) or (0, 1). In the limit, i.e. when one coefficient becomes −∞, we end
up with an ordinary classical line whose Newton polytope is just a seg-
ment (see Figure 2.20). In total, there are a single non-degenerated and
3 degenerated types of tropical lines, and two lines of the same type are
translations of each other. Let us stress that the only classical lines which
show up as “tropical lines” are the lines of slope (1, 1), (−1, 0) or (0, −1).
Let us now decrease a second coefficient of our linear polynomial, such that
in the limit two coefficients are −∞. Geometrically, the line vanishes at in-
finity and V ( f ) is empty. This reminds us of the fact that we work with the
tropical algebraic torus Rn = (T× )n here, a non-compact space on which
single monomials are “non-vanishing” functions. Later on, we will consider
various compactifications of Rn , in particular tropical projective space TP2 .
In this space, when decreasing two of the coefficients, the corresponding
moving line will attain one of the coordinate lines x = −∞, y = −∞ or
z = −∞ as limit. 1
For now, let us stick to the non-compact picture. There are two elemen-
tary properties of classical planar lines which we want to study tropically
now: Two generic lines intersect in a single point, and, dually, through two
different points in the plane there passes a unique line.
First, let us consider the intersection of two tropical lines. Indeed, for
most pairs of lines, we get exactly one point of intersection, but there are
two notable exceptions, as illustrated in the Figure 2.22. While in the first
1
In particular, we see that the dual space of lines in TP2 is itself a copy of TP2 and inherits
a stratification into affine spaces corresponding to the seven types of lines.
Figure 2.20: Moving a tropical line
58
Figure 2.21: The types of tropical lines
Figure 2.22: Intersections of tropical lines
two pictures we get a perfectly nice single point of intersection, the third
picture shows two parallel degenerated lines with no intersection. As in
classical geometry, we will get rid of this special case by compactifying R2 .
However, a classical geometer would very likely compactify to the real pro-
jective plane RP2 to make the lines intersect at infinity. This is different
from the tropical compactifications we will study later, e.g. TP2 . The fourth
picture shows a more interesting way of failure. Here, the two tropical
lines have a whole ray in common. This type of abnormality does not have
a classical counterpart (except for taking the same line twice) and will later
encourage us to introduce the concept of stable intersection. The stable in-
tersection of two curves consists only of those points in the set-theoretic in-
tersection which are stable under small deformations — small translations,
in our case. E.g. for lines, a small translation of one of them will yield a
unique intersection point close to the apex of the common ray. When mov-
ing the translation back to the original line, the limit of intersection points
59
is just this single point, which we therefore call the stable intersection of
the two lines. The dual problem of finding a line through two points shows
a completely similar behaviour.
Proposition 2.5.1
Any pair of points p1 , p2 ∈ R2 can be joined by a tropical line. Furthermore,
this line is unique if and only if the points do not lie on an ordinary classical
line of slope (1, 1), (−1, 0) or (0, −1).
Proof. Even though the statement is completely elementary, let us give a

short proof here. First, we set L to be the tropical line whose single vertex is
p1 . If p2 lies on one of the rays of L, we are already done (with existence). If
not, p2 is contained in one of the sectors of R2 \ L corresponding to a certain
monomial being maximal. We move L into this sector by decreasing the
coefficient of the corresponding monomial. While doing so, the two rays
of L that bound the sector will sweep it out completely. Hence we just
stop when the moving line meets p2 (see Figure 2.24). The uniqueness
statement follows from our previous discussion of the intersection of two
lines. Namely, if we can find two different lines containing p1 and p2 , then
the two lines have a ray of slope (1, 1), (−1, 0) or (0, −1) in common, and
the statement follows.
We now consider curves of degree 2 given by polynomials of the form
f (x, y) = “ax 2 + bx y + c y 2 + d x + e y + f ”.
stable
intersection
point
Figure 2.23: The stable intersection point of two lines
60
p1
p2
Figure 2.24: Finding a line passing through two points
Contrary to the case of lines, where we found only one non-degenerated

example up to translations, the variety of conics is already more interest-
ing: There exist several combinatorial types of non-degenerated conics and
two curves of the same type are not necessarily just translation of each
other. Figure 2.25 gives a list of the four “smooth” combinatorial types
of smooth non-degenerated conics, given by the four unimodular subdi-
visions of ∆2 . In all four cases the complement R2 \ V ( f ) consists of six
Figure 2.25: Smooth (non-degenerated) conics
connected components — the linearity domains of the six monomials in

f . Let us again deform a given conic by changing the coefficient of just
one monomial. This will enlarge or shrink the corresponding connected
component, depending on whether we increase or decrease the coefficient.
More precisely, the deformation will move all edges adjacent to this com-
ponent while all other edges rest in the sense that they stay in the same
affine line (they might change length, however). This follows from the fact
that the position of each of these edges is given by an equation of the form
61
ai + i1 x + i2 y = a j + j1 x + j2 y with i, j ∈ ∆2 ∩ Z2 , and this affine line moves

if and only if ai or a j is the coefficient being changed (see Figure 2.26).
This describes small changes of a single coefficient. However, if we keep
Figure 2.26: Increasing the x-coefficient
deforming, a couple of things can eventually happen. Of particular interest

is the case when the combinatorial type of the conic changes. This happens
if one of the edges shrinks to length zero and vanishes. In this case also the
corresponding edge in the dual subdivision vanishes which merges together
two of the triangles to a 2-cell of volume 2. Such a 2-cell can either be a
parallelogram or a triangle with one side of length 2. Both situations can
be illustrated by deforming the conic from the previous picture. Increasing,
as before, the coefficient of the monomial “x”, we run into a conic with a
singular point whose dual cell is a square. Increasing the coefficient further,
we get a smooth conic again, but the combinatorial type has changed (see
Figure 2.27). If we instead increase the monomial “x 2 ”, we run into a sin-
gular conic with a ray of multiplicity 2 (see Figure 2.28). Note that in this
case, increasing the coefficient further does not remove the weight 2 edge.
The parallelogram degeneration corresponds to a reducible conic which de-
composes into two lines. The second degeneration may be interpreted as a
conic that is tangent to the coordinate axis y = −∞ at infinity.
For higher degree curves, the number of combinatorial types becomes
large very quickly. Already, for curves of degree 3, there are 79 smooth non-
degenerate combinatorial types. Figure 2.29 depicts particular examples
of a smooth and a singular cubic. In general, a planar curve is smooth
if and only if each of its vertices has exactly 3 adjacent edges (we say all
62
Figure 2.27: Passing through a singular conic
vertices are 3-valent) and each pair of the primitive integer vectors v1 , v2 , v3
spanning the 3 edges of a vertex form a Z-basis of Z2 . We now present a
few special cases of combinatorial types for arbitrary degrees.
The honeycomb triangulation of ∆d is defined by the property that all
its edges are parallel to one of the three boundeary edges of ∆d . In other
words, it is obtained from the collection of lines x = i, y = j, x + y = k
with i, j, k ∈ {0, . . . , d}. Our figure depicts the case d = 5. The curves of
this combinatorial type are called honeycombs. They proved to be useful in
Figure 2.28: A conic with a “fat” edge
63
Figure 2.29: A smooth and a singular cubic
the context of the Horn problem (e.g. [KT01]).

The bathroom tiling curves form a similar type of curves. Their dual tri-
angulation of ∆d is given by x = i, y = j with i, j ∈ {0, . . . , d}, x + y =
k, x − y = l with k, l ≡ d mod 2. Again, our picture shows an example of
degree 5. These curves are of interest for example when constructing real
Figure 2.30: A honeycomb curve of degree 5
64
Figure 2.31: A bathroom tiling curve of degree 5
algebraic curves.
Let us also recall an example of a special tropical curve which we men-
tioned in the introduction. Figure 2.32 shows the curve of degree 10 to-
gether with its subdivision. It is this curve (together with the extra data
of suitable signs) that was used by Itenberg (cf. [IV96]) to disprove the fa-
mous Ragsdale conjecture (cf. [Rag06]; the conjecture was an inequality
involving the numbers of “odd” and “even” ovals of a real planar curve of
given degree).
A fundamental theorem in the study of classical planar curves is Bézout’s
Figure 2.32: The Itenberg-Ragsdale curve of degree 10
65
Figure 2.33: The 6 intersection points of a conic and a cubic
theorem which states that two projective curves of degree d and e have
d e points of intersections (in various meanings and under various assump-
tions). Previously we discussed the case of lines which are supposed to
intersect in a single point. It is easy to convince oneself that a more gen-
eral statement should also be true tropically. For example, a smooth non-
degenerate curve of degree d has exactly d unbounded rays in each of the
directions (1, 1), (−1, 0) and (0, −1) corresponding two the d dual segments
in the boundary of ∆d . Thus two such tropical curves of degree d and e can
be translated in such a way that they only intersect in such rays of fixed
direction (for each curve) — in this case there are exactly de intersection
points (see Figure 2.33). We can actually prove a general version of Bézout’s
theorem based on our understanding of dual subdivisions. For this purpose,
let us assume that C and D are non-degenerate tropical P planarj curves of
degree d resp. e and given by the polynomials f = “ j∈∆d a j x y ” resp.
1 j2
g = “ j∈∆e b j x j1 y j2 ”. The non-degeneracy condition ensures that no in-

P
tersections “at infinity” occur, so it is enough to work in Rn here. Let us

furthermore assume that C and D intersect transversally. This means that
that C ∩ D is finite and each intersection point lies in the relative interior
of an edge of C and an edge of D. We count each such point with an inter-
section multiplicity defined as follows. Assume the intersecting edges are σ
and τ, then we define
mult(p) := ω(σ)ω(τ)[Z2 : LZ (σ) + LZ (τ)]

= ω(σ)ω(τ)| det(v, w)|,
66
where v and w are primitive integer vectors describing the slope of σ resp.
τ. This multiplicity is one if and only if both weights are one and v, w form
a lattice basis of Z2 .
v w v w
mult(p) = 1 mult(p) = 2
Figure 2.34: Intersection points of multiplicity 1 and 2
Theorem 2.5.2
Let C = V ( f ) and D = V (g) be two non-degenerate tropical curves of degree
d and e, respectively, which intersect transversally. The number of intersection
points, counted with multiplicities, is equal to de.
X
de = mult(p)
p∈C∩D
Proof. The main trick is to consider the union of the two curves B := C ∪D =
V (“ f g”). Note that each intersection point p ∈ C ∩ D is a vertex of C ∪ D
and thus has a corresponding dual cell σ in SD(“ f g”). This dual cell must
be a parallelogram whose pairs of parallel edges correspond to the edges
of C resp. D that intersect. Moreover, we see that the multiplicity of p can
be computed in terms of the volume of σ, namely
Vol(σ)
mult(p) = .
2
If a vertex of B is not an intersection point, it is just a vertex of C or D.
Note that the corresponding dual triangle in SD(“ f g”) is just a shift of the
corresponding triangle in SD( f ) resp. SD(g) (see Figure 2.35). Indeed,
assume that v is a vertex of C and therefore by assumption v ∈ / D. Then
locally around v the function g is affine-linear, say b j + j x, and therefore
“ f g” is locally equal to f + b j + j x = “ f b j x j ”, which corresponds to a
shift of the dual picture by j. Hence the maximal cells of SD(“ f g”) are
67
either triangles, which are in (volume-preserving) bijection to the triangles

of SD( f ) and SD(g), or parallelograms, which are in bijection to C ∩ D. This
implies
Vol(parallelograms of SD(“ f g”))

= Vol(∆d+e ) − Vol(∆d ) − Vol(∆e )
= (d + e)2 − d 2 − e2 = 2de.
Together with the above formula, this proves the result.
Our proof is adapted from a paper of Vigeland (cf. [Vig09]).
Exercise 2.5.3
In this exercise we give an alternative proof of tropical Bézout’s theorem
(cf. [RST05]). Let C = V ( f ) and D = V (g) be two non-degenerate tropical
curves of degree d and e, respectively. We denote deg(C·D) = p∈C∩D mult(p)
P
whenever C and D intersect transversally.
(a) Show that the set of vectors v ∈ R2 such that C and D + v do not have
a vertex in common is connected.
(b) Assume that C is a fan and Y is an (usual) affine line. Show that
deg(C · (D + v)) is constant for all v such that the intersection is
transversal.
(c) Using the previous two items, show that deg(X · (Y + v)) = const. also
holds in the general case (for all “generic” v).
Figure 2.35: Bézout’s theorem in the case of two tropical conics
68
(d) Find a particular v such that C and D + v intersect transversally in de

points, all of multiplicity 1 (Hint in Figure 2.33).
Exercise 2.5.4
Let f : R2 → R be a tropical polynomial and C = V ( f ) the associated
planar tropical curve. The (embedded) genus g(C) of C is the first Betti
number b1 (C) = dim H1 (C, Z) (i.e. the number of independent 1-cycles in
C). Show the following statements.
(a) The genus g(C) is equal to the number of vertices of SD( f ) contained
in NP( f ) .
(b) Assume that C is a non-degnerated smooth curve of degree d. Then
(d − 1)(d − 2)
g(C) = .
2
Exercise 2.5.5
Let C ⊂ R2 be a planar tropical curve of degree d (possibly degnerated).
For each ray ρ of C with primitive generator v P = (v1 , v2 ) ∈ Z2 , we set
ord∞ (ρ) := ω(ρ) max{v1 , v2 , 0}. Show that d = ρ ord∞ (ρ), where the
sum runs through all rays of C.
2.6 Floor decompositions

Floor decomposition is a combinatorial tool allowing to construct/describe
a hypersurface in Rn using several hypersurfaces in Rn−1 . The hypersurfaces
which admit such a description (as well as their dual subdivisions) are of
rather special type and are called floor-decomposed. Particular types of floor-
decomposed subdivisions appeared in the context of Viro’s patchworking
method, see [IV96]. See Figure 2.39 for an example. In tropical geometry,
they are particularly useful in the context of enumerative geometry [BM07;
BM09]. We present the basic ideas here.
Floor decompositions can be defined with respect to any surjective linear
map of tropical vector spaces V → W of dimension n and n − 1, respec-
tively. For simplicity, we choose explicit coordinates here and only consider
the fixed projection π : Rn → Rn−1 forgetting the last coordinate. We will
69
Figure 2.36: Floor-decomposed subdivisions in dimension 2 and 3
also need the dual projection q : Rn → R which forgets all except the n-th
coordinate.
Definition 2.6.1
Let V ( f ) ⊂ Rn be a hypersurface with dual subdivision SD( f ). The V ( f ) is
called floor-decomposed (with respect to π) if for each cell D ∈ SD( f ) the
image q(σ), which is an integer interval in R, is of length at most 1 (i.e. of
length 0 or 1).
Examples of floor-decomposed subdivisions are given in Figure 2.36. The
surface corresponding to the three-dimensional subdivision is displayed in
Figure 2.37.
A floor-decomposed hypersurface splits naturally into floors and elevators
as follows. We assume q(NP( f )) = [0, m]. For a cell σ of V ( f ), we denote
the dual cell in SD( f ) by Dσ . For i = 1, . . . , m, we define
[
Fi := σ,
σ cell of V ( f )
q(Dσ )=[i−1,i]
and call this closed set a floor of V ( f ). For i = 0, . . . , m, we define

[
Ei := σ ,
σ cell of V ( f )
q(Dσ )={i}
70
Figure 2.37: Two views of a floor-decomposed surface
and call this set an elevator of V ( f ). (In a more refined setting, one might
call the connected components of the Ei elevators.) The floors and elevators
in Figure 2.37 are shown in gray and red, respectively.
Accordingly, we can split the polynomial f , i.e. we can write f as a
polynomial of the last coordinate x n
m
X
f =“ f i x ni ”,
i=0
with f i ∈ T[x 1 , . . . , x n−1 ]. We have the following relation.
Proposition 2.6.2
The projection π(Ei ) of an each elevator is a tropical hypersurface in Rn−1
given by V ( f i ).
Each floor Fi projects one-to-one to Rn−1 and is equal to the graph of the
function “ f i−1 / f i ” = f i−1 − f i on Rn−1 .
The floors and elevator projections of our example surface are displayed
in Figure 2.38.
Proof. For the first part, let us fix an elevator Ei . We want to show that
π(Ei ) = V ( f i ) as sets. First, take x ∈ Ei . By definitions, the dual cell Dx
71
F2
π(E0 )
π(E1 )
F1
Figure 2.38: The elevators and floors of our example surface
of x in SD( f ) satisfies q(Dx ) = {i}. It follows that locally at x we have

f ≡ f Dx ≡ “ f i x ni ”. In particular, f i (π(x)) attains its maximum at least twice
and thus π(x) ∈ V ( f i ). For the Pother direction, pick y ∈ Rn−1 . We may
consider the polynomial g = “ i f i ( y)x ni ” in one variable x n obtained by
restricting f to π−1 (x). Note that g is of degree m and its subdivision SD(g)
can be computed as
SD(g) = {q(Dσ ) : σ ∈ S ( f ) such that σ ∩ π−1 ( y) 6= ;}.
As f is floor-decomposed, all cells of SD(g) have length 1, i.e. [0, m] is max-

imally subdivided. It follows that for any i, there exists x n such that the only
maximal term in g(x n ) is “ f i ( y)x ni ”. If we further assume y ∈ V ( f i ), the
point x = ( y, x n ) ∈ Rn is in V ( f ) and the maximum of f (x) is only attained
on the i-th level. It follows x ∈ Ei and therefore y ∈ π(Ei ). This proves
π(Ei ) = V ( f i ). Note that also the multiplicities of the two sets are compat-
ible in the sense that ωV ( f ) (x) = ωV ( fi ) (π(x)). This is true because we de-
fined the multiplicity to be the volume of the dual cell in SD( f ) resp. SD( f i ),
which is the same (up to being embedded in different ambient spaces).
Now let us prove the second part of the assertion. First note that x ∈ Fi
is equivalent to
f (x) = “ f i (π(x))x ni ” = “ f i−1 (π(x))x ni−1 ”.
Transforming the second equality according to the rules of tropical arith-

metics, we get x n = f i−1 (π(x)) − f i (π(x)). So Fi is contained in the graph
72
of the function f i−1 − f i and therefore π| Fi is injective. For surjectivity, we

use again the factP that, for any y ∈ Rn−1 and corresponding one variable
polynomial g = “ f i ( y)x ni ”, SD(g) maximally subdivides [0, m]. Hence,
there exists a (unique) “zero” z of g with g(z) = “ f i ( y)z i ” = “ f i−1 ( y)z i−1 ”
and therefore ( y, z) ∈ Fi .
The nice feature of floor-decomposed hypersurfaces is that they are com-
pletely described by (the projection of) their elevators and the “height” of
their floors. This is the statement of the following proposition. For simplic-
ity, we restrict ourselves to the case of non-degenerated hypersurfaces, i.e.
NP( f ) = ∆d , the simplex of size d. Of course, for other toric varieties and
different shapes of Newton polytopes, one can proceed similarly.
We make the following convention. For a non-degenerated hypersurface
V in TPn , we denote by f V the unique tropical polynomial (not Laurent
polynomials) in n variables such that V = V ( f V ) and the constant term of
f V is 0 (not −∞).
Proposition 2.6.3
Let V0 , . . . , Vd−1 be non-degenerated tropical hypersurfaces in Rn−1 of degree
deg(Vi ) = d −P i. Moreover, let h1 < h2 < · · · < hd be real numbers and
set f i := f Vi − 0< j≤i h j for all i = 0, . . . , d (with and f Vd ≡ 0). Under the
assumption
(2 f i − f i−1 − f i+1 )(x) > 0
for all x ∈ Rn−1 and i = 1, . . . , m − 1, there exists a unique generic floor-
decomposed hypersurface V ⊂ TPn such that π(Ei ) = Vi for all elevators and
each floor Fi intersects the x n -coordinate
Pm axis at height hi . This hypersurface
is given by V = V ( f ), where f = i=0 f i x ni . In particular, V has degree d.
Pm
Proof. Let us study V := V ( f ) with f := i=0 f i x ni . The assumption (2 f i −
f i−1 − f i+1 )(x) > 0 makes sure that the various floors (given by the graphs of
f i−1 − f i ) do not intersect and therefore V is indeed floor-decomposed. By
proposition 2.6.2, the elevators of V satisfy π(Ei ) = V ( f i ) = Vi . Moreover,
again by proposition 2.6.2, the height of the intersection of Fi with the x n -
coordiante axis is given by the the difference of the constant terms of f i−1
and f i , which is −h1 −· · ·−hi−1 +h1 +· · ·+hi−1 +hm = hm . For the uniqueness
of V , we note that the condition π(Ei ) = Vi already implies f i − f Vi ≡ const.
These constants are uniquely fixed (up to adding a global constant) by the
“heights” of the floors Fi .
73
Figure 2.39: The Itenberg-Viro subdivision in two dimensions
Example 2.6.4
Figure 2.39 shows an example of a floordecomposed curve that appears
in the context of Viro’s patchworking method. An simple example for the
construction of a stacking floors on top of each other is given in Figure 2.40.
References
The main themes of this chapter — tropical polynomials, tropical hypersur-
faces, dual subdivisions and the balancing condition — appear it many of
the early papers on tropical geometry and we do not attempt to describe
the history of the development of these notions and its diverse predecessors
in other areas of mathematics. Among the earliest “tropical” papers treat-
ing (or at least mentioning) some of this material, let us mention [Mik01;
Mik04; Stu02; SS04; Mik05; EKL06]. Note, in particular, that [EKL06]
is based on an earlier unpublished preprint by Kapranov treating tropical
hypersurfaces (under the name “non-archimedean amoebas”).
74
Figure 2.40: Stacking floors on top of each other
75
3 Projective space and other
tropical toric varieties
In the previous chapter we studied the tropical space Rn and its hypersur-
faces. Let us recall once more that R = T× is the set of units for tropical
multiplication on T = R ∪ {−∞}. Hence, Rn = (T× )n is the tropical alge-
braic torus with classical analogue (C× )n and we treated what is sometimes
called very affine varieties so far.
The goal of this chapter is to study tropical geometry for certain com-
pactifications of Rn , in particular, tropical projective space. In fact, most
of the new features that will show up are already present in tropical affine
space Tn , a partial compactification of Rn . The set of new points Tn \ Rn at
“infinity” behaves differently from the points in Rn , even on a topological
level. In particular, the number of coordinates equal to −∞, called the
sedentarity of a point, is an invariant under tropical automorphisms of Tn .
This is of course in contrast to the classical situation, with Cn and CPn being
fully homogeneous spaces.
A natural class of compactifications of Rn = (T× )n are toric varieties. The
main feature of toric varieties is that they can be constructed by monomial
maps and binomial equations, respectively, which is to say, without using
addition. Since the peculiar idempotent nature of tropical arithmetics ac-
curs only in the context of tropical additions, it turns out that toric varieties
can be pushed to the tropical world instantly.
3.1 Tropical affine space Tn

The natural first step when compactifying Rn is, of course, to consider
tropical affine space Tn — with classical analogue Cn . Here, we always
equip T with the topology generated by intervals of the type [−∞, a), (b, c),
a, b, c ∈ R, and use the product topology on Tn .
76
3 Projective space and other tropical toric varieties
An element of Tn is an n-tuple of numbers x = (x 1 , . . . , x n ), but now some

of these coordinates might be equal to x i = −∞. If all coordinate entries
x i are finite, i.e. x ∈ Rn , we call x a points at finite distance. The additional
points x ∈ Tn \ Rn , with one or more coordinate entries equal to −∞, are
called points at infinity. Of course, we can refine this distinction by setting
for each subset I ⊆ [n]
R I := Rn\I := {x ∈ Tn : x i = −∞∀i ∈ I, x i 6= −∞∀i ∈

/ I}.
This gives us a natural stratification

G
Tn = RI .
I⊆[n]
Note that for all I we have an identification R I = Rn\I ∼

= Rn−|I| , which means
n n
that all the strata in T \ R are tropical algebraic tori of smaller dimension
(see Figure 3.1). Analogously, we set
T I := Tn\I := {x ∈ Tn : x i = −∞∀i ∈ I}.
Each T I is again a tropical affine space T I = Tn\I ∼

= Tn−|I| .
(−∞, b), b ∈ R (a, b) ∈ R2
(−∞, −∞) (a, −∞), a ∈ R
Figure 3.1: T2 and its stratification
Definition 3.1.1
Let x be a point in Tn . The sedentarity sed(x) is the number of coordinates
of x equal to −∞. Namely, if x ∈ R I , then sed(x) = |I|.
77
How do the points of different sedentarity interact? How are the various
strata glued together geometrically? To understand this better, let x, v ∈ Rn
be vectors consider the ray R = x + R≥0 v ⊆ Rn . We are interested in the
closure R ∈ Tn . A few examples of rays in T2 are depicted in Figure 3.2 (the
vectors labelling each ray describe the particular choice of v).
(−2, 1)
(1, 1)
(−1, 0)
(0, −1)
(−1, −1) (−1, −2)
Figure 3.2: Rays in T2 labelled by their direction vectors
We have to distinguish two cases. Whenever v contains a strictly positive

entry, no limit point exists, i.e. R = R. Conversely, if all entries of v are non-
positive, a limit point x = x(x, v) is added to R = R ∪ {x}. However, figure
3.2 might give us a wrong impression of where this point x is located. Its
coordinates are ¨
xi if vi = 0,
xi =
−∞ if vi < 0.
In particular, whenever vi < 0, the coordinate x i does not depend on the
starting point x. In our examples, it follows that two of the rays, though
having different directions, run into the same limit point (−∞, −∞), in-
dependent of their exact starting point (see Figure 3.3). Only the two
rays with direction (−1, 0) resp. (0, −1) have limit points different from
(−∞, −∞).
Let σ ⊆ Rn be the cone spanned by the negative standard basis vectors
−e1 , . . . , −en . Its faces are of the form σ I , where I ⊆ [n] and σ I is the cone
spanned by −ei , i ∈ I. Then the above discussion can be formalized and
summarized as follows. If v ∈ / σ, then no limit point is added and R = R.
Conversely, if v ∈ σ, let σ I be the minimal face of σ containing v. Then
78
(a, b)
(a, −∞)
Figure 3.3: The closures of the rays in T2
x ∈ R I and its finite coordinates are just given by the projection
Rn → Rn−|I| ∼
= RI ,
(x i )1≤i≤n 7→ (x i )i ∈I
/ .
We see that the stratum containing x is determined by the minimal face

containing v. For example, the codimension one strata of Tn can only be
reached along the n special directions −e1 , . . . , −en .
x3
(−∞, −∞, −∞)
x2
x1
Figure 3.4: Tropical affine space T3
79
Exercise 3.1.2
Let A = (ai j ) ∈ Mat(m× n, Z) be an integer matrix and let φ : Rn → Rm , x 7→
Ax be the associated S linear map. Set I(A) := {i : aki ≥ 0 ∀ k} and
integer
U(A) = T \ i ∈I(A) T{i} = I⊂I(A) R I . Show that the following statements.
n
S
/
(a) There exists a unique continuous extension of φA to a map φ : Tn ⊃

U(A) → Tm .
(b) For I ⊂ I(A), we set K = K(I) = {k : aki > 0 for some i ∈ I}. Then the
extension φ maps R I ⊂ Tn to RK ⊂ Tm .
(c) For K = K(I), the map φ : R I → RK is the integer linear map associ-
ated to the matrix A with K-rows and I-columns removed.
3.2 Tropical toric varieties

A complex monomial map is a map of
ΦA : (C× )n → (C× )m ,
Y Y
a a
(z1 , . . . , zn ) 7→ ( zi 1i , . . . , zi mi ),
i i
where A = (ai j ) ∈ Mat(m × n, Z) is an integer matrix. Complex toric va-

rieties are glued via (invertible) monomial maps. Obviously, the concept
can
Q beakitransferred to any other group, and in particular, to (T, “ · ”). Since
“ x i ” = aki x i , a tropical monomial map is nothing but an integer lin-
P
ear map
φA : R n → R m ,
X X
x = (x 1 , . . . , x n ) 7→ Ax = ( a1i x i , . . . , ani x i ).
i i
We are mostly interested in invertible maps, i.e., in the case m = n and

det(A) = ±1 (again, independent of the group). In other words: Tropical
monomial transformations of Rn are Z-invertible integer linear maps of Rn .
We are would like to extend φA to (parts of) Tn — the problemF that may
occur is that aki · (−∞) is not defined when aki < 0. Let Tn = R I be the
stratification of Tn as before. Then φA can be extended to the stratum R I if
80
and only if aki ≥ 0 for all i ∈ I (see Exercise 3.1.2). Again, the analogous
statements hold in the complex setting, which means that the rules for glu-
ing affine patches together are exactly the same. For simplicity, we restrict
to smooth toric varieties here.
Recall that a pointed fan Ξ in Rn is called unimodular if every cone σ
of Ξ can be generated by dim(σ) integer vectors that form a lattice basis
of LZ (σ). Let Ξ be a unimodular pointed fan in Rn . For any cone σ ∈ Ξ
of dimension k, fix a lattice basis e1σ , . . . , enσ of Zn such that −e1σ , . . . , −ekσ
generate σ and set
Uσ := Tk × (T× )n−k = Tk × Rn−k .

{0}
By convention, we use the standard basis for σ = {0}, i.e., ei = ei . Given
0
two cones σ, σ0 ∈ Ξ, let Aσ,σ ∈ GL(n, Z) be the coordinate change matrix
0
from basis {eiσ } to basis {eiσ }. Explicitly, if M σ ∈ GL(n, Z) is the matrix
0 0
whose columns are given by eiσ , then Aσ,σ = (M σ )−1 M σ . We obviously
0 0 00 00 0 0
have Aσ,σ Aσ ,σ = Aσ,σ and (Aσ,σ )−1 = Aσ ,σ . It is straightforward to check
that the corresponding maps φAσ,σ0 extend to boundary strata as summa-
rized in the following lemma.
Lemma 3.2.1
Let τ be a face of σ ∈ Ξ. Then the map φAσ,τ : Rn → Rn extends to an open
embedding φσ,τ = φAσ,τ : Uτ → Uσ . We denote the image by Uστ ⊂ Uσ .
Let σ, σ0 ∈ Ξ two cones with τ = σ ∩ σ0 6= ;. Then the map φAσ,σ0 :
Rn → Rn extends to a homeomorphism φσ,σ0 = φAσ,σ : Uστ0 → Uστ . Moreover,
φσ,σ0 ◦ φσ0 ,σ00 = φσ,σ00 whenever defined.
Exercise 3.2.2
Prove the preceding lemma (e.g. using Exercise 3.1.2).
Definition 3.2.3
Let Ξ be a unimodular pointed fan in Rn . The tropical smooth toric variety
associated to Ξ is the topological space
[
X Ξ := Uσ / ∼,
σ∈Ξ
where ∼ is given by x ∼ φσ,σ0 (x) for all σ, σ0 ∈ Ξ, x ∈ Uστ0 .
81
The following properties follow easily from the previous discussion.
Proposition 3.2.4
Let X Ξ be a tropical smooth toric variety. Then the following properties hold.
(a) X (Ξ) contains Rn = U{0} as an open dense subset and addition on Rn

extends to an action of Rn on X (Ξ),
(b) The orbits of the tropical torus action are in bijection to the cones of Ξ.
For each σ ∈ Ξ of dimension k the corresponding orbit is given by
Rσ = {−∞} × Rn−k ⊂ Tk × Rn−k = Uσ , (3.1)
or rather, its image in X Ξ . It can be canonically identified with the quo-

tient Rσ ∼
= Rn / L(σ).
(c) Let X τ denote the closure of Rτ in X Ξ , τ ∈ Ξ. Then X τ is the tropical

smooth toric variety associated to StarΞ (τ) and can be written as disjoint
union of orbits G
X τ = X StarΞ (τ) = Rσ .
τ⊂σ
(d) Let x ∈ Rn be a point, and let v ∈ Rn be a direction vector. Then

x + λv converges for λ ∈ R, λ → ∞ to a point x if and only if v is
contained in the support of Ξ. Moreover, if v lies in the relative interior
of σ, then x ∈ Rσ and is equal to the image of x under the projection
πσ : Rn → Rn / L(σ) ∼
= Rσ (using the canonical identification mentioned
previously).
(e) The topological space X Ξ is compact if and only if the fan Ξ is complete,
i.e., |Ξ| = Rn .
There is a corresponding complex smooth toric variety CX (Ξ) which is

obtained from exact same gluing procedure, but using complex monomial
maps on (C× )n instead. Properties (a), (b), (c), (e) can be literally trans-
lated to the complex case (replacing R, T by C× , C). Property (d) reflects
the corresponding statement about limits of (translations of) one-parameter
subgroups in (C× )n (which are of the form (z1 t v1 , . . . , zn t vn ) with z ∈ (C× )n , t ∈
C× , v ∈ Zn ).
82
Remark 3.2.5
Property (d) together with the requirement that the embeddings Rσ ,→ X Ξ
are continuous fixes the topology on X Ξ uniquely. Hence an F alternative
construction of X Ξ could be based on defining a topology on σ∈Ξ R/ L(σ)
satisfying these two properties. Note that this description extends to arbi-
trary (not necessarily unimodular) pointed fans Ξ, providing (not neces-
sarily smooth) tropical toric varieties. Proposition 3.2.4 still holds for such
general toric varieties, except for the description of orbits given in Equa-
tion (3.2.4) (in the sense that the open charts Uσ are no longer of the form
Tk × Rn−k in general). The description of toric varieties in terms of semi-
group homomorphisms (see for example [Ful93]) can also be immediately
“tropicalized”.
For later use, we extend the discussion of limit points of rays to arbitrary
polyhedra P ⊂ Rn . Recall that the recession cone rc(P) of P is the set of
direction vectors of all rays contained in P.
Proposition 3.2.6
Let X Ξ be a tropical toric variety, σ ∈ Ξ a cone and P ⊂ Rn a polyhedron. Let
P denote the closure of P in X Ξ . Then the following holds.
(a) P ∩ Rσ = ; ⇐⇒ rc(P) ∩ σ = ;.
(b) If P ∩ Rσ 6= ;, then P ∩ Rσ = πσ (P). In particular, P ∩ Rσ is a rational

polyhedron in Rσ and dim(P) = dim(P ∩ Rσ ) + dim(L(P) ∩ L(σ)).
Example 3.2.7
Before treating “serious” examples of tropical toric varieties in the next
section, let us look back at tropical affine space from Section 3.1. Let
σ I , I ⊂ {1, . . . , n} be the cones generated by the negative standard basis
vectors −ei , i ∈ I and let Ξ be the fan formed by all such cones. Then ob-
viously the tropical toric variety X Ξ is canonically isomorphic to Uσ[n] = Tn .
Indeed, choosing the lattice basis e1 , . . . , en globally (i.e., M σ = Id for all
σ), all gluing maps are identity maps on the subsets UσI ⊂ Tn . Moreover,
the torus orbit RσI agrees with the sedentarity stratum R I , and the closure
X σI is equal to T I .
Exercise 3.2.8
83
Prove Proposition 3.2.4, for example, by copying the proofs of the corre-
sponding facts from your favorite textbook on classical toric varieties (e.g.
[Ful93]).
Exercise 3.2.9
Prove Proposition 3.2.6.
3.3 Tropical projective space

The most basic compact toric variety is, of course, projective space. We start
with the symmetric description of TPn as a quotient and then recover the
fan by using property (d) above.
Definition 3.3.1
Tropical projective space is
TPn := Tn+1 \ {−∞}/ ∼,
where the equivalence relation ∼ is given by x ∼ “λ · x” for all λ ∈ T× . In

other words,
(x 0 , . . . , x n ) ∼ (x 0 + λ, . . . , x n + λ)
for all λ ∈ R. We equip TPn with the quotient topology.
As usual, projective coordinates are denoted by (x 0 : . . . : x n ). Projective

space TPn can be covered by affine patches Ui = {x ∈ TPn : x i 6= −∞}
which can be identified with Tn via
(x 0 : . . . : x n ) 7→ (x 0 − x i , . . . , x i−1 − x i , x i+1 − x i , . . . , x n − x i ).
On the overlaps Ui ∩ U j we get induced maps glueing two copies of Tn via
φi j : ( y0 , . . . , yi−1 , yi+1 , . . . , yn ) 7→
( y0 − y j , . . . , yi−1 − y j , − y j , yi+1 − y j , . . . , Ù
y j − y j , . . . , yn − y j ). (3.2)
Thus TPn is obtained by glueing n + 1 copies of Tn along these integer lin-

ear transformations (whenever defined). The gluing process is visualized
for the tropical projective plane TP2 in Figure 3.5. Note that TPn contains
84
U2
TP2

1 −1
0 −1 −1 1 0 −1
1 −1 −1 0

−1 0
−1 1
U0 U1
Figure 3.5: Glueing TP2 from its three affine patches
Rn+1 /R(1, . . . , 1) ∼
= Rn as an open dense subset and this torus acts on TPn
by coordinatewise addition. The orbits of this action are of the form
R I = {(x 0 : . . . : x n ) ∈ TPn : x i = −∞∀i ∈ I, x i 6= −∞∀i ∈
/ I}
with I ( {0, . . . , n}.
Let us review the construction in the language of toric variety. For any
proper subset I ( {0, . . . , n}, we define σ I ⊂ Rn as the cone generated by
the vector −ei , i ∈ I. Here, e1 , . . . , en denotes the standard basis of Rn and
−e0 = e1 + · · · + en = (1, . . . , 1). Let Ξ be the fan consisting of all such cones.
We denote the maximal cones by σi := σ{0,...,î,...n} , i = 0, . . . , n. For each
such cone, we take as lattice basis the one given by e0 , . . . , êi , . . . , en . Then
it is straightforward to check that linear maps described Pn by Aσ j ,σi are equal
to the gluing maps φi j from above. Indeed, since k=0 y j ek = 0, by setting
yi = 0 we change coordinates via the computation
X n
X n
X X
yk ek = yk ek = ( yk − y j )ek = ( yk − y j )ek ,
k6=i k=0 k=0 k6= j
which is in agreement with φi j . Hence X Ξ = TP . Figure 3.6 shows the fan

n
and orbit stratification for TP2 .
Remark 3.3.2
Note that the fans we associate to tropical toric varieties are the reflection at
85
pt
(1, 1)
R (−1, 0)
R
R2
(0, −1)
pt R pt
Figure 3.6: The stratification and the fan of TP2
the origin of fans commonly used in classical toric geometry. This because it
seems more natural tropically to follow v in positive direction, i.e. λ → ∞,
whereas in the classical case one usually performs t → 0 (for t ∈ C× ).
By Proposition 3.2.4 part (d), the torus Rn ⊂ TPn is equipped with n + 1

divisorial directions (−1, . . . , 0), (0, −1, . . . , 0), . . . , (0, . . . , −1), (1, . . . , 1). In
homogeneous coordinates, we can rewrite them more symmetrically as
(−1 : · · · : 0), (0 : −1 : · · · : 0), . . . , (0 : · · · : −1). Among all rays in Rn ,
only those whose direction vector is one of the divisorial directions end up
at point in an (n − 1)-dimensional stratum of TPn \ Rn . All other rays have
limit points in strata of higher codimension. Examples for TP2 are given in
Figure 3.7.
Note, in particular, that our illustration of TP2 as a triangle, while reflect-
ing correctly the stratification into torus orbits, does not exhibit the metric
properties of TPn : Of course, the points on the boundary of the triangle
are infinitely far away from all points in the interior, and points in the inte-
rior of edges can only be reached via the corresponding divisorial direction.
Note also that the divisorial directions showed up before, for example, in
our treatment of non-degenerated planar curves in Section 2.5. In the same
way, TPn is topologically an n-simplex (with matching stratifications). The
n + 1 divisorial directions are in correspondence to the n + 1 maximal faces
of the simplex (see Figure 3.8).
86
Figure 3.7: Some rays and their limit points in TP2
(1, 1, 1)
(−1, 0, 0)
R3
R
(0, −1, 0)
R2
(0, 0, −1)
Figure 3.8: TP3 and its distinguished directions
87
3.4 Projective hypersurfaces

In chapter 2, we used tropical Laurent polynomials to describe hypersur-
faces in Rn . We will now see how a homogeneous tropical polynomial de-
scribes a hypersurface in TPn . Let us start with Tn first. Let
X
f (x) = “ aj x j”
j∈Nn
be a tropical polynomial in n variables (with multi-index notation). The

coefficients a j ∈ T are tropical numbers and the set A = { j : a j 6= −∞} is
finite (and non-empty). As all exponents are positive now, this describes a
function
f : Tn → T,
and similar to the case of Rn , the locus of points where two terms in f attain
the maximum simultaneously is independent of the polynomial represen-
tation.
Definition 3.4.1
The affine hypersurface V ( f ) ⊂ Tn given by the tropical polynomial f ∈
T[x 1 , . . . , x n ] is (as a set)
V ( f ) := {x ∈ Tn : ∃i 6= j ∈ Nn such that f (x) = ai x i = a j x j }. (3.3)
Note that since we allow any i, j ∈ Nn , the hypersurface V ( f ) contains

the “honest” zero-set {x : f (x) = −∞} of f , a possibly non-empty set now.
Clearly, V ( f ) ∩ Rn is equal to the very affine hypersurface V ( f |Rn ). Since
V ( f ) is closed, the closure of V ( f |Rn ) is also contained in V ( f ). We denote it
by V ( f )fin and call it the finite part or sedentarity zero part of V ( f ). Beyond
V ( f )fin , we claim that V ( f ) consists only of coordinate hyperplanes
H i := V (x i ) = {x ∈ Tn : x i = −∞}.
Let ord x i ( f ) ≥ 0 denote the maximal integer k such that x i divides f (i.e. f
can be written as f = “x k f 0 ” for some other tropical polynomial f 0 ). Note
that k only depends on the Newton polytope of f and hence on the func-
tion f (not on the polynomial representation). We say f is monomialfree if
ord x i ( f ) = 0 for all i = 1, . . . , n (equivalently, x i 6 | f ).
88
Theorem 3.4.2
Let f ∈ T[x 1 , . . . , x n ] be a non-zero tropical polynomial. Then
V ( f ) = V ( f )fin ∪ H i1 ∪ · · · ∪ H ik ,
where i1 , . . . , ik ∈ {1, . . . , n} are exactly those indices for which ord x i ( f ) > 0.
Proof. Given two polynomials g, h ∈ T[x 1 , . . . , x n ], we can check easily that
V (“gh”) = V (g) ∪ V (h). Let us write f = “x j g”, where j = ( j1 , . . . , jn ) ∈ Nn
is the exponent vector of a monomial S and g is monomialfree. Then V ( f ) =
V (x ) ∪ V (g). Clearly, V (x ) = i: ji >0 H i . Moreover, note that f |Rn and
j j
g|Rn only differ by the integer linear function x 7→ j x and thus V ( f |Rn ) =
V (g|Rn ) and V ( f )fin = V (g)fin . So it suffices to prove the statement in the
monomialfree case.
So from now on we assume f is monomialfree and it remains to show
that V ( f ) = V ( f )fin . The inclusion “⊃” is obvious since V ( f ) is closed and
V ( f |Rn ) ⊂ V ( f ). For the other inclusion, let p ∈ V ( f ) be point, say, in the
stratum R I ⊂ Tn . Without loss of generality, we may assume I = {1, . . . , s},
s = |I|. Let us first consider the case f (p) 6= −∞. Then there exist two
0
terms “a j x j ”, “a j 0 x j ” attaining the (finite) maximum at p. In particular,
this implies ji = ji0 = 0 for all i = 1, . . . , s. Considering the points p(λ) :=
(−λ, . . . , −λ, ps+1 , . . . , pn ), we see that for sufficiently large λ ∈ R
0
f (p(λ)) = “a j p(λ) j ” = “a j 0 p(λ) j ” = f (p).
Hence p(λ) ∈ V ( f |Rn ) for large λ and therefore p ∈ V ( f )fin .

Let us consider the case f (p) = ∞. We set Tsp := Ts ×{(ps+1 , . . . , pn )} ⊂ Tn
and consider the restriction
f p := f |Tsp = f (x i = pi : i = s + 1, . . . , n) ∈ T[x 1 , . . . , x s ].
Obviously, f p is monomialfree, f p (p) = f (p) = −∞ and

V ( f p ) × {(ps+1 , . . . , pn )} ⊂ V ( f ) ∩ Tsp .
Hence the claim follows from the following lemma.
Lemma 3.4.3
Let f ∈ T[x 1 , . . . , x n ] be a non-zero monomialfree tropical polynomial such
that f ((−∞)n ) = −∞, (−∞)n = (−∞, . . . , −∞) ∈ Tn . Then (−∞)n ∈
V ( f )fin .
89
Proof. Let P := NP( f ) be the Newton polytope of f and let N (P) be its
normal fan. By our assumptions, P lies in the positive orthant of Rn but
is touching all the hyperplanes { ji = 0}, i = 1, . . . , n. Consider the open
negative orthant N := (R<0 )n . We want to show that N intersects one of
the non-maximal cells of N (P). Assuming the contrary, it follows that N is
contained in a n-cone of N (P) corresponding to a vertex V of P. Choosing
a sequence of rays in N converging to −ei , it follows by continuity that V
is contained in { ji = 0}. Running through all i = 1, . . . , n, we find V =
(0, . . . , 0) ∈ P, which is a contradiction to f ((−∞)n ) = −∞. Hence N
intersects a non-maximal cone σ D of N (P) dual to a positive-dimensional
face F of P. Let D be a positive-dimensional cell of SD( f ) such that D ⊂ F .
Let σ D ⊂ V ( f |Rn ) be the cell of S ( f )(n−1) dual to D. By Exercise 2.3.13 we
have rc(σ D ) = σ F , which intersects N . Hence by Proposition 3.2.6 we see
that (−∞)n ∈ σ D ⊂ Tn , which proves the claim.
So far, we described V ( f ) as a set. Clearly, as in the case of Rn , we

would like to add a weighted polyhedral structure to V ( f ). Postponing the
definition of polyhedral complexes in the compactified setting to Chapter
5, we restrict ourselves here to an adhoc description of the weight function
based on the description of V ( f ) as a set.
Definition 3.4.4
Let X Ξ be a tropical toric variety with torus Rn . A tropical divisor or tropical
n − 1-cycle D is a formal sum of the form
D = D0 + k1 D1 + · · · + kl Dl ,
where
• D0 is a (single) n − 1-cycle in Rn ,
• k1 , . . . , kn ∈ Z,
• Di = X ρi , where ρ1 , . . . , ρl denote the rays of Ξ.
The support of D is the union of |D0 | with Di for all i such that ki 6= 0.
Definition 3.4.5
The affine hypersurface V ( f ) ⊂ Tn defined by the tropical polynomial f ∈
90
T[x 1 , . . . , x n ] is (as a divisor)

n
X
V ( f ) := V ( f |Rn ) + ord x i ( f )H i .
i=1
By Theorem 3.4.2 the support of V ( f ) agrees with the set-theoretic defi-

nition in Equation (3.3).
Once again, given our treatment of the case Tn , we can easily extrapolate
to arbitrary tropical toric varieties. Since the procedure is completely anal-
ogous to the case of classical toric varieties, we will only discuss the case
TPn in more details and restrict ourselves to a brief outline of the general
case in Remark 3.4.8.
Let X
F (x) = “ aj x j”
j∈Nn+1
be a tropical homogeneous polynomial of degree d ∈ Z in n + 1 variables.

In more details, the coefficients a j ∈ T are tropical numbers, the set A =
{ j : a j 6= −∞} is finite and for all j ∈ A we have | j| = j0 + . . . + jn = d. In
projective coordinates, this does not quite define a function to T as we have
F (λ · x) = “λd F (x)” = λd + F (x). However, the set
V (F ) = {x ∈ TPn : ∃i 6= j ∈ Nn such that F (x) = ai x i = a j x j } (3.4)
is still well-defined. Fixing an affine chart Ui ∼

= Tn ⊂ TPn corresponds to
the dehomogenization
f i = F ( y0 , . . . , 1, . . . , yn ) ∈ T[ y0 , . . . , ŷi , . . . , yn ].
Under the coordinate change (3.2), these polynomial follow the rule
f i ( y) − f j (φi j ( y) = d y j .
In particular, since the difference is an integer linear function very affine

hypersurface V ( f i |Rn ) ⊂ Rn ⊂ Ui is independent on the affine chart (under
the coordinate changes φi j ). We call its closure in TPn the finite part or
sedentarity zero part of V (F ), denoted by V (F )fin .
91
Definition 3.4.6
The projective hypersurface V (F ) ⊂ TPn defined by the tropical homoge-
neous polynomial F is the tropical divisor
n
X
V (F ) := V (F ) +
fin
ord x i (F )H i ,
i=1
where H i = {x ∈ TPn : x i = −∞} is the i-th coordinate hyperplane.
Remark 3.4.7
Note that ord x k (F ) = ord yk ( f i ) for all i 6= k. In particular, defining V (F )∩ Ui
in the obvious way, we have V (F ) ∩ Ui = V ( f i ) for all i = 0, . . . , n. Together
with Theorem 3.4.2, this also shows that the support of V (F ) agrees with
the set-theoretic definition in Equation (3.4). Moreover, we may define the
dual subdivision SD(F ) of NP(F ) as before. The canonical identification
NP(F ) ∼= NP( f i ) for all i also identifies the subdivisions SD(F ) ∼ = SD( f i ).
Hence we get a duality between V (F ) and SD(F ) as in the very affine
fin
case.
Remark 3.4.8
The definitions can be immediately generalized to arbitrary tropical toric
varieties following the standard representation of Cox rings for (classical)
toric varieties. More precisely, let Ξ be a pointed fan in Rn and let v1 , . . . , vl
be the primitive generators of its rays. Let A be the kernel of the map Zl →
Zn , ei 7→ vi . Set B = A∗ and let δ : Zl → B be the map dual to A ,→ Zl . The
toric degree of a monomial x j in the variables x 1 , . . . , x l is given by δ( j) ∈ B.
Fixing d ∈ B, a homogenuous polynomial of degree d is a tropical sum
X
F (x) = “ a j x j ” ∈ T[x 1 , . . . , x l ]
j∈Nl
such that A = { j : a j 6= −∞} is finite and for all j ∈ A we have δ( j) = d.

Let c : Zn → Zl be the map dual to Zl → Zn from above. Fixing j0 with
δ( j0 ) = d, we can define the dehomogenization
X
f (x) = “ ac(i)+ j0 y i ” ∈ T[ y1± , . . . , yn± ].
i∈Nn
Choosing different j0 only results in multiplying f by an monomial y i , so

the very affine hypersurface V ( f ) ∈ Rn is well-defined. We call its closure
92
in TPn the finite part or sedentarity zero part of V (F ), denoted by V (F )fin .

Finally, we define the toric hypersurface V (F ) ⊂ X Ξ as the tropical divisor
n
X
V (F ) := V (F ) +
fin
ord x i (F )Di ,
i=1
where Di is the toric boundary divisor spanned by the vector vi . Again,

for a given affine chart Ui ∼ = Tn , there exists a unique choice of j0 such
that V (F ) ∩ Ui = V ( f˜), where f˜ is obtained from f via the corresponding
coordinate change. The statements about the dual subdivision SD(F ) of
NP(F ) and its duality with V (F )fin extend without difficulties.
Proposition 3.4.9
Let X Ξ be a tropical toric variety and let
D = D0 + k1 D1 + · · · + kl Dl
be a tropical divisor in the notation of Definition 3.4.4. Assume that all

weights of D0 are positive and that ki ≥ 0 for all i = 1, . . . , l. Then there
exists a Ξ-homogeneous polynomial F such that V (F ) = D.
Proof. We content ourselves to discuss the projective case, the general case
can be easily adapted. First, note that by Theorem 2.4.10 there exists a
polynomial g ∈ T[ y1 , . . . , yn ] such that V ( f ) = D0 , and we may assume
that f is monomialfree. Let G ∈ T[x 0 , . . . , x n ] denote the homogenization
of f (still mononialfree). Then the polynomial F = x j G with ji = ki satisfies
V (F ) = D.
Let us close this section with a short digression on the homotopy type of
tropical hypersurfaces.
Proposition 3.4.10
Let V ( f ) be a hypersurface in Rn whose Newton polytope NP( f ) is full-dimensional.
Then V ( f ) is homotopy-equivalent to a bouquet of k (n−1)-spheres. Here k is
the number of vertices of the dual subdivision SD( f ) which are interior points
of NP( f ).
Proof. Recall that each vertex of SD( f ) corresponds to a connected compo-

nent of Rn \ V ( f ). For vertices in the boundary of NP( f ), the boundary of
93
this component is contractible and the homotopy type after removing the
component is unchanged. For vertices in the interior of NP( f ), the bound-
ary of the corresponding component is homeomorphic to an (n − 1)-sphere.
Up to homotopy equivalence, removing such a component is equivalent to
removing a point of Rn . But Rn minus k points is a bouquet of k (n − 1)-
spheres, as claimed.
Remark 3.4.11
Assume NP( f ) is not full-dimensional, but generates the affine subspace
A ⊂ Rn in dual space. Then V ( f ) is tranlation-invariant along A⊥ and the
projection to Rn /A⊥ provides a contraction of V ( f ) to a hypersurface in
a smaller-dimensional space whose Newton polytope is full-dimensional
(namely NP( f ) ⊂ A). Therefore the proposition is basically valid in this
case, too. The only difference is that now we count vertices in the relative
interior of NP( f ) and V ( f ) is a bouquet of (m − 1)-spheres, where m is the
dimension of NP( f ).
When considering hypersurfaces in compact toric varieties, the statement

remains nearly unchanged. All we have to do is to use the right notion of
being an interior point now. We state the result for projective hypersur-
faces here, although the generalization to other (compact) toric varieties is
straightforward.
Proposition 3.4.12
Let F be a homogeneous polynomial of degree d such that x i 6 |F for all i and de-
fine ∆d to be the simplex obtained as convex hull of (d, 0, . . . , 0), . . . , (0, . . . , d) ∈
Nn+1 . Let k denote the number of vertices of SD(F ) contained in ∆d . Then the
projective hypersurface V (F ) ⊂ TPn is homotopy-equivalent to a bouquet of k
(n − 1)-spheres.
Proof. Note that TPn is homeomorphic to a closed ball of dimension n. As in

the proof for Rn , each vertex of SD( f ) corresponds to a connected compo-
nent of TPn \ V ( f ). Again, these components differ, depending on whether
the vertex is in the boundary of ∆d or not. Accordingly, the boundary of
the connected component is contractible or homeomorphic to an (n − 1)-
sphere. Thus again, V (F ) is homotopy equivalent to an n-ball minus k
interior points, which is a bouquet of k (n − 1)-spheres.
94
V ( f ) ⊂ TP2 SD( f ) ⊂ ∆3
Figure 3.9: The homotopy type of a plane projective cubic (k = 1)
Exercise 3.4.13
Let f : Tn → T be a tropical polynomial. Show that V ( f ) ⊂ Tn is closed.
Conclude that projective hypersurfaces V (F ) ⊂ TPn are closed.
ExerciseP 3.4.14
Let f = j∈A a j x j be a tropical polynomial. For I ⊂ {1, . . . , n}, set A I ⊂ A
to be the subset of exponent vectors j = ( j1 , . . . , jn ) with ji = 0 whenever
i ∈ I. We call X
f I := / I]
a j x j ∈ T[x i : i ∈
j∈A I
the truncation of f to I. Show that V ( f ) ∩ T I = V ( f I ), including the special

case T I ∈ V ( f ) ⇔ f I = −∞, and an agreement of the weights otherwise.
3.5 Projective curves

A tropical homogeneous polynomial in 3 variables describes a tropical curve
in TP2 . Such a curve always splits into its sedentarity zero part and possi-
bly a union of coordinate lines. We focus our attention to sedentarity zero
curves here. Such curves may equivalently be described by a not necessarily
homogeneous polynomial f in 2 variables which is not divisible by mono-
mials. Here we work with the fixed (very) affine chart R2 = {0} × R2 ⊂ TP2 .
Then f defines the sedentarity zero projective curve V ( f ) ⊂ TP2 . In consis-
tency with section 2.5, the degree of V ( f ) is the smallest number such that
95
Figure 3.10: A generic and a non-generic projective conic in TP2
NP( f ) ⊆ ∆d . Here ∆d is the d-fold standard simplex, i.e. the convex hull
of the points (0, 0), (d, 0), (0, d). In section 2.5, we mostly considered non-
degenerated curves with NP( f ) = ∆d . In projective world, they are distin-
guished by not containing any of the torus fixed points p0 = (0, −∞, −∞),
p1 = (−∞, 0, −∞), p2 = (−∞, −∞, 0) ∈ TP2 . Vice versa, any curve V ( f )
with NP( f ) 6= ∆d contains at least one of these points (see Figure 3.10).
Our goal is to extend Bézout theorem from Section 2.5 to projective
curves. To do so, we need to define intersection multiplicities for curves
intersecting at torus fixed points. We will do this in the following adhoc
manner following [BS11]. Let X , Y be two curves in T2 and let ρ1 , ρ2
be (unbounded) rays of X 1 resp. X 2 . These rays intersect at (−∞, −∞)
if and only if both primitive generators v1 , v2 have only negative entries,
say vi = (x i , yi ), x i , yi ∈ Z<0 (see Figure 3.11). In this case, the contribu-
tion of the intersection of these two rays to the intersection multiplicity at
(−∞, −∞) is defined to be
mult(ρ1 , ρ2 ) = ω(ρ1 )ω(ρ2 ) min{x 1 y2 , x 2 y1 }.
The general definition is then just a summation over all such pairs of rays.
Definition 3.5.1
96
T2
ρ1
ρ2
v1
v2
Figure 3.11: Two rays intersecting at (−∞, −∞)
Let X , Y be two affine tropical curves in T2 . The intersection multiplicity of

two X and Y at (−∞, −∞) is
X
mult((−∞, −∞)) = ω(ρ1 )ω(ρ2 ) min{x 1 y2 , x 2 y1 },
ρ1 ,ρ2
where the sum is taken over all pairs of rays ρ1 and ρ2 of X and Y , respec-
tively, with primitive generators (x i , yi ) ∈ Z2<0 , i = 1, 2.
Let X , Y ⊂ TP2 be two projective tropical curves which intersect transver-
sally in R2 . For the three torus fix points p0 , p1 , p2 , we use the multiplic-
ity mult((−∞, −∞)) in the corresponding charts to define multiplicities
mult(p0 ), mult(p1 ), mult(p2 ). The intersection product X · Y is the formal
sum of points
X
X ·Y = mult(p)p + mult(p0 )p0 + mult(p1 )p1 + mult(p2 )p2 ,
p∈R2
where mult(p) denotes intersection multiplicity in R2 from Section 2.5 (for

the curves X ∩ R2 and Y ∩ R2 ).
Our justification for this ad-hoc definition is that Bézout’s theorem holds.
In the course of the proof, we will deform the rays going to the torus fixed
points to “non-degnerated” rays, providing more precise evidence as to why
this is the correct definition (see Figure 3.12).
Theorem 3.5.2 (Tropical Bézout’s theorem)

Let X and Y be two tropical sedentarity zero curves in TP2 of degree d and e,
97
respectively. Then
deg(X · Y ) = deg(X ) · deg(Y )
holds.
Proof. First note that in the non-degenerated case, i.e. when both X and Y
do not contain torus fixed points, the statement is already proven in Theo-
rem 2.5.2.
In the degenerated case, let σ be a ray of X containing a torus fixed point.
In the corresponding affine chart T2 , let v = (x 1 , y1 ) be the primitive gen-
erator of this ray, x 1 , y1 < 0. Now we perform the following “deformation”
of σ. We pick a point p ∈ σ ∩ R2 such that the remaining unbounded part
p + R≥0 v does not intersect Y . We replace this part of σ by the two rays
p + R≥0 (−1, 0) and p + R≥0 (0, −1) with weights ω(σ)x 1 and ω(σ) y1 , re-
spectively. By construction, this deformed curve X 0 is still balanced, hence
a tropical curve. By Exercise 2.5.5, it has the same degree as X .
By choosing p close enough to (−∞, −∞), we can ensure that the two
new rays only intersect Y in rays also going to (−∞, −∞). Then, each
such ray τ of Y , with primitive generator (x 2 , y2 ), intersects exactly one of
the two rays. Moreover, by our assumption that p + R≥0 v does not intersect
Y , it follows that each τ intersects the (−1, 0)-ray only if x 1 y2 ≤ x 2 y1 , and
vice versa for the (0, −1)-ray. Hence the intersection multiplicity at this new
intersection point is
ω(σ)ω(τ) min{x 1 y2 , x 2 y1 }.
In other words, the new finite intersection points of X 0 · Y exactly compen-
sate the loss of intersection multiplicities at (−∞, −∞), compared to X ·Y ,
and hence deg(X · Y ) = deg(X 0 · Y ). Repeating this procedure for each “de-
generated” ray of X and Y , we eventually end up with two non-degenerated
tropical curves and the assertion follows.
Exercise 3.5.3
Let X , Y ⊂ TP2 be two tropical curves of sedentarity zero such that X ∩ R2
and Y ∩ R2 intersect transversally. Show that X and Y intersect only in R2
and some of the torus fixed points.
Exercise 3.5.4
Show that any two tropical projective curves X , Y ⊂ TP2 intersect.
98
σ σ
p τ (−1, 0)
v1
v v2
u (0, −1)
Figure 3.12: Replacing a degenerated ray by two standard ends
Exercise 3.5.5
Let (TP2 )∗ be the set of tropical lines in TP2 . Show that (TP2 )∗ can be
canonically identified with TP2 . We call (TP2 )∗ the dual projective plane.
99
4 Tropical cycles in Rn
In this chapter, we jump back to case of very affine tropical geometry in
Rn . We will define tropical cycles as balanced poylhedral sets of any codi-
mension and introduce the concept of stable intersection of such cycles.
We also extend the construction of associating a hypersurface to a tropical
polynomial to arbitrary piecewise affine functions on tropical cycles. This
is strongly related to the concept of tropical modifications, an important
construction in tropical geometry which does not have a direct counterpart
in classical geometry.
4.1 Polyhedral geometry dictionary II

A polyhedral set X (of pure dimension k) in Rn is a finite union of polyhedra
X = σ1 ∪· · ·∪σl (with dim(σi ) = k). Obviously, the support of a polyhedral
complex X is a polyhedral set. Vice versa, a polyhedral structure of X is
a polyhedral complex X such that |X | = X . Let us collect a few useful
facts concerning polyhedral sets. We use the following notations. Given an
integer affine function κ : Rn → R, the halfspace complex associated to κ is
the polyhedral subdivision of Rn
Gκ := {Hκ , H−κ , Hκ ∩ H−κ }.
Given a polyhedral complex X in Rn and an abritrary set S ⊂ Rn , the re-
striction of X to S is X |S := {σ ∈ X : σ ⊂ S}. It is clear that X |S is a
polyhedral complex.
Proposition 4.1.1
The following statements hold true.
(a) Let X = σ1 ∪ · · · ∪ σl be a polyhedral set. Then there exists a polyhedral
structure X of X such that X |σi is a subdivision of σi for all i = 1, . . . , l.
In particular, any polyhedral set admits a polyhedral structure.
100
(b) Let X and Y be two polyhedral complexes in Rn and set X = |X |,

Y = |Y |. Then there exists a polyhedral structure Z of X ∪ Y such that
Z |X and ZY are refinements of X and Y , respectively. In particular, if
X = Y , then X and Y have a common refinement.
(c) In both previous items, the promised polyhedral structure can be com-
pleted to a polyhedral subdivision of Rn .
Proof. To prove (a), let κ1 , . . . , κm : Rn → R be a collection of integer affine

forms such that each σi , i = 1, . . . , l can be represented by intersecting
some of the halfspaces Hκ1 , . . . , Hκm . Consider the subdivision of Rn
S = Gκ1 ∩ · · · ∩ G κ m
(see Exercise 2.1.3). Then the polyhedral complex X = S |X satisfies the

conditions in part (a).
To prove (b), apply the construction for part (a) to the union of polyhedra
[ [
X ∪Y = σ∪ σ.
σ∈X σ∈Y
Part (b) is obvious, the completion is given by S .
In other words, polyhedral sets and polyhedral complexes up to common

refinements are essentially the same thing.
Let us turn to fan sets now. A fan set is polyhedral set X such that for all
x ∈ X the ray R≥0 x is contained in X . It follows from Exercise 4.1.3 that
such an X can be written as the union of polyhedral cones. A fan structure
of X is a polyhedral fan X such that X . Clearly, Proposition 4.1.1 and its
proof can be adapted such that all involved sets, polyhedra, complexes are
fan sets, cones, fans, respectively. In particular, any fan set admits a fan
structure and two such structures have a common (fan) refinement.
For later use, let us recall a theorem about simplicial and unimodular
refinements of fans. A cone σ of dimension k is simplicial if it can be gener-
ated by k vectors v1 , . . . , vk . This implies that σ is pointed. Moreover, σ is
unimodular if it can be generated by a part of a lattive basis (equivalently,
if 〈v1 , . . . , vk 〉Z = LZ (σ). A fan F is simplicial or unimodular if all its cones
are simplicial or unimodular, respectively. A fan F is complete if |F | = Rn .
101
Theorem 4.1.2
Let F be a fan in Rn . Then the following statements are true.
(a) There exists a simplicial fan G which is a refinement of F containing all

its simplicial cones.
(b) There exists a unimodular fan G which is a refinement of F containing

all its unimodular cones.
(c) There exists a complete fan E containing all cones of F .
Proof. For (a) and (b), see for example [CLS11, Theorems 11.1.7 and 11.1.9].
For (c), see [Ewa96, Theorem III.2.8] (or [EI06]).
Exercise 4.1.3
Let σ ∈ Rn be a (rational) polyhedron. Show that R≥0 σ = {λx : λ ∈
R≥0 , x ∈ σ} is a (polyhedral) cone.
4.2 Tropical cycles and subspaces in Rn

In chapter 2 we explained that a weighted polyhedral complex of pure di-
mension k in Rn is called balanced if it satisfies the balancing condition
X
ω(σ)vσ/τ = 0 mod L(τ)
σ facet
τ⊂σ
for all k − 1 cells τ (see Definition 2.4.6). Tropical hypersurfaces are exam-
ples of balanced polyhedral complexes of dimension n−1, but note that the
definition makes sense in any codimension. Indeed, such balanced polyhe-
dral complexes of arbitrary codimension are the basic objects in tropical
geometry.
Compared to the case of hypersurfaces, we will sligthly shift our point of
view by focusing on the underlying support sets and not insisting on a fixed
subdivision into cells. Here are the relevant definitions. Recall that for a
weighted polyhedral complex X , the support |X | is the union of cells of
non-zero weight.
102
Definition 4.2.1
Two weighted polyhedral complexes X and Y in Rn are equivalent if |X | =
|Y | and for any pair of facets σ ∈ X , σ0 ∈ Y with σ ∩ τ 6= ; the weights
agree. A weighted polyhedral complex Z is a (weighted) refinement of X
if |Z | = |X |, any σ ∈ Z , σ ⊂ |Z |, is contained in some σ0 ∈ X , and if σ is
a facet, then ωZ (σ) = ωX (σ0 ).
Clearly, any (unweighted) polyhedral complex Z such that Z ||X | is a

refinement of X ||X | inherits a unique weight function such that Z is a
weighted refinement of X . It is easy to check that a weighted refinement of
X is balanced if and only if X is balanced (see Ecercise 4.2.9). Moreover,
using Proposition 4.1.1 we see that X and Y are equivalent if and only
if they have common weighted refinement (see Exercise 4.2.10). Conse-
quently, if the balancing condition holds for X , then also for all equivalent
complexes. This leads to the following definition.
Definition 4.2.2
A tropical k-cycle X in Rn is an equivalence class of balanced polyhedral
complexes in Rn of pure dimension k. A tropical subspace is a tropical cycle
that can be represented by a complex with only positive weights. A tropical
fan cycle is a tropical cycle X that can be represented by a fan X .
A reformulation emphasizing the underlying support set can be given as

follows. Let X ⊂ Rn be a polyhedral set. A point x ∈ X is called generic
if there exists a polyhedral structure for X such that x is contained in the
relative interior of a facet. Equivalently, there exists an affine subspace
A ⊂ Rn and a neighbourhood x ∈ U ⊂ Rn such that X ∩ U = A ∩ U. The set
of generic points is an open dense subset of X and is denoted by X gen . A
weighted polyhedral set is a polyhedral set X equipped with a locally constant
function ω : X gen → Z \ {0}, called weight function of X .
Any polyhedral structure X of X inherits a weight function for the facets
by setting ω(σ) = ω(x) for any x ∈ σ . Clearly, two polyhedral structures
equipped with the inherited weights are equivalent in the sense of definition
4.2.1. Vice versa, a balanced polyhedral complex X also induces a weight
function on |X |gen by setting ω(x) = ω(σ) for x ∈ σ and extending the
function to all of |X |gen (cf. Exercise 4.2.11). This leads to the following
equivalent reformulation of Definition 4.2.2.
103
Definition 4.2.3
A tropical k-cycle X (or balanced polyhedral subset) in Rn is a weighted poly-
hedral set in Rn of pure dimension k such that for any polyhedral structure
of X , equipped with the inherited weights, the balancing condition is satis-
fied. A tropical subspace is a tropical cycle with positive weight function. A
tropical fan cycle is a tropical cycle X supported on a fan set.
In the following we will take the freedom to take both viewpoints simul-
taneously. For example, will regard a tropical cycle X as a subset of Rn and
use the notation |X | only when we want to emphasize that we are forgetting
the weight function.
The set of all k-cycles of Rn is denoted by Zk (Rn ) and forms a group under
taking unions and adding weights. That is to say, to obtain X + Y we first
equip (X ∪ Y )gen ⊂ X gen ∪ Y gen with the weight function ωX (x) + ωY (x) (for
x ∈ (X ∪ Y )gen ⊂ X gen ∪ Y gen ), extending ωX and ωY by zero if necessary,
and secondly we restrict to the closure of the set of points with non-zero
weight. In particular, X + Y ⊂ X ∪ Y , but in general equality does not hold.
For a subset X ⊂ Rn , we denote by Zk (X ) groups of tropical k-cycles whose
support is contained in X .
Note that in tropical geometry the distinction between between algebraic
subvarieties as honest geometric objects and algebraic cycles as formal sums
of irreducible subvarieties does not really exist. In particular, our definition
of tropical subspaces as “effective” tropical cycles is somewhat arbitrary.
Rather, one should think of tropical cycles as a mixture between the two
concepts, carrying features of honest geometric subobjects (weights already
occur for hypersurfaces) as well as algebraic properties of cycles (such as
they form a group). For later use, let us prove the following lemma.
Example 4.2.4
Let A ⊂ Rn be a rational affine subspace of dimension k. We denote by [A]
the k-cycle supported on A with constant weight function 1.
Lemma 4.2.5
Any k-cycle X in Rn can be written as a difference X = X 1 − X 2 of tropical
k-subspaces X 1 , X 2 .
Proof. We set X 2 = σ:ω(σ)<0 −ω(σ)[Aσ ]. Here, the sum runs through all
P
facets of (a polyhedral structure of) X with negative weights and Aσ denotes
104
the affine space spanned by σ. By definition, X 2 and X + X 2 are tropical

subspaces, and hence the claim follows.
Example 4.2.6
Choose non-negative P integers k ≤ n. Let e1 , . . . , en denote the standard basis
n
of R and set e0 = − i=1 ei = (−1, . . . , −1). For any subset I ⊂ {0, . . . , n},
n
let C I denote the rational polyhedral cone in Rn spanned by the vectors

−ei , i ∈ I. We consider the polyhedral fan
Lk := {C I : I ⊂ {0, . . . , n}, |I| ≤ k}.
Equipped with trivial weights all equal to one, Lk is balanced. Indeed, fix
I ⊂ {0, . . . , n} with |I| = k − 1, the the corresponding codimension one cone
C := C I has adjacent facets C j := C I∪{ j} , j ∈ {0, . . . , n} \ I. We can use as
primitive generators
vC j /C := −e j .
Hence the balancing condition is equivalent to
X
−e j ∈ L(C) = 〈ei : i ∈ I〉R ,
j∈{0,...,n}\I
which is obvious since e0 + · · · + en = 0. The associated k-cycle L k is called

the standard tropical k-plane.
For later use, we close this section with a brief discussion of recession fans
in the one-dimensional case (see Section 6.6 for the general treatment).
Definition 4.2.7
Let C ∈ Z1 (Rn ) be a one-dimensional tropical cycle with polyhedral struc-
ture C . The recession fan of C is the fan RF(C ) = {rc(σ) : σ ∈ C } equipped
with the following weights. For a ray σ of RF(C ), let σ1 , . . . , σk be the rays
Pk
of C with rc(σi ) = σ. Then ω(σ) := i=1 ω(σi ). The recession fan cycle
RF(C) is the weighted polyhedral set represented by RF(C ).
It is clear that RF(C) is independent of the chosen polyhedral structure
C.
Lemma 4.2.8
The weighted fan RF(C ) is balanced. Hence RF(C) ∈ Z1 (Rn ) is a tropical fan
cycle.
105
Proof. We sum up the balancing conditions for all vertices of C . Every

bounded edge σ with endpoints p1 , p2 contributes twice, and the contri-
butions cancel since vσ/p1 = −vσ/p2 . They remaining terms are exactly the
terms in the balancing condition for RF(C ) at the origin.
Exercise 4.2.9
Let X be a weighted polyhedral complex and let Z be a weighted refine-
ment. Show that Z is balanced if and only if X is balanced
Exercise 4.2.10
Let X and Y be two weighted polyhedral complexes. Show that X and Y
are equivalent if and only if they have common weighted refinement.
Exercise 4.2.11
Let X be a polyhedral complex of dimension k. A cell τ ∈ X (k−1) is called
redundant if τ ⊂ |X |gen . Show that X the weight functions for |X | are in
bijection with the non-zero weight functions for X which are balanced for
all redundant (k − 1)-cells τ.
Exercise 4.2.12
Let L k be the standard k-plane in Rn defined in Example 4.2.6. Let C I be a
cone of Lk , the standard fan structure of L k . Show that StarLk (C I ) is again
(the fan structure of) a standard plane, namely the standard (k| I|)-plane
in Rn / L(C I ) (where the standard directions in Rn / L(C I ) are given by the
/ I).
images of e j , j ∈
4.3 Stable intersection

In this Section we treat the concept of stable intersection of tropical cycles
alluded to in Section 2.5. The adjective “stable” refers to the fact that this
type of intersection behaves continuously under small deformations of the
two cycles. In fact, we will use this property in order to define stable inter-
sections. The main feature of the construction is that it always produces as
output a specific cycle of the correct codimension, even in the case of say
self-intersections (without passing to equivalence relations such as rational
equivalence).
106
Given two cycles X and Y in Rn of pure dimension k resp. l, our goal is

to construct a cycle X · Y of pure dimension k + l − n which is supported on
X ∩ Y . Note that X ∩ Y is obviously a polyhedral set again, but might have
parts of dimension bigger than k + l − n. So our approach is as follows: First
we define X · Y in nice cases, namely when the intersection is transversal.
Then, in the general case, we translate one of the cycles slightly. For generic
translations, the resulting cycles are transversal and we can define the stable
intersection as the limit of the transversal intersections when moving the
translated cycles back to the original one.
4.3.1 Transversal intersections

Let us start with the definition of the transversal case. In the following, we
always assume that X ∩ Y is equipped with the polyhedral structure X ∩ Y
for two (sufficiently nice) polyhedral structures of X and Y , respectively
(see Exercise 2.1.3).
Definition 4.3.1 (Transversal intersection)

Let X and Y be two cycles in Rn of pure dimension k resp. l. We say X and
Y intersect transversally if X ∩ Y is of dimension k + l − n and if every facet τ
of X ∩ Y can be written uniquely as τ = σ ∩ σ0 with facets σ, σ0 of X resp.
Y (for suitable polyhedral structures X , Y ).
In this case, we define the transversal intersection X · Y to be the poly-
hedral complex X ∩ Y with weights
ω(τ) = ω(σ) · ω(σ0 ) · [Zn : LZ (σ) + LZ (σ0 )],
(where τ, σ, σ0 are as before). The transversal intersection X · Y is the

weighted polyhedral set represented by X · Y .
Proposition 4.3.2
In the transversal case, the transversal intersection X · Y is balanced and hence
forms a tropical (k + l − n)-cycle.
Proof. As X and Y intersect transversally, any codimension one cell of X ∩

Y lies in the codimension one skeleton of either X or Y . We may assume
the former, i.e. the codimension one cell can be written uniquely as τ ∩ σ0 ,
107
Figure 4.1: A tranversal and two non-transversal intersections
where τ is a codimension one cell of X and σ0 is a facet of Y . Let σ ⊃ τ

be a facet of X . We may compare primitive generators of σ ∩ σ0 (modulo
τ ∩ σ0 ) and σ (modulo τ). By definition we have
Zvσ∩σ0 /τ∩σ0 + LZ (τ) = (LZ (σ) ∩ LZ (σ0 )) + LZ (τ),

Zvσ/τ + LZ (τ) = LZ (σ),
and therefore
vσ∩σ0 /τ∩σ0 = [LZ (σ) : (LZ (σ) ∩ LZ (σ0 )) + LZ (τ)] · vσ/τ mod LZ (τ).
Plugging in the identity
[LZ (σ) : (LZ (σ) ∩ LZ (σ0 )) + LZ (τ)] = [LZ (σ) + LZ (σ0 ) : LZ (τ) + LZ (σ0 )].
and multiplying by [Zn : LZ (σ) + LZ (σ0 )] we get
[Zn : LZ (σ) + LZ (σ0 )] · vσ∩σ0 /τ∩σ0 = [Zn : LZ (τ) + LZ (σ0 )] · vσ/τ mod LZ (τ).
(4.1)
Now let σ1 , . . . , σm be the collection of facets containing τ (with primitive
vectors vi ). Then the facets of X ∩ Y containing τ ∩ σ0 are σi ∩ σ0 (with
primitive vectors w i ) and the above relation gives
m
X m
X
ω(σi ∩ σ )w i = ω(σ )
0 0
ω(σi )[Zn : LZ (σi ) + LZ (σ0 )]w i
i=1 i=1
m
X (4.2)
= ω(σ0 ) · [Zn : LZ (τ) + LZ (σ0 )] · ω(σi )vi = 0,
i=1
i.e. the balancing condition around τ implies the balancing condition around
τ ∩ σ0 .
108
v→0
Figure 4.2: A non-transversal intersection and two small deformations
4.3.2 The Moving Lemma

To extend the intersection of two cycles to the non-transversal case, we need
the following “moving lemma”. As usual, the degree of a zero-dimensional
P formal sum of points X = ωi pi , is the
P
cycle X , which is just a weighted
sum of all weights deg(X ) = ωi .
Proposition 4.3.3
Let X and Y be two cycles in Rn of pure dimensions k resp. l. Then the following
holds.
(a) For a generic vector v ∈ Rn , the intersection of X and the translation

Y + v is transversal.
(b) Assume n = k + l. Then deg(X · (Y + v)) is the same for all such generic
v ∈ Rn .
Remark 4.3.4
In the following, a set of generic vectors will always be the complement of a
polyhedral subset of dimension n − 1 (or even a union of affine subspaces).
Proof. Choose arbitrary polyhedral structures X and Y . For the first state-
ment, we consider all pairs of cells σ ∈ X , σ0 ∈ Y . If L(σ) + L(σ0 ) 6= Rn ,
then for any v ∈ Rn \ (L(σ) + L(σ0 )), we have L(σ) ∩ (L(σ0 ) + v) = ;. It
109
follows that σ ∩ (σ0 + v) = ; for all v not contained in a certain translation

of the linear subspace L(σ) + L(σ0 ). Hence for any v not contained in any
of these affine subpaces, all intersections σ ∩ (σ0 + v) are either empty or of
“expected dimension”, which implies that X and Y + v intersect transver-
sally.
For the second statement, let us first refine the previous consideration.
For any vector v ∈ Rn , the data of pairs of cells σ ∈ X , σ0 ∈ Y such that
σ ∩ (σ0 + v) 6= ; is called the intersection type of v. It is easy to check
that the set of vectors with given intersection type form (the interior of)
a polyhedron, and this subdivides Rn into a complete polyhedral complex.
X and Y + v intersect transversally if v is contained in the interior of a
maximal cell of this subdivision.
Now assume we are given two such generic vectors. If they are contained
in the same cell, the degree is obviously constant (as it is determined by X ,
Y and the intersection type). If not, we can connect the two vectors by
passing through at most codimension one cells of the subdivision of Rn . In
other words, it suffices to study intersection types where σ ∩ (σ0 + v) = ;
whenever the sum of the codimensions of σ and σ0 is greater than one
(otherwise, the intersection type is of higher codimension).
Therefore let τ be a codimension one cell, say of X , such that there is
a (unique) facet σ0 of Y with τ ∩ (σ0 + v) 6= ;. It is enough to show the
invariance of the degree locally at τ. That is to say, we may assume that
Y = {σ0 } is a linear space and X = {τ, σ1 , . . . , σm } is a fan with exactly
one codimension one cell τ (a linear space of dimension k − 1) and facets
σi . If dim(τ + σ0 ) < n − 1, for generic v we have X ∩ (Y + v) = ; and
hence deg(X · (Y + v)) = 0 is invariant. So we may assume that H = τ + σ0
is a hyperplane. Let vi be primitive generators for σi ⊃ τ. The balancing
condition states m
X
ω(σi )vi = 0 mod τ.
i=1
Choose a primitive generator w of Zn /(H ∩ Zn ) and write vi = λi w mod H

for (unique) λi ∈ Z. It follows
m
X
ω(σi )λi = 0. (4.3)
i=1
110
σi
λi > 0
λi < 0
σ0
Figure 4.3: Passing a codimension one case
On the other hand, using the shorthand
[H : τ + σ0 ] := [(H ∩ Zn ) : (τ ∩ Zn ) + (σ0 ∩ Zn )]
we may write
[Zn : LZ (σi ) + LZ (σ0 )] = |λi | · [H : τ + σ0 ].
Now, if we move Y towards w, (i.e. v is a translation vector which points

in the same direction as w modulo H), then Y + v intersects the facets σi
with λi > 0 and
X
deg(X · (Y + v)) = ω(σ0 )[H : τ + σ0 ] ω(σi )λi .
i
λi >0
Contrary, when moving Y towards −w, we must sum over i with λi < 0
instead. By equation (4.3) both sums are equal.
4.3.3 Stable intersections

We are now ready to define the stable intersection of two cycles in the gen-
eral case.
Definition 4.3.5
Let X and Y be two tropical cycles in Rn of pure dimension k resp. l. The
stable intersection X · Y is
X · Y := lim X · (Y + εv),
ε→0
111
where v ∈ Rn is a vector such that X and (Y + εv) intersect transversally for

small ε > 0.
More precisely, let X and Y be polyhedral structure for X and Y , re-

spectively. The stable intersection X · Y is weighted polyhedral complex
X ∩ Y (k+l−n) with weights
X
ω(τ) = ω(σ)ω(σ0 )[Zn : LZ (σ) + LZ (σ0 )], (4.4)
σ,σ0
where the sum runs through the pairs of facets σ ∈ X , σ0 ∈ Y with τ =

σ ∩ σ0 and σ ∩ (σ0 + εv) 6= ; for small ε > 0. The weighted polyhedral set
associated to X · Y is X · Y .
We have to show that X · Y does not depend on v and is balanced. It is
convenient to prove these facts together with a fact that stable intersections
can be computed locally. This is summarized in the following Proposition.
Proposition 4.3.6
Let X and Y be two cycles in Rn of pure dimension k resp. l with polyhedral
structures X and Y . Then the following holds.
(a) The stable intersection X · Y is well-defined, i.e. does not depend on the
choice of v ∈ Rn .
(b) The stable intersection can be computed locally. In formulas, we have
ω(τ) = deg(StarX (τ) · StarY (τ)),
for any k + l − n-cell τ of X ∩ Y and, more general, for any cell ρ
StarX ·Y (ρ) = StarX (ρ) · StarY (ρ).
(c) The stable intersection X · Y is balanced and therefore defines a tropical

(k + l − n)-cycle in Rn .
(d) Assume k + l = n. Then deg(X · Y ) = deg(X · (Y + v)) for all v ∈ Rn .
(e) Assume k + l = n + 1 and X , Y fan cycles. Then X · Y = RF(X · (Y + v))

for all v ∈ Rn .
112
Proof. We start by noting that the weight of τ in X · Y , computed with

respect to a generic v, is
ω(τ) = deg(StarX (τ) · (StarY (τ) + [v])).
This follows directly from Equation 4.4 and the observation that dividing by
L(τ) does not affect any of the lattice indices. With the help of our earlier
invariance result 4.3.3, this proves (a) and, using the definition of stable
intersection again, (b) and (d).
Similarly, to prove (e) it suffices to consider generic vectors v. We choose
a generic v and set C = X · (Y + v). In order to prove RF(C ) = X · Y ,
we note that by Exercise 2.1.1 we have rc(σ ∩ (σ0 + v)) = rc(σ ∩ σ0 ) for
all facets σ ∈ X , σ0 ∈ Y such that σ ∩ (σ0 + v) 6= ;. Hence for a given
ray τ ∈ X ∩ Y , the terms in the sum 4.4 correspond exactly to the rays
in C whose recession cone is equal to τ. It follows that τ ∈ RF(C ) if
ωX ·Y (τ) 6= 0 and moreover ωX ·Y (τ) = ωRF(C ) (τ). This proves (e).
It remains to prove (c). By (b), if suffices to consider the case where X
and Y are fans and k + l = n + 1. By (e), in this case we have X · Y =
RF(X ·(Y + v)) for all v. For generic v, we know that X ·(Y + v) is balanced
by Proposition 4.3.2. Hence X · Y is balanced by Lemma 4.2.8.
Stable intersection provides a map
· : Zk (Rn ) × Zl (Rn ) → Zk+l−n (Rn ).
Let us emphasize that this intersection product is constructed on the non-

compact space Rn and without passing to equivalence classes of any kind
— even in the case of non-tranversal (or self-)intersections. The product ·
turns Z∗ (Rn ) into a graded commutative R-algebra with unit 1 = Rn , as the
following proposition shows.
Proposition 4.3.7
Stable intersection is associative, commutative, bilinear and its neutral ele-
ment is given by the “fundamental” cycle [Rn ]. We have X · Y ⊆ X ∩ Y .
Proof. Commutativity, Bilinearity, the neutral element assertion and X · Y ⊆

X ∩ Y follow directly from the definitions.
It remains to show associativity. We first study the situation where all
of the involved intersections are transversal. In this case, each facet τ of
113
X ∩ Y ∩ Z can be written uniquely as τ = σ1 ∩ σ2 ∩ σ3 for facets of X , Y

and Z respectively. The weight of τ in (X · Y ) · Z is given by
ω(τ) =ω(σ1 )ω(σ2 )ω(σ3 )

· [Zn : LZ (σ1 ) + LZ (σ2 )] [Zn : (LZ (σ1 ) ∩ LZ (σ2 )) + LZ (σ3 )].
After plugging in the identity
[Zn : (LZ (σ1 ) ∩ LZ (σ2 )) + LZ (σ3 )]

= [Zn : LZ (σ2 ) + LZ (σ3 )] [LZ (σ2 ) : (LZ (σ1 ) ∩ LZ (σ2 )) + (LZ (σ2 ) ∩ LZ (σ3 ))],
we obtain an expression which is symmetric in σ1 and σ3 and thus is equal

to the weight in X ·(Y · Z). For the general case, we just have to unwind our
definition and observe that a triple stable intersection can be computed by
moving two of the cycles simultanously, i.e.
(X · Y ) · Z = lim(X · (Y + v)) · (Z + u).

v→0
u→0
Therefore the previous argument is sufficient.
Example 4.3.8
Let us consider the stable intersection of the standard planes L k defined in
Example 4.2.6. We claim
L k · L l = L n−k−l .
We set m = n − k − l. By definitions, the m-skeleton of L k ∩ L l is exactly the

support of L m , so we just have to show that all facets appear with weight 1
in L k · L l . To do this, we use locality and compute the weight of a facet by in-
tersecting the corresponding stars instead. We mentioned before that these
stars are standard planes again (cf. example 4.2.12), so in fact it suffices to
show that
L k · L n−k = {0}
with weight 1. This may be verified explicitly by choosing the translation
vector
v := (−1, . . . , −1, 1, . . . , 1),
| {z } | {z }
k times n−k times
114
Figure 4.4: Two planes in R3 which intersect in a line with a bounded edge
for example. Indeed, in this case, the only facets of L k and L n−k + v that
intersect are C{1,...,k} and C{k+1,...,n} + v. The intersection point
k
X n
X
ei = ei + v.
i=1 i=k+1
has weight 1, as LZ C{1,...,k} + LZ C{k+1,...,n} = Ze1 + · · · + Zen = Zn . So the claim

follows. In particular, if we denote by H := L n−1 the standard hyperplane in
Rn , then
L n−k = H k = H · · · H (k times).
Figure 4.4 show the intersection of two copies of H ⊂ R3 moved to transver-
sal position.
Exercise 4.3.9
Let X and Y be tropical cycles in Rn that intersect transversally. Show that
X · Y does not depend on the chosen polyhedral structures X and Y .
115
4.4 The divisor of a piecewise affine function

In Chapter 2 we discussed how a tropical polynomial gives rise to a tropical
hypersurface which, in turn, represents a tropical subspace of dimension
n − 1. In this section, we want to extend this construction to more general
functions on tropical cycles and describe its relationship to stable intersec-
tions. In the following, we use the same letter X for the support of a tropical
cycle X ⊂ Rn (forgetting the weights).
4.4.1 Piecewise integer affine, regular and rational

functions
Definition 4.4.1
Let X ⊂ Rn be a tropical cycle. A function f : X → R is a piecewise integer
affine function if there exists a polyhedral structure X for X such that for
any cell σ ∈ X , the restriction f |σ is the restriction of an integer affine func-
tion on Rn to σ. Any polyhedral structure satisfying this property is called
sufficiently fine for f . We denote the set of piecewise integer affine functions
on X by PAZ (X ). The function f is piecewise integer linear if furthermore X
can be chosen to be a fan and f (0) = 0.
Note that a piecewise integer affine function is always continuous. It is

clear that (PAZ (X ), +) forms an abelian group under ordinary addition, i.e.,
tropical multiplication. It is equally straightforward to check that PAZ (X )
is closed under tropical addition. In particular, after adding the constant
function −∞, we obtain a semigroup (PAZ (X ) ∪ {−∞}, max).
Example 4.4.2
Let f : Rn → R be a tropical polynomial. Then f is a piecewise integer
affine function. A sufficiently fine polyhedral structure is given by S ( f ).
Let g : Rn → R be a second polynomial. Then the tropical quotient
“ f /g” : x 7→ f (x) − g(x) is a piecewise integer affine function as well,
with sufficiently fine subdivision given by the intersection of the subdivi-
sions S ( f ) and S (g).
Note that our definition includes a finiteness condition since polyhedral
structures are assumed to be finite here. For example, a tropical Laurent
116
series in n variables is in general not included in our definition since its

domains of linearity may produce an infinite polyhedral structure in Rn .
We may turn these two basic examples into general definitions.
Definition 4.4.3
Let X ⊂ Rn be a tropical cycle. A polynomial function f : X → R is the re-
striction of a tropical polynomial Rn → R to f . A (tropical) regular function
f : X → R is a piecewise integer affine function which is locally polynomial.
We denote the set of regular functions by O (X ).
A (tropical) rational function f : X → R is the restriction of a tropical quo-
tient of two tropical polynomials “g/h” to X . We denote the set of rational
functions by R(X ). A locally rational function f : X → R is a piecewise
integer affine function which is locally rational.
Note that the definitions, even though imposing a local condition, still
contain the finiteness condition inherited from our definition of piecewise
integer affine function. In particular, with the given definition neither reg-
ular functions nor locally rational functions form a sheaf on X .
4.4.2 The divisor construction

Definition 4.4.4
Let f : X → R be a piecewise integer affine function on a tropical cycle X ∈
Zm (Rn ) and let X be sufficiently fine polyhedral structure. The divisor of f
on X is the weighted polyhedral complex divX ( f ) (or f · X ) constructed
as follows.
• The underlying polyhedral complex is X m−1 .
• For a facet τ ∈ X m−1 , the weight of τ in divX ( f ) is given by

k k
X X
ω(τ) := ω(σi )d f |σi (vσi /τ ) − d f |τ ω(σi )vσi /τ . (4.5)
i=1 i=1
Here, σ1 , . . . , σk denote the facets of X containing τ, vσi /τ is a choice

of primitive generator for facet, and d f |σi denotes the differential of
the (affine) function f |σi .
117
Remark 4.4.5
A few remarks are in order.
• First note that by the balancing condition for X around τ, the sum
Pk
i=1
ω(σi )vσi /τ is a vector in L(τ) and its value under d f |τ is well-
defined. Second, let us check that the definition of weights is inde-
pendent of the chosen primitive generators vσi /τ . Indeed, for an al-
ternative choice vσ0 /τ we have w i := vσ0 /τ − vσi /τ ∈ L(τ). It follows
i i
d f |σi (vσ0 i /τ ) = d f |σi (vσi /τ ) + d f |τ (w i ).

Pk
Hence we get a correction term ω(σi )d f |τ (w i ) in the first sum
i=1 Pk
which cancels, by linearity of d f |τ , with the correction term d f |τ ( i=1 ω(σi )w i ).
Pk
• We can always choose primitive vectors such that i=1 ω(σi )vσi /τ =
0. In this case the weight formula (4.5) simplifies to
k
X
ω(τ) := ω(σi )d f |σi (vσi /τ ).
i=1
To verify that such a choice of primitive vectors is possible, we set

Pk
g := gcd{ω(σi )} and write g = i=1 ai ω(σi ), ai ∈ Z. Then for any
choice of primitive vectors we have
k
1X
w := ω(σi )vσ0 i /τ ∈ PAZ (τ).
g i=1
Setting vσi /τ := vσ0 /τ − ai w we obtain a collection of primitive vectors

i
whose sum up to zero.
• Note that the construction may produce zero weights ω(τ) = 0. In

this case, the (weighted) support | divX ( f )| is strictly contained in
|X m−1 |. In particular, ω(τ) = 0 if f can be written as the restriction
of an integer affine function g in a neighbourhood of τ , since then
d f |σi = d g for all i and the two terms cancel by linearity of d g (as
above).
118
Example 4.4.6
Let f : Rn → R be a tropical polynomial with associated subdivision S ( f ).
Then divS ( f ) ( f ) = V ( f ). Indeed, let P be an (n − 1)-cell of S ( f ) and let
h1 , h2 the integer linear functions associated to the endpoints of the dual
cell σ P . Let h be the integer linear form such that h2 − h1 = Vol(σ P ) · h
and let F1 , F2 be the facets of S ( f ) on which h1 , h2 attain the maximum,
respectively. The primitive generators v of F2 modulo P are characterized
by the property (v) = 1. Hence
ω(τ) = h2 (v) + h1 (−v) = Vol(σ P )h(v) = Vol(σ P ).
Proposition 4.4.7
The weighted polyhedral complex divX ( f ) is balanced.
Proof. We need to check the balancing condition for each m−2-cell ρ ∈ X .
Let fρ : Rn → R be an integer affine function restricting to f |ρ on ρ. Then
the function f − fρ descends to a piecewise integer linear function fe on
S := StarX (ρ), a polyhedral fan in Rn / L(ρ). By the previous remark, the
weights of divX ( f ) are invariant under adding a (globally) integer affine
function. Hence, the balancing condition at ρ is equivalent to the balancing
condition of divS ( fe) at the origin. In other words, we may restrict to the
case where m = 2 and X is a balanced polyhedral fan (with 0-cell {0}).
We may additionally assume that X is unimodular, that is to say, each 2-
cone σ ∈ X is generated by two integer vectors which form a lattice basis
of LZ (σ). To see this, recall that any X admits a unimodular refinement
and note that refining X does not affect the balancing condition. Indeed,
any newly added ray gets assigned weight 0 by the previous remark. So let
us assume that X is unimodular.
Let τ1 , . . . , τl denote the rays of X , and let v1 , . . . , vl be their primitive
generators. If a 2-cone σ is spanned by τi and τ j , then unimodularity
implies that we can use v j as a primitive generator for σ modulo τi (and
vice versa). We use the following notation. For i 6= j ∈ {1, . . . , l}, we set
ωi j = ω(σi j ) if τi and τ j span a 2-cone σi j of X , and ωi j = 0 otherwise.
Note that ωi j = ω ji . The balancing condition of X implies that there is an
integer αi ∈ Z such that
l
X
ωi j v j = αi vi ∈ L(τi ). (4.6)
j=1
119
Then the weight of τi in divX ( f ) can be computed as

l
X
ω(τi ) = ωi j f (v j ) − αi f (vi ).
j=1
Finally, the balancing condition of divX ( f ) at the origin can be checked by

computing
l
l l
X X X
ω(τi )vi = ωi j f (v j )vi − αi f (vi )vi
i=1 i, j=1 i=1
l
l
X X
= ωi j f (vi )v j − αi f (vi )vi
i, j=1 i=1
l
l

X X
= f (vi ) ωi j v j − αi vi .
i=1 j=1
But the term in big brackets in the last line is zero by Equation (4.6).
Definition 4.4.8
Let f : X → R be a piecewise integer affine function on a tropical cycle
X ∈ Zm (Rn ) and let X be sufficiently fine polyhedral structure. The divisor
of f is the tropical cycle of dimension m − 1 defined by divX ( f ). We denote
it by divX ( f ) or f · X .
Remark 4.4.9
By the previous proposition, divX ( f ) is balanced and hence defines a tropi-
cal cycle. Note that this tropical cycle does not depend on the chosen poly-
hedral structure X . Indeed, let X 0 be a refinement of X and let τ be m−1-
cell of X 0 which is not contained in any m − 1-cell of X . Then f is integer
affine in a neighbourhood of τ and therefore the weight of τ in divX 0 ( f )
is zero. For all other m − 1-cells of X , obviously the weight agrees with the
weight of the (unique) m − 1-cell of X containing τ. Hence divX 0 ( f ) and
divX ( f ) are equivalent as weighted polyhedral complexes and induce the
same tropical cycle.
Example 4.4.10
Let f : Rn → R be a tropical polynomial. Then div( f ) = V ( f ).
120
Example 4.4.11
Let X = L2 = V (“0 + x + y + z”) ⊂ R3 be the standard plane in R3 from
Example 4.2.6. Note that X contains the (honest) line L = R(e0 + e3 ) =
R(e1 + e2 ) as a subcycle. We want to describe a function f : X → R such
that f · X = L.
We start by subdividing the cones σ{0,3} and σ{1,2} by intersecting rays in
the direction −(e0 +e3 ) and −(e1 +e2 ), respectively. Note that the refined fan
X is (still) unimodular. In particular, there exists a unique function g which
is linear on the cones of X and maps the primitive generators according to
−ei 7→ 0 (i = 0, 1, 2, 3), −(e0 + e3 ), −(e1 + e2 ) 7→ −1.
The divisor of this function carries the weights

ω(R≤0 e0 ) = g(−e1 ) + g(−e2 ) + g(−(e0 + e3 )) − g(0) = −1,
ω(R≤0 (e0 + e3 )) = g(−e0 ) + g(−e3 ) − g(−(e0 + e3 )) = 1,
and symmetrically for all other rays. In other words, g · X = L − L1 , where
L1 denotes the standard line in R3 . But h · X = L1 for h = “0 + x + y + z”
(see Examples 4.3.8 and 4.4.20, or just calculate it directly). Therefore the
function f = g + h satisfies f · X = L (see Remark 4.4.15).
We may generalize Theorem 2.4.10 to arbitrary tropical n − 1-cycles al-
lowing negative weights.
Corollary 4.4.12
Let X ⊂ Rn be a tropical n − 1-cycle. Then there exists a tropical rational func-
tion f : Rn → R such that X = div( f ). The function f is unique up to adding
an integer affine function (among all piecewise integer affine functions). In
particular, any piecewise integer affine function on Rn is rational.
Proof. By Lemma 4.2.5, we can write X = X 1 − X 2 with X 1 , X 2 tropical n−1-
subspaces. By Theorem 2.4.10 there exist tropical polynomials f1 , f2 such
that X i = V ( f i ) = div( f i ), i = 1, 2. Hence the rational function f = f1 −
f2 satisfies V ( f ) = X . To show uniqueness and the last claim, let f be
a piecewise affine function such that V ( f ) = 0 (as a cycle). Let X be
a sufficiently fine subdivision of Rn . Then for any n − 1-cell τ ∈ X the
condition ωdiv( f ) (τ) = 0 implies that the two affine functions on both sides
of τ agree. Hence f is a globally integer affine function.
121
Corollary 4.4.13
Any piecewise integer affine function f : X → R is locally rational.
Proof. Since this is a local statement, we may restrict to the case where f
is piecewise linear function on a fan cycle X . Let X be a sufficiently fine
polyhedral fan structure for f on X . By refining X further if necessary, we
may assume that X ⊂ F where F is unimodular complete fan of Rn . Then
by choosing arbitrary integer values on the primitive generators of rays not
contained in X and extending by linearity, we may construct an extension
fe : Rn → R of f which is integer linear on cones of F . Then by Corollary
4.4.12, fe is a rational function, and the claim follows.
Exercise 4.4.14
y
X = R≥0 (−1, 2) ∪ R≥0 (−1, −2) ∪ 2 · R≥0 (1, 0), f : X → R, (x, y) 7→ 2 . Show
that f is piecewise integer linear on X and div( f ) = 0. Show that f is not
regular. In particular, it is not true that any locally convex piecewise integer
affine function is regular.
The statement holds, however, if we allow Laurent polynomials with ra-
tional exponents: Show that every locally convex piecewise integer affine
function f : X → R is locally the restriction of a function x 7→ max j∈A{a j +
j x} with A ∈ Qn finite and a j ∈ R (for arbitrary X ).
4.4.3 Further properties of the divisor construction

Remark 4.4.15
For further reference, let us collect some of the properties of the divisor
construction (most of which appeared in the previous proofs). Let f : X →
R be a piecewise integer affine function on a tropical cycle X ∈ Zm (Rn ). We
say f is locally affine at x ∈ X if there is a neigbourhood x ∈ U ⊂ X and a
affine function F : Rn → R such that f |U = F |U . We denote by V ( f ) ⊂ X
the set of points where f is not locally affine. The following holds.
(a) The support of div( f ) is contained in V ( f ). In particular, if f is the

restriction of an integer affine function, then div( f ) = 0.
(b) Let X be a sufficiently fine polyhedral structure for f and let τ be

a cell of X (m−1) . Then d f induces piecewise integer linear function
122
f τ : StarX (τ) → R which is unique up to adding an integer affine

function. Moreover, we have
f τ · StarX (τ) = Star f ·X (τ).
In other words, f · X can be computed locally.
(c) By the linearity property of Equation (4.5) it follows that the map
PAZ (X ) × Zk (X ) → Zk−1 (X ),
( f , Z) 7→ f · Z := f | Z · Z
is linear in both arguments.
Recall that a tropical subspace is a tropical cycle with only positive weights.
Proposition 4.4.16
Let f : X → R be a regular function on a tropical subspace X ∈ Zm (Rn ). Then
f · X is a tropical subspace and its support is equal to V ( f ).
In fact, the statement can be generalized to locally convex functions. A

function f is called locally convex at x if it is equal to the restriction of a
convex function on Rn in a neighbourhood of x. Clearly, regular functions
are locally convex functions.
Proof. By locality we can restrict to the case where X is a one-dimensional

fan and f is linear on each ray. Let ρ1 , . . . , ρl be the rays with with weights
ωi > 0 and primitive generators vi . Let ω denote the weight of the origin
in div( f ). Then
l
X Xl
ω= ωi f (vi ) ≥ f ( ωi vi ) = 0,
i=1 i=1
where the inequality follows from the local convexity of f . This proves the
first claim.
For the second claim, it remains to show that if ω = 0 then f is locally
Pl
affine. Assumimg the contrary, there exists a linear combination i=1 λi vi =
Pl Pl
0, λi ∈ Z, such that i=1 λi f (vi ) 6= 0. We may assume i=1 λi f (vi ) < 0
by replacing λi with −λi , if necessary. Moreover, replacing λi by λi + mωi
123
for sufficiently large m ∈ N we may assume λi > 0. Hence the λi define

another tropical subspace X 0 (with |X 0 | ⊂ |X |) such that f · X 0 carries a
negative weight the origin. This is a contradiction to the first claim of the
proposition, and therefore proves the second claim.
Example 4.4.17
Let X = L2 ⊂ R3 be the standard plane and L ∈ X the (honest) line from Ex-
ample 4.4.11. Note that L is rigid in X : No non-trivial translation of L lies
in X . Philosophically, we might hope to detect this by computing a negative
self-intersection of L in X . Since in Example 4.4.11 we constructed a func-
tion f : X → R such that f · X = L, we can interpret this self-intersection
as f 2 · X . By defnition, this 0-cycle is supported {0} and the weight of the
origin is given by
f (−(e0 + e3 )) + f (−(e1 + e2 )) = 0 − 1 = −1.
So indeed, the self-intersection f 2 · X is negative. In particular, f is not a
regular (or locally convex) function.
Proposition 4.4.18
Let f , g : X → R be a two piecewise integer affine functions on a tropical cycle
X ∈ Zm (Rn ). Then
f · (g · X ) = g · ( f · X ).
Proof. The proof is very similar to the proof of 4.4.7. Again by locality, we
may restrict to the case of a two-dimensional unimodular fan X and f , g
piecewise integer linear functions on X . Let ω and ω0 be the weight of the
origin in f · (g · X ) and g · ( f · X ), respectively. Let τ1 , . . . , τl be the rays of
X and let ωi and ω0i be the weights of ρi in g · X and f · X , respectively.
Then
l
l l
X X X
ω= ωi f (vi ) = ωi j g(v j ) f (vi ) − αi g(vi ) f (vi )
i=1 i, j=1 i=1
l
l
X X
= ωi j f (v j )g(vi ) − αi f (vi )g(vi )
i, j=1 i=1
l
X
= ω0i g(vi ) = ω0 ,
i=1
124
where ωi j , vi , αi are defined as before in 4.4.7.
4.4.4 Compatibility with stable intersection

Our next goal is to show that the divisor construction is compatible with
stable intersection. More precisely, we want to prove the following theorem.
Theorem 4.4.19
Let X and Y be tropical cycles in Rn . Let f : X → R be a piecewise integer
affine function on X . Then
f · (X · Y ) = ( f · X ) · Y.
Example 4.4.20
Let X ⊂ Rn be a tropical cycle and let f : Rn → R be a tropical polynomial.
Then
f · X = V(f ) · X.
Let f1 , . . . , f k be a collection of tropical polynomials. Then
f1 · · · f k · Rn = V ( f1 ) · · · V ( f k ).
We need the following lemma.
Lemma 4.4.21
Let X be a tropical fan cycle, f : X → R a piecewise integer linear function on
X and C ⊂ X a one-dimensional tropical cycle. Then
deg( f · C) = deg( f · RF(C)).

Pl
Proof. First note that deg( f · C) = i=1 ω(σi )d f |σi (vi ), where σ1 , . . . , σl
are the rays of C (for some polyhedral structure) and vi is a primitive gen-
erator for σi . As in the proof of Lemma 4.2.8, this follows easily from
the fact that each bounded edge of C produces two contributions to the
degree which cancel each other. Since X is a fan and f is linear on a poly-
hedral fan structure for X , we have f (vi ). Since the vi are also primitive
generators for rc(σi ), we see that deg( f · C) is equal to deg( f · RF(C)) =
Pl
i=1
ω(σi ) f (vi ).
125
Proof of Theorem 4.4.19. By locality of both sides, it suffices to restrict to

the case where X , Y are tropical fan cycles of dimension k, l, respectively,
k + l = n + 1, and f is piecewise linear on X . Let Y + v be a generic
translation of Y which is transverse both to X and f · X . Then by Lemma
4.4.21 and Proposition 4.3.6 (d) and (e), it suffices to show f ·(X ·(Y +v)) =
( f · X ) · (Y + v), i.e., we may restrict to the transversal case.
So let us assume that Y = {L} is an ordinary linear space of dimension l
and that X is a polyhedral fan of dimension k with a single (k − 1)-cell Q,
linear space transversal to L, and facets σ1 , . . . , σl . We denote by vi and w i
the primitive generators of σi modulo Q and σi ∩ L modulo Q ∩ L = {0},
Pl
respectively. We may assume i=1 ω(σi )vi = 0 for simplicity. With the
notations ai := [Zn : LZ (σi ) + (L ∩ Zn )] and a = [Zn : (Q ∩ Zn ) + (L ∩ Zn )],
we get
ai w i = avi mod Q
for all i (see Equation 4.1). Finally, let ω and ω0 denote the weight of the
origin in f · (X · Y ) and ( f · X ) · Y , respectively. Then
l
X l
X
ω= ωX ·Y (σi ∩ L) f (w i ) = ωY (L) ωX (σi )ai f (w i )
i=1 i=1
l
X
= ωY (L) · a · ω(σi ) f (vi ) = ωY (L) · a · ω f ·X (Q) = ω0 .
i=1
Exercise 4.4.22
Let X be a tropical 1-cycle in Rn and f : X → R be a piecewise integer affine
function. Let X be a sufficiently fine (pointed) polyhedral structure and
let ρ1 , . . . , ρl denote the rays (i.e. unbounded edges) of X , with primitive
generators v1 , . . . , vl . Then
l
X
deg( f · X ) = ω(ρi )d f |ρi (vi ).
i=1
126
4.5 Tropical modifications

In classical algebraic geometry, it is a basic feature of any regular function
f : X → K that its graph Γ (X , f ) ⊂ X × K is Zariski closed. In Remark 2.4.9
we constructed a balanced completion of the graph of a tropical polynomial
f which served as a motiviation for the definition of V ( f ). We will extend
this graph completion to arbitrary piecewise integer affine functions now.
Beyond giving some justification of the divisor construction, these graph
completions are also of interest in their own right and are called tropical
modifications.
The graph of a function f : X × R is denoted by Γ (X , f ) or just Γ ( f ).
If X ⊂ Rn is a polyhedral set and f is a piecewise integer affine function,
then Γ (X , f ) ⊂ Rn × R is a polyhedral set. Let π : Rn × R → Rn denote
the projection. If X is weighted, we equip Γ (X , f ) with weights by setting
ω(x̃) = ω(π(x̃)) for all x̃ ∈ Γ (X , f )gen . Note that the weighted polyhedral
set Γ (X , f ) is in general not balanced. However, in a sense to be made
precise, there is a unique balanced completion Γ (X , f ). Let us discuss this
in more details.
Let X be a polyhedral structure of X sufficiently fine for f . For σ ∈ X ,
we denote by σ e := Γ (σ, f |σ ) the lift of σ under f . The polyhedra σ e form a
polyhedral structure X for Γ (X , f ). Let τ be an k−1-cell of X , let σ1 , . . . , σl
f
be the facets of X adjacent to τ and let v1 , . . . , vl be corresponding primitive
Pl
generators in Zn . For simplicity, we choose them such that i=1 ω(σi )vi =
0. Then the facets around τ e in Xf are σ e 1, . . . , σ
e l , and as primitive generators
we can use the lifts
vi := (vi , d f |σi (vi )) ∈ Zn × Z.

e
The weighted sum of primitive generators around τ

e gives
l
X l
X
ω(e
σi )e
vi = ω(σi )(vi , d f |σi (vi ))
i=1 i=1
l
X (4.7)
= (0, ω(σi )d f |σi (vi ))
i=1
= (0, ω(τ)) ∈ Zn × Z,
127
Figure 4.5: The modification of R2 along “(−1)x 2 +1x y +(−1) y 2 + x + y +0”
where ω(τ) denotes the weight of τ in divX ( f ) defined in Equation (4.5).

Hence a natural way to restore the balancing condition around τ e is to add
another facet
τ≤ = τ
e + ({0} × R≤0 ) = {(x, y) ∈ Rn × R : x ∈ τ, y ≤ f (x)}
equipped with weight ω(τ≤ ) = ω(τ) (see Figure 4.5). Let us summarize
the discussion in the following definition.
Definition 4.5.1
Let f : X → R be a piecewise integer affine function on tropical k-cycle and
let X be a sufficiently fine polyhedral structure. The (open) modification of
X along f is the weighted polyhedral complex
Mod(X , f ) := {e
σ : σ ∈ X } ∪ {τ≤ : τ ∈ divX ( f )},
where σe = Γ (σ, f |σ ) and τ≤ = τ

e + ({0} × R≤0 ). The weights are given by
ω(eσ) = ωX (σ) (for facets σ ∈ X ) and ω(τ≤ ) = ωdivX ( f ) (τ) (for facets
τ ∈ divX ( f )).
Lemma 4.5.2
The weighted polyhedral complex Mod(X , f ) in Rn × R is balanced.
Proof. The k − 1-cells of Mod(X , f ) are of two types. For τ

e, τ an k − 1-cell
of X , the balancing condition follows from the computation in Equation
(4.7). Note that (0, −1) ∈ Zn × Z is a primitive generator for τ≤ modulo τ e.
128
The second type of codimension cells is of the form ρ≤ for a k − 2-cell

ρ ∈ X . In this case, obviously the equality
StarMod(X , f ) (ρ≤ ) = StardivX ( f ) (ρ)
holds. Hence the balancing condition around ρ≤ is equivalent to the bal-

ancing condition around ρ in divX ( f ), which was proven in Proposition
4.4.7.
Definition 4.5.3
Let f : X → R be a piecewise integer affine function on tropical k-cycle
X . The (open) modification Mod(X , f ) of X along f is the tropical cycle in
Rn × R defined by Mod(X , f ) for any sufficiently fine polyhedral structure
X . The map π : Mod(X , f ) → X is the contraction of the modification.
Remark 4.5.4
Let us make a few remarks here.
• It follows from Equation (4.7) that Mod(X , f ) is the unique balanced

completion of the graph Γ (X , f ) under the condition that we only al-
low adding cells in the “downward” direction (more precisely, only
polyhedra P such that (0, −1) ∈ rc(P)).
• Let us, for the moment, consider the closure Mod(X , f ) of Mod(X , f )
in Rn × T. Let us denote by H−∞ = Rn × {−∞} ∼ = Rn the hyperplane
added at infinity. Then by construction the intersection Mod(X , f ) ∩
H−∞ is equal to divX ( f ), and this equality is respecting the weights
(in an obvious sense). See Figure 4.6 for an example. Hence, in
agreement with Remark 2.4.9, the modification construction justifies
to some extent our definition of div( f ) as locus of (tropical) zeros
and poles of f . Here, poles are actually described in terms of zeros
with negative weight. It might be tempting to use facets in upward
direction (0, 1) instead of negative weights. Note, however, that for a
general piecewise integer affine function there is no globally consis-
tent construction which does so.
• Let X ⊂ (C× )n be a classical very affine variety and f : X → C a regular

(or rational) function. Then taking the very affine graph Γ (X , f ) ∩
129
Mod(X , f )
div( f )
X
Figure 4.6: The modification of R2 along the standard line is the standard
plane.
(C× )n+1 is isomorphic X after removing the zero locus V ( f ) (or the
divisor of zeros and poles of f , respectively). Hence, the process of
tropical open modifications can be regarded as the result of a very
affine embedding of a principally open subset of X .
Assume that f = “g/h” : X → R is a rational function for two polynomial
functions g, h on X . In this case there exists an alternative version of the
modification along f which does distinguish between zeros and poles, in
contrast to the previous remark.
Definition 4.5.5
Let g, h : X → R be two polynomial functions on a tropical subspace X . Let
X be a polyhedral structure sufficiently fine for both g and h. The (open)
modification of X along the quotient of g by h is the weighted polyhedral
complex
σ : σ ∈ X } ∪ {τ≤ : τ ∈ divX (g)} ∪ {τ≥ : τ ∈ divX (h)},
Mod(X , g, h) := {e
where σ e = Γ (σ, (g−h)|σ ), τ≤ = τe +({0}×R≤0 ) and τ≥ = τe +({0}×R≤0 ) The
weights are given by ω(e σ) = ωX (σ), ω(τ≤ ) = ωdivX (g) (τ) and ω(τ≥ ) =
ωdivX (h) (τ), respectively. The (open) modification Mod(X , g, h) of X along
the quotient of g by h is the tropical cycle defined by Mod(X , g, h).
130
Remark 4.5.6
Again, a few remarks.
• It is straightforward to adapt the proof of Lemma 4.5.2 to show that
Mod(X , g, h) is balanced and hence Mod(X , g, h) is well-defined as
a tropical cycle.
• The construction does depend on the choice of g, h and not only on

“g/h”. Also, in general no canonical minimal choice for g, h given
“g/h” might be available.
• Let us identify the open torus in TP1 with R by mapping (x : y) 7→

x − y. Then the map g − h : X → R can be identified with map (g, h) :
X → TP1 . This alternative description shows that Mod(X , g, h) can
be regarded as a tropical blow-up of X at V (g) ∩ V (h).
• We have met the construction of Mod(X , g, h) before in our discus-

sion of floor decompositions. Let us recall the notations from Propo-
Pm 2.6.i Let V ( f ) be a floor-decomposed hypersurface in R , f =
n
sition
“ i=0 f i x n ”, 0 6= f i ∈ T[x 1 , . . . , x n−1 ]. Let F1 , . . . , Fm and E0 , . . . , Em
denote the corresponding floors and elevators, respectively. Then for
all m = 1, . . . , m we have
Ei−1 , Fi , Ei ⊂ Mod(Rn , f i−1 , f i )
by Proposition 2.6.2. In other words, Mod(Rn , f i−1 , f i ) is equal to the
floor Fi with elevators Ei−1 and Ei attached and extended upwards
and downwards to infinity. The operation of “gluing together” the
various modifications Mod(Rn , f i−1 , f i ) to obtain V ( f ) is analogous to
the symplectic sum operation in symplectic geometry.
Exercise 4.5.7
Show that both Mod(X , F ) and Mod(X , g, h) are independent of the chosen
(sufficiently fine) polyhedral structure X .
4.6 The projection formula

The goal of this section is to establish a simple type of projection formula
which relates divisors of pull-backs of piecewise integer functions with push
131
forwards of tropical cycles. We start by introducing the necessary terminol-

ogy.
4.6.1 Integer affine maps

Recall that an integer affine map Φ : Rn → Rm is a map of the form x 7→
Ax + b where A ∈ Mat(m × n, Z) and b ∈ Rm . Let X ⊂ Rn and Y ⊂ Rm be
subsets. A map ϕ : X → Y is integer affine if it is the restriction of an integer
affine map Φ : Rn → Rm .
Lemma 4.6.1
Let X ⊂ Rn and Y ⊂ Rm be polyhedral sets and ϕ : X → Y be an integer affine
map. Then there exist polyhedral structures X of X and Y of Y such that
ϕ(σ) ∈ Y for all σ ∈ X .
Proof. We choose polyhedral structures X 0 of X and Y 0 of Y and Φ : Rn →

Rm an integer affine map with Φ|X = ϕ. Since ϕ(σ0 ) ⊂ Rm is a polyhedron
for all σ0 ∈ X 0 , we can write |Y | as union of polyhedra
[ [
|Y | = τ0 ∪ ϕ(σ0 ). (4.8)
τ0 ∈Y 0 σ0 ∈X 0
Applying Proposition 4.1.1 (a) to this union, we obtain a polyhedral subdi-

vision
S = Gκ1 ∩ . . . ∩ Gκl ,
of Rm such that the polyhedra in Equation (4.8) are unions of cells in S .
The pull back subdivision of Rn is
T = Gκ1 ◦Φ ∩ . . . ∩ Gκl ◦Φ
A simple computation shows that Y = S ∩ Y 0 = S |Y and X = T ∩ X 0

fulfil the required condition.
IfX and Y satisfy the conditions of the lemma, they are called sufficiently
fine for ϕ. We denote by ϕ∗ (X ) the part of the subcomplex {ϕ(σ) : σ ∈
X } ⊂ Y of pure dimension dim(X ). Note that the facets of ϕ∗ (X ) are of
the form ϕ(σ) for a facet σ ∈ X for which ϕ|σ is injective.
132
4.6.2 Push forward of tropical cycles

Definition 4.6.2
Let ϕ : X → Y be an integer affine map between tropical cycles and let X
and Y be sufficiently fine polyhedral structures. The push forward of X
along ϕ is the polyhedral complex ϕ∗ (X ) whose facets σ0 carry the weight
X
ωϕ∗ (X ) (σ0 ) = ωX (σ) · [LZ (σ0 ) : dϕ|σ (LZ (σ)]. (4.9)
σ∈X
ϕ(σ)=σ0
The push forward ϕ∗ (X ) is the tropical cycle defined by ϕ∗ (X ).

Note that since σ0 is a facet of ϕ∗ (X ) and ϕ(σ) = σ0 , it follows that
ϕ|σ is injective and hence [LZ (σ0 ) : dϕ|σ (LZ (σ)] ∈ Z is finite. To show that
ϕ∗ (X ) defines a tropical cycle, we need to prove the following lemma.
Lemma 4.6.3
The weighted polyhedral complex ϕ∗ (X ) is balanced.
Proof. By Exercise 4.6.10 it suffices to prove the local case. More precisely,
we can restrict to the case where X is a one-dimensional fan. Let ρ1 , . . . , ρl
denote the rays of X with primitive generators v1 , . . . , vl . We may assume
ϕ(vi ) = 0 if and only if i = k + 1, . . . , l for some 0 ≤ k ≤ l. Then the rays
of ϕ∗ (X ) are spanned by the vectors ϕ(vi ), i = 1, . . . , k, and ϕ(vi ) is equal
to [LZ (ϕ(ρi )) : ϕ(LZ (ρi )] · w i , where w i is primitive. Hence the balancing
condition for ϕ∗ (X ) turns into
k
X k
X
ω(ϕ(ρi ))w i = ω(ρi )ϕ(vi )
i=1 i=1
l
l

X X
= ω(ρi )ϕ(vi ) = ϕ ω(ρi )vi = 0,
i=1 i=1
which finishes the proof.
Remark 4.6.4
Let us make a few remarks.
• Note that the support of a push forward satisfies |ϕ∗ (X )| ⊂ ϕ(|X |) ⊂
|Y |.
133
• If ϕ(|X |) is a polyhedral set of dimension strictly lower than dim(X ),

then ϕ∗ (X ) = 0.
• An integer affine map ϕ : X → Y induces a group homomorphism

ϕ∗ : Zk (X ) → Zk (Y ) given by Z 7→ ϕ∗ (Z) := (ϕ| Z )∗ (Z). This follows
directly from the linearity of the weight formula 4.9.
• Push forwards are functorial. This is a straightforward computation

using the multiplicativity of indices
[Zn : AB(Zn )] = [Zn : A(Zn )] · [Zn : B(Zn )]
for two (injective) integer linear maps A, B : Rn → Rn (or, equivalently,

| det(AB)| = | det(A)| · | det(B)|).
Example 4.6.5
Let us consider the standard line X = L1 = V (“0+ x + y”) ⊂ R2 and the map
ϕ : R2 → R, (x, y) 7→ x+ y. The standard structure on L1 and the polyhedral
structure {R≥0 , R≤0 , {0}} are sufficiently fine for ϕ|X . Let ρ0 , ρ1 , ρ2 be the
rays of X spanned by (1, 1), (−1, 0), (0 − 1). The generators are mapped
under ϕ to
(−1, 0), (0, −1) 7→ −1 (1, 1) 7→ 2.
Therefore the weights of ϕ∗ (X ) are given by
ω(R≤0 ) = [Z : ϕ(L Z (ρ1 ))] + [Z : ϕ(L Z (ρ1 ))] = 1 + 1 = 2,

ω(R≥0 ) = [Z : ϕ(L Z (ρ0 ))] = 2,
and so ϕ∗ (X ) = 2[R].
4.6.3 The projection formula

Definition 4.6.6
Let ϕ : X → Y be an integer affine map between tropical cycles and let
f : Y → R be a piecewise integer affine function on Y . The pull back of f
along ϕ is the piecewise integer affine function ϕ ∗ f = f ◦ ϕ on X .
134
Remark 4.6.7
In order to construct a sufficiently fine polyhedral structure for ϕ ∗ f , it suf-
fices to start with a sufficiently fine structure Y 0 for f and then apply the
construction of Lemma 4.6.1.
We are now ready to state the main the projection formula.
Theorem 4.6.8 (Projection formula)

Let ϕ : X → Y be a tropical morphism and let f : Y → R be a piecewise integer
affine function on Y . Then the following formula holds:
f · ϕ∗ (X ) = ϕ∗ (ϕ ∗ ( f ) · X ) ∈ Zdim X −1 (Y )
Proof. We choose polyhedral structures X and Y which are sufficiently fine

for ϕ, f and hence also ϕ ∗ ( f ) (see Remark 4.6.7). Set k = dim(X ). We
have to compare the weights of a cell σ0 ∈ Y (k−1) on both sides. Note that
this weights can be computed locally on both sides. Therefore it suffices
to treat the case where X is a one-dimensional fan, ϕ : X → Y is integer
linear and f is a piecewise linear function on f∗ (X ). Given a ray ρ, we use
the shorthand vρ/{0} = vρ . Denoting the rays of X and ϕ∗ (X ) by ρ and ρ 0 ,
respectively, we obtain
ωϕ∗ (ϕ∗ ( f )·X ) ({0}) = ωϕ∗ ( f )·X ({0})

X
= ωX (ρ) f (ϕ(vρ ))
ρ∈X
X X
= ωX (ρ)[LZ (ρ 0 ) : ϕ(LZ (ρ))] f (vρ0 )
ρ 0 ∈ f∗ (X ) ρ∈X
ϕ(ρ)=ρ 0
X
= ωϕ∗ (X ) (ρ 0 ) f (vρ0 )
ρ 0 ∈ f∗ (X )
= ω f ·ϕ∗ (X ) ({0}).
Exercise 4.6.9
Prove that the definition of the push forward ϕ∗ (X ) is independent of the
(sufficiently fine) polyhedral structures X and Y .
135
Exercise 4.6.10 (Locality of push forwards)

Let ϕ : X → Y be an integer affine map between tropical cycles and let X
and Y be sufficiently fine polyhedral structures.
(a) For any cell τ ∈ X , show that (the linear part of) ϕ induces a canon-
ical “local” integer affine map ϕ τ : StarX (τ) → StarY (ϕ(τ)) for which
StarX (τ) and StarY (ϕ(τ)) are sufficiently fine.
(b) Show that for any cell τ0 ∈ Y of dimension m, we have

X
Starϕ∗ (X ) (τ) = [LZ (τ0 ) : dϕ|τ (LZ (τ))] · ϕ∗τ (StarX (τ)).
τ∈X (m)
ϕ(τ)=τ0
4.7 Examples and further topics

4.7.1 Tropical Bernstein-Kouchnirenko Theorem
The Bernstein-Kouchnirenko Theorem states that the number of common
zeros in (C× )n of a system of n generic Laurent polynomials in n variables is
equal to the mixed volume of the associated Newton polytopes (see [Ber76;
Kou76]). The mixed volume MV(P1 , . . . , Pn ) can be defined as the unique
function evaluated on n-tuples of (lattice) polytopes which is
(a) symmetric,
(b) multilinear, i.e. for all λ, µ ∈ N
MV(λP1 + µP10 , . . . , Pn ) = λ MV(P1 , . . . , Pn ) + µ MV(P10 , . . . , Pn ),
(c) normalized by MV(P, . . . , P) = Vol(P).
In this subsection, we want to prove the following tropical analogue.
Theorem 4.7.1
Let f1 , . . . , f n ∈ T[x 1± , . . . , x n± ] be tropical Laurent polynomials with associated
Newton polytopes P1 , . . . , Pn . Then
deg( f1 · · · f n · Rn ) = MV(P1 , . . . , Pn ).
136
We start by showing that the degree on the left hand side only depends on
the Newton polytopes of the polynomials, not the polynomials themselves.
We put the details of this verification in exercises below. For a lattice poly-
tope P with vertices j1 , . . . , jk , we set
k
X
fP = “ x ji ” T[x 1± , . . . , x n± ].
i=1
Lemma 4.7.2
The degree of tropical 0-cycle f1 · · · f n ·Rn only depends on the Newton polytopes
P1 , . . . , Pn .
Proof. We set g i := f Pi , i = 1, . . . , n. Since NP( f i ) = Pi = NP(g i ), by Exercise

4.7.4 (a) the functions f i − g i are bounded. Applying Exercise 4.7.4 (b) to
the 1-cycle X = f2 · · · f n · Rn , we obtain
deg( f1 · · · f n · Rn ) = deg(g1 · f2 · · · f n · Rn ) = deg( f2 · · · f n · g1 · Rn ),
where we used Proposition 4.4.18 in the second equality. Repeating this

procedure we arrive at deg( f1 · · · f n · Rn ) = deg(g1 · · · g n · Rn ), and the claim
follows.
Proof of Theorem 4.7.1. By what we have said before, it suffices to prove

the following: Given lattice polytopes P1 , . . . , Pn and setting f i = f Pi , the
function deg( f1 · · · f n ·Rn ) satisfies the characteristic properties of the mixed
volume function on page 136. Symmetry follows from Proposition 4.4.18,
multilinearity follows from
fλP+µP 0 = λ f P + µ f P 0
and the linearity of f 7→ f · X (see Remark 4.4.15). It remains to prove

normalization, which we put as an extra theorem for clarity.
Theorem 4.7.3
Let f T[x 1± , . . . , x n± ] be a tropical Laurent polynomial with associated Newton
polytope P. Then deg( f n · Rn ) = Vol(P).
Proof. To prove the theorem, we proceed as follows. First, recall that any
lattice polytope P admits a regular triangulation T (choose generic lift of
137
the vertices to Rn+1 , see e.g. [GKZ08, Proposition 1.5]). Let g be a tropical
polynomial with SD(g) = T . Using Lemma 4.7.2 again, it suffices to show
deg(g n · Rn ). But g n · Rn consists of a finite number of points dual to the
full-dimensional simplices of T . Since the weight of these points in g n · Rn
can be computed locally, we reduced the statement to the case of simplices.
Let us assume that P = ∆ is a lattice simplex and f = f∆ . Let us first
consider the case f = f1 := “0 + x 1 + · · · + x n ”, i.e. ∆ = ∆1 is equal to the
minimal standard simplex. Then the claim deg( f1n · Rn ) = 1 follows from a
direct calculation (or from Examples 4.3.8 and 4.4.20).
Now let us go back to the case of an arbitrary lattice simplex ∆. Note
that there exists a unique integer affine map ψ : (Rn )∗ → (Rn )∗ such that
ψ(∆1 ) = ∆. Let ϕ : Rn → Rn denote the dual map such that ϕ ∗ ( f1 ) = f .
Then the volume of ∆ can be expressed as Vol(∆) = [Zn : dψ(Zn )] = [Zn :
dϕ(Zn )]. Finally, using the projection formula 4.6.8 we get
deg( f n · Rn ) = deg(ϕ∗ (ϕ ∗ ( f1 )n · Rn )) = deg( f1n · ϕ∗ (Rn ))

= Vol(∆) deg( f1n · Rn ) = Vol(∆),
Exercise 4.7.4
Show that the following statements hold.
(a) Let f , g ∈ T[x 1± , . . . , x n± ] be tropical Laurent polynomials. If NP( f ) =

NP(g), then the function “ f /g” = f − g is bounded on Rn .
(b) Let X ⊂ Rn be a one-dimensional tropical cycle and let h : X → R be a

bounded rational function, then deg(h · X ) = 0.
4.7.2 Volumes of faces of polytopes

There is an interesting equation satisfied by the volumes of the faces of a
(lattice) polytope which is can be expressed in terms of the tropical balanc-
ing condition.
Theorem 4.7.5
Let P ∈ Rn be a lattice polytope and let N = N (P) be its normal fan. For a
138
face F of P, we denote the dual cone in N by σ F . Then for any k = 0, . . . , n,

the polyhedral fan N (k) with weights
ω(σ F ) = Vol(F )
is balanced.
We may generalize this by considering the subdivision S ( f ) of Rn as-

sociated to a tropical Laurent polynomial f . Recall from Section 2.4 that
we can equip each cell σ ∈ S ( f ) with the weight ω(σ) = Vol(Dσ ), where
Dσ ∈ SD( f ) is the dual cell to σ. In Theorem 2.4.7 we proved that the hy-
persurface complex S ( f )(n−1) is balanced. We now extend this to arbitrary
codimensions.
Theorem 4.7.6
The polyhedral complex S ( f )(k) with weights ω(σ) = Vol(Dσ ) is balanced
for all k = 0, . . . , n.
It is clear that Theorem 4.7.6 follows from Theorem 4.7.5 applied locally.
On the other hand, both statements have a nice conceptual explanation in
terms of tropical intersection products.
Theorem 4.7.7
Let f : Rn → R be a tropical Laurent poylnomial with associated subdivision
S ( f ). Then for all k = 0, . . . , n, the intersection product f n−k · S ( f ) is equal
to the polyhedral complex S ( f )(k) with weights ω(σ) = Vol(Dσ ).
Proof. Using the local computation of ω(σ) in f n−k · S ( f ), the statement

follows immediately from Theorem 4.7.3.
Proof of Theorems 4.7.5 and 4.7.6. We know that f n−k · S ( f ) is balanced

by Proposition 4.4.7. This implies Theorem 4.7.6. The special case f = f P
proves Theorem 4.7.5.
4.7.3 Intersecting with the diagonal

An alternative way for constructing stable intersections is given in [AR10].
The approach is based on intersecting with the diagonal ∆ ⊂ Rn × Rn and
only uses the divisor construction while circumventing the moving lemma.
139
The approach is more formal but sometimes simplifies arguments (for ex-
ample, for showing the compatibility of the divisor construction with stable
intersection) can be generalized to situations where usual stable intersec-
tion (i.e. moving cycles around) is not available. We briefly state the main
ideas here.
Given tropical X ⊂ Rn and Y ⊂ Rm , the cartesian product X × Y ⊂
R n+m
with weights ω((x, y)) = ω(x)ω( y) is a tropical cycle of dimension
dim(X ) + dim(Y ). We are particularly interested in the case n = m. Let us
denote by x 1 , . . . , x n and y1 , . . . , yn the coordinates of the first and second
factor of Rn × Rn , respectively.
Definition 4.7.8
Given a tropical cycle Z ⊂ Rn × Rn of dimension k, the intersection with the
diagonal is
∆ · Z := max{x 1 , y1 } · · · max{x n , yn } · Z ∈ Zk−n (R2n ).
Given two tropical cycles X , Y ⊂ Rn of dimension k and l, respectively, the

∆-intersection product of X and Y is
X .Y := π∗ (∆ · (X × Y )) ∈ Zk+l−n (Rn ),
where π : Rn × Rn → Rn denotes the projection to the first factor.

Note that ∆ · R2n = {(x, x) : x ∈ R} with constant weight 1. It is easy
to check that the definitions are invariant under (integer invertible) coordi-
nate changes and switching the two factors of Rn . Moreover, X .Y coincides
with stable intersection.
Theorem 4.7.9
For any two tropical cycles X , Y ⊂ Rn we have X .Y = X · Y .
Proof. It is easy to check that X .Y satisfies the same locality formula as
X · Y (see Exercise 4.7.10). It therefore suffices to prove the case k + l = n
and X , Y fan cycles. Since by Exercise 4.7.11 X .Y and X · Y agree in the
transversal case, it suffices to show deg(X .Y ) = deg(X .(Y + v)) for any v =
(v1 , . . . , vn ) ∈ Rn . But note that X .(Y +v) can be described as π∗ (∆ v ·(X ×Y ))
for the modified diagonal
∆ v · Z := max{x 1 , y1 + v1 } · · · max{x n , yn + vn } · Z.
140
Obviously, the functions max{x i , yi + vi } − max{x i , yi } are bounded for all

i = 1, . . . , n. Using Exercise 4.7.4 and the trick in the proof of Lemma 4.7.2
this implies deg(∆·(X ×Y )) = deg(∆ v ·(X ×Y )), which proves the claim.
Exercise 4.7.10
Let X , Y ⊂ Rn be two tropical cycles with polyhedral structures X and Y ,
respectively. Then for any cell τ ∈ (X ∩ Y )(k+l−n) we have
StarX .Y (τ) = StarX (τ). StarY (τ).
Exercise 4.7.11
Show the following statements.
(a) Let λ1 , . . . , λn : Rn → R be integer linear forms such that 〈λ1 , . . . , λn 〉R =

(Rn )∗ . Then
max{λ1 , 0} · · · max{λn , 0} · Rn
is equal to the origin equipped with weight [(Zn )∗ : 〈λ1 , . . . , λn 〉Z ].
(b) Assume that X and Y are affine subspaces of Rn (equipped with con-
stant weight 1) such that L(X ) + L(Y ) = Rn . Then X .Y is equal to the
affine space X ∩ Y equipped with weight [Zn : LZ (X ) + LZ (Y )].
(c) Let X , Y ⊂ Rn be two tropical cycles intersecting transversally. Then

X .Y = X · Y .
4.7.4 Irreducible tropical cycles

A tropical k-cycle X is irreducible if Zk (X ) = Z[X ]. A tropical cycle X of di-
mension k is Q-irreducible if Zk (X ) ⊗ Q = Q[X ]. In this case, the underlying
polyhedral set |X | is called irreducible. In other words, there is a unique
balanced weight function on |X | up to multiplying with a constant. Clearly,
if X is Q-irreducible and gcd{ωX (x) : x ∈ X gen } = 1, then X is irreducible.
A tropical k-cycle X is (Q-)irreducible in codimension one if for one (and
then for any) polyhedral structure X the stars StarX (τ) are (Q-)irreducible
for all k − 1-cells τ ∈ X . We call X connected in codimension one if for one
(and then for any) polyhedral structure X the set |X |\|X (k−2) | is connected.
141
Exercise 4.7.12
Show the two “and then for any” parts. Show the following statements.
If X is (Q-)irreducible and connected in codimension one, then X is (Q-
)irreducible.
Exercise 4.7.13
Show that if X is a k-cycle irreducible in codimension one and f : X → R is
a piecewise integer affine function, then | div( f )| = V ( f ). Let us denote by
Aff (X ) ⊂ PAZ (X ) the locally affine functions on X . Show that the sequence
div
0 −→ Aff (X ) −→ PAZ (X ) −→ Zk−1 (X ) −→ 0
is exact. Give counterexamples to these statements for X irreducible, but

not irreducible in codimension one.
Exercise 4.7.14
Construct an irreducible tropical fan k-cycle X and a fan (k − 1)-subcycle D
such that there exists no f ∈ PAZ (X ) such that div( f ) = D. Can you go on
to X irreducible in codimension one and all weights are 1?
4.7.5 Toric intersection theory

Predecessors of some the constructions of this chapter can be found in
intersection theory for (classical) toric varieties. Let us outline some of
these connections. For simplicity, we restrict to compact varieties over the
base field C. For a more details on the presented material and generaliza-
tions, we refer the reader to [Rau16; Kat09; Rau09; Kat12] and references
therein. For more background on toric (co-)homology theory, please con-
sult [Dan78, Chapter 10], [Ful93, Chapter 5], [CLS11, Chapter 12].
Let Ξ be a complete pointed fan in Rn . We denote by Ξk the subset of k-
cones and by Fk (Ξ) the free abelian group generated by Ξk , i.e. the group of
formal linear combinations of k-cones with integer coefficients. We denote
by Zk (Ξ) the group of balanced weight functions on the k-cones of Ξ. By
the discussion in Section 4.1, Zk (Ξ) is equal to the group of tropical k-
cycles supported on |Ξ(k) |. Regarding Zk (Ξ) as a sublattice of Fk (Ξ)∗ , we
can replace the balancing condition by its dual version as follows. For any
(k − 1)-cell τ, we regard the lattice LZ (τ)⊥ of integer linear functions λ that
142
vanish on LZ (τ). Evaluation of λ at primitive generators provides a map
evτ : LZ (τ)⊥ → Fk (Ξ),

X
λ 7→ λ(vσ/τ )[σ].
σ∈Ξk
τ⊂σ
Note that the choice of primitive generators does not matter since λ|LZ (τ) =
0. Obviously, a weight function ω ∈ Fk (Ξ)∗ is balanced around τ if and only
ω ∈ Im(evτ )⊥ . Let R k (Ξ) = ⊗τ∈Ξk−1 Im(evτ ) the sum of all these sublattices.
It follows that
Zk (Ξ) = R k (Ξ)⊥ = Hom(Fk (Ξ)/R k (Ξ), Z). (4.10)
Let now X = CX Ξ be the associated complex compact toric variety to Ξ.

Then the objects from above have the following (mostly straightforward)
reinterpretation. We denote the complex algebraic torus by T = (C× )n .
• Ξk : closures of (n − k)-dimensional torus orbits, Xσ
• Fk (Ξ) : torus-invariant algebraic (n − k)-cycles, generated by [Xσ ]
• LZ (τ)⊥ : (monomial) rational functions x λ on X which are regular

invertible on CUτ
• evτ (λ) : principal divisor of x λ restricted to Xτ
• R k (Ξ) : torus-invariant cycles rationally equivalent to zero
Let Ak (X) denote the k-th Chow group of X. In our case, it can be com-
pletely described in terms of torus-invariant cycles and relations, hence
A(n−k) (X) = Fk (Ξ)/R k (Ξ). Let Ak (X ) denote the “operational” Chow co-
homology groups (see [Ful98]). For compact X, these can be described as
Ak (X) = Hom(Ak (X), Z). In view of Equation 4.10, we find An−k (X) = Zk (Ξ).
In other words, Chow cohomology classes correspond to tropical fan cycles
on |Ξ(k) |, and the identification is given by ωα (σ) = deg(α ∩ [Xσ ]) for any
α ∈ An−k (X). Chow cohomology carries a ring structure given by the cup
product ∪ : Ak (X) × Al (X) → Ak+l (X). It turns out that under the identifica-
tion An−k (X) = Zk (Ξ), the cup product is replaced by stable intersection of
tropical cycles (the “fan displacement rule”, see [FS97]).
143
We denote by PLZ (Ξ) the group of piecewise integer linear functions ϕ :

Rn → R which are linear on the cones of Ξ (i.e., Ξ is sufficiently fine). The
integer linear functions form a sublattice (Zn )∗ ⊂ PLZ (Ξ). Again, there are
toric wordings for that.
• PLZ (Ξ) : torus-invariant Cartier divisors
• (Zn )∗ : principal torus-invariant Cartier divisors
Let Pic(X) denote the Picard group of X. Again, in the toric case it suf-
fices to consider torus-invariant Cartier divisors and relations and therefore
Pic(X) = PLZ (Ξ)/(Zn )∗ . There are natural maps Pic(X) → A1 (X) and more
generally Pic(X) × Ak (X) → Ak+1 (X) given by intersecting with Cartier divi-
sors. It turns out that under the identification Pic(X) = PLZ (Ξ)/(Zn )∗ , these
maps are replaced by the tropical divisor construction PLZ (Ξ) × Zn−k (Ξ) →
Zn−k−1 (Ξ), (ϕ, X ) 7→ ϕ · X . Note that in particular λ · X = 0 for λ ∈ (Zn )∗ .
In codimension one, some interesting special cases occur. For exam-
ple, the fact that the map Pic(X) → A1 (X) is an isomorphism can be de-
duced from Theorem 2.4.10 and Corollary 4.4.12 using the previous iden-
tifications. Indeed, a simple adaptation of the proofs shows that the map
PLZ (Ξ) → Zn−1 (Ξ), ϕ 7→ div(ϕ), is surjective with kernel (Zn )∗ , which im-
plies Pic(X) ∼
= A1 (X). Another interesting map is Pic(X) → An−1 (X) given by
sending a Cartier divisor to its associated Weil divisor. This relates to the
map ev : PLZ (Ξ) → F1 (Ξ), ϕ 7→ ρ∈Ξ1 ϕ(vρ )[ρ]. Clearly, this map is injec-
P
tive. Moreover, if Ξ is unimodular (or at least simplicial), then the map is

bijective (bijective after tensoring with Q, respectively. In this case, the in-
verse map is given by linear extension of ϕ on cones given its values on the
generators. It follows that Pic(X) ∼ = An−1 (X) and Pic(X) ⊗ Q → An−1 (X) ⊗ Q
if X is smooth or Q-smooth, respectively.
References
Parts of the theory presented here appeared in many papers in the early
days of tropical geometry. Stable intersection in the setting of tropical cy-
cles was introduced in [Mik06]. A formal approach to tropical intersec-
tion theory covering most of the material presented here was developed in
[AR10; Rau16] (see also [Rau09; Kat09]).
144
5 Projective tropical geometry
5.1 Tropical cycles in toric varieties
5.1.1 The quick and formal definition
Our first objective is to define tropical cycles in tropical toric varieties, in
particular, in Tn and TPn . Again, there are several equivalent definitions,
depending on our preferred viewpoint. Let us start with the quick and
formal definition.
Let Ξ be a fan in Rn . In this section, we denote its cones by ϑ, ρ ∈ Ξ, to
distinguish them from the cells τ, σ ∈ X of polyhedral complexes to appear
soon. We denote by W = WΞ the tropical toric variety associated to Ξ. Let us
recall that to any ϑ ∈ Ξ we can associate a torus orbit Rϑ as well as its closure
Wϑ = Rϑ ⊂ W . They have dimension dim(Wϑ ) = dim(Rϑ ) = n − dim(ϑ). In
particular, recall that Rϑ ∼
= Rn−dim(ϑ) .
Definition 5.1.1
A tropical k-cycle in WΞ is a formal sum
X
X= Xϑ,
ϑ∈Ξ
where each X ϑ is a tropical k-cycle in Rϑ . Thus the group of tropical k-cycles

in WΞ is M
Zk (WΞ ) := Zk (Rϑ ).
ϑ∈Ξ
The support of X is |X | = ϑ∈Ξ |X ϑ |, where the closure is taken in W .

S
A k-cycle X is of pure sedentarity ϑ if it consists of a single summand

X = X ϑ (i.e. X ρ = 0 for all ϑ 6= ρ ∈ Ξ. Moreover, if ρ = {0}, X is of
sedentarity zero.
Note that this definition, in the special case of k = n − 1, agrees with the
definition of tropical divisors in Definition3.4.4.
145
5.1.2 Polyhedra and their peers in toric varieties

It will be useful later to also extend the language of polyhedral complexes
and sets to the toric setting. This will lead to other, equivalent descriptions
of tropical k-cycles. A (generalized) polyhedron σ in W is the closure in
W of a (usual) polyhedron σ0 ⊂ Rϑ for some ϑ ∈ Ξ. We refer to ϑ as the
sedentarity of σ (and σm ). Moreover, we call σm := σ0 = σ ∩ Rϑ the mobile
part of σ, while σs := σ \ σm is the sedentary part of σ. The dimension
of σ is the dimension of σm . Analogously, we define the linear span L(σ),
recession cone rc(σ), etc., by referring to the corresponding notions for σm .
There are two types of faces of σ: For any face τ0 ⊂ σm , we call the
closure τ0 ⊂ σ ⊂ W a mobile face of σ (or face of same sedentarity). On
the other hand, for all ρ ∈ Ξ the intersection σ ∩ Rρ is either empty or
a (usual) polyhedron in Rρ , see Proposition 3.2.6. If non-empty, we call
σ ∩ Rρ ⊂ σ ⊂ W a sedentary face of σ (or face of higher sedentarity). Note
that in general σ ∩ Wρ is a disjoint union of sedentary faces. Figure 5.1 an
Example of a generalized polyhedron with its sedentary faces.
R2 T2
σ0 σ
Figure 5.1: The closures of a standard polyhedron in T2 with its new faces
We can directly generalize the definitions of polyhedral complexes, equiv-

alence of (weigthed) complexes, polyhedral sets, etc. to the toric case by
repeating the definitions word by word, only replacing polyhedra (and its
faces) by its generalized versions of polyhedra in W . A polyhedral set
X ⊂ W is of pure sedentarity ϑ if X = X ∩ Rϑ . In this case, the mobile and
sedentary part of X are X m = X ∩Rϑ and X s = X \X m , respectively. A polyhe-
dral complex X is of pure sedentarity ϑ is |X | is so. The mobile and seden-
tary part of X are X m = {σ ∩ Rϑ : σ ∈ X } and X s = {σ ∈ X : σ ⊂ |X |s },
146
respectively.
Given a weighted polyhedral complex X in W (a polyhedral structure of
a weighted polyhedral set X in W , respectively), the balancing condition in
W requires that for all τ ∈ X (k−1) we have
X
ω(σ)vσ/τ = 0 mod LZ (τ). (5.1)
σ⊃τ
same sed.
Here, the sum runs through all facets σ ∈ X containing τ and of same
sedentarity, in which case the notion of primitive generator vσ/τ := vσm /τm
is well-defined. Note that if τ is of sedentarity ϑ, then the computation
takes place completely inside Zϑ ⊂ Rϑ . We can now reformulate Definition
5.1.1 in the following equivalent ways (cf. 4.2.2 and 4.2.3).
Definition 5.1.2
A tropical k-cycle X in W is the equivalence class of balanced polyhedral
complex X in W of pure dimension k. Equivalently, X is a balanced poly-
hedral set in W of pure dimension k. If all weights are positive, X is a
tropical subspace in W . Moreover, X is of pure sedentarity ϑ if |X | is so.
Remark 5.1.3
Having rephrased our definition P like this, when writing a cycle X as sum
of its pure sedentarity parts X = ϑ X ϑ , the X ϑ will from now on be closed
cycles in Wϑ . The mobile parts X ϑ ∩ Rϑ that were used in Definition 5.1.1
will be denoted by X ϑm .
Example 5.1.4
For each k with 0 ≤ k ≤ n, let L km ⊂ Rn be the standard tropical k-plane
defined in Example 4.2.6. Its closure L k ⊂ TPn is a sedentarity zero k-
cycle in TPn and is called the standard tropical projective k-plane. Figure
5.2 illustrates the example L2 ⊂ TP3 . Note that L2s consists of four tropical
lines of higher sedentarity, one for each coordinate hyperplane at infinity.
More generally, for any boundary stratum TPnI ∼ = TPn−|I| , we have L k ∩TPnI =
L k−|I| ⊂ TPn−|I| . On a set-theoretic level, this follows immediately from the
recursive structure of the stars of L km , see Exercise 4.2.12. Regarding the
weights, we will be more precise in Section .
147
Figure 5.2: A tropical plane in TP3
5.1.3 Push forwards along toric morphisms

Let Ξ and Ξ0 be pointed fans in Rn and Rm , respectively. Let ϕ : Rn → Rm
be an integer affine map such that for all ϑ ∈ Ξ there exists ϑ0 ∈ Ξ0 with
dϕ(ϑ) ⊂ ϑ0 (we may assume dϕ(ϑ ) ⊂ ϑ0 ). Then there exists a unique

continuous extension to a map between toric varieties ϕ : WΞ → WΞ0 . More

concretely, for each ϑ ∈ Ξ and ϑ0 ∈ Ξ0 such that dϕ(ϑ ) ⊂ ϑ0 we have

ϕ(Rϑ ) ⊂ Rϑ0 and the diagram

ϕ
Rn Rm
πϑ πϑ0
ϕ
Rϑ Rϑ0
commutes. Such a map ϕ : WΞ → WΞ0 is called an integer affine map (or

monomial map) between toric varieties. If there exists an inverse integer
affine map ψ : WΞ0 → WΞ , ϕ is an integer affine isomorphism.
Definition 5.1.5
Let ϕP: WΞ → WΞ0 be an integer affine map between toric varieties and let
X = ϑ X ϑ be a tropical k-cycle in WΞ . The push forward of X along ϕ is
148
the tropical k-cycle in WΞ0 given by

X
ϕ∗ (X ) := ϕ∗ (X ϑm ).
ϑ
Here, ϕ∗ (X ϑm ) denotes the push forward of along ϕ : Rϑ → Rϑ0 , where ϑ

and ϑ0 are such that dϕ(ϑ ) ⊂ ϑ0 .

5.2 Projective degree

5.2.1 Intersecting with the standard projective
hyperplane
In this section, we want to define the degree of tropical k-cycle in TPn .
We will use the following notations. First recall that the cones σ I of the fan
defining TPn are labelled by subsets I ( {0, . . . , n}. In the following, we will
always use I instead of σ I at the appropriate places. For example, given a
tropical cycle X ⊂ TPn , we will refer to its pure sedentarity parts by X I ⊂
TPnI . For any I ( {0, . . . , n}, we denote by H I ⊂ TPnI the standard projective
(n − |I| − 1)-plane (via the canonical identification TPnI ∼ = TPn−|I| ). In the
case I = ;, we use the shorthand H := H; ⊂ TPn for the standard projective
hyperplane in TPn . As explained in Example 5.1.4, we have H ∩TPnI = H I for
all I (at least set-theoretically). By slight abuse of notation, we will define
the stable intersection of any cycle with H as follows.
Definition
P 5.2.1
Let X = I X I be a tropical k-cycle in TPn . The stable intersection of H and
X is X
H · X := H Im · X Im ,
I
where each product H Im · X Im refers to the stable intersection of cycles in R I .
In analogy to the classical case, we define the projective degree of a k-

cycle as the number of intersection points with k hyperplanes. In contrast to
the classical case, where these hyperplanes are typically chosen generically
such as to obtain a simple count of (transversal) intersection points, we will
only use the fixed standard hyperplane H ⊂ TPn . Recall that the degree of
149
0-cycle A, regarded as a formal sum of points A = a p · [p], is deg(A) =

P
P p
p a p ∈ Z.
Definition 5.2.2
Let X be a tropical k-cycle in TPn . The (projective) degree of X is
deg(X ) := deg(H k · X ).
where H denotes the standard projective hyperplane in TPn .
Example 5.2.3
From our computations in example 4.3.8 it follows that the standard pro-
jective k-planes L k ∈ TPn have degree 1.
In the following Proposition, we show that a local part of tropical sub-
space X can only have smaller degree than X itself.
Proposition 5.2.4
Let X be a tropical subspace of Rn and let X be a polyhedral structure. Then
the followings holds.
(a) For every τ ∈ X , we have deg(StarX (τ)) ≤ deg(X ).
(b) For every facet σ ∈ X , we have ω(σ) ≤ deg(X ).
(c) If deg(X ) = 0, then X = 0.

Here, all closures are taken in TPn ⊃ Rn .
Proof. Obviously, (a) ⇒ (b) ⇒ (c), hence it remains to prove (a). Note that
by 4.3.6 the degree is invariant under translations, e.g. deg(X ) = deg(X ·
((H m )k + p)) for some p ∈ τ . As both factors in this product are subspaces,
all points in X · (H k + p) have positive weight. But by the locality of the
intersection product (again 4.3.6), the weight of p in this product is equal
to deg(StarX (τ)), and hence the claim follows.
5.2.2 The degree of projective hypersurfaces

As a more general example, let us compute the degree of projective hyper-
surfaces.
150
Proposition 5.2.5
Let F be a homogeneous tropical polynomial of degree d and V (F ) ⊂ TPn its
associated hypersurface (see Definition 3.4.6). Then deg(V (F )) = d.
Proof. As in Definition 3.4.6, let H i = V (x i ) denote the i-th coordinate hy-

perplane. Obviously deg(H i ) = 1. On the other hand, dividing F by x i
decreases the degree of F by 1. Hence we may assume that F is monomi-
alfree.
Let f = F (x 0 = 0) ∈ T[x 1 , . . . , x n ] be the dehomogenization of F and let
L = L1 = (H m )n−1 ⊂ Rn be the standard line. It remains to show deg( f · L) =
deg( f ). Since f is monomialfree, the function will be eventually constant
on each ray R≤0 ei , i = 1, . . . , n, of L. On the other hand, on the ray R≤0 ei
the slope of f will ultimately be d = deg( f ). It follows by Exercise 4.4.22
that deg( f · L) = d as required.
5.2.3 Projective tropical lines

Definition 5.2.6
A (projective) tropical line is a tropical subspace of TPn of dimension and
degree 1.
Let us recall a trivial fact from classical algebraic geometry: Any line in
CPn is isomorphic to CP1 . In tropical geometry, however, there exist (topo-
logically) different projective lines. Of course, the archetype is still TP1 . But
already the standard line L1 in TP2 is different from TP1 . L contains 3 infi-
nite points and one vertex, whereas TP1 has 2 infinite points and no vertex
at all. Figure 5.3 depicts a tropical line in TP3 containing a bounded edge.
On the other hand, tropical lines are of course very special 1-cycles. Some
properties are summarized in the following Proposition. Note that any 1-
cycle carries a canonical coarsest polyhedral structure. In the following, all
statements about edges and vertices refer to this polyhedral structure.
Lemma 5.2.7
Let L be a tropical line in TPn of sedentarity zero. Then L has the following
properties:
(a) All weights of L are 1.
151
(1, 1, 1)
(−1, 0, 0)
(0, −1, 0)
(0, 0, −1)
Figure 5.3: A tropical line in TP3
(b) L is contractible. In other words, it is a rational graph.
(c) Let v0 , . . . , vm be the primitive generators of the rays of L m (pointing

towards infinity). Then m ≤ n and
X
vj = − ei
i∈I j
for a suitable subset I j ⊂ {0, . . . , n}. Moreover, {0, . . . , n} is the disjoint

union of I0 , . . . , I m .
(d) The graph L m has at most m − 3 bounded edges and at most m − 2 finite
vertices. Equalities holds when all vertices of L m are 3-valent.
Proof. Statement (a) follows from Proposition 5.2.4.

For statement (b), assume that L contains a cycle C. We can find a vector
v ∈ Rn such that the translated hyperplane H and L only intersect in Rn and
such that H m + v intersects C in a transversal intersection point p. Since the
facets of H + v subdivide TPn into n + 1 connected components, the cycle
C must intersect H + v in a second point in Rn . It follows that deg((H m +
v) · L) ≥ 2, a contradiction.
Let us now prove the assertion (c). We start be showing that each v j
is a simple sum of the standard primitive vectors e0 , . . . , en . By relabelling
the coordinates we may assume v j ∈ (Z≤0 )n . We have to show that all
152
coordinate entries of v j are either 0 or −1. Assume contrarily that the first
coordinate entry is lower than −1. Then we take a hyperplane H transversal
to L and such that the facet of H spanned by −e2 , . . . , −en and the ray of
L with direction v j intersect. As the lattice index contributing to the local
intersection multiplicity is equal to the first coordinate entry of v j , we get a
contradiction again. Thus we showed
X
vj = − ei
i∈I j
for a suitable subset I j ⊂ {0, . . . , n}. Next, we show that these sets are
pairwise disjoint. Assume i ∈ I j ∩ I j 0 . Then any hyperplane H which we
move far towards −ei will intersect both rays generated by v j and v j 0 , which
Sm Pm
is impossible. It remains to show {0, . . . , n} = j=0 . But note j=0 vk = 0
by Part (a) and Lemma 4.2.8. Since i∈I ei = 0 if and only if I = {0, . . . , n},
P
the claim follows.

Statement (d) follows immediately from the fact established so far that L
is a rational graph with m + 1 1-valent vertices (the points of higher seden-
tarity).
A line L ⊂ TPn of sedentarity zero is called non-degenerate if L m it has

n + 1 rays (with direction vectors −e0 , . . . , −en ). We will come back to the
relationship between tropical lines (or more generally, rational curves) and
TP1 in Section 5.5.
5.3 Modifications in toric varieties

In this section, we want to extend the divisor and modification constructions
from Chapter 4 to cycles in toric varieties. Recall that both constructions are
closely related. The modification along a function f is the completion of its
graph by adding “downward” faces projecting to the divisor of f . Vice versa,
the divisor of f can be obtained from the modification by “intersecting with
∞”. In the toric case, it is useful to study both constructions simultaneously
right away. We start, however, with formalizing the concept of “intersecting
with ∞”.
153
5.3.1 Intersections with toric boundary divisors

Let Ξ be a fan in Rn . We denote the primitive generator of a ray ρ ∈ Ξ(1)
by vρ . In order to intersect with a specific divisor D = Wρ of WΞ , we need
a partial smoothness or unimodularity assumption with respect to ρ.
Definition 5.3.1
A ray ρ is Cartier in Ξ if there exists a function f : |Ξ| → R which integer
linear on each cone of Ξ and such that f (vρ ) = −1 and f (vρ0 ) for all other
rays ρ 0 6= ρ. A divisor D = Wρ is Cartier in WΞ if ρ is Cartier in Ξ.
In such function f exists, it is obviously unique and we denote it by fρ or
fD.
Remark 5.3.2
Note that ρ is Cartier in Ξ if and only if the following holds. Any cone
ϑ ∈ Ξ containing ρ has a unique face ϑ0 not containing ρ and LZ (ϑ) =
LZ (ϑ0 ) + Zvρ . The terminology is adapted from classical toric geometry,
where the condition is equivalent to CWρ being a Cartier divisor in CWΞ (in
the classical sense).
In order to define intersections with toric divisors, we need to introduce
the notion of primitive generators for faces of higher sedentarity. Let σ ⊂ W
be a polyhedron of dimension k and sedentarity zero and let τ ⊂ σ be a
(k − 1)-face of higher sedentarity, say ϑ. By Proposition 3.2.6, it follows
that rc(σm ) ∩ ϑ 6= ; and that πϑ (σm ) = τm , where πϑ : Rn → Rϑ = Rn / L(ϑ)
is the canonical projection. In particular,
dim(rc(σm ) ∩ ϑ) = dim(L(rc(σm )) ∩ L(ϑ)) = dim(L(σm ) ∩ L(ϑ)) = 1,
and hence rc(σm ) ∩ ϑ is a (one-dimensional) ray.
Definition 5.3.3
Let σ ⊂ W be a polyhedron of dimension k and sedentarity zero and let
τ ⊂ σ be a (k − 1)-face of higher sedentarity, say ϑ. Let −vσ/τ ∈ Zn be the
primitive generator of rc(σm )∩ϑ. Then its negative, vσ/τ , is the (sedentary)
primitive generator of σ modulo τ.
If σ is of sedentarity ϑ0 , we replace Rn by Rϑ0 and Ξ by StarΞ (ϑ0 ) and
obtain the sedentary primitive generator vσ/τ ∈ Zϑ0 = Zn / LZ (ϑ0 ).
154
We are now ready to define intersections with toric divisors D.
Definition 5.3.4
Let X ⊂ W be a k-cycle of sedentarity zero and let X be a polyhedral struc-
ture. Let D = Wρ be a toric divisor which is Cartier, with associated function
f D : |Ξ| → R. The (sedentary) intersection D · X is the weighted polyhedral
complex X | D with weights
X
ω(τ) = ω(σ) f D (vσ/τ ).
σ⊃τ
Here, τ of (k − 1)-cell of X | D and the sum runs through the k-cells σ ∈ X

(hence of sedentarity zero) containing τ. The (sedentary) intersection D · X
is the tropical (k − 1)-cycle represented by D · X .
Getting routine by now, it remains to show that D · X is balanced.
Proposition 5.3.5
The weighted polyhedral complex D · X in W is a balanced.
Proof. Let α be a (k − 2)-cell of D · X of pure sedentarity ϑ. We denote

by π = πϑ : Rn → Rϑ the sedentary projection. Recall that the balancing
condition around α only involves the (k − 1)-cells β ⊃ α of D · X of same
sedentarity ϑ. Throughout this proof, we use β to denote such cells and
use σ to denote k-cells in X (hence of sedentarity zero) that contain α.
For any flag α ⊂ β ⊂ σ, recall that π(σm ) = β m and αm is a (usual) face of
β m . It follows that σm ∩ π−1 (αm ) is a face of σm ; we denote its closure by τ.
Note that rc(τm ) ∩ ϑ = rc(σm ) ∩ ϑ. Thus τ is (k − 1)-cell of X of sedentarity
zero such that τ∩Rϑ = αm and such that the sedentary primitive generators
vτ/α = vσ/β coincide. Moreover, given a (mobile) primitive generator vσ/τ ,
the vector vβ/α = π(vσ/τ ) is a primitive generator for β modulo α. This
follows directly from LZ (β m ) = π(LZ (σm )) and LZ (αm ) = π(LZ (τm )).
We add the convention that τ always denotes a (k − 1)-cells of X of
sedentarity zero containing α. Given a facet σ ⊃ α two cases can occur.
Either σ ∩Rϑ is of dimension k −1. Then the previous considerations apply.
Otherwise, σ ∩ Rϑ = αm is of dimension k − 2. In this case, π(L(σ)) = L(α)
and hence π(vσ/τ ) ∈ L(α) for for any τ ⊂ σ. We are now ready to combine
our considerations in the following calculation, keeping in mind the various
155
conventions for β, σ and τ. We get

X X
ω D·X (β)vβ/α = ωX (σ) f D (vσ/β )vβ/α
β σ⊃β
X X
= f D (vτ/α )π ωX (σ)vσ/τ mod L(α).
τ σ⊃τ
Since for all τ the argument of π in the last expression is in L(τ) (balancing
around τ), it follows that the total sum is contained in π(L(τ)) = L(α),
which proves the claim.
Remark 5.3.6
Clearly, the definition of D·X can be extended to cycles X of pure sedentarity
ϑ as long as ρ 6⊂ ϑ. More precisely, if ϑ + ρ is a cone of Ξ, we consider X
as a sedentarity zero cycle in W 0 = WStarΞ (ϑ) and let D0 be the toric divisor
in W 0 associated to the ray (ϑ + ρ)/ϑ. We set D · X := D0 · X , using the
previously defined construction in W 0 . Otherwise, we set D · X = 0. Note,
however, that in the case ρ ⊂ ϑ the intersection D · X is not well-defined (at
least, on the cycle level).
Exercise 5.3.7
Prove that the multiplicities introduced in Exercise 2.5.5 correspond to the
intersection of a curve C ⊂ TP2 of sedentarity zero with the toric divisor
L∞ corresponding to the ray R≥0 (1, 1).
5.3.2 Modifications and divisors

Let us first introduce the class of functions that we want to consider. In the
following, W = WΞ is a fixed tropical toric variety.
Definition
P 5.3.8
Let X = ϑ∈Ξ X ϑ be tropical k-cycle in W . A piecewise integer affine function
on X is a collection of piecewise integer affine functions fϑ : X ϑm → R for
each ϑ ∈ Ξ. In particular, the group of piecewise integer linear functions
on X is M
PAZ (X ) = PAZ (X ϑm ).
ϑ∈Ξ
156
If X is of pure sedentarity, f is just a piecewise integer affine function

f : X m → R.
For general cycles X = ϑ X ϑ we set X m = ϑ X ϑm ⊂ W . Then a piecewise
P F
integer affine function is determined by the induced function f : X m → R,

which we usually identify with f . One should keep in mind, however, that
this function can be rather peculiar. For example, in the case of mixed
sedentarity it will in general not be continuous.
DefinitionP5.3.9
Let X = ϑ∈Ξ X ϑ be tropical k-cycle in W and f : X m → R a piecewise
integer affine function on X . The modification of X along f is the k-cycle in
W × TP1 given by X
Mod(X , f ) = Mod(X ϑm , fϑ ).
ϑ
On a formal level, we could think of Mod(X , f ) as a cycle in W ×R as well,

staying closer to the definition in Section 4.5. However, we already gave
some motivation for compactifying the second factor in Remark 4.5.4, and
will find more reasons shortly. Note that piecewise integer affine functions
on X may have points of indeterminancy at higher sedentarity even after
extending the range of values from R to TP1 . Here is a typical example.
Example 5.3.10
Consider the Laurent monomial f = “x 2 /x 1 ” = x 2 − x 1 : R2 → R. We con-
sider it here as a element in PAZ (T2 ). Note that f obviously be extended to
continuous function on T2 \ {(−∞, −∞)} → TP1 , the value at (−∞, −∞)
is not well-defined. Let us compute Mod(T2 , f ). Note that f |2R is just a lin-
ear function. Taking the closure of its graph {x ∈ R3 : x 2 = x 1 + x 3 } in
T2 × TP1 , we see easily that the extra added points form the union of three
coordiante lines
Mod(T2 , f ) \ R3 = T × {−∞} × {−∞}

∪ {−∞} × T × {+∞}
∪ {−∞} × {−∞} × TP1
(see Figure 5.4). While the first two lines are due to the extension of f to
T2 \ {(−∞, −∞)} → TP1 , the last line {−∞} × {−∞} × TP1 reflects the
157
T2 × TP1
Mod(T2 , f )
Figure 5.4: The modification Mod(T2 , x 1 − x 2 ) with its 3 boundary lines
fact that f is not well-defined at (−∞, −∞). In fact, Mod(T2 , f ) can be

understood as the blow up of T2 at (−∞, −∞).
We denote by H− = W × {−∞} and H+ = W × {+∞} the two canonical

toric divisors in W × TP1 . Clearly, the projection π : W × TP1 → W induces
an isomorphism of cycle groups π∗ : Zk (H± ) → Zk (W ). By definition, the
modification Mod(X , f ) is a tropical cycle in W × TP1 . It has parts of pure
sedentarity ϑ × {0}, where ϑ is a cone in Ξ. Hence by Remark 5.3.6, the
sedentary intersections H− ·Mod(X , f ) and H+ ·Mod(X , f ) are well-defined.
DefinitionP5.3.11
Let X = ϑ∈Ξ X ϑ be tropical k-cycle in W and f : X m → R a piecewise
integer affine function on X . The divisor of f , denoted by div( f ) or f · X , is
the tropical k − 1-cycle
div( f ) = π∗ (H− · Mod(X , f ) − H+ · Mod(X , f )) ∈ Zk−1 (W ).
Note that this time the balancing condition for div( f ) follows from Propo-
sition 5.3.5.
Remark 5.3.12
Let us assume that X is a tropical subspace. Then the domains of div( f ) of
positive and negative weight can be regarded as “zeros” and “poles” of f ,
respectively. Note that poles can arise in two ways: From H+ · Mod(X , f )),
158
obviously, but also from facets in Mod(X , f ) pointing downwards but car-
rying negative weights. As discussed in Remark 4.5.4, in Rn -case only the
poles of the second type occur and it is therefore sufficient to intersect with
H− . Again, for general piecewise integer affine functions there seems to be
no canonical way to repair this asymmetry between zeros and poles.
For rational functions, on the other hand, we can of course adapt Def-
inition 4.5.5 as follows. For simplicity, let us assume that X is a tropical
subspace of W of sedentarity zero. Let f = “g/h” : X m → R be a ratio-
nal function given as the quotient of two polynomial functions g, h. Then
modification of the quotient “ g/h " is given by
Mod(X , g, h) := Mod(Xm , g, h) ⊂ W × TP1 .
It is easy to check that div( f ) = div(g)−div(h), div(g) = π∗ (H− ·Mod(X , g, h))

and div(h) = π∗ (H+ · Mod(X , g, h)). Hence this time the zeros and poles of
f come as two seperate cycles with positive weights. We will consider an-
other generalization of this in Section 5.3.4.
We would like to give an explicit formula for the weights of div( f ) avoid-
ing the detour via Mod(X , f ). For the sedentarity zero part we can obviously
still use the weight formula from Equation (4.5). Actually, something sim-
ilar is true at the boundary. Actually, the formula can be easily adapted to
cover faces of higher sedentarity as well. It enough to consider the seden-
tarity zero case.
Let X ⊂ W be a tropical k-cycle of sedentarity zero and let f : X m → R
be a piecewise integer affine function. A polyhedral structure X is called
sufficiently fine for f if X m is sufficiently fine.
Corollary 5.3.13
Let X ⊂ W be a tropical k-cycle of sedentarity zero and let f : X m → R be
a piecewise integer affine function. Let X be a sufficiently fine polyhedral
structure and τ ∈ X a (k − 1)-cell of higher sedentarity. Then the weight of
τ in div( f ) is equal to
k
X
ω(τ) = ω(σi )d f |σmi (vσi /τ ), (5.2)
i=1
where σ1 , . . . , σk denote the facets of X containing τ.
159
Recall that if τ is of sedentarity zero and the primitive generators around

Pk
τ are chosen such that i=1 ω(σi )vσi /τ = 0, then the formula for ω(τ) is
identical to the one presented here (see Remark 4.4.5).
Proof. Each pair τ ⊂ σi corresponds to a pair τ
e⊂σ
e i of faces in Mod(X , f ),
and primitive generators are given by
vσe /eτ = (vσ/τ , d f |σmi (vσi /τ )) ∈ Zn × Z.
We set si = d f |σmi (vσi /τ ) ∈ Z. Let ϑ be the sedentarity of τ. Depending

on the cases si > 0, si = 0, or si < 0, the sedentarity of τ e is ϑ × R≤0 en+1 ,
ϑ × {0} or ϑ × R≥0 en+1 , respectively. In the middle case, the pair τ e⊂σ ei
neither contributes to Equation (5.2) nor to either of the intersections H± ·
Mod(X , f ). If si > 0, the pair τ e⊂σ e i contributes to H− · mod (X , f ) with
weight ω(e σi ) f H− (vσe /eτ ). But note that ω(e σi ) = ω(σi ) and f H− (vσe /eτ ) = si
(see Example ). So the contribution of τ e⊂σ e equals the contribution of
τ ⊂ σi to Equation (5.2). The case si < 0 works symmetrically.
Example 5.3.14
H± Cartier in W × TP1 .
Example 5.3.15
The Laurent monomial f = “x 2 /x 1 ” = x 2 − x 1 from Example 5.3.10 has
divisor div( f ) = V (x 2 ) − V (x 1 ) ⊂ T2 .
Example 5.3.16
If D is Cartier then D = div( f D ).
Explain subdivision of W induced by Ξ??
5.3.3 Line bundles of tropical toric varieties

In the previous section, we considered modifications in W × TP1 , the trivial
TP1 -bundle over W , provided we are given a piecewise integer affine func-
tion on X . Starting from homogeneous polynomials F , we can similarly
construct a modification Mod(X , F ) which this time sits naturally in a non-
trivial line bundle over W . In this subsection we introduce these objects in
the generality needed here, postponing a general discussion of tropical line
bundles to .
160
Let Ξ be a pointed fan in Rn and let f : |Ξ| → R be a function which

integer linear on every cone of Ξ. In other words, f is piecewise integer
linear and Ξ is sufficiently fine. For any ϑ ∈ Ξ, we denote the by ϑ
e = Γ (ϑ, f )
n+1
the cone in R equal to the graph of f |ϑ . Furthermore, we set
ϑ≤ := ϑ
e + R≤0 en+1 , ϑ≥ := ϑ
e + R≥0 en+1 ,
and consider the fans in Rn+1
Ξ≤ ( f ) := {ϑ,
e ϑ≤ : ϑ ∈ Ξ}, (5.3)
Ξ( f ) := {ϑ,
e ϑ≤ , ϑ≥ : ϑ ∈ Ξ}. (5.4)
Exercise 5.3.17
Show that Ξ≤ ( f ) and Ξ( f ) are fans.
Definition 5.3.18
The tropical line bundle OW ( f ) associated to f on W = WΞ is the toric variety
associated to Ξ≤ ( f ) together with the canonical projection π : OW ( f ) → W .
The tropical projective line bundle (or TP1 -bundle) PW ( f ) associated to f
on W = WΞ is the toric variety associated to Ξ( f ) together with the canon-
ical projection π : PW ( f ) → W .
Exercise 5.3.19
Show that a fiber of π : OW ( f ) → W can be identified with T. Show that a
fiber of π : PW ( f ) → W can be identified with TP1 .
Example 5.3.20
If f ≡ 0 is constant zero, we have OW ( f ) = W × T and PW ( f ) = W × TP1 .
Example 5.3.21
For W = TPn , f = “0 + x 1 + · · · + x n ” and d ∈ Z, we obtain the tropical lines
bundles O (d) := OTPn (d · f ) (sometimes called the twisting bundles) their
projective counterparts and P (d) := PTPn (d · f ). Let us describe the gluing
maps for O (d) explicitly. We denote by Ui = {x ∈ TPn : x i 6= −∞} the
standard affine charts of TPn . On the overlap Ui ∩ U j the function ai j (x) :=
d(x j − x i ) is a well-defined Z-invertible integer linear function (in particular,
it extends to the sedentary points of Ui ∩ U j ). The result of gluing the affine
161
charts Ui × T along the maps
φi j : (Ui ∩ U j ) × T → (Ui ∩ U j ) × T
(x, t) 7→ (x, t + ai j (x))
can be canonically identified with O (d). The same argument applies to

P (d).
Remark 5.3.22
Let us assume that Ξ is complete and f : |Ξ| → R is strongly convex, i.e.
f is strictly convex on any line in Rn . Then the convex hull Conv(Γ ( f )) of
the graph of f is a pointed cone in Rn+1 . Adding this cone to Ξ≤ ( f ), we
obtain a complete fan Ξ≤ ( f ) of Rn+1 . The associated tropical toric variety
is denoted by O W ( f ).
Example 5.3.23
For W = TPn , the variety O W (d) is called a tropical weighted projective space
TP(1, . . . , 1, d).
Example 5.3.24
Let λ : Rn → R be an integer linear function. Then the fans Ξ≤ ( f ) and
Ξ≤ ( f + λ) are canonically isomorphic under the action of GL(n + 1, Z). The
same is true for Ξ( f ) and Ξ( f + λ). Hence we can canonically identify
OW ( f ) ∼
= OW ( f + λ) and PW ( f ) ∼
= PW ( f + λ).
Remark 5.3.25
Recall the toric degree map δ : Zl → B from Remark 3.4.8. Fixing a toric
degree d ∈ B, we consider the rational polyhedron P(d) := (δ−1 (d) ⊗ R) ∩
l
R≥0 . Up to translation Pd determines a polyhedron in Rn via the map Z n →
Zl . On the dual side, the support function f d of Pd (i.e. the Legendre dual of
the characteristic function of Pd ) is a function on |Ξ| which is linear on the
cones of Ξ. We call d Cartier if f (d) is integer linear (if Ξ is complete, this
is equivalent to Pd being a lattice polytope). Under this assumption, we get
associated line bundles OW (d) := OW ( f d ) and PW (d) := PW ( f d ). Moreover,
if Ξ is complete and Pd is an n-dimensional polytope, then f d is strongly
convex and we set O W (d) := O W ( f d ).
162
5.3.4 Homogeneous modifications

Let F ∈ T[x 0 , . . . , x n ] be a homogeneous polynomial of degree d. Recall
that F does not define a function on TPn . On the standard affine chart
Ui = {x ∈ TPn : x i 6= −∞}, we can consider the dehomogenization f i of
F . In homogeneous coordinates, f i is given by the quotient Fi := “F /x id ” =
F − d x i . On the overlap Ui ∩ U j , we have
Fi (x) − F j (x) = d(x j − x i ) = ai j (x), (5.5)
where ai j denotes the transition functions from Example 5.3.21. Let now
X ⊂ TPn be a tropical k-cycle of sedentarity zero. We set X i = X ∩ Ui . On
each chart, we may consider the modification Mi := Mod(X i , f i |X i ). Since
f i is a polynomial, Mi is contained in Ui × T (versus Ui × TP1 ). Moreover, it
follows from Equation (5.5) that restricted to (Ui ∩ U j ) × T, Mi and M j are
identified via φi j . In more precise terms,
φi j∗ (M j |(Ui ∩U j )×T ) = Mi |(Ui ∩U j )×T .
It follows that the various modifications can be glued to a tropical k-cycle

Mod(X , F ) in O (d).
Definition 5.3.26
Let F ∈ T[x 0 , . . . , x n ] be a homogeneous polynomial of degree d and X ⊂
TPn a tropical k-cycle of sedentarity zero. We denote by X i = X ∩ Ui the
affine pieces of X . The modification Mod(X , F ) of X along F is the tropical
k-cycle in O (d) such that for any i = 0, . . . , n we have
Mod(X , F ) ∩ (Ui × T) = Mod(X i , f i ).
The previous discussion shows that Mod(X , F ) exists and is unique. We

can immediately extend the construction to a quotient of two homogeneous
polynomials.
Definition 5.3.27
Let G, H ∈ T[x 0 , . . . , x n ] be homogeneous polynomials and set d = deg(G)−
deg(H). Let X ⊂ TPn a tropical k-cycle of sedentarity zero. We denote by
X i = X ∩ Ui the affine pieces of X . The modification Mod(X , G, H) of X
163
along the quotient “G/H” is the tropical k-cycle in P (d) such that for any
i = 0, . . . , n we have
Mod(X , F ) ∩ (Ui × TP1 ) = Mod(X i , g i , hi ).
Example 5.3.28
Let us look at two examples of modifications of TP2 along linear polynomi-
als. Let us first consider the standard line given by f = “x 0 + x 1 + x 2 ”.
Then Mod(TP2 , f ) = V (“x 0 + x 1 + x 2 + x 3 ”) is just the standard plane
in TP3 (cf. Figure 5.4). Let us now consider the degenerated polynomial
g = “x 1 + x 2 ”, hence (0 : −∞ : −∞) ∈ V (g). Correspondingly, the mod-
ification Mod(TP2 , g) is still a plane in TP3 , but degenerated as well. In
particular, it contains the torus fixed points (0 : −∞ : −∞ : −∞). Both
examples are illustrated in Figure 5.5.
Figure 5.5: Modifying TP2 along two different lines
Remark 5.3.29
Clearly, both construction can be extended to arbitrary toric varieties W
as follows. Let d ∈ B be a toric degree which is Cartier (see Remark
5.3.25). Recall that OW (d) and PW (d) denote the corresponding (projec-
tive) line bundles. Let F be a homogeneous polynomial of degree d (see Re-
mark 3.4.8). There exist well-defined dehomogenizations f i ∈ T[ y1 , . . . , yn ]
164
(the translation part is fixed by requiring that if F has full support Pd ,

then all f i are monomialfree polynomials), and corresponding shifts Fi ∈
T[x 1± , . . . , x l± ]. It is easy to check that on overlaps Uϑ ∩ Uϑ0 , the transition
functions for the Fi coincide with the transition functions for OW (d). Hence
we can again define Mod(X , F ) ⊂ OW (d) as the tropical k-cycle such that
Mod(X , F ) ∩ (Uϑ × T) = Mod(X ∩ Uϑ , f i ) for all ϑ ∈ Ξ. In exactly the same
way we define Mod(X , G, H) ⊂ PW (d) for two homogeneous polynomials
G, H such that d = deg(G) − deg(H) is Cartier.
Definition 5.3.30
Two tropical k-cycles X ⊂ W ,Y ⊂ W 0 are isomorphic, denoted X ∼ = Y , if
there exist open toric subvarieties U ⊂ W and U 0 ⊂ W 0 such that X ⊂ U and
Y ⊂ U 0 and an integer affine isomorphism ϕ : U → U 0 such that ϕ∗ (X ) = Y .
We say X and Y are (modification) equivalent, denoted by X ≡ Y , if
there exists a chain of tropical k-cycles X = X 0 , X 1 , . . . , X l = Y , X i ⊂ Wi
such for each consecutive pair X i , X i+1 either X i ∼
= Mod(X i+1 , Fi ) or X i+1 ∼=
Mod(X i , Fi ) for suitable homogeneous polynomials Fi on Wi or Wi+1 , respec-
tively.
Exercise 5.3.31
Show that the tropical cycle D·X does not depend on the chosen polyhedral
structure X .
Exercise 5.3.32
Show that Mod(W, F ) = V (“ y + F (x)”). Show that Mod(X , F ) = “ y + F (x)”·
π−1 (X ) with π : OW (d) → W . Show that div(F ) = V (F ).
Remark 5.3.33
Cycles of higher sedentarity
5.3.5 Linear modifications

Theorem 5.3.34
deg(Mod(X , F )) = deg(X ), deg(div( f )) = deg
165
5.4 Configurations of hyperplanes in

projective space
A difference between classical and tropical varieties is that tropical varieties
always are equipped with a divisor of special points, the points of higher
sedentarity. The main meaning of modifcations and contractions is to add
or remove (sedentarity zero) hypersurfaces to this divisor if necessary.
As an example, let us regard tuples of three distinct points in TP1 . For
the classical projective line CP1 , all such tuples are equivalent, as we can
always find an automorphism of CP1 which maps the tuple to 0, 1, ∞.
In contrast, tropical TP1 has two distinguished points ±∞. They are
topologically different from all finite points and therefore the classical state-
ments can obviously not be translated directly. Indeed, as we will see later,
the automorphisms of TP1 are translations of the finite part by a real con-
stant (which keeps the infinite points fixed) and reflections at any real num-
ber (which exchanges the infinite points). We see that automorphisms do
not change the sedentarity of points. Instead, we are forced to use modifi-
cations here. Let us study this in more details.
First, we may fix 3 points in TP1 as reference points (like 0, 1, ∞ for CP1 ).
A natural choice is −∞, 0, ∞. After what we just said above, it is even
more natural to modify TP1 along 0 to get the line L = V (“x 0 + x 1 + x 2 ”)
in TP2 . Now all our three reference points p0 , p1 , p2 are infinite and are
symmetrically given by x 0 = −∞, x 1 = −∞ resp. x 2 = −∞. Now let
q0 , q1 , q2 ∈ L be any other configuration of three distinct points. Then we
can formulate the following statement.
Lemma 5.4.1
There exists a series of modifications and contractions of L to an isomorphic
line L 0 which transforms the points q0 , q1 , q2 to the infinite points p0 , p1 , p2 .
Proof. We modify along the points q0 , q1 , q2 , obtaining a line in TP5 with

coordinates x 0 , x 1 , x 2 , y0 , y1 , y2 . We now contract three times, namely we
forget the original coordinates x 0 , x 1 , x 2 of TP2 . We end up with a line L 0
in TP2 whose infinite points q00 , q10 , q20 correspond to q0 , q1 , q2 . L 0 is non-
degenerated as the points q0 , q1 , q2 are distinct, hence also q00 , q10 , q20 . Hence
166
q1
p0 p0
q1 q1 p0
q2 p2 p2
p2 modifying contracting q2
along the qi q2
q0 q0 the pi p1
p1
p1 q0
Figure 5.6: Arrangements of three points in TP1 via modifications
L and L 0 are isomorphic and the identification pi = qi0 satisfies the required
properties.
Basically the same is true for higher dimensions. Again as a reference we

may fix the standard hyperplane H ∈ TPn+1 with its n+2 planes P0 , . . . , Pn+1
at infinity. Let Q 0 , . . . Q n+1 be a another collection of hyperplanes in H. We
assume that this collection is generic, i.e. the intersection of any choice of
n + 1 of these hyperplanes is empty. Then the following is true.
Lemma 5.4.2
There exists a sequence of modifications and contractions of H to an isomorphic
hyperplane H 0 which transforms the chosen hyperplanes Q 0 , . . . , Q n+1 to the
infinite hyperplanes the infinite hyperplanes P0 , . . . , Pn+1 .
Under the assumption that each Q i is obtained as the stable intersection

H · H i of H with another hyperplane H i of TPn+1 , the proof of this statement
is completely analogous to the one-dimensional case. The assumption is
not necessarily satisfied, but these issues will be adressed more thoroughly
in the following chapters.
Instead of hyperplanes, we may also consider points in TPn . Let us con-
sider the case n = 2. Let p0 , p1 ∈ TP2 be two distinct points. Then there
exists a line L ⊂ TP2 containing p0 and p1 . The modification along L gives a
hyperplane H ⊂ TP3 , and we identify L with the line L = H ∩{x 3 = −∞} of
sedentarity (at least) 1. Consistently we identify p0 and p1 with the points
p0 , p1 ∈ L which are mapped to p0 resp. p1 by the contraction map. For any
point p ∈ TP2 \ L, there is a unique lift to H, denoted by p. Note that H can
be contracted in any of the 4 standard directions −e0 , . . . − e3 to projective
167
space TP2 . For a point configuration p0 , p1 , p2 , . . . , pn and a chosen contrac-

tion of H, the images of p0 , p1 , p2 , . . . , p n give a new point configuration in
TP2 . Two point configurations are called projectively equivalent if they can
be connected by a series of the above construction and automorphisms of
TP2 (generated by translations and permutations of the coordinates).
Lemma 5.4.3
Let p0 , . . . , p3 be 4 generic points in TP2 (i.e. no three of them are contained
in a line). Then this point configuration is tropically equivalent to the config-
uration (−∞ : 0 : 0), (0 : −∞ : 0), (0 : 0 : −∞), (0 : 0 : 0).
Proof. Let s = sed(p0 ) + sed(p1 ) + sed(p2 ) be the sum of the sedentari-

ties of the first three points. We are done if s = 6, because this implies
{p0 , p1 , p2 } = {(−∞ : 0 : 0), (0 : −∞ : 0), (0 : 0 : −∞)} and, as the pi are
generic, p3 must be a finite point in R2 . Thus, we can use an automorphism
of TP2 to reorder the first 3 points correctly and to move p3 to (0, 0, 0).
It remains to show that, if s < 6, there is always a choice of modification-
contraction as above such that the obtained tropically equivalent point con-
figuration is still generic and has sedentarity s0 > s. Then the claim follows.
So let us assume s < 6. Let L0 , L1 , L2 be the three lines passing through
{p1 , p2 }, {p0 , p2 } resp. {p0 , p1 }. Note that s < 6 implies
{L0 , L1 , L2 } =
6 {V (x 0 ), V (x 1 ), V (x 2 )}.
In other words, one of the lines, say L0 , must be of sedentarity zero. More-
over, there exists at least one coordinate lines, say V (x 0 ), which contains
at most one of the points p0 , p1 , p2 . We modify along L0 and then con-
tract V (x 0 ) (to be precise, we contract by forgetting the coordinate x 0 ).
Let H = Mod(TP2 , L0 ) be the modified plane in TP3 . By construction we
have sed(p0 ) = sed(p0 ) + 1, sed(p1 ) = sed(p1 ) + 1 and sed(p2 ) = sed(p2 ).
When contracting (let us call the contraction π), the sedentarity of a point
drops (by one) if and only if it is contained in the contracted line at infinity
V (x 0 ) = Mod(V (x 0 ), L0 ∩ V (x 0 )). By our choice, V (x 0 ) contains at most
one of the points p0 , p1 , p2 . It follows
sed(π(p0 )) + sed(π(p1 )) + sed(π(p2 )) ≥ s + 2 − 1 > s,
so indeed, we increased the sedentarity of the new point configuration.
168
s=3 s0 = 4
s=4 s0 = 5
Figure 5.7: Two examples of the modification-contraction process
It remains to check that π(p0 ), . . . , π(p3 ) are still generic. By construc-

tion, π(p0 ), π(p1 ), π(p2 ) ∈ / π(L ∞ ). Let M be a line passing through at least
3 of the points π(p0 ), . . . , π(p3 ). For any x ∈ M ∩ π(L ∞ ), we can choose
the “height” of Mod(M , x) such that Mod(M , x) ⊂ H. We choose x = π(p3 )
if π(p3 ) ∈ M , or any other x ∈ M ∩ π(L ∞ ) otherwise. Then Mod(M , x)
is a line in H which contains at least 3 of the points p0 , . . . , p3 . Contract-
ing along the original modification gives a line containing at least 3 of the
points p0 , . . . , p3 . This contradicts to the assumption that the p0 , . . . , p3 are
generic.
Similar arguments also work in higher dimensions.
5.5 Equivalence of rational curves

Let us anticipate the definition of an abstract tropical variety in its easiest
case, namely for (rational) tropical curves. An abstract tropical smooth curve
is a tuple C = (Γ , d), where Γ is a graph (i.e. a topological space homeo-
morphic to a one-dimensional simplicial complex) and d is a complete inner
metric on the “finite part” Γ ◦ := Γ \ ∂ Γ . Here ∂ Γ denotes the set of 1-valent
vertices of Γ (which is clearly independent of the chosen simplicial struc-
169
ture). It follows that, for a chosen simplicial structure, we call the edges
containing an 1-vertex ends of C. Other edges are called bounded edges. A
bounded edge is homeomorphic to an interval [0, l] ⊂ R, l > 0, and we call
l the length of the edge. An end is homeomorphic to [−∞, 0], where −∞
is identified with the 1-valent vertex. Therefore, C is completely described
its combinatorial graph Γ and by a positive real length l(E) for any bounded
edge E. Finally, C is called rational if Γ is a tree.
The reader might be surprised that the definition of an abstract curve does
not require any balancing condition. In fact, as we will later see, given the
valence of a point, there is only one smooth local model in dimension 1. For
a point of valence n+1, this local model is given by the standard line in TPn
(i.e. the 1-dimensional fan with n+1 rays pointing to the standard directions
−e0 , . . . , −en ) or T, if n = 0. Hence the tropical structure of a smooth curve
is completely determined by its graph and the balancing condition at each
(finite) point is somewhat hidden by the uniquness of local models.
An isomorphism of two abstract tropical curves C, D is a continuous map
C → D which restricts to an isometry C ◦ → D◦ .
Lemma 5.5.1
Any smooth rational tropical curve C is isomorphic to a sequence of modifica-
tion of TP1 along single points.
Proof. Assume that C = (Γ , d) has more than two 1-valent vertices. Then
we can contract one of them, i.e. we just remove the vertex and the interior
of the adjacent edge E to obtain a new abstract curve C 0 = C \ E. Let x ∈ C 0
be the point where E was attached to C 0 . When we modify C 0 along x, we
just glue an interval [−∞, 0] to C 0 (identifying 0 to x). Hence Mod(C 0 , x)
is isomorphic to C. Now we repeat this process until we end up with a curve
with only two ends. Clearly its graph must be linear and hence the curve is
isomorphic to TP1 , which proves the claim.
5.6 Equivalence of linear spaces

In classical algebraic geometry, a linear space of dimension n (embedded
in some big CPN ) is always isomorphic to CPn . In tropical geometry, lin-
ear spaces of a fixed dimension n, i.e. positive tropical n-cycles of degree 1
170
in some TPN ), can look quite differently. This is already happens for lines
where, aside from the most natural ”model“ TP1 , we may consider lines in
TPN with a bigger number of infinite points and more complicated (though
rational) graphs. Again, tropical equivalence generated by tropical modifi-
cations is the right notion to identify all these different models/embeddings
of tropical lines and linear spaces.
Lemma 5.6.1
Let L ⊆ TPN be a tropical linear space of dimension n. Then L is a multi-
ple modification of TPn . In particular, all linear spaces of dimension n are
tropically equivalent to each other.
Proof. We choose a torus fixed point p ∈ TPN such that p ∈ / L and project
along p. More precisely, we consider the projection map π : TPN \ {p} →
TPN −1 which in each affine chart not containing p is just given by TN → TN −1
forgetting one coordinate (say, after reordering, the last one). Let L 0 = π(L)
be the image of L, For each point x ∈ L 0 , the preimage π−1 (x) ∩ L is either
a single point or an interval of the form [−∞, a]. In any other case, (e.g.
an interval) one can the existence of a nearby fiber which contains at least
two isolated points which contradicts the assumption that L is of degree
1. It follows that L, equipped with trivial weights, is a linear space again.
Moreover, a generic fiber π−1 (x)∩L consists of only one point, say x̃, and we
may define, in each affine chart, a piecewise linear function f on L 0 whose
value at a generic x is given by the last coordinate of x̃. Let us now modify
L 0 along f (in each chart). By construction, the graph of f is contained
in L and ... Therefore L is in fact equal to Mod(L 0 , f ). If L 0 ( TPN −1 , we
continue until we finally end up with TPn .
171
6 Tropical cycles and the
Chow group
6.1 Tropical cycles
We defined tropical cycles in toric varieties in section 5.1. The main idea
was that tropical cycles should be polyhedral sets (with rational slopes)
whose generic points carry multiplicities such that around each codimen-
sion one cell the balancing condition holds. As all these requirements are
of a local nature, it is easy to extend the definition to arbitrary tropical
varieties.
Let V be a tropical variety. A subset X ⊆ V is called a polyhedral set
if it is (finite closed) polyhedral in any chart. Recall that in charts with
points of higher sedentarity, we define a polyhedron to be the closure of a
usual polyhedron in Rn . Let x ∈ X be a point. We define the speciality of
x, denoted by specX (x) or just spec(x), to be the minimal codimension of
the polyhedra P such that x ∈ P ⊆ X . Points with specX (x) = 0 are called
generic points. Let X be of pure dimension m. Then the closure of all points
of speciality m − k is called the k-skeleton of X , denoted by X (k) .
X (k) := {x ∈ X : spec(x) = m − k} = {x ∈ X : spec(x) ≥ m − k}.
A polyhedral set is called weighted if it is equipped with a locally constant

function mult : X gen → Z \ {0}, where X gen denotes the set of generic points
of X . If P ⊆ X is a polyhedron of maximal dimension m, then we also write
mult(P) for the number mult(x) for any x ∈ P .
Definition 6.1.1
Let V be a tropical variety. A tropical k-cycle X of V is a weighted polyhedral
set X ⊆ V of pure dimension k such that for any chart U ∈ V , any polyhedral
172
6 Tropical cycles and the Chow group
structure of X ∩ U and any codimension one cell τ ⊂ X ∩ U the balancing

condition X
mult(σ)vσ/τ = 0 mod Rτ
τ⊂σ facet
sed(τ)=sed(σ)
is satisfied.
X is called of pure sedentarity if all its generic points have the same seden-
tarity (in V ).
X is called effective if all its weights are positive, i.e. mult : X gen → N.
Remark 6.1.2
Note that an effective tropical cycle X ⊆ V is a tropical space itself by re-
stricting the charts of V to X . In other words, in our terminology effective
tropical cycles are just the closed tropical subvarieties of V . In particular, V
satisfies the requirement of a cycle itself. Hence V is the fundamental cycle
of itself.
Given two tropical cycles X 1 and X 2 , we can form the sum X 1 +X 2 . We just
take the union X 1 ∪ X 2 (which is again a polyhedral set) and add weights.
This means for a generic point of x of X 1 ∪ X 2 we set multX 1 +X 2 (x) =
multX 1 (x) + multX 2 (x) (where multX (x) = 0 if x ∈ / X ). If this sum turns
out to be zero, we just remove the point. Thus, in general, X 1 + X 2 is only
supported on a subset of X 1 ∪ X 2 . It is straightforward to check that X 1 + X 2
still satisfiesL
the balancing condition. So the set of all cycles in V , denoted
by Z∗ (V ) = k Zk (V ), forms a group under addition with neutral element
the empty cycle 0 := ;.
One further remark: As above, we will always ignore points of weight
zero. This is to say, whenever a construction (like summing two cycles)
produces zero weights, we just discard these points.
6.2 Push-forwards of tropical cycles

Let V and W be tropical varieties and let f : X → Y be a tropical morphism.
Given a tropical cycle X ∈ Zk (V ), we define its push-forward f∗ (X ) ∈ Zk (W )
as follows. Let y ∈ f (X ) be a point such that X y := f −1 ( y) ∩ X is isolated
and generic in X (in particular, y is generic in f (X ) and X y is finite). Fix
173
x ∈ X y and choose charts around x and y. Then f induces a map of lattices

d f xZ : TxZ X → T yZ f (X ), and we define
X
ω f∗ (X ) ( y) := [T yZ f (X ) : Im(d f xZ )] · ωX (x). (6.1)
x∈X y
Proposition 6.2.1
In the situation above, there is a unique k-cycle supported on f (X ) whose
weight function agrees with (6.1) for sufficiently generic points. This cycle is
called the push-forward of X , denoted by f∗ (X ) ∈ Zk (W ).
Proof. f (X ) is a polyhedral set in W . If dim( f (X )) < k, then f∗ (X ) = 0. If
dim( f (X )) = k, let Y denote its k-dimensional part. Then equation (6.1)
defines a locally constant weight function on a open polyhedral dense sub-
set U ∈ Y . We have to show that this weight function satisfies the balancing
condition.
Choose y ∈ Y and let S denote the part of StarY ( y) given by points of
the sedentarity (in Y ) . We have to show that S with weight function (6.1)
is balanced. We use the following locality statement. Let X y := f −1 ( y) ∩ X
and let Z ⊆ X y be the set of vertices of X y . For each z ∈ Z, let Sz denote the
sedentarity part of StarX (z) as above. Then, for y 0 ∈ S, we have
X
ω f∗ (X ) ( y ) =
0
ω f∗ (Sz ) ( y 0 ).
z∈Z
This follows form the fact that when we let y 0 converge to y, then the
preimage points in X y 0 have to converge to points in Z. Using this equa-
tion, we can assume that X is a fan, and f is integer linear. The balancing
condition is a condition for the ridges of f (X ). Hence, by applying locality
one more time and using the fact that dividing by a lineality space is com-
patible with the lattice index showing up in (6.1), we can assume that X is
one-dimensional. We denote by uρ the primitive generator of a ray ρ. For
any ray ρ 0 of Y we have
X
ω f∗ (X ) (ρ 0 ) = [TρZ0 f (X ) : Im(d fρZ )]ωX (ρ).
ρ⊆X
f (ρ)=ρ 0
Note that the primitive generators are related by

f (uρ ) = [TρZ0 f (X ) : Im(d fρZ )]uρ0 .
174
Hence the balancing condition for Y follows from

X X X
ω f∗ (X ) (ρ 0 )uρ0 = [TρZ0 f (X ) : Im(d fρZ )]ωX (ρ)uρ0
ρ 0 ⊆Y ρ 0 ⊆Y ρ⊆X
f (ρ)=ρ 0
X
= ωX (ρ) f (uρ ) (6.2)
ρ⊆X

X
=f ωX (ρ)uρ = f (0) = 0.
ρ⊆X
6.3 Linear equivalence of cycles

Let V be a tropical space of pure dimension n and consider the variety
V × TP1 . Let Z be a cycle in V × TP1 which is the closure of a cycle in
V × R Then Z can be intersected with V−∞ := V × {−∞}. Namely, in every
chart U × T we can use Definition 5.3.4. It is easy to check that the results
agree on the overlaps and therefore can be glued together to give the cycle
Z−∞ := Z · V−∞ . As V × {−∞} = V , we think of Z−∞ as a cycle in V
(of dimension dim(Z) − 1). In the same way, we can construct Z+∞ :=
Z · Z+∞ . This construction suffices to translate the classical definition of
linear equivalence to the tropical world.
Definition 6.3.1
Let X 1 , X 2 be two k-cycles in the tropical space V . Then X 1 and X 2 are
called linearly equivalent, denoted by X 1 ∼ X 2 , if there exists a (k + 1)-cycle
Z ⊆ V × TP1 such that
• Z is the closure of a cycle in V × R,
• X 1 = Z−∞ , and
• X 2 = Z+∞ .
Lemma 6.3.2
The relation ∼ defined in the previous definition is an equivalence relation.
175
Furthermore, we have
X 1 ∼ X 2 , Y1 ∼ Y2 ⇒ X 1 + Y1 ∼ X 2 + Y2 .
Proof. To show X ∼ X , we take Z = X × TP1 . The symmetry of ∼ follows

from the symmetry TP1 → TP1 : x 7→ −x. To show the compatibility with
sums, let Z be the cycle in V × TP1 showing X 1 ∼ X 2 (and, analogously, Z 0
for Y1 ∼ Y2 ). Then the sums Z + Z 0 shows X 1 + Y1 ∼ X 2 + Y2 . Finally, using
the additivity twice, we find
X 1 ∼ X 2, X 2 ∼ X 3 ⇒ X 1 + X 2 ∼ X 2 + X 3 ⇒ X 1 ∼ X 3,
6.4 Algebraic equivalence of cycles

As in classical algebraic geometry, we may replace TP1 by any other smooth
tropical curve to obtain another, weaker equivalence relation for tropical
cycles. Let C be a smooth tropical curve, and let V be a tropical space. For
a cycle Z ⊆ V × C, we have to make sense of
Z p = Z · (V × {p})
for any p ∈ C. If sed(p) = 1, C looks like T near p and we can use Definition
5.3.4 again (assuming that Z is the closure of a cycle in V × C \ {p}). If
sed(p) = 0, then C locally near p looks like the line L ⊆ Rm with single
vertex sitting at 0. So we can pull-back the function max{x 1 , . . . , x m , 0} to
V × U, where U is a neighbourhood of p. Let us denote this pull-back by
ϕ p . Then we define Z p = div(ϕ p | Z∩V ×U ).
Definition 6.4.1
Let X 1 , X 2 be two k-cycles in the tropical space V . Then X 1 and X 2 are called
alg
algebraically equivalent, denoted by X 1 ∼ X 2 , if there exists a tropical curve
C, two points p1 , p2 ∈ C and a (k + 1)-cycle Z ⊆ V × C such that
• Z is the closure of a cycle in V × C \ {p1 , p2 },
• X 1 = Z p1 , and
176
• X 2 = Z p2 .
Lemma 6.4.2
alg
The relation ∼ defined in the previous definition is an equivalence relation.
Furthermore, we have
alg alg alg
X 1 ∼ X 2 , Y1 ∼ Y2 ⇒ X 1 + Y1 ∼ X 2 + Y2 .
alg
Proof. It suffices to show that the cycles X such that X ∼ 0 form a sub-
group of Z∗ (V ). So let X 1 , X 2 be two tropical cycles which are algebraically
alg
equivalent to zero. We want to show that X 1 − X 2 ∼ 0. According to our
definition, there exists a curve C1 , two points p, p0 and a cycle Z1 ⊆ V × C1
such that (Z1 ) p = X 1 and (Z1 ) p0 = 0. Analogously, we find C2 , q, q0 ∈ C2 and
Z1 ⊆ V × C1 such that (Z2 )q = X 2 and (Z2 )q0 = 0.
6.5 Linear systems

Let V be a tropical space of pure dimension n. Then a (Weil-)divisor D ⊂ V
is a subcycle of dimension n − 1. Therefore, the group of all divisors is
Zn−1 (V ).
Definition 6.5.1
Let V be a tropical variety and let D ⊂ V be a divisor. Then the set of all
effective linearly equivalent divisors
L(D) = |D| := {D0 ≥ 0 : D0 ∼ D}
is called the complete linear system of D.
6.6 Recession Fans

Exercise 6.6.1
Let σ ⊂ Rn be a polyhedron. Show that the set of faces of rc(σ) is equal to
the set of cones rc(τ) for all faces τ ⊂ σ.
177
Exercise 6.6.2
Let S be a polyhedral subdivision of Rn . A path of cells in S is a sequence
σ0 , . . . , σn , σi ∈ S , such that for any i either σi ⊂ σi+1 or σi+1 ⊂ σi . Let σ
and σ be celly of S such that v ∈ rc(σ) ∩ rc(σ0 ) 6= ;.
(a) Show that there is a path of cells connecting σ and σ0 such that v ∈
rc(σi ) for all i. You might want to reduce to the 2-dimensional case
by choosing two generic points in σ and σ0 .
(b) Assume that v ∈ rc(σ) . Conclude that the rc(σ) is equal to the face
of rc(σi ) containing v for all i.
(c) Assume that v ∈ rc(σ) ∩ rc(σ0 ) . Conclude that rc(σ) = rc(σ0 ).
(d) Prove that {rc(σ) : σ ∈ X } forms a fan.
178
7 Tropical manifolds
Lemma 7.0.1
Let U ⊂ Rn be open and connected and let f : U → R be a smooth function.
Then the following conditions are equivalent.
(a) For all x ∈ U and v ∈ Zn , we have d f x (v) ∈ Z.
(b) f is integer affine, i.e., of the form
f (x) = a + j x,
with a ∈ R and j ∈ Zn .
Proof. Let f be a function satisfying the first condition. This means d f x ∈

(Zn )∗ = Zn for all x ∈ U. As U is connected and Zn is discrete, d f x is constant
to, say, j ∈ Zn . It follows that f is of the form a + j x. The other implication
is trivial.
In the previous chapters all the objects we considered were embedded in

some ambient variety, mostly Rn or one of its toric compactifications, e.g.
Tn or TPn . We will now proceed to define what we mean by an abstract
tropical space. We will do this by following the standard strategy using
an atlas of charts and gluing maps. We will proceed step by step, starting
with a very general definition allowing rather strangely behaved spaces and
ending with the quite restrictive definition of smooth tropical spaces.
7.1 Tropical spaces

To get started, we have to describe the local building blocks we want to
use in order to create abstract tropical spaces. We will give the relevant
definitions here, noting that many of the notions we considered before are
179
of purely local nature. For the reader’s convenience, we repeat some ter-
minology which has appeared before. Recall that for any index set I ⊂ [n],
we denote by R I ⊂ Tn the torus orbit of points with coordinates x i = ∞ for
all i ∈ I.
Definition 7.1.1 (Open polyhedral sets)

A polyhedron in Tn is the closure of a (usual) polyhedron in one of the torus
orbits R I ≡ Rn−|I| . A union of polyhedra in Tn is called a polyhedral set in Tn .
An open polyhedral set X is defined to be the open subset of a polyhedral
set Y in Tn .
A point p ∈ X is called generic if there exists an open neighbourhood of
p homeomorphic to an open subset in affine space Rm . The set of generic
points is denoted by X gen . The integer m is called the dimension of X at p,
denoted dimX (p). Hence dimX : X gen → N is a locally constant function.
Its maximum is called the dimension of X , denoted by dim(X ). If dimX is
globally constant, we say X is of pure dimension. Let X be pure-dimensional
and p ∈ X . The codimension of p in X , denoted by codimX (p), is the min-
imum of the codimension of the cell whose relative interior contains p for
any polyhedral structure of X .
As before, we now equip open polyhedral sets with weights.
Definition 7.1.2 (Weighted open polyhedral sets)

Let X be an open polyhedral set. A weight function on X is a locally constant
function ω : X gen → Z, i.e. a choice of integer weight for every connected
component of X gen . The set X equipped with a weight function is called
a weighted open polyhedral set. A weighted open polyhedral set is called
effective if all weights are non-negative.
In the following we will always assume that all weights are non-zero. If
zero weights show up (for example when adding to weighted open polyhe-
dral sets), we remove them by taking the closure of ω−1 (Z \ {0}) in X . This
is again an open polyhedral set.
We continue with the local formulation of the balancing condition.
Definition 7.1.3 (Open tropical cycles)

Let X be an open polyhedral set of pure dimension m and let p ∈ X ∩R I be a
point of sedentarity I. We define the star of X at p, denoted by StarX (p), to
180
the fan in R I containing all direction vectors v for which p + εv ∈ X holds

for small ε > 0. The m-dimensional part of StarX (p), i.e. the closure of
the set dim−1
StarX (p)
(m), is denoted by StarX (p)(m) . If non-empty, StarX (p)(m)
inherits a weight function from the weights around p. We call X a balanced
open polyhedral set or open tropical cycle if StarX (p)(m) is empty or balanced
(according to 2.4.6 and 4.2.2) for all p ∈ X . In other words, after choosing
a (open) polyhedral structure for X , we may check the balancing condition
as in Equation (5.1).
Note that the balancing condition at point p ∈ R I only takes into account
the full-dimensional parts of X near p contained in R I . In other words,
balancing is checked in the locus of points of same sedentarity.
Balanced open polyhedral sets are the local building blocks for tropical
spaces. Let us now consider the maps we use to glue these blocks together.
Recall from chapter 3 that in tropical arithmetics, monomial maps from
R to Rm correspond simply to linear maps which map Zn to Zm . We call such
n
maps Z-linear maps. In chapter 2 we discussed that the tropical structure

of Rn manifests through the data of integer tangent vectors TxZ Rn ⊂ Tx Rn at
each point x ∈ Rn (cf. Definition ). Obviously, Z-linear maps preserve this
tropical structure. However, as we only fix the lattices of integer tangent
vectors and not the lattice Zn ⊂ Rn itself, we may combine the monomial
maps with arbitrary translations with translation vector v ∈ Rm . Still, the
tropical structure of integer tangent vectors is preserved. Such maps are
called affine Z-linear maps. Conversely, any smooth map F : Rn → Rm
whose differential is Z-linear at each point is affine Z-linear, as Mat(n, m, Z)
is discrete (cf. Lemma 7.0.1).
Definition 7.1.4 (Tropical morphisms for open tropical cycles)

A map F : Rn → Rm is called affine Z-linear if it is the sum of a Z-linear map
Rn → Rm and a translation by an arbitrary vector v ∈ Rm .
A map f : Rn → Tm is called affine Z-linear if it is the composition of
an affine Z-linear map Rn → R I and the inclusion map R I ,→ Tm , for a
suitable torus orbit R I . Note that we also include the constant map f ≡
(−∞, . . . , −∞) corresponding to I = [n].
Let X ⊆ Tn and Y ⊆ Tm be two open polyhedral sets. A map f : X → Y
is called a tropical morphism if it is continuous and locally affine Z-linear.
This means that for each point p ∈ X there exists an open neighbourhood
181
U of p in X such that for all I ⊂ [n] with U ∩ R I 6= ;, the map f |U∩RI is the
restriction of an affine Z-linear map R I → Tm .
Assume moreover that X and Y are weighted. Then f is called a tropical
isomorphism if there exists an inverse tropical morphism g : Y → X and,
moreover, if
ω(x) = ω( f (x))
for all x ∈ X gen . Note that the existence of g implies that f (x) is a generic
point of Y .
Example 7.1.5
The group of tropical automorphisms of Rn is the semidirect product
Aut(Rn ) = GL(n, Z) n Rn ,
where the first factor represents Z-invertible linear maps and the second
factor parameterizes translations.
Example 7.1.6
The group of tropical automorphisms of Tn is the semidirect product
Aut(Tn ) = Sn n Rn ,
where the symmetric group represents permutations of the variables and

the second factor parameterizes translations.
Note that there are some subtleties hidden in the definition of tropical
morpshisms when higher sedentarity points are involved. The following
example shows some of that behaviour.
Example 7.1.7
We will compare four polyhedral sets in T2 . Let u, v, w ∈ Z2+ be integer vec-
tors with all entries positive and assume that v, w are linearly independent.
We define X 1 = T{1} ∪ T{2} (union of the two coordinate axes at infinity),
X 2 = T{1} ∪ Tu (one coordinate axis plus a ray), X 3 = Tu ∪ (Tu + p) for some
p∈ / Tu (two parallel rays) and X 4 = Tv ∪ Tw (two non-parallel rays). All of
these four sets contain the corner point o = (−∞, −∞) and we are inter-
ested in whether or not two such sets are isomorphic in a neighbourhood
of o. It turns out that such isomorphic neighbourhoods exist for X 1 and X 2 ,
182
but for none of the remaining pairs. Note that some authors in other con-
texts use different definitions for tropical morphisms which may discard or
allow for extra isomorphisms in this example.
We are now ready to introduce the rather general notion of a tropical

space, obtained from glueing balanced open polyhedral sets. Later on, we
will add more and more additional requirements to these spaces to make
them more manageable.
Definition 7.1.8
Let X be a topological space. A tropical atlas or tropical structure on X is a
collection of tuples (Ui , ψi , Vi )i subject to the following constraints:
• X = i Ui is an open covering of X .
S
• For each i, Vi ⊂ TN (for suitable N ) is an effective balanced open

polyhedral set.
• For each i, ψi : Ui → Vi is a homeomorphism.
• For each pair i, j with Ui ∩ U j 6= ;, the composition map
ψi ◦ ψ−1
j
: ψ j (Ui ∩ U j ) → ψi (Ui ∩ U j )
is a tropical morphism with ω(x) = ω(ψi ◦ ψ−1 j

(x)) whenever x and
ψi ◦ ψ j (x) are generic. In other words, ψi ◦ ψ j is a tropical isomor-
−1 −1
phism with inverse ψ j ◦ ψ−1i

.
Two atlases on X are called equivalent if their union also forms an at-
las. A tropical space is a topological space X together with the choice of an
equivalence class of atlases.
If all Vi are of pure dimension n, we say that X is of dimension n.
Let us list some obvious properties of a tropical space X .
• X contains an open dense subset of generic points X gen .
• The weights on the various charts are compatible and therefore glue
to give a locally constant weight function on X gen .
183
• For each point p ∈ X , the codimension codimX (p) is well-defined as

the codimension of ψi (p) in Vi for any chart containing p. A point
p ∈ X is generic if and only if codimX (p) = 0.
• For each point p ∈ X , we can consider StarX (p) ⊆ RN (for suitable N ).

If non-empty, its n-dimensional part StarX (p)(n) is a balanced fan of
dimension n. It is well-defined only up to tropical isomorphisms of
fans given by Z-linear maps in Mat(N , N 0 , Z).
By definition, tropical spaces can locally be stratified as rational polyhe-

dral fans. Note however that the definition does not imply the existence
of global stratification (e.g. as a CW complex) into rational polyhedral cells
(i.e. cells which, locally in each chart, are rational polyhedral). Here are is
a example.
Example 7.1.9
Let Λ ⊂ R2 be the lattice generated by the vectors (1, π) and (1, e). Then
X = R2 /Λ is a tropical space, with charts given by restrictions of the pro-
jection map R2 → X . Note that here we stick to integer affine structure on
R2 (and hence also on X ) given by the standard lattice Z2 ⊂ R2 . It follows
that X cannot be stratified using rational polyhedra (it can, however, be
triangulated by non-rational triangles).
Definition 7.1.10
Let X be a tropical space and let Y ⊂ X be a subset. Then Y is called a
polyhedral subset of X if Y is closed in X and if the image set ψ(Y ∩ U)
under every chart (U, ψ, V ) of X is an open polyhedral set in TN .
Polyhedral subsets often arise as the sub- resp. superlevel sets of the fol-
lowing class of functions.
Definition 7.1.11
Let X be a tropical space and let f : X → M be a function to some set M .
Then f is called semiconstant if it is constant on the relative interior of every
cell of every polyhedral structure of every chart of X . Moreover, let ≤ be
a total order on M . Then f is called lower (resp. upper) semiconstant if for
every containment of cells τ ⊂ σ we have f (τ ) ≤ f (σ ) (resp. f (τ ) ≥
f (σ )).
184
We have the following straightforward statement.
Proposition 7.1.12
Let f : X → M be a lower (resp. upper) semiconstant function on at tropical
space X and choose m ∈ M . Then the sublevel (resp. superlevel) set
{p ∈ X : f (x) ≤ m} (resp. {p ∈ X : f (x) ≥ m})
is a polyhedral subset of X .
A typical example of a upper semiconstant function on X is the codi-

mension function codimX . It gives raise to a corresponding filtration and
stratification of X , which we collect in the following definition.
Definition 7.1.13
For k ∈ N, the superlevel set
X (k) := {x ∈ X | codimX (p) ≥ n − k}
is called the k-skeleton of X . By the previous proposition, X (k) is a polyhedral

subset of X . A connected component of
X (k) \ X (k−1) = {x ∈ X | codimX (p) = n − k}
is called a combinatorial stratum of X . The collection of these strata is called

the combinatorial stratification of X .
In most situations, we would like to impose some kind of finite type con-
dition on tropical spaces. This regards the local polyhedral sets Vi (a priori,
we might allow polyhedral sets which have an infinite number of cells), the
number of charts needed to give an atlas of X and, finally, the completeness
of X (open polyhedral sets can just stop at some bounded distance — this
is usually undesirable for tropical spaces). We will address all these issues
in the following definition.
Definition 7.1.14
A tropical space X is called of finite type if it admits a tropical atlas {(Ui , ψi , Vi )}i
subject to the following conditions.
185
(a) The number of charts is finite, i.e. i ∈ I is taken from a finite index set
I.
(b) Each Vi is the open subset of a finite polyhedral set, i.e. the union of
finitely many polyhedra. This was implicit in the general definition,
but we want to emphasize it here.
(c) Each chart can be extended, i.e. there exists a chart (Ui0 , ψ0i , Vi0 ) such
that Vi ⊂ Vi0 for all i. The closure is taken in TN .
A Hausdorff tropical space of pure dimension and of finite type is called

a tropical variety.
Example 7.1.15
A (finite) balanced effective polyhedral set in Tn is a tropical space of finite
type. More general, tropical cycles, considered as polyhedral sets, in any
toric variety are examples of tropical spaces of finite type. Note however
that in order to satisfy the third condition from above, it might be neces-
sary to split a chart in Tn into several smaller ones. For example, tropical
cycles in Rn are of finite type, but in general we have to use more than one
embedding Rn ⊂ Tn to satisfy the third condition.
Example 7.1.16
A simple example of a tropical space which is not of finite type is the unit
interval (0, 1) ⊂ T.
Example 7.1.17
Here is a strange example of a tropical space of finite type — we call it ρ
space, denoted X ρ . It is obtained from TP1 simply by glueing the points
+∞ and 0. The glued point is denoted O ∈ X ρ . The result is depicted in
Figure 7.1. This quotient space can be made a tropical space by using a
neighbourhood of (−∞, 0) ∈ V (“x 1 (x 2 + 0)”) ⊂ T2 as a chart for O. Note
that in a certain sense the point O has sedentarity 0 and 1 at the same time,
so for general tropical spaces we need to be careful when we speak about
sedentarity.
186
−∞
Figure 7.1: The so-called ρ-space X ρ
7.2 Regularity at Infinity

As we know from chapter 5, tropical varieties come naturally in classes
given by the process of modification. This reflects the fact that, when com-
paring tropical to classical varieties, they correspond not only to a variety
but to a tuple (X , D) where X is a variety and D ⊂ X is a distinguished divi-
sor. In the tropical world, this divisor is given by points of higher sedentarity
and can be changed by modifications. Unlike in the classical world, we can-
not completely forget this additional choice of divisor. As a consequence,
when we want to define a good class of tropical spaces, we are forced to
think about requirements imposed on the divisor at infinity. Probably the
most important property to impose is regularity at infinity. It should be
compared to the normal crossing divisor property when considering a tuple
(X , D) in the classical world.
We start by proving the local version of the fact that the locus of points
of sedentarity k has codimension at most k.
Proposition 7.2.1
Let X ⊆ TN be an open polyhedral set of pure dimension n. Let R I be a torus
orbit of TN such that X ∩ R I 6= ;. Then it holds
dim(X ∩ R I ) ≥ n − |I|.
Proof. Choose a polyhedral structure for X . Then the cells contained in

R I form a polyhedral structure for X ∩ R I . Choose a cell τ of X ∩ R I of
maximal dimension, i.e. dim(τ) = dim(X ∩ R I ). As X is pure-dimensional,
there exists a cell σ of X of dimension n containing τ. The maximality of τ
187
implies σ ∩ R I = τ. Let RJ be the torus orbit containing the relative interior

of σ (we have J ⊂ I). Then we can apply 3.2.6 to σ ⊂ TJ and obtain
n = dim(σ) = dim(τ) + dim(Rσ ∩ 〈ei : i ∈ I \ J〉) ≤ dim(X ∩ R I ) + |I|.
Remark 7.2.2
Note that when replacing R I by T I , the previous proposition is note true
anymore (at least as long as we do not impose the balancing condition or
something similar). A simple counterexample is given in Figure 7.2. For
the same reason, the proof of the proposition can not proceed inductively,
i.e. can not be reduced to the case |I| = 1.
R2 T2
X ∩ R2 X
Figure 7.2: A 2-dimensional polyhedron with a single vertex at infinity
We now consider the case when the inequality of proposition 7.2.1 is

sharp.
Proposition 7.2.3
Let X ⊆ TN be an open polyhedral set of pure dimension n. Then the following
conditions are equivalent.
(a) For each torus orbit R I ⊂ TN such that X ∩ R I 6= ;, we have
dim(X ∩ R I ) = n − |I|.
(b) For each stratum T I ⊂ TN such that X ∩ T I 6= ;, we have

dim(X ∩ T I ) ≤ n − |I|.
188
(c) For each point p ∈ X , we have

sedTN (p) ≤ codimX (p).
Proof. (a) ⇒ (c). Let p ∈ X be an arbitrary point. Assume p ∈ R I . Note that

a polyhedron in TN has the same sedentarity as any of point in its relative
interior, hence dim(X ∩ R I ) ≥ n − codimX (p). We get
codimX (p) ≥ n − dim(X ∩ R I ) = |I| = sedTN (p).
(c) ⇒ (b). Assume X ∩ T I 6= ;. Choose a polyhedral structure of X and
a point p ∈ P , where P is a cell of X ∩ T I of dimension dim(X ∩ T I ). It
follows dim(X ∩ T I ) = n − codimX (p). But codimX (p) ≥ sedTN (p) ≥ |I| by
assumption, and hence dim(X ∩ T I ) ≤ n − |I|.
(b) ⇒ (a). Assume X ∩ R I 6= ;. We have X ∩ R I ⊂ X ∩ T I and hence
dim(X ∩ R I ) ≤ dim(X ∩ T I ) ≤ n − |I|.
But by 7.2.1 we also have dim(X ∩ R I ) ≥ n − |I|, and hence equality holds.
Definition 7.2.4
Let X ⊆ TN be an open polyhedral set of pure dimension n. Then X is called
regular at infinity if the equivalent conditions of 7.2.3 hold. A tropical space
X is called regular at infinity if it admits a tropical atlas {(Ui , ψi , Vi )}i where
all Vi are regular at infinity.
Example 7.2.5
Let us give a few examples.
(a) The polyhedron depicted in Figure 7.2 is regular at infinity.
(b) The hyperplane H := V (“x 1 + x 2 + x 3 ”) ⊂ T3 is not regular at infinity,

because at p = (−∞, −∞, −∞) ∈ H we have sedT3 (p) = 3 > 2 =
codimH (p). Note that the “divisor” of points of higher sedentarity
consists of the 3 lines H ∩ Ti , which meet at p.
(c) Consider the line L = V (“x 1 + x 2 ”) ⊂ T2 . As a open polyhedral

set, L is not regular at infinity, as sedT2 ((−∞, −∞)) = 2 > 1 =
codim L ((−∞, −∞)). However, as tropical space, we may instead use
the chart L ∼
= T1 and hence the tropical space L is regular at infinity.
189
(d) Let Y be a tropical toric surface and let X ⊂ Y be a curve (i.e. an

effective 1-cycle) such that X ∩ R2 6= ;. Then the tropical space X is
regular at infinity if and only if it does not contain any of the boundary
curves and if for each torus fixed point q there is at most one ray
in X ∩ R2 converging to q. In this case, any point p ∈ X of higher
sedentarity admits a chart homeomorphic to a neighbourhood of −∞
in T.
(e) Let C be a one-dimensional tropical space. Then C is regular at infinity

if and only if can be covered by charts which are isomorphic to an open
subset of T or a 1-cycle in RN .
(f) The ρ-space constructed in 7.1.17 is not regular at infinity.
(g) The integer-affine manifolds defined in 1.3.2 are examples of tropical

spaces which are regular at infinity. Indeed, all charts are open subsets
of Rn and hence trivially regular at infinity.
As suggested by these examples, one property of tropical spaces regular
at infinity is that the notion of sedentarity makes sense. Indeed, note that
for an open polyhedral set X ⊂ TN regular at infinity, we have X gen ⊂ RN .
Given this it follows easily that the sedentarity sedTN (p) does not change
under an isomorphism of open polyhedral sets regular at infinity. We can
hence make the following definition.
Definition 7.2.6
Let X be a tropical space which is regular at infinity. For p ∈ X we define
the sedentarity of p in X to be
sedX (p) := sedTN (ψ(p)),
where (U, ψ, V ) is a chart with p ∈ U and V regular at infinity.
Example 7.2.7
The class of tropical spaces X which are regular at infinity and such that
every point satisfies codimX (p) = sedX (p) = 0 coincides with the class of
integer-affine manifolds defined in 1.3.2.
The main feature of regularity at infinity is that each point of sedentarity
k admits a chart which factors into a product of a sedentarity 0 part and
190
a Tk -part. Before we formulate this, let us fix some notations. Let X be a

tropical space which is regular at infinity. Note that the sedentarity function
sedX is another example of an upper semiconstant function on X . For k ∈ N,
the superlevel set
X [k] := {p ∈ X : sedX (p) ≥ k}
is called the k-sedentarity locus of X . By 7.1.12, X [k] is a polyhedral subset
of X . We also use the shorthand X ∞ := X [1] .
Theorem 7.2.8
Let X be a tropical space which is regular at infinity. Then each point p ∈ X
of sedentarity k has a chart isomorphic to U × V ⊆ X [k] × Tk , where U and V
are open neighbourhoods of p resp. (−∞, . . . , −∞) in X [k] resp. Tk .
Proof. As we want to prove a local statement, it we may restrict to the case

where X is an open polyhedral set in TN which is regular at infinity. Choose
I such that X ∩ R I 6= ; and consider X 0 := (X ∩ R I ) × T I ⊂ T I × T I = TN . Reg-
ularity at infinity implies that X 0 is a polyhedral set of dimension dim(X ).
It follows from 3.2.6 that the intersection X ∩ X 0 contains an open neigh-
bourhood of P ∩R I both in X and X 0 . As the topology on TN agrees with the
product topology on T I ×T I , this implies that every p ∈ X ∩R I admits a open
neighbourhood in X of the form U × V where U is a open neighbourhood of
p in X ∩ R I and V is a open neighbourhood of (−∞, . . . , −∞) in T I . This
is what we asked for, as X ∩ R I is obviously an open subset of X [|I|] .
Let Y ⊂ RN be a open polyhedral set. Note that a weight function on

the open polyhedral set Y × Tk induces a weight function on Y (and vice
versa). Moreover, if this weight function satisfies the balancing condition
(on Y × Tk ), the induced weight function is also balanced (on Y ). Hence,
with the help of the previous theorem we can turn the k-skeletons into
tropical spaces in their own right.
Definition 7.2.9
Let X be a tropical space which is regular at infinity. Then we define a
weight function on the k-skeleton X [k] by defining the weight of a generic
point p ∈ X [k] to the weight of a nearby generic point of X . This turns X [k]
into a tropical space of pure dimension n − k. We call X ∞ the divisor at
infinity of X .
191
Theorem 7.2.8 can be refined as follows. Note that by our convention

StarX (p) only takes into account points of same sedentarity, i.e. we have
StarX (p) = StarX [sed(p)] (p) ⊂ RN . Hence by 7.2.8 every point in X admits a
chart isomorphic to an open neighbourhood of (0, . . . , 0, −∞, . . . , −∞) ∈
StarX (p) × TsedX (p) . On the other hand, every tropical space which can be
covered by such charts is obviously regular at infinity. We reformulate this
as follows.
Definition 7.2.10
Let F be a tropical fan cycle in RN , U ∈ RN a convex open set containing
the origin and k ∈ N. The open neighbourhood
(0, . . . , 0, −∞, . . . , −∞) ∈ V := (F ∩ U) × [−∞, 0)k ⊂ F × Tk .
is called a standard neighbourhood.

Let X be a tropical space and p ∈ X . A chart of p to V is called a standard
chart of p.
Corollary 7.2.11
A tropical space X is regular at infinity if and only if it can be covered by
standard charts.
In view of these considerations, the following generalization suggests it-

self and will be useful later.
Definition 7.2.12
Let F be a tropical fan cycle in RN , U ∈ RN a convex open set containing the
origin and k, M ∈ N with k ≤ M . We set Snk to be the union of coordinate
planes of dimension k in Tn , i.e.
[
Snk := (Tn )[n−k] = TI .
I⊂[n]
|I|=n−k
The open neighbourhood
(0, . . . , 0, −∞, . . . , −∞) ∈ V := (F ∩ U) × (Snk ∩ [−∞, 0)n ) ⊂ F × Snk
is called a normal crossing neighbourhood.
192
Let X be a tropical space and p ∈ X . A chart of p to V is called a normal

crossing chart of p. If X can be covered by normal crossing charts, X is called
normal crossing.
Note that k is an invariant of all normal crossing charts of a point p, hence

sedX (p) = k is again well-defined.
Example 7.2.13
Let X be a tropical space which is regular at infinity. Then all its skeletons
0 0
X [k] are normal crossing tropical spaces. It holds (X [k] )[k ] = X [k+k ] .
7.3 Tropical morphisms and structure

sheaves
After the preceding discussions, the notion of a tropical morphism between
tropical spaces is be clear. We state it here for reference.
Definition 7.3.1
Let X and Y be two tropical spaces. A continuous map f : X → Y is called a
tropical morphism if for each pair of charts (U, ψ, V ) for X and (U 0 , ψ0 , V 0 )
for Y the composition map
ψ0 ◦ f ◦ ψ−1 : ψ( f −1 (U 0 ) ∩ U) → V 0
is a tropical morphism of open polyhedral sets in the sense of definition

7.1.4.
Which classes of functions are preserved under tropical morphisms? Let

us give some examples. Recall that for an open polyhedral set X a tropical
morphism f : X → T is just a locally affine Z-linear function (see 7.1.4).
Definition 7.3.2
Let X be a tropical space. We define the sheaf of locally affine Z-linear func-
tions, denoted Aff X , by setting
Aff X (U) := { f : U → T : f tropical morphism}

(7.1)
= { f : U → T : f locally affine Z-linear}.
193
Note that the invertible sections Aff ∗X (U) of this sheaf are given by tropi-
cal morphisms/locally affine Z-linear functions to R. The sheaf Aff X as well
as most of the sheaves we will introduce has very special properties, e.g. it
is semiconstant (the stalk map p 7→ Aff X ,p is semiconstant in the sense of
7.1.11) and hence constructible.
Example 7.3.3
Let us give a few examples.
(a) Let us start with X = Tn . We obviously have
= (Zn )∨ × R ∼
Aff X (Rn ) \ {−∞} ∼ = Zn × R,
(7.2)
f 7→ (d f , f (0)).
In order to extend f to R I , we need d f (−ei ) ≤ 0 (or equivalently

d f (ei ) ≥ 0) for all i ∈ I. In particular, the the space of non-zero
global sections is Aff X (Tn ) \ {−∞} ∼= Nn × R. The same holds for
every connected open neighbourbood of O = (−∞, . . . , −∞).
(b) Let F ⊂ Rn be a tropical fan cycle of sedentarity zero. Let us assume

that F is non-degenerate, i.e. not contained in a proper subspace of
Rn . Let U be a connected open neighbourhood of 0. Then
Aff F (U) \ {−∞} = Aff F,0 \{−∞} = Aff Rn (Rn ) \ {−∞} ∼

= Zn × R.
(c) Let X be a tropical space and let (U, ψ, V ) be a standard chart for
p ∈ X . Hence V is a connected open subset of StarX (p) × Tk . We may
assume StarX (p) ⊂ Rn is non-degenerate. Then
Aff X ,p \{−∞} = Aff X (U) \ {−∞} ∼

= Zn × Nk × R.
Here, the first factor encodes the differential in Rn -directions (the di-
rections of “same” sedentarity), the second factor encodes the Rk -
directions (the directions to smaller sedentarity) and the last factor
again just parameterizes shifts by a constant. Be careful, however,
that in general the (Zn )∨ -coordinates are not determined by the re-
striction of f to the points of k-sedentarity X [k] ∩ U. Namely, if one
of the Nk -coordinates is positive, this restriction takes constant value
194
−∞, but the (Zn )∨ -coordinates are still well-defined and might be
non-zero. The sheaf Aff ∗X of invertible sections can be described, via
the above identification, as follows.
Aff ∗X ,p = Aff ∗X (U) ∼

= Zn × {0} × R
(d) Let X ρ be the ρ-space constructed in 7.1.17. We denote by O the

point obtained from glueing 0 and +∞. Then the stalk at O can be
described as
Aff X ρ ,O \{−∞} ∼
= (Z × R) t (Z>0 × R).
The first part describes functions with f (O) 6= −∞ using the slope
along the 0-branch and the value at O. The second part describes
functions with f (O) = −∞ using the slope along the +∞-branch
and a shift parameter.
These examples suggest to make a few more definitions related to the

stalks of Aff X .
Definition 7.3.4
Let X be a tropical space and p ∈ X be a point. The cotangent space lattice
of P at p is defined to be
ZTp∗ := { f ∈ Aff ∗X ,p | f (p) = 0} (∼ ∗

= Aff X ,p /R). (7.3)
ZTp∗ forms a group with respect to usual addition (tropical multiplication) of

germs. Moreover, ZTp∗ is free abelian and finitely generated. The cotangent
space of P at p is defined to be the R-vectorspace
Tp∗ := ZTp∗ ⊗ R. (7.4)
By dualizing, we obtain the tangent space (lattice) ZTp ⊂ Tp .

Given a morphism of tropical spaces f : X → Y , pulling back functions
induces a linear map (d f )∗x : ZTY,∗ f (x) → ZTX∗,x for all x ∈ X . The dual linear
map
d f x : ZTX ,x → ZTY, f (x)
is called the differential of f at x.
195
Remark 7.3.5
One nice feature of the tangent space Tp is that it contains StarX (p) canoni-
cally and non-degenerately. Hence we can remove the choice of representa-
tion of StarX (p) and regard it as embedded in the (abstract) vector space Tp .
If X is regular/normal crossing at infinity, StarX (p) ⊂ Tp is a well-defined
tropical fan cycle of sedentarity zero.
For completeness, we add the following complementary definition.
Definition 7.3.6
Let X be at tropical space and let p ∈ X be a point. We define the slope
semigroup of X at p to be
S p := Aff X ,p \{−∞}/ Aff ∗X ,p . (7.5)
If X is regular at infinity, by 7.3.3 S p is isomorphic to Nsed(p) . On a stan-

dard neighbourhood of p, the isomorphism is given by the slopes of a germ
in the ei -direction (where ei is the standard basis vector in Rk ⊆ Tk ). All the
notions defined above fit together in the exact sequence
0 → ZTp∗ × R → Aff X ,p \{−∞} → S p → 0.
7.4 Smooth tropical spaces

The category of tropical spaces of finite type and regular at infinity forms a
reasonable class of objects for tropical geometry. However, so far we have
not imposed any local condition which might reflect the classical concept of
smoothness. In particular, we can not expect to get fully-fledged homology
and intersection theory in such a generality. In the past, different authors
used different concepts of smoothness and it seems that there is no “best”
definition of smoothness in the tropical world which reflects all the features
of the classical theory.
There is one definition, however, which seems especially useful. This is
due to the fact that, on one hand, the definition is simple and fits naturally
in the tropical theory developed so far. On the other hand, while being
sufficiently restrictive to provide us with many of the properties we would
expect from smooth spaces, it is at the same time interesting and flexible
196
enough — even locally, where it leads to interesting combinatorics related

to matroid theory and hyperplane arrangements. In fact, the definition es-
sentially says that a smooth tropical space should locally look like a tropical
plane in TPn .
Recall that in section 5.2 we defined the degree of a k-cycle in Rn (resp.
TPn ) to be deg(X ) = deg(X · H k ), where the product is given by stable
intersection and H denotes the standard hyperplane of Rn (resp. TPn ). Let
X ⊂ Rn be an effective tropical fan with deg(X ) = 1. A balanced open
polyhedral set obtained as a open subset of X ⊂ Tn is said to be of degree 1.
Definition 7.4.1
A tropical space X is called smooth if it can be covered by charts to degree
1 open polyhedral sets. We call X a tropical manifold if it is of finite type,
regular at infinity and smooth.
Note that smoothness does not in general imply regularity at infinity, as
example 7.2.5 (b) shows.
Let Λ be a lattice and Λ × R the associated vector space. Let X ⊂ V be a
tropical fan of pure dimension k. The choice of a basis B = (v1 , . . . , vn ) of
Zn gives us a tropical isomorphism ΦB : V → Rn .
Definition 7.4.2
We define the degree of X with respect to the basis B to be
degB (X ) := deg(ΦB (X )).
A tropical fan X ⊂ V is called degree 1 fan if it is effective and if there exists
a basis B such that degB (X ) = 1.
Corollary 7.4.3
Let X be a tropical space regular at infinity. Then X is smooth if and only if
for all p ∈ X the star StarX (p) ⊂ Tp is a degree 1 fan.
We will now have a closer look at the local building blocks of tropical
manifolds, degree 1 fans. Our goal is to show that degree 1 fans satisfy
some very special properties and, summarizing this, can be described in
terms of matroids. This correspondence depends on the chosen basis B.
In the following we fix as ambient space Rn with its standard basis e1 , . . . , en
(i.e. we compute the degree of a fan with respect to the standard hyperplane
197
H = V (“x 1 + . . . + x n + 0”)). Additionally, we set e0 = (−1, . . . , −1)

P and
E = {0, 1, . . . , n}. For any subset S ⊂ E, we define the vector eS = i∈S ei .
In particular, e E = 0. To any chain of subsets S = (; ( S1 ( . . . ( Sl ( E),
we assign the cone σS = R≥0 eS1 + . . . + R≥0 eSl . Here, l = dim σS is called
the length of S . The collection of σS for all possible chains of subsets
of E forms a unimodular polyhedral fan covering Rn . It is called the fine
subdivision of Rn and denoted by FS(Rn ). The first theorem we want to
prove is that all degree 1 fans are supported on this subdivision (i.e. can be
represented as a union of cones of the form σS ).
Theorem 7.4.4
Let X ⊂ Rn be a degree 1 fan of dimension k. Then X is supported on FS(Rn )(k) .
Proof. Note that the subdivision FS(Rn ) can be obtained from intersecting
all hyperplane subdivisions given by the hyperplanes x i = x j . Hence it
suffices to show that the linear span Rσ of every facet σ of X is the inter-
section of n − k such hyperplanes. Note that Rσ, considered as a tropical
cycle, is itself of degree 1 by 5.2.4. Hence the following lemma finishes the
proof.
Lemma 7.4.5
Let V ⊂ Rn be a linear space which, considered as a tropical cycle with weight
1, has degree 1. Then there exists a chain ; ( S1 ( . . . ( Sk ( E such that
V = 〈vS1 , . . . , vSk 〉.
Proof. We want to use induction on k. The case k = 0 is trivial. Now let V

be of dimension k + 1. Choose a hyperplane x i = x j which does not contain
V . Then the stable intersection V 0 := V (“x i + x j ”) · V is a again of degree
1. Indeed, in H k−1 · V (“x i + x j ”) · V we may always replace V (“x i + x j ”) by
a translated copy of H such that all points in the intersection are contained
in the interior of the facet of H contained in the hyperplane x i = x j . Hence
deg(V 0 ) = deg(H k ·V ) = 1. So by induction hypothesis V 0 = 〈eS1 , . . . , eSk 〉 for
a suitable chain S . Choose an integer vector v that completes eS1 , . . . , eSk to
a lattice basis of V . Let ai be projective coordinates for v. Adding a suitable
linear combination of eS1 , . . . , eSk , e E , we may assume that ai ≥ 0 for all i and
that for each 0 ≤ j ≤ k there exists an index i ∈ S j+1 \ S j such that ai = 0.
To simplify notation, we permute the coordinates such that S j = {0, . . . , r j }
198
with 0 ≤ r1 < . . . < rk < rk+1 = n and such that a r j = 0 for all 1 ≤ j ≤ k + 1.
Here is a schematic picture of the matrix obtained from this basis of V .
 
eS1 1 . . . 1 0 . . . 0 0 . . . 0
eS2 1 . . . 1 1 . . . 1 0 . . . 0
..  .. .. .. .. . . 
. 
 . . . . . 

e 1 . . . 1 1 . . . 1 1 . . . 1
E
v ∗ ... 0 ∗ ... 0 ∗ ... 0
Our next claim is that v has only 0 and 1 entries. Namely, consider the entry
/ {r1 , . . . , rk+1 }. Let I be the complement of {i, r1 , . . . , rk+1 }. Since
ai for i ∈
|I| = n − k − 1, σ I is a facet of H k+1 . Now deg(H k+1 · V ) = 1 implies (after
choosing a suitable translation) that if Rσ I and V intersect transversally, the
corresponding lattice index should be 1. In the language of determinants
this means that after deleting the columns of M labeled by indices in I the
square matrix has determinant 0 or ±1. However, after the deletion the
only non-zero entry in the v-row is ai . Moreover, after developing this row,
the matrix becomes triangular with 1’s on the diagonal. Hence ai ∈ 0, 1. It
remains to show that v has non-zero entries in exactly one of the segments
0 ≤ r1 < . . . < rk < n, say between r j and r j+1 . This finishes the proof, as
this allows us to replace v by eS , where S = S j ∪ {i : ai = 1}. So let us prove
this. Assume by contradiction that we find r j < i1 < r j+1 < i2 such that
ai1 = ai2 = 1. We set
I = E \ {i1 , r1 , . . . , rk+1 },
J = E \ {i1 , i2 , r1 , . . . , rd
j+1 , . . . , rk+1 }.
Again, deleting the columns labeled by I resp. J, we see easily that the
determinant is non-zero in both cases (note that I is exactly of the form as
above). Hence Rσ I and RσJ both intersect V transversely. We now move
H k+1 such that the translated vertex p has coordinates satisfying
i = i1 ,

1
pi = 0 i ∈ {i2 , r1 , . . . , rk+1 },
< 0 otherwise.

Note that pi = ai for all i ∈

/ I and pi < ai for all i ∈ I, hence v − p ∈ σ I or
equivalently v ∈ p + σ I . One also checks easily that eS j+1 − eS j − p has non-
negative entries with zeros labeled by {i1 , i2 , r1 , . . . , rd
j+1 , . . . , rk+1 }, hence
199
eS j+1 − eS j ∈ p +σJ . It follows that both v and eS j+1 − eS j occur in (H k+1 + p)· V
and therefore deg(V ) ≥ 2, a contradiction.
Up to now we know that every degree 1 fan is a union of cones σS for

some chains S , with trivial weights 1. Which chains S can occur together?
To answer this question, let us first give a criterion describing which trans-
lations of H k intersects a given cone σS .
Lemma 7.4.6
Let p ∈ Rn be a point with pairwise distinct projective coordinates pi . Let O
be the induced complete order on E such that pi > p j ⇔ i > j. Let S = (; (
S1 ( . . . ( Sk ( E) be a chain and set m j := maxO (S j+1 \ S j ), with respect to
O. Then the following are equivalent:
(a) m0 > m1 > . . . > mk forms a decreasing sequence with respect to O.
(b) (H k + p) ∩ σS 6= ;.
In this case the intersection is transversal and we say O is compatible with S .
Proof. It follows from the definitions that (H k + p) ∩ RσS contains a single

transversal intersection point x given by x i = pm j for all i ∈ S j+1 \ S j . Fur-
thermore, x ∈ σS if and only if the m j form a decreasing sequence. Hence
the claim follows.
Corollary 7.4.7
Let X ⊂ Rn be of degree 1 and let σF , σG ⊂ X be two facets. If F and G admit
a common compatible order, then F = G .
The next lemma describes the situation around a codimension 1 face.
Lemma 7.4.8
Let X ⊂ Rn be a degree 1 fan of dimension k. Let S = (; ( S1 ( . . . ( Sk−1 (
E) be a chain of length k − 1 such that σS ⊂ X . Let σS1 , . . . , σSl be the facets
of X containing σS . Then all chains Si are of the form ; ( . . . S j ( Ti (
S j+1 ( . . . ( E for a fixed 0 ≤ j ≤ k − 1 and the Ti \ S j form a partition of
S j+1 \ S j .
200
Proof. Each chain Si is obtained from S by inserting some set Ti at some

position ji . The balancing condition around σS is equivalent to the condi-
tion that the Ti \ S ji cover each element in S ji +1 \ S ji the same number of
times. If follows that the subcollection of facets with some fixed value ji sat-
isfies the balancing condition independently. However, it follows from 5.2.4
that degree 1 fans are locally irreducible (i.e. StarX (p) = E1 + E2 for two
effective cycles E1 , E2 implies E1 = 0 or E2 = 0). Hence only one value for
ji can occur, say j. It remains to show that the Ti \ S j cover each element of
S j+1 \S j exactly once. If an element is covered twice, the two corresponding
chains admit a common order, which contradicts to 7.4.7.
After these preliminary works, we can now prove that X can be com-
pletely recovered from its rays, i.e. from its intersection with FS(Rn )(1) .
Theorem 7.4.9
Let X ⊂ Rn be a degree 1 fan and let S = (; ( S1 ( . . . ( Sl ( E) be a chain.
Then σS ⊂ X if and only if eS j ∈ X for all 1 ≤ j ≤ l.
The proof uses a certain reordering construction which we choose to
present seperately in the following lemma.
Lemma 7.4.10
Let S = (; ( S1 ( . . . ( Sk ( E) be a chain such that σS ⊂ X is a facet of X .
Fix 1 ≤ l ≤ k and let O be a complete order on Sl . Then there exists another
chain S 0 = (; ( S10 ( . . . ( Sk0 ( E) such that
(a) S 0j = S j for all j ≥ l,
(b) O is compatible with ; ( S10 ( . . . ( Sl0 = Sl ,
(c) σS 0 ⊂ X .
We call S 0 the reordering of S by O.
Proof. Let i be the maximal element in Sl with respect to O. Take 1 ≤ j ≤ l
minimal such that i ∈ S j . By 7.4.8 we can replace S j−1 by a (unique) set
such i ∈ S j−1 . Proceeding in this way, we may assume i ∈ S1 . Now consider
the maximal element i 0 in Sl \ S1 and proceed analogously to obtain a new
chain with i 0 ∈ S2 . Continuing in this way, we end up with a chain as
described in the statement.
201
Proof of theorem 7.4.9. The “only if” direction is clear. Hence let us assume
that S is a chain of length l with eS j ∈ X for all j. We prove σS ⊂ X by
induction on l. By induction hypothesis we assume σS ⊂ X and show for
given G with Sl ( G and eG ∈ X the concatenated chain also describes a
cone of X . The containments in X guarantee that there are maximal chains
F resp. G completing S resp. containing G. Choose a complete order of Sl
compatible with S and complete it to an order O on E such that elements
in Sl are greater than elements in G \ Sl which in turn are greater than
elements in E \ G. Let F 0 resp. G 0 be the reorderings of F resp. G with
respect to O. It follows from the construction that the lower part of F
remains unchanged, hence F 0 is still a completion of S . At the same time
G 0 still contains G by construction. However, F 0 and G 0 have common
order O, hence by 7.4.7 they have to be equal. Therefore F 0 = G 0 is a
completion of ; ( S1 ( . . . ( Sl ( G ( E, which proves the claim.
We will now show that there is a one-to-one correspondence between

degree 1 fans of Rn (with respect to the chosen basis) and loopfree matroids
supported on the E = {0, . . . , n}. Recall that a matroid M can for example
be given by a set F ⊂ 2 E whose elements are called flats (or closed sets) of
M and which satisfies the following three axioms.
(A) E is a flat.
(B) If F and G are flats, then so is F ∩ G.
(C) Let F be a flat and let G1 , . . . , Gl be all flats which cover F (i.e. F ( Gi
and F ⊂ H ⊂ Gi implies F = H or F = Gi ). Then the sets Gi \ F form
a partition of E \ F .
The matroid M is called loopfree if the empty set is a flat as well. Here is
how a degree 1 fan can be turned into a matroid.
Theorem 7.4.11
Let X be a degree 1 fan of dimension k. Let F be the set of subsets S ⊂ {0, . . . , n}
such that eS ∈ X . Then F satisfies the axioms of the lattice of flats of a matroid.
It is called the matroid associated to X and denoted by M (X ).
Proof. Axiom (A) follows from e E = 0 ∈ X .
202
For Axiom (B) choose any order O of E such that the elements in F ∩G are
greater than the rest. Let σF and σG be facets of X such that F contains
F and G contains G. Reorder F by O and let F 0 be smallest set in the
new chain such that F ∩ G ⊂ F 0 . By construction of O, we have F 0 ⊂ F .
Analogously, reordering G we get a minimal set G 0 ⊂ G. The two new chains
have common order O and hence are equal, and by minimality F ∩ G ⊂ F 0 =
G 0 ⊂ F ∩ G. Hence F ∩ G = F 0 = G 0 is a flat.
For axiom (C), first note that by 7.4.9 it is clear that the minimal flats Gi
containing F are exactly those appearing behind F in some chain F such
σF is a facet of X . Let i ∈ E\F . Let F be any such chain that contains F and
let O be an order on E such that the largest elements are in F , then i, then
the rest. Reordering F by O we obtain a chain such that the set G behind
F contains i. Hence the Gi cover E \ F . Assume there was a second such G 0
containing i. Then reordering a chain containing F ⊂ G 0 by O produces the
same chain as before, hence G = G 0 . This proves the claim.
By theorem 7.4.9, the inverse to X 7→ M (X ) should look as follows.
Definition 7.4.12
Let M be a loopfree matroid on E. The associated matroid fan B(M ) consists
of the collection of cones σF , where F = (; ( F1 ( . . . ( Fl ( E) is a chain
of flats of M . In particular, B(M ) is a subfan of the fine subdivision of
dimension rank(M ) − 1 (which is the maximal length of chains of flats in
M ).
Theorem 7.4.13
Let M be a loopfree matroid. Then B(M ) is a degree 1 fan.
Proof. First note that axiom (C) of the matroid axioms guarantees that the
statement of lemma 7.4.8 holds and hence B(M ) is balanced. So B(M )
forms indeed a tropical fan in Rn and it remains to compute its degree. Our
approach is as follows. Let H ⊂ Rn be the standard hyperplane. We will
show that the stable intersection H · B(M ) is the fan of a loopfree matroid
again. Therefore, by induction it suffices to prove the claim for the single
loopfree matroid of rank 1 on E. Its only flats are ; and E itself. Therefore
B(M ) is just the origin which is obviously of degree 1.
So, let us now prove that when rank(M ) > 1, H · B(M ) is a matroid fan
again. Indeed, we will show H · B(M ) = B(M 0 ) where M 0 is the matroid
203
obtained from M by removing all flats of rank rank(M ) − 1 (which can

be easily checked to be an admissible collection of flats again). To do so,
we recall that H is given by the homogeneous linear tropical polynomial
l = max{x 0 , . . . , x n }, where the x i are the homogeneous coordinates on
TPn . Therefore H · B(M ) can be computed as the divisor of l on B(M ). As l
is obviously linear on all cones of the fine subdivision, H ·B(M ) is supported
on the codimension one skeleton of B(M ). The weight of a codimension one
cone σF of B(M ) in H · B(M ) can be computed by
k
X
ω(σF ) = l(eHi ) − l(eG ) − (k − 1)l(e F ),
i=1
where k is the number of facets containing σF , F ⊂ G is where the rank

gap appears in F and H i are the filled-in flats corresponding to a facet. To
be precise, in this expression l should be replaced by an arbitrary dehomog-
enization of l. Alternatively, we may set
¨
−1 if S = E,
l(eS ) =
0 otherwise.
In any case, it is easy to compute

¨
1 if G = E,
ω(σF ) =
0 otherwise.
Thus H · B(M ) consists of all the cones σF where F is a chain of flats in M

not containing a flat of rank rank(M )−1. But this agrees with the definition
of B(M 0 ).
We can summarize this circle of ideas in the following statement.
Corollary 7.4.14
There is a one-to-one correspondence between degree 1 fans in Rn and loopfree
matroids on E given by
{degree 1 fans in Rn } ←→ {loopfree matroids on E},

X 7−→ M (X ),
B(M ) ←− M .
204
Proof. Theorems 7.4.11 and 7.4.13 show that both maps are well-defined.
Theorem 7.4.9 proves that B(M (X )) = X . Finally, M (B(M )) = M follows
directly from the definitions.
Remark 7.4.15
Note that the map X 7→ M (X ) has a completely analogous counterpart in
classical algebraic geometry. An embedding of a linear space L in CPn is
basically equivalent (up to translating TPn by the torus action) to a hyper-
plane arrangement on L. The hyperplanes are given by the intersection of L
with the coordinate hyperplanes of CPn . In this situation, it is easy to check
that the above definition of M (L) is equivalent to the standard definition of
the matroid of the hyperplane arrangement.
In fact, we may consider a slight generalization of our statement.
A degree 1 fan X in TPn is the closure of a degree 1 fan in an orbit O F ,
F ⊂ E. We think of M (X ), which a priori is a matroid on the ground set
E \F , as a matroid on E by summing with the rank zero matroid on F (hence
F is the minimal flat of M (X )).
Correspondingly, let M be a (not necessarily loopfree) matroid M and let
F = ; be its minimal flat. Then M \ F is a loopfree matroid and we define
B(M ) := B(M \ F ) ⊂ O F .
The philosophy here is that loops in the matroid correspond to fans con-
tained in the corresponding coordinate hyperplane at infinity.
With these generalizations, the one-to-one correspondence can be ex-
tended to degree 1 fans in TPn and (arbitrary) matroids on E.
{degree 1 fans in TPn } ←→ {matroids on E}
Theorem 7.4.16
Let Y ⊂ Rn be a matroid fan of dimension k and let V ⊂ Y be a balanced
polyhedral complex of dimension k−1 contained in Y . Then there exist tropical
polynomials f , g such
V = div(“ f /g”|Y ).
Proof. Use FrancoisRau and Alexander Esterov: We can find X 0 ⊂ Rn of
dimension n − 1 such that X = Y · X 0 (stable intersection). Use the previous
205
theorem to find X 0 = V (“ f /g”) and use compatibility of stable intersection

and Cartier divisors.
Theorem 7.4.17
In the above two theorems, the function ϕ = “ f /g” is unique up to adding an
affine linear function.
Proof. In the Rn -case, it follows from the proof that once we fix f on one
connected component, it is uniquely determined everywhere else. For a
general matroid fan, we can use the fact that it is a (multiple) modifica-
tion of Rk along matroid divisors and show that the property survives such
a modification. So let us assume V V is a matroid modification of V and
f · V V = 0 for some rational function f . We want to show that f is affine
linear. We prove first that f is the pull-back of a function g on X . Then the
claim follows as the function g has to be affine linear by induction assump-
tion.
We check the claim by proving that f is affine linear on every half-ray
V V ⊂ {p} × R, where p is a generic point of divisor. Assume there is a
point
We write f = δ∗ g + h, where g is a function on X and h == 0 away from
d i visor × R. It follows δ∗ h · V V ⊂ divisor, and, as divisor is irreducible,
δ∗ h · V V = a · divisor. This implies g · V = −adivisor and therefore g =
−aϕ, where ϕ is the modification function, by uniqueness in X . From this
follows f = −a y.
Example 7.4.18
Example: Smooth curves, stable curves, M0,n , M 0,n
7.5 Rational functions, Cartier divisors and

modifications on abstract spaces
As an example of constructions that generalize naturally to abstract tropical
spaces, let us consider Cartier divisors and modifications of abstract spaces.
Definition 7.5.1
Let X ⊆ Tn be an open polyhedral set of pure dimension. A function f :
206
X → T is called a polynomial function on X if it is the restriction f = F |X

of a tropical polynomial F ∈ T[x 1 , . . . , x n ]. A function f : X → T is called
piecewise Z-linear if f is continuous and if there exists a subdivision X of
X such that f restricts to an affine Z-linear function (or is constant −∞)
on the interior of each cell. Assume that X is of sedentarity zero (i.e. X =
X ∩ Rn ). A piecewise Z-linear function f : X ∩Rn → T is also called a rational
function on X , denoted by f : X ¹¹Ë T. A general X admits a decomposition
X = ∪ I X I such that X I is a open polyhedral set of sedentarity zero in T I . A
rational function f : X ¹¹Ë T is defined to be a tuple of rational functions
( f I : X I ¹¹Ë T) I . Since open sets of X are again open polyhedral sets, this
obtain the presheaves OX0 and RX0 of polynomial resp. rational functions on
X . The associated sheaves of locally polynomial resp. rational functions
are denoted by OX resp. RX . Note that all these sheaves are sheaves of
T-algebras.
Remark 7.5.2
Both OX0 and RX0 are only presheaves in general (recall that we require sub-
divisions X of X to be finite). An example is given by the tropical Laurent
series
∞
X i i
f =“ − x ” ∈ OT (T) (7.6)
i=0
2
with infinitely many tropical zeros.
Exercise 7.5.3
Let ϕ : X → Y be a tropical morphism of open polyhedral sets. Show that
local polynomiality resp. rationality is preserved under pull-back along ϕ.
This leads to sheaf homomorphisms
ϕ ∗ : OY → OX and ϕ ∗ : RY → RX . (7.7)
Find an example for the fact that the pull-back of a polynomial function is
not necessarily polynomial again (only locally).
The exercise shows that the sheaves OX and RX can be glued along trop-
ical isomorphisms to give sheaves on tropical spaces.
Definition 7.5.4
Let X be a tropical space with atlas (Ui , ψi , Vi ). The unique sheaf OX on
207
X such that OVi ∼ = OX |Ui along ψ∗i is called the sheaf of locally polynomial
functions on X . Analogously, we define the sheaf of locally rational functions
RX . Finally, a Cartier divisor on X is a section of the quotient sheaf CDiv =
RX∗ / Aff ∗ . Hence it can be described by a collection of non-vanishing ra-
tional functions whose domains cover X and which differ by affine Z-linear
functions on the overlaps.
Let X ⊂ TN be an open tropical cycle of pure dimension n and sedentarity
zero and let f : X ¹¹Ë T be a rational function. We may choose a subdivision
X of X such that f is defined and affine Z-linear on all cells of sedentarity
zero. Let τ ∈ X be a cell of codimension one. We define the order of
vanishing ω f (τ) of f at τ in analogy to the constructions in by
k
X
ω f (τ) = ω(σi ) fσi (vi ), (7.8)
i=1
where σ1 , . . . , σk denote the facets of X containing τ and fσi = d f |σi de-

notes the Z-linear part of f restricted to σi (note that the sedentarity of a
facet is zero by assumption). The meaning of vi depends on the sedentar-
ity of τ. If sed(τ) =P0, the vector vi denotes a primitive generator of σi
modulo τ such that i ω(σi )vi = 0 ∈ Zn . If Sed(τ) = I 6= ;, the projec-
tion π : Rσi → R I has a one-dimensional kernel K (since dim(π(Rσi )) =
dim(τ) = dim(σi ) − 1) and we define vi to be the unique primitive inte-
ger vector in K with non-negative entries (intuitively, −vi is the primitive
integer vector which points to τ (at infinity)).
Definition 7.5.5
Let X ⊂ TN be an open tropical cycle of pure dimension n and sedentarity
zero and let f : X ¹¹Ë T be a rational function. We define the divisor (of
zeros and poles) of f , denoted by div( f ), to be the weighted open polyhedral
set obtained (for a chosen subdivision X as above) as the union of the
codimension one faces τ of X with weights ω f (τ).
Proposition 7.5.6
The divisor construction satisfies the following properties.
(a) The weighted open polyhedral set div( f ) is balanced, hence gives a open
tropical cycle.
208
(b) The open tropical cycle div( f ) does not depend on the chosen subdivision
X.
(c) Let l ∈ Aff ∗ (X ) be a locally affine Z-linear function on X . Then
div( f ) = div( f + l). (7.9)
The last property implies that we can extend the definition to Cartier
divisors.
Definition 7.5.7
Let X be a tropical space and let ϕ ∈ CDiv(X ) be a Cartier divisor. Let
(Ui )i be a open cover of X such that ϕ|Ui = [ f i ] for some rational function
f i ∈ R ∗ (Ui ). Then the (Weil) divisor div(ϕ) of ϕ is the unique tropical
subspace of X such that
div(ϕ)|Ui = div( f i ). (7.10)
As we recall from Chapter 5, the divisor construction
209
8 Tropical curves
In the previous chapters we presented the general concepts of tropical ge-
ometry. Instead of developing the general theory further, in the present
chapter we want to bring these concepts into action in the easiest non-
trivial concrete class of examples. Unsurprisingly, this class is formed by
tropical curves.
8.1 Smooth tropical curves

Definition 8.1.1
A tropical curve C is a connected tropical space of dimension one. If the
tropical space is smooth, we call C a smooth tropical curve. If C is a tropical
manifold, we call it a regular tropical curve.
Let us unwind the definition of tropical manifold to see what that actually
means in the curve case. The classification of local models is very easy. We
set L(n + 1) = H n−1 ⊂ Rn , the fan in Rn consisting of the n + 1 rays in the
directions −e1 , . . . , −en , e1 + . . . en , for suitable n. In the following, we call
L(n) the n-valent line.
Exercise 8.1.2
Let F ⊂ Rm be a one-dimensional tropical fan of degree 1. Show that F
is isomorphic to L(n) for suitable n ≥ 2. Show that F ∈ Tm is regular at
infinity. You can either proceed directly and use the definition of projective
degree, or classify matroids of rank 2.
In particular, we observe that smooth tropical curves are automatically

regular at infinity and hence a tropical curve is regular if and only if it is
smooth and of finite type.
Let C be a smooth tropical curve. The points of higher sedentarity p ∈
[1]
C are called infinite points for short. All other points are called finite
210
8 Tropical curves
points of C. Note that by regularity at infinity the only smooth local model
at an infinite point is −∞ in T. It is convenient to extend our notation
by setting L(1) := T (with "vertex" −∞). Then it follows from the above
exercise that at any point p the curve C admits a chart to L(n), for suitable
n ∈ N, sending p to the vertex. The number n =: val(p) is unique and called
the valence of p. We have val(p) = 1 if and only if sed(p) = 1.
We will now show that smooth tropical curves are the same thing as met-
ric graphs. The first ingredient to this statement is is the uniqueness for the
local model for each valence n as discussed above. The second ingredient is
the fact that in dimension one an integral affine structure is the same thing
as an (inner) metric. Let us be more precise.
Definition 8.1.3
A metric graph (with open ends) G is a connected one-dimensional finite CW-
complex with some 1-valent vertices removed and equipped with a com-
plete inner metric on G \ {1-valent vertices}. To avoid some tautologies, we
exclude the case of a single edge with both endpoints removed (we may
subdivide it by a 2-valent vertex, however). We distinguish three kinds of
edges. The edges adjacent to an erased resp. remaining 1-valent vertex are
called open resp. closed ends. All other edges are called inner edges.
Two metric graphs G, G 0 are called isomorphic if there exists a homeo-
morphism which is an isometry after removing 1-valent vertices.
Note that the open ends, closed ends, resp. inner edges e of G are iso-
metric to (−∞, 0], [−∞, 0], resp. [0, l e ] (for some l e ∈ R> ). It follows that
given a fixed cell decomposition of G, the choice of a complete inner metric
is equivalent to specifying a positive real number l e ∈ R> for each inner
edge e of G.
Let L(n + 1) be the n + 1-valent line. We equip L(n + 1) with the unique
complete inner metric such that all primitive generators −e1 , . . . , −en , e1 +
. . . en have length 1. Let p ∈ G be the vertex of a metric graph and let U v be
the open neighbourhood consisting of p and the interior of all edges adja-
cent to p (in the presence of loop edges, we subdivide them into two pieces).
Then there exists a unique distance-preserving map U p → L(val(p)) up to
permuting the rays of L(val(p)). We include the case val(p) = 1 with map
Up ∼ = [−∞, 0] ,→ T). The collection of these maps forms a tropical atlas
and give G the structure of a regular tropical curve, denoted (presently) by
211
8 Tropical curves
tc(G).
Exercise 8.1.4
Check that the collection of maps distance-preserving maps U p → L(val(p))
forms a tropical atlas which is of finite type and smooth.
Proposition 8.1.5
The map G 7→ tc(G) given by the above construction induces a bijection be-
tween the set of isomorphism classes of metric graphs G and the set of isomor-
phism classes of regular tropical curves (i.e., smooth curves of finite type).
We leave this as an exercise.
Exercise 8.1.6
Let C be a smooth tropical curve.
(a) Show that C has the structure of a connected finite CW-complex with
some 1-valent vertices removed.
(b) Show that given the metric on L(n), there is a unique inner metric on
C \ C [1] such that the charts to L(n) are distance-preserving.
(c) Show that this metric is complete.
(d) Let C 0 be another smooth tropical curve and let ϕ : C → C 0 be a

homeomorphism of the underlying topological spaces. Show that ϕ
is a tropical isomorphism if and only if it is distance-preserving (after
removing the points at infinity).
Note that along this identification, the ends of the graph correspond to
rays of the tropical curve, with sedentarity 1 endpoint in the case of closed
ends. The inner edges correspond to bounded line segments. In the follow-
ing, we will use the same letters and jump back and forth freely between
the two points of view.
Definition 8.1.7
We define the genus of a smooth tropical curve to be g(C) = b1 (C) =
dim H1 (C, Z), the first Betti number of the underlying graph.
212
8 Tropical curves
Remark 8.1.8
We may drop the condition of finite type and extend the above one-to-one
correspondence to any smooth tropical curve. In this case we should relax
the completeness condition in the definition of metric graph to the property
that the metric cannot be extended to a metric on G. It corresponds to
allowing a finite length parameter l e ∈ (0, ∞] for open ends.
8.2 Divisors and linear systems

Definition 8.2.1
Let X be a tropical
Pn curve. A divisor on X is formal finite linear combination
of points D = i=1 ai pi , ai ∈ Z, pi ∈ X . We say D is effective ifPall ai ≥ 0
(denoted by D ≥ 0). The degree of D is defined by deg(D) := i ai . The
support of D is the set supp(D) := {pi : ai 6= 0} ⊂ C. We say D is a divisor
at infinity if supp(D) ⊂ X [1] . Formal addition turns the set of all divisors
into an abelian group
L graded by degree and called the divisor group of X ,
denoted Div(X ) = d∈Z Divd (X ).
Let Rat(X ) be the group of rational functions of finite type on X (with

classical addition). Recall that to any such function f we can associate a
divisor div( f ) ∈ Div(X ). The coefficient of div( f ) at p is called the order
(of vanishing) at p and denoted by ord p ( f ) ∈ Z. Moreover, the map div :
Rat(C) → Div(C) is group homomorphism.
Definition 8.2.2
Two divisors D, D0 on C are called linearly equivalent, denoted by D ∼ D0 ,
if D − D0 = div( f ) for some f ∈ Rat(X ). The corresponding quotient group
Div(X )/ Im(div) is called the divisor class group of X . The divisors of the
form div( f ), f ∈ Rat(X ) are called principal divisors of X .
Proposition 8.2.3
Let X be a compact tropical curve. Then deg(div( f )) = 0 holds for all f ∈
Rat(X ).
Proof. Fix a subdivision of X without loop edges on which f is piecewise

linear. As X is compact, each edge has two endpoints. The slope of f
on an edge contributes to the coefficients of both endpoints with opposite
213
8 Tropical curves
signs. Hence, when adding up the coefficients the contribution of each

eadge cancels and the degree vanishes.
Let us also note the following fact.
Proposition 8.2.4
Let X be a compact tropical curve and f ∈ Rat(X ). Then it holds
div( f ) = 0 =⇒ f ≡ const,
i.e., the kernel of div consists of constant functions only.

Proof. Let M ⊂ X be the locus in C where f attains its minimum. If M ( X ,
the boundary ∂ M is non-empty and we may choose some p ∈ ∂ M . At this
point, all outgoing slopes are positive and at least along one edge the slope
must be strictly positive (otherwise p lies in the interior of M ). This implies
ord p ( f ) > 0, which contradicts div( f ) = 0.
Definition 8.2.5
Let D be a divisor on a tropical curve X . The set of effective divisors linearly
equivalent to D
|D| := {D0 ∈ Div(X ) : D ∼ D0 ≥ 0}
is called the complete linear system of D. We also use the notation
Γ (D) := { f ∈ Rat(X ) : div( f ) + D ≥ 0} ∪ {−∞}.
Note that Rat(X ) ∪ {−∞} carries the structure of a semiring by setting
“ f + g”(x) := max( f (x), g(x)),

“ f g”(x) := f (x) + g(x),
for all x ∈ X \ X [1] and extending to X by continuity. Moreover, Rat(X )

carries a T-alegbra structure by identifying T with the constant functions in
Rat(X ). By Proposition 8.2.4, if X is compact, |D| is just the projectivization
of Γ (D) with respect to this multiplication by T, i.e.,
|D| = P(Γ (D)) := Γ (D) \ {−∞}/R.
Proposition 8.2.6
Let D be an effective divisor on a tropical curve X . Then the following holds.
214
8 Tropical curves
(a) The set Γ (D) is a T-subalgebra of Rat(X ).
(b) The set Γ (D) is finitely generated as a T-module.
Proof. Cite LINEAR SYSTEMS ON TROPICAL CURVES, CHRISTIAN HAASE,

GREGG MUSIKER, AND JOSEPHINE YU
8.3 Line bundles on curves

Definition 8.3.1
Let X be a tropical curve. A line bundle (of finite type) on X consists of a
tropical space L together with a tropical morphism π : L → X such that
there exists a finite collection of tuples (Ui , ψi )i , where
• the Ui form an open covering of X ,
• the maps ψi : π−1 (Ui ) → Ui × T are tropical isomorphisms such that

the diagram
ψi
π−1 (Ui ) Ui × T
π
Ui
commutes.
The collection (Ui , ψi ) is called a local trivialization for L. Let π0 : L 0 → X

be a second line bundle on X . An isomorphism of line bundles is a tropical
isomorphism Ψ : L → L 0 which commutes with projections, i.e., π0 ◦ Ψ = π.
The set of isomorphism classes of line bundles on X is called the Picard group
of X (we will introduce the group structure in a moment) and denoted by
Pic(X ).
Remark 8.3.2
Note that the tropical definition is slightly simpler than in the classical case.
In the classical case, in way or the other, the vector space structure on the
fibers has to be included in the definition and the trivialization is required
to respect this structure. (In practice, this is often done by fixing a triv-
ialization and asking the transition maps to be linear on fibers.) As we
215
8 Tropical curves
will see in a moment, in tropical geometry we get this for free. In some
sense, this boils down to the fact that the classical automorphism group
Aut(C) contains many non-C-linear maps, whereas a tropical automorphism
φ ∈ Aut(T) ∼ = R is automatically T-linear. Similar remarks apply to the def-
inition of isomorphisms of line bundles.
Remark 8.3.3
A line bundle L is smooth if and only if X is smooth. It is a tropical manifold
if and only if X is regular.
Let (Ui , ψi ) be a local trivialization of a line bundle L. Consider the in-
duced automorphisms on the overlaps
ψ
ei j := ψ j ◦ ψ−1 : (Ui ∩ U j ) × T → (Ui ∩ U j ) × T,
i
which acts trivially in the first component. It follows that ψ ei j induces auto-
∼
morphisms on {x}×T for each x ∈ Ui ∩ U j . Since Aut(T) = R, this gives rise
to functions ψi j : Ui ∩ U j → R such that ψ ei j (x, y) = (x, “ψi j (x) y”). More-
over, the functions ψi j are tropical morphisms (write them as ψi j (x) =
π2 (ψei j (x, 0)), where π2 is the projection to the second component) and
hence are sections of Aff ∗X (Ui ∩ U j ). They are called the transition func-
tions associated to the trivialization (Ui , ψi ). The identity ψ e jk ◦ ψ
ei j = ψ
eik
translates to the cocycle identity
ψi j − ψik + ψ jk ≡ 0 on Ui ∩ U j ∩ Uk .
The other way around, any collection of functions ψi j ∈ Aff ∗X (Ui ∩ U j ) sat-
isfying the cocycle identity occur as transition functions of a line bundle,
and the following statement summarizes this in terms of the first Čech co-
homology group H 1 (X , Aff ∗X ).
Proposition 8.3.4
Let L be a line bundle on X . The transition functions of any trivialization of
L induce a cocycle class [(ψi j )i j ] ∈ H 1 (X , Aff ∗X ). The map
Pic(X ) → H 1 (X , Aff ∗X ),
L 7→ [(ψi j )i j ]
is well-defined and bijective. In particular, it induces a group structure on
Pic(X ).
216
8 Tropical curves
Exercise 8.3.5
Prove the proposition (for example, by copying the proof of the classical
case). Recall that Aff ∗X is a constructible sheaf and hence Čech cohomology
group can be computed on a suitably fine open covering of X .
Exercise 8.3.6
Let us fix a subdivision of X without loops and double edges and let U x
denote the open neighbourhood of the vertex x consisting of x and the
interior of the adjacent edges. Show that the U v form an admissible open
covering for Aff ∗X and compute the cohomology groups with respect to this
covering.
We denote by L the TP1 -bundle obtained by extending the gluing maps

from (Ui ∩ U j ) × T to (Ui ∩ U j ) × TP1 .
Definition 8.3.7
Let L be a line bundle on X and (Ui , ψi )i be a local trivialization of L. A
map s : X → L is called a rational section (of finite type) of L if π ◦ s =
idX and si := π2 ◦ ψi ◦ s|Ui ∈ Rat(Ui ) for all i. We call the si the local
parts of s. WePdefine the order of s at x ∈ Ui by ord x (s) := ord x (si ) and
call div(s) := x ord x (s)x ∈ Div(X ) the divisor of s. The T-semifield of all
rational sections of L is denoted by Rat(L).
Exercise 8.3.8
Check that if s is a rational section for some local trivialization of X , then
it is so for any trivialization. Check that ord x (s) does not depend on the
choices made and hence div(s) is well-defined.
Remark 8.3.9
Every line bundle L has a canonical trivial section which is constant −∞
in every local trivialization. By abuse of notation, let {−∞} ⊂ L denote
the image of this section. Note that any non-trivial rational section s can be
restricted to a map X \ X [1] → L \ {−∞}, and this restriction determines
s uniquely. By abuse of notation, we will often write s : X → L in the
following.
Proposition 8.3.10
217
8 Tropical curves
Let L be a line bundle on X and let ψi j be the transition functions of a given

trivialization (Ui )i . Then the following holds.
(a) A collection of rational functions si ∈ Rat(Ui ) occurs as the local parts of

a rational section of L if and only if si − s j = ψi j .
(b) Let s, s0 be two rational sections of L with local parts si resp. si0 . Then there
exists a unique global rational function f ∈ Rat(X ) such that f |Ui =
si − si0 . We use the notation f = “s/s0 ” = s − s0 .
(c) There exists a rational section s of L.
Proof. Part (a) is clear. For part (b), set f i := si − si0 . Then part (a) implies
f i − f j ≡ 0 on Ui ∩ U j , hence we can glue the rational functions f i to a global
rational function f . We leave (c) as an exercise.
Exercise 8.3.11
Show that every line bundle L on a tropical curve admits a rational section.
You can proceed as follows:
(a) Reduce to the case where C has no points of higher sedentarity.
(b) Show that every line bundle admits a trivialization (U x , ψ x ) of the

form given in Exercise 8.3.6 for a suitable graph structure.
(c) For each U x and c ∈ R, set s x := −c dist(x, .) : U x → R. Show that for

sufficiently large c, both functions s x and s y + ψ x y dominate the other
somewhere on U x ∩ U y (if non-empty).
(d) Take an appropriate “maximum” of the s x to obtain a section of L.
The previous proposition should bring to mind the very closely related
concept of Cartier divisors. Recall that a Cartier divisor (of finite type) on X
is a global section of the (pre-)sheaf quotient Rat / Aff ∗ . In other words, it
can be described by a finite collection (Ui , f i )i where the Ui form an open
covering of X and f i ∈ Rat(Ui ) subject to the condition f i − f j ∈ Aff ∗X (Ui ∩U j )
for all i, j. Two such collections are considered equal if the merge of the
two still satisfies this condition. Any rational function f ∈ Rat(X ) gives rise
to a Cartier divisor given by the single tuple (X , f ). Such Cartier divisors
are called linearly equivalent to zero. They form a subgroup of the group
218
8 Tropical curves
of Cartier divisors CDiv(X ). Two Cartier divisors D, D0 are called linearly

equivalent if their difference is linearly equivalent to zero, denoted by D ∼
D0 .
Let (Ui , f i )i be the representative of a Cartier divisor on X . Clearly, the
functions f i j = f i − f j ∈ Aff ∗X (Ui ∩ U j ) satisfy the cocycle identity. Hence by
8.3.4 they form the transition functions of a line bundle L and by 8.3.10
the f i define a rational section s ∈ Rat(L). Vice versa, let (L, s) be a tuple of
a line bundle L and a section s ∈ Rat(L). Then for a given trivialization of
the local parts (Ui , si )i of s determine a Cartier divisor (again by 8.3.10). A
second tuple (L 0 , s0 ) is called isomorphic to (L, s) if there exists an isomor-
phism between the line bundles which identifies the sections. We can now
generalize 8.3.4 in the following way.
Proposition 8.3.12
The map
{(L, s) : L line bundle , s ∈ Rat(L)}/isom. → CDiv(X ),

(L, s) 7→ (Ui , si )i ,
is a bijection. In particular, it induces a group structure on the source. The

induced map
Pic(X ) → CDiv(X )/ ∼,
L 7→ [(Ui , si )i ] for some s ∈ Rat(X ),
(which is well-defined by 8.3.10) is a bijection. The induced group structure

on Pic(X ) coincides with the group structure obtained from 8.3.4.
Exercise 8.3.13
Check the details of the proof.
In total, we get the following commutative diagram of groups.

∼
=
{(L, s)}/isom. CDiv(X )
π
∼
= ∼
=
H 1 (X , Aff ∗X ) Pic(X ) CDiv(X )/ ∼
219
8 Tropical curves
relation between line bundles and locally free sheaves of rank 1

Let us finally have a look at the map
{(L, s)}/isom. → Div(X ),

(L, s) 7→ div(s),
induced by taking divisors of rational sections of line bundles. It is obviously

a group homomorphism, but in general neither injective nor surjective.
However, in the smooth case we have the following statement.
Proposition 8.3.14
Let X be a smooth tropical curve. Then the map
{(L, s)}/isom. → Div(X ),

(L, s) 7→ div(s),
and the induced map
Pic(X ) → Div(X )/ ∼,
L 7→ div(s) for some s ∈ Rat(X ),
are group isomorphisms.
We first prove the following lemma covering the local situation.
Lemma 8.3.15
Let L = L(n+1) be the n-valent line and let R(L) be the set of rational functions
f : L → R which are affine linear on each ray of L(n). Then the map ord0 :
R(L) → Z fits into the following exact sequence.
0 → Aff ∗ (L) → R(L) → Z → 0
Proof. The sequence of the statement, when restricted to the case f (0) = 0,
is dual to the sequence
0 → Z → Zn+1 → Zn → 0
given by 1 7→ (1, . . . , 1) and (ω0 , . . . , ωn ) 7→ i ωi (−ei ).

P
220
8 Tropical curves
Proof of 8.3.14. By 8.3.10 it suffices to establish the statement for the first
map. We prove injectivity first. Let L be a line bundle and s ∈ Rat(L) such
that div(s) = 0. Let si be the local parts of s with respect to some trivializa-
tion of L. By 8.3.15 we get si ∈ Aff ∗ (Ui ). By 8.3.4 this implies the existence
of an isomorphism L ∼ = X ×T which identifies s with the constant zero func-
tion. This proves injectivity. Let now D ∈ Div(X ). Choose a subdivision of X
without loops whose vertices contain the support of D. Let U v be the open
set containing the vertex v and the interior of the adjacent edges. By 8.3.15
there exists a collection of functions (U v , f v ) v such that ord v ( f v ) equals the
coefficient of v in D and such that each f v is affine linear on the interior
of edges. Hence, the collection determines a Cartier divisor and by 8.3.12
also a tuple (L, s) with div(s) = D. This proves surjectivity.
8.4 Riemann-Roch theorem

Let us now turn to the Riemann-Roch theorem for tropical curves. In this
section, all tropical curves are compact and smooth.
Definition 8.4.1
Let X be a compact smooth tropical curve and let L ⊆ |D| be a subset of a
complete linear system. We define the rank of L to be the maximal integer r
such that for all effective divisors D0 of degree r there exists a divisor D ∈ L
such that D − D0 ≥ 0. The rank of L is denoted by rk(L). We set rk(L) = −1
if L = ;. We write rk(D) = rk(|D|) for short.
Obviously, we have rk(D) = −1 if deg(D) ≤ 0. Moreover, if deg(D) = 0,
then rk(D) = 0 if and only if D ∼ 0 — otherwise rk(D) = −1.
Let us make a quick comparison to the classical situation. Given a com-
plete linear system and a given point on the curve which is not a base point,
the set of divisors "containing" the point forms a hyperplane. Hence our
definition requires that the intersection of L with rk(L) such hyperplanes is
non-empty. It easy to check that in the classical situation we just get back
the dimension of L.
Remark 8.4.2
Note that there are other non-equivalent definitions for the rank of a trop-
ical linear system |D|, for example
221
8 Tropical curves
• the minimal number of generators of the T-module Γ (D) minus one,
• the (maximal) dimension of the CW-coplex |D|.

To convince yourself that these numbers are in general different, take a
hyperelliptic genus 3 curve with two bridges and let D be a hyperelliptic
divisor. Then these numbers are 3 resp. 2, whereas rk(D) = 1. The above
definition was introduced in and is distinguished in particular by the fact
that Riemann-Roch holds.
We will present the Riemann-Roch theorem in the form using Serre du-
ality, i.e. with "error" term rk(K − D).
Definition 8.4.3
Let X be a smooth tropical curve. The canonical divisor class [K] ∈ Pic(X )
is given by X
K= (val(p) − 2)p.
p∈C
Remark 8.4.4
The motivation behind this definition is as follows. Let X be a classical
nodal curve whose whose irreducible components are P1 and whose dual
graph is equal to X . Then the dualizing sheaf of X (which is, in some appro-
priate sense, the limit of the canonical sheaf of a smoothing of X ) restricted
to the irreducible component X p , p ∈ X is isomorphic to OP1 (val(p) − 2).
Hence K just reflects the degree of the dualizing sheaf restricted to irre-
ducible components. Note however, that K is not “realizable” in general.
We can now state the theorem.
Theorem 8.4.5
Let X be a compact smooth tropical curve of genus g. Let D be a divisor of
degree d. Then the following equality holds.
rk(D) − rk(K − D) = d − g + 1
Before proving the statement, we need a few basic facts concerning the
rank of a linear system. As some of these facts can be proven most conve-
niently using the Abel-Jacobi map, we postpone the proof of them to the
next chapter.
222
8 Tropical curves
Exercise 8.4.6
Show that
rk(D) + rk(D0 ) ≤ rk(D + D0 ) ≤ rk(D) + deg(D0 )
for all divisors D, D0 ∈ Div(X ). In particular, −1 ≤ rk(D) ≤ deg(D).
Proposition 8.4.7
Let D be a divisor of degree d on a compact smooth tropical curve X of genus
g.
(a) If d ≥ g, then |D| =

6 ;.
(b) The rank of D is bounded from below by rk(D) ≥ d − g.
(c) If d ≥ 2g − 1, then rk(D) = d − g.
Proof. Part (a) follows from the Jacobi inversion theorem (cf. Theorem
9.3.12 and Remark 9.3.15 (a)). Part (b) follows immediately from (a).
For part (c), assume conversely that d − rk(D) < g. Choose a divisor
D0 of degree d − rk(D) with |D0 | = ; (such a choice exists by the Abel-
Jacobi theorem, cf. (cf. Theorem 9.3.6 and Remark 9.3.15 (c)). We have
deg(D − D0 ) ≥ g and hence by part (a) there exists an effective divisor
E ∈ |D−D0 |. But deg(E) = deg(D−D0 ) = rk(D) and hence |D0 | = |D−E| =
6 ;,
a contradiction.
In the following, it will be useful to consider some rational functions

based on the distance function on X . Let us denote the distance of two
points in the metric space X [0] by dist(x, y) ∈ R. Let A ⊆ X [0] be a non-
empty subset. Then we define dist(A, x) := infa∈A dist(a, x).
Let us now fix a point p ∈ X [0] . We introduce a total preorder on the set
of effective divisors on X \ {p} as follows. We write such a divisor as D =
p1 +· · ·+pn with dist(p, p1 ) ≤ . . . ≤ dist(p, p1 ) (where we set dist(p, pi ) = ∞
if pi ∈ X [1] ). The increasing sequence of distances is denoted by dist(p, D).
Given two such divisors D, D0 , we declare D lower than D0 if deg(D) <
deg(D0 ) or if deg(D) = deg(D0 ) and dist(p, D) is smaller than dist(p, D0 )
with respect to the lexicographic order. We call this preorder the p-distance.
223
8 Tropical curves
Definition 8.4.8
A divisor D on X is called p-reduced if its restriction to X \ {p} is effective
and minimal with respect to the p-order among such restrictions in the
equivalence class of D.
Exercise 8.4.9
Show that if D is p-reduced and |D| 6= ;, then D ≤ 0.
Here is another characterization of p-reduced divisors using subgraphs
A ⊆ X (i.e., closed subsets obtained as the union of cells for some subdivi-
sion of X ).
Proposition 8.4.10
A divisor D is p-reduced if and only if it is effective on X \ {p} and for any
subgraph A ⊆ X with p ∈/ A there exists a point q 6= p such that the coefficient
aq of D at q satisfies
aq + valA(q) < valX (q). (8.1)
In intuitive words: Any A is leaking into X and cannot be sealed off using
D.
Proof. The characterization based on (8.1) will be referred to as leaking
property in the following. Let us first show that p-reducedness implies the
leaking property. First note that a p-reduced divisor does not contain infi-
nite points. This is true since an infinite point is linearly equivalent to any
other point on the corresponding leaf edge and hence can be replaced by
any finite point, thereby decreasing the p-distance. Hence the leaking prop-
erty holds for any A ⊆ X [1] and it remains to check the case A[0] = A∩ X [0] 6=
;. In this case the function fA : X [0] → R, x 7→= dist(A[0] , x) is a well-defined
rational function on X . We set mA = max{ fA, −ε} where ε > 0 is chosen
such that the ε-ball around A does not contain p nor any vertices outside of
A. For all q ∈ A[0] we have ordmA (q) = ord fA (q) = valA(q) − valX (q). More
precisely, we get
X X
div(mA) = x+ (valA(q) − valX (q))q.
dist(A,x)=ε q∈A[0]
It follows that if the condition aq +valA(q) ≥ valX (q) holds for all q ∈ A, then
D0 := D + div(mA) is effective. We can see that D0 is strictly smaller with
224
8 Tropical curves
respect to p-distance by looking at a minimizing path from p to A. For points

x on this path we have dist(A, x)+dist(x, p) = dist(A, p). Hence D0 contains
a (new) point with dist(x, p) = dist(A, p) − ε, and (D0 − D)|Bp (dist(A,p)) > 0.
Hence D0 is strictly smaller than D, a contradiction.
Let us now assume D satisfies the leaking property for all A. First of all,
note that D does not contain an infinite point q since A = {q} implies aq < 1.
Let D0 be any other divisor equivalent to D and effective away from p and
assume that D0 . Let f be rational function with div( f ) = D0 −D. Let A be the
maximality domain of f (note that f is bounded since D does not contain
/ A, then by the leaking property there exists a point
infinite vertices). If p ∈
q ∈ A with aq < valX (q) − valX (q) ≤ − ord f (q). Hence D + div( f ) = D0 is
not effective at q, a contradiction. It follows p ∈ A. Now, choose a point
of minimal distance to p in Supp(div( f )) (including the case q = p). A
shortest path from p to q is disjoint from Supp(div( f )) except for q. Since
A ∩ Supp(div( f )) = ∂ A and p ∈ A, this implies q ∈ ∂ A and hence ord f (q) <
0. It follows that the coefficient of q in D is greater than in D0 showing that
D has smaller p-distance. (If q = p, we use deg(D) = deg(D0 ) and hence the
restriction of D to X \ {p} has lower degree than the restriction of D0 .)
Proposition 8.4.11
Fix p ∈ X [0] . Then every divisor class [D] ∈ Pic(X ) has a unique p-reduced
representative.
Proof. First we show existence. It follows from Proposition 8.4.7 that there
exists m ∈ Z such that |D + mp| 6= ; (choose m ≥ g − deg(D)). Pick the
minimal such m ∈ Z. Since |D+mp| is compact , we can choose E ∈ |D+mp|
with minimal p-distance. For any other D0 = −m0 p+E 0 ∼ D, E ≥ 0, we have
m0 ≤ m, by minimality of m, and hence deg(E 0 ) ≥ deg(E). This together
with the minimality of E implies that D00 = −mp + E is p-reduced.
For uniqueness, let D ∼ D0 be two p-reduced divisors in the same class.
Let f be a rational function such that div( f ) = D0 − D. The leaking property
for D implies that p is contained in the maximality domain of f (cf., the
proof of the previous proposition). Reversing the roles of D and D0 , the
same argument shows that p is contained in the minimality domain of f .
Hence f is constant and D = D0 .
We continue by defining an important class of divisors of “intermediate”
degree g − 1.
225
8 Tropical curves
Definition 8.4.12
An acyclic orientation O of X is a choice of directions of the edges of a sub-
division of X which does not admit oriented cycles. We denote by val+ (x)
the number of outgoing edges at x. The divisor
X
K+ := (val+ (x) − 1)x
x∈X
is called the moderator with respect to O .
PTaking the opposite orientation −O gives rise to the moderator K− =

x∈X (val− (x) − 1)x counting the incoming edges of O .
Exercise 8.4.13
Check deg(K+ ) = deg(K− ) = g − 1 and K+ + K− = K.
Note that the points with negative coefficient in K+ are exactly the sinks
of O . Since O is acyclic, there always exists at least one sink, hence K+ is
never effective. This is even true for the full equivalence class of K+ .
Lemma 8.4.14
For any moderator K+ it holds |K+ | = ;.
Proof. Let f be a rational function on X with E = K+ +div( f ) ≥ 0. First note

that f cannot have poles at infinity since the coefficient of infinite points
in K+ is at most 0. Let A be the maximality domain of f . The moderator
K+ is produced by some acyclic orientation O of X . Choose a sink q of the
restriction of O to A (i.e., a point in A such that all adjacent edges in A are
incoming). It follows val+ (q) ≤ valX (q) − valA(q) ≤ − ord f (q), which shows
that E is not effective at q.
The following lemma shows in which sense the divisors K+ "moderate"
between the effective and non-effective divisor classes in Pic(X ).
Lemma 8.4.15
For any divisor D ∈ Div(X ) exactly one of the following two options holds:
Either |D| =
6 ; or there exists a moderator K+ such that |K+ − D| =
6 ;.
Proof. By lemma 8.4.14 at most one of the two options holds true. It re-
mains to show that if |D| = ;, there exists a moderator K+ with |K+ − D| =
6 ;.
226
8 Tropical curves
Without loss of generality we may assume that D is p-reduced with respect

to some finite point p ∈ X . Fix a subdivision of X without loops such that p
and the points in Supp(D) are vertices. We will define an acyclic orientation
on X by ordering the vertices and then orienting the edges. from higher to
lower vertices. The order of the vertices is given inductively as follows. The
lowest vertex is v1 = p. Now assume v1 , . . . , vk have already been chosen.
We set A to be the full subgraph of X with vertices X (0) \ {v1 , . . . , vk }. By the
leaking property there exists a vertex v of A where equation (8.1) holds,
and we pick vk+1 = v. Note that the number of outgoing edges at v in the
acyclic orientation obtained that way is exactly valX (v) − valA(v), hence the
construction ensures that K+ − D is effective away from p. At p the con-
structed orientation has a sink, so the coefficient in K+ is −1. However,
since D = ;, the coefficient of p in D is strictly negative and hence K+ − D
is effective on all of X , proving the statement.
A corollary of this proof is the inverse of lemma 8.4.14.
Corollary 8.4.16
Let D be a divisor of degree g − 1. If |D| = ; then D is linearly equivalent
to a moderator. Moreover, if D is p-reduced, then D is equal to a moderator.
Finally, we have |D| = ; if and only if |K − D| = ;.
Proof. By lemma 8.4.15 there exists a moderator with |K+ − D| = 6 ;. Since

deg(K+ − D) = 0, this implies K+ − D ∼ 0. If D is p-reduced, the construction
from 8.4.15 even gives K+ −D ≥ 0 and hence K+ −D = 0. The final statement
follows from the fact that K− = K − K+ is also a moderator.
The above computations lead to a useful criterion for testing non-emptyness

of linear systems.
Corollary 8.4.17
Let D be a divisor of degree d < g − 1. If |D + q| =
6 ; for all q then |D| =
6 ;.
Proof. If D = ;, there exists a moderator K+ and effective E ∼ K+ − D.

Choose q ∈ Supp(E), then |K+ − D − q| 6= ; and therefore |D + q| = ;.
A divisor D is called special if rk(D) > max d − g, −1. By 8.4.7 D can only
be special if 0 ≤ d ≤ 2g − 2. In the range 0 ≤ d ≤ g − 1, specialness is
equivalent to |D| 6= ;.
227
8 Tropical curves
Lemma 8.4.18
A divisor D is special if and only if K − D is special.
Proof. By symmetry we may assume 0 ≤ deg(D) ≤ g − 1. We set e :=

g − 1 − d = deg(K − D) − g + 1. If K − D is special, then for all effective
divisors of degree e we have |K − D − E| = 6 ;. This implies |D + E| = 6 ;
by 8.4.16 (since these linear systems have degree g − 1). Applying 8.4.17
several times, we obtain D 6= ;. Conversely, assume K − D is non-special.
Then there exists effective E of degree e with |K − D − E| = ;. Again this
implies |D + E| = ; and hence |D| = ;.
We are now ready to prove Riemann-Roch theorem.
Proof of 8.4.5. By exchanging D and K − D it is enough to prove the in-

equality
rk(D) ≥ rk(K − D) + d − g + 1 =: e.
The case rk(K − D) = −1 is contained in 8.4.7, hence we may assume rk(K −
D) > −1. Let E be an effective divisor of degree e. We have
rk(K − D + E) ≥ rk(K − D) = deg(K − D + E) − g + 1,
hence K − D + E is special. It follows that D − E is special by 8.4.18 and

in particular |D − E| =
6 ;. This implies rk(D) ≥ e and hence finishes the
proof.
228
9 Tropical jacobians
9.1 (Coarse) differential forms, integration,
and the Jacobian
The canonical line bundle plays a very special role in the theory of curves.
Its sections are differential forms which have special properties and allow
for special constructions which are not present in sections of arbitrary line
bundles. These peculiarities of the canonical bundle, which manifest clas-
sically in the theory of jacobians, have a tropical counterpart, which is pre-
sented here.
In the following, we will always assume C to be a smooth tropical curve
(and add finite type or compactness when necessary). Hence C can be
represented by a metric graph with open and closed ends (where the latter
are closed up by infinite vertices). Recall from Definition 7.3.4 that we can
define the cotangent sheaf on C in terms of the exact sequence
0 → R → R ⊗Z Aff ∗C → TC∗ → 0,
where the first map is just the inclusion of constant functions. Note that the
tensor product in the middle term is a product over Z, where Z acts on Aff ∗C
by (non-tropical) multiplication. In other words, R ⊗Z Aff ∗C is the sheaf of
affine linear functions where we also allow non-integer slopes.
Definition 9.1.1
A residue form ω is a global section of the cotangent sheaf TC∗ . The set of
all residue forms constitutes a real vector space which we denote by Ω(C).
Remark 9.1.2
Note that residue forms are (locally constant) classical differential forms on
the interior of edges. The name reflects the fact that these forms appear
229
as limits of the residues of classical differential forms when tropicalizing

a complex curve to a tropical curve. In a sense that can be made precise,
residue forms are therefore only “half-tropicalized” objects and should be
distinguished from honest tropical differential forms, which can be defined
as sections of the canonical line bundle. While the latter are arguably the
more intrisically tropical objects, residue forms play an important role in
some contexts, in particular, in the context of tropical jacobians.
Since the function in Aff ∗C are locally affine linear, any residue form is
locally constant (in particular, constant on the interior of edges). Moreover,
it is zero on any closed end (since Tp∗ = 0 for any infinite point p). For a
given edge e with endpoint p, we can evaluate ω on the primitive vector
ve/p ∈ ZTp . If we fix vertex p ∈ C and let v1 , . . . , vk denote the primitive
generators of all the adjacent edges, then the balancing condition translates
to
Xk
ω(vi ) = 0.
i=1
(Note that the statements holds trivially for any infinite endpoint by the
previous remark). Hence, a residue form can alternatively described by the
following data.
Proposition 9.1.3
A residue form ω on C is uniquely described by a assignment of numbers to
each oriented edge
ω : {oriented edge of C} → R,
such that
• if e+ and e− are the two orientations of an edge, then ω(e+ ) = −ω(e− ),
• for any vertex p, if e1 , . . . , ek denote all the outgoing edges from p, then
k
X
ω(ei ) = 0.
i=1
Of course, due to the equation in the proposition, the values of ω on ori-

ented edges are still largely redundant. In fact, Ω(C) is a g +m-dimensional
230
vector space, where
g := dim H1 (C, R), m := max(#{open ends} − 1, 0).
To prove this, we need some terminology. A collection of 2-valent points

c1 , . . . , ck ∈ C is called a tree cut if C \ c1 , . . . , ck is simply connected.
Exercise 9.1.4
Show that if c1 , . . . , ck is a tree cut, then k = g.
Proposition 9.1.5
Let c1 , . . . , c g be a tree cut of C. Let e1 , . . . , e g be oriented edges with ci ∈ ei ,
and let e g+1 , . . . , e g+m+1 be the set of open ends of C (oriented away from the
end). Then the map
Ω(C) → R g × Rm ,
ω 7→ (ω(e1 ), . . . , ω(e g+m )),
is an isomorphism of real vector spaces. In particular, the dimension of Ω(C)

is equal to g + m.
Proof. The map is obviously linear. Note that the contraction of a closed
infinite end does neither affect Ω(C) nor the map above. We can hence
contract all such ends and obtain a curve without such ends (or T1 or TP1 ,
which obviously satisfy the statement).
Let us now assume that C is a tree, i.e., g = 0. Let us fix some vector
v ∈ R g+m . If C does not contain an inner edge, then v obviously has a unique
preimage residue form by the equations in Proposition 9.1.3. Otherwise, let
e be an inner edge and C1 and C2 be the two curves obtained by cutting C
along some interior point of e. We may assume that C2 contains the end
em+1 . By induction, the statement is true for C1 and there exists a unique
residue form ω1 on C1 determined by the values of v (restricted to the ends
of C1 ). In particular, ω1 has some value on the cut open end e. If we
assign −ω1 (e) to the opposite end in C2 , the remaining values of v, again
by induction, fix a unique residue form ω2 on C2 . Since the two residue
forms are compatible on e, they can be combined to a residue form ω on
C, and the uniqueness on C1 , C2 implies the uniqueness on C. This finishes
the tree case.
231
Finally, let us assume g > 0. We denote by C 0 := C \ c1 the curve of genus

g − 1 obtained by removing the first cut point. If we denote by e+ , e− the
two freshly created open ends of C 0 (we may assume that e+ and e1 are
oriented equally), we can identify the sets
Ω(C) = {ω0 ∈ Ω(C 0 ) : ω(e+ ) = −ω(e− )}. (9.1)
By induction on g, we may assume that the map Ω(C 0 ) → R g+m+1 is an iso-

morphism (for the tree cut c1 , . . . , c g−1 and ends ordered as e+ , e− , e1 , . . . , em+1 ).
The linear map is question is then given by the concatenation
Ω(C) → Ω(C 0 ) → R g+m+1 → R g+m ,
where the last map is the projection which forgets the e− -coordinate (and
moves the e+ -coordinate to the right place). It is hence still an isomorphism
by Equation (9.1).
We can restrict the aforementioned exact sequence to the integers,
0 → R → Aff ∗C → ZTC∗ → 0,
and obtain a canonical integer lattice in Ω(C).
Definition 9.1.6
A integer residue form ω is a global section of the integer cotangent sheaf
ZTC∗ , or, in the description of Proposition 9.1.3, satisfies ω(e) ∈ Z for any
oriented edge e. The set of all integer residue forms, denoted by ΩZ (C),
constitutes a lattice in Ω(C) and hence turns Ω(C) into a tropical vector
space in the sense of Section 2.1.
As explained above, residue forms are honest classical differential forms
on the interior of edges and therefore can be integrated along (compact)
paths. To fix notation, let us explicitly define this integration here.
Definition 9.1.7
Let ω be a residue form on C and let e be an oriented bounded edge. We
define the edge integral of ω along e to be
Z
ω := l(e)ω(e).
e
232
Note that this is can be extended to closed leavess, since then ω(e) = 0.
Given an oriented 1-chain α = e1 + . . . + ek (i.e., a formal sum of oriented
bounded edges), we can extend integration by linearity
Z k
X
ω := l(ei )ω(ei ).
α i=1
Given the opposite orientation e+ and e− of an edge e, we obviously get

Z
ω = l(e)(ω(e+ ) + ω(e− )) = 0.
e+ +e−
It follows that integration gives a well-defined pairing on homology.
Definition 9.1.8
We define integration pairing between Ω(C) and H1 (C, R) by
Z
: Ω(C) × H1 (C, R) → R,
Z
(ω, [α]) 7→ ω.
α
Proposition 9.1.9 R
Let C be a compact smooth tropical curve. Then the integration pairing is
non-degenerate.
Proof. We have toR construct for any non-zero residue form ω a suitable
cycle α such that α ω 6= 0 — and vice versa. Let us start with a non-zero
R
ω. Assume that α ω = 0 for all cycles α of C. This would imply that ω can
be integrated to a affine linear function
Z p
f (p) := ω ∈ Aff (C).
p0
By Proposition 8.2.4 we know that f must be a constant function, which

implies ω = d f = 0.
Vice versa, let α be a cycle determining a non-zero homology class [α] ∈
H1 (C, R). Let ωα (e) be the coefficient of the oriented edge e in α (when
233
writing α as a linear combination of oriented edges using only the given

orientation of e). The fact that α is closed translates directly to the zero di-
vergence condition in Proposition 9.1.3, hence this way we obtain a residue
form ωα on C. Integration gives
Z X
ωα = ωα (e)2 l(e) > 0,
α e
where the sum runs through the edges of C (with arbitrary orientation).
This proves the claim. which proves the claim.
Let Ω(C)∗ be the dual vector space of Ω(C). Equipped with the dual
lattice ΩZ (C)∗R
, this is again a tropical vector space. By Proposition 9.1.9, the
isomorphism : H1 (C, R) → Ω(C)∗ allows us allows us to regard H1 (C, Z)
as another lattice in Ω(C)∗ . The quotient of Ω(C)∗ by this second lattice
carries a canonical structure as smooth tropical manifold and as group (see
Example 7.1.9 and the next section).
Definition 9.1.10
Let C be a compact smooth tropical curve. We define the Jacobian of C as
the quotient of the tropical vector space Ω(C)∗ by the lattice H1 (C, Z),
Jac(C) := Ω(C)∗ /H1 (C, Z).
Before we proceed to analyze the specific properties of jacobians, it is

worthwhile to make a quick stop and have a look at the general features of
this construction.
9.2 Abelian varieties, polarization, and theta

functions
Analogous to the classical case, the jacobian of a tropical curve is a spe-
cial example of more general objects called tropical abelian varieties and
tropical tori.
The ingredients needed to define s tropical torus are
• a real vector space V ,
234
• two lattices L, Λ ⊂ V .
However, the lattices play different roles in the construction. Namely, we
use L to determine the tropical structure of V (i.e., the integral tangent
vectors), whereas we use Λ to quotient by (i.e., to fix the metric structure).
Here is the exact definition.
Definition 9.2.1
A tropical torus A is the quotient of a tropical vector space V = R ⊗ L by a
lattice Λ ⊂ V ,
A = (R ⊗ L)/Λ.
A tropical torus is compact regular tropical variety in a canonical way.
Exercise 9.2.2
Show that the restrictions of the quotient map V → A to sufficiently small
domains provide a regular tropical atlas for A.
Example 9.2.3
The jacobian of a compact smooth tropical curve of genus g is tropical torus
of dimension g, with V = Ω(C)∗ , L = ΩZ (C)∗ and Λ = H1 (C, Z) (embedded
via integration).
Remark 9.2.4
Let A be tropical variety homeomorphic to a topological torus (S 1 )n (this
implies that A is regular), and let p ∈ A be a fixed point. Parallel transport
of integer tangent vectors produces a monodromy action
γ : H1 (A, Z) → GL(ZTA,p ).
Remark 9.2.5
tori up to isom are equal to (V, L, Λ) up to isom
Tropical tori often come equipped with additional structure. In the fol-
lowing, we use the identification
H 1 (A, ZTA∗ ) = Λ∗ ⊗ L ∗
and think of an element in this space either as a bilinear form on V or as
a linear map P : Λ → L ∗ . We also denote the space of symmetric bilinear
forms on V by Sym2 V ∗ ⊂ V ∗ ⊗ V ∗ .
235
Definition 9.2.6
A polarization P of the tropical torus A is an element
P ∈ (Λ∗ ⊗ L ∗ ) ∩ Sym2 V ∗ ,
such that the corresponding symmetric bilinear form is positive definite. In

other words, P is a scalar product on V such that
P(v, w) ∈ Z for all v ∈ Λ, w ∈ L.
In this case, the pair (A, P) is called a tropical abelian variety.

A polarization P is called principal if the linear map P : Λ → L ∗ is an iso-
morphism of lattices. In terms of lattice bases v1 , . . . , vn of Λ and w1 , . . . , w n
of L, this is equivalent to the integer matrix P(vi , w j ) to be invertible over Z.
A tropical torus together with a principal polarization is called principally
polarized.
Remark 9.2.7
Given a principal polarization P, we can dualize the representation of A and
obtain a canonical isomorphism
A∼
= (R ⊗ Λ∗ )/L ∗ . (9.2)
Note that the roles of Λ and L get exchanged.
To any polarization of a tropical torus, we can associate a theta function

and a theta divisor.
Definition 9.2.8
Let A be a tropical abelian variety with polarization P. We define the theta
function ϑ : V → R by
1
ϑ(x) := max{P(λ, x) − P(λ, λ)}. (9.3)
λ∈Λ 2
Note that the maximum always exists since P is positive definite (and hence
−P(λ, λ) is a negative definite quadratic functional).
The theta function satisfies ϑ(x) = ϑ(−x) and is quasi-periodic in Λ in

the following sense.
236
Lemma 9.2.9
For any x ∈ V and µ ∈ Λ the theta function satisfies
1
ϑ(x + µ) = ϑ(x) + P(µ, x) + P(µ, µ). (9.4)
2
In particular, for fixed µ the difference function ϑ(x + µ) − ϑ(x) is affine Z-
linear in x.
Proof. By shifting the λ to λ + µ we can evaluate the theta function as
1
ϑ(x + µ) := max{P(λ + µ, x + µ) − P(λ + µ, λ + µ)}. (9.5)
λ∈Λ 2
Multiplying out this type of term we get
1
P(λ + µ, x + µ) − P(λ + µ, λ + µ)
2
1
= P(λ, x) + P(µ, x) + P(λ, µ) + P(µ, µ) − P(λ, λ) + 2P(λ, µ) + P(µ, µ)
2
1 1
= P(λ, x) − P(λ, λ) + P(µ, x) − P(µ, µ). (9.6)
2 2
Since the first two terms in the last expression coincide with the terms in
Equation (9.3), the first statement follows. The second statement holds
since P ∈ Λ∗ ⊗ L ∗ . Hence for fixed µ ∈ Λ, we have P(µ, .) ∈ L ∗ and P(µ, x)−
1
2 P(µ, µ) is indeed affine Z-linear on R ⊗ L.
The quasi-periodicity of ϑ has some interesting consequences. In partic-

ular, it follows that div(ϑ) ⊂ V is a Λ-periodic divisor and hence induces a
well-defined divisor Θ ⊂ A. In fact, ϑ defines a Cartier divisor on A (hence
a line bundle), and Θ is the Weil divisor associated to this Cartier divisor.
Definition 9.2.10
Given a tropical abelian variety A, we call
Θ = div(ϑ)/Λ ⊂ A (9.7)
the Theta divisor of A.
Remark 9.2.11
The definition depends upt to translation on the fixed group structure of A
(alternatively, on a fixed marked point 0 ∈ A).
237
9.3 The Abel-Jacobi theorem

Recall that given a compact smooth tropical curve C, we define the jacobian
of C as the tropical torus
Jac(C) := Ω(C)∗ /H1 (C, Z). (9.8)
Also recall that this notation implicitly uses the isomorphism

Z
: H (C, R) ∼
1 = Ω(C)∗ . (9.9)
We continue to use this identification in the following definition of a canon-

ical polarization of Jac(C). In the following, we fix arbitrary orientations
on the edges of C (for some graph structure) and think of H1 (C, R) as a
subspace of the space of inner 1-chains (i.e., the real vector space formally
generated by the inner edges of C). As before, the length of an inner edge
is denoted by l(e).
Definition 9.3.1
We define the bilinear form P on the space of inner 1-chains as bilinear
extension of ¨
l(e) if e = f ,
P(e, f ) = (9.10)
0 otherwise,
where e, f are two inner edges of C.
Proposition 9.3.2
The bilinear form P on H1 (C, R) is symmetric, positive definite and lies in
Z −1
P ∈ H1 (C, Z) ⊗
∗
ΩZ (C). (9.11)
Moreover, P is principal and hence turns Jac(C) into a principally polarized

abelian variety.
Proof. Let R{inner edges} denote the space of inner 1−chains. By defini-
tion, P has diagonal form with respect to the standard basis of R{inner edges}
238
and the diagonal entries l(e) are strictly positive. Hence P is symmetric and
positive definite on R{inner edges}, thus also on H1 (C, R).
For the integrality statement, note we can think of Ω(C) as a subspace
of R{inner edges} by Proposition 9.1.5. In fact, this is exactly the same
subspace as H1 (C, R) ⊂ R{inner edges}, since both spaces are given by the
condition that the total in and out flow at each vertex is zero. The induced
identification
H1 (C, R) = Ω(C) (9.12)
is denoted by ω(α) for α ∈ H1 (C, R) in the following. Note that identifies
H1 (C, Z) with ΩZ (C). Unwrapping the definition, it turns out that P is noth-
ing else but the bilinear form associated to the isomorphism H1 (C, R) ∼ =
H1 (C, R)∗ obtained by concatenating the above identification (9.12) with
integration. In formulas, this just means
Z
P(α, β) = ω(β). (9.13)
α
The isomorphism H1 (C, R) ∼

= H1 (C, R)∗ descends to an isomorphism of lat-
tices Z −1
H1 (C, Z) ∼
= ΩZ (C), (9.14)
R −1
which shows that P ∈ H1 (C, Z)∗ ⊗ ΩZ (C) is indeed a principal polariza-
tion.
Remark 9.3.3
Using the canonical isomorphisms
H 1 (C, Z) ∼
= H1 (C, Z)∗ , H 1 (C, R) ∼
= R ⊗ H 1 (C, Z), (9.15)
and the principal polarization P, by Remark 9.2.7 we obtain a second de-

scription of the jacobian as
Jac(C) ∼
= (R ⊗ H 1 (C, Z))/ΩZ (C), (9.16)
where the embedding ΩZ (C) ,→ H 1 (C, R) is given by integration.
239
Recall from Proposition 8.3.14 that on a smooth curve C, line bundles are
in one-to-one correspondence to rational equivalence classes of divisors
Pic(C) ∼
= Div(C)/ ∼ . (9.17)
Moreover, by Proposition 8.2.3 we know that two rationally equivalent di-

visors on a compact curve C have the same degree. This gives rise to a
well-defined well-defined surjective group homomorphism
deg : Pic(C) ∼
= Div(C)/ ∼→ Z. (9.18)
The fibers of the degree map split the Picard group into isomorphic pieces
X
Pic(C) = Picd (C) (9.19)
d∈Z
parameterizing line bundles of a given degree d. The Abel-Jacobi theorem

states that Pic0 (C) (and hence any Picd (C)) is isomorphic to Jac(C). Let us
proceed to prove this.
As before, we assume in the following that C is a compact smooth curve.
We fix a base point p0 ∈ C. Let p ∈ P be another
R point, and choose some
path γ in C from p0 to p. By integration, γ defines an element in Ω(C)∗ .
R
Choosing another γ0 from p0 to p results in another linear functional γ0 ,
but since γ − γ0 ∈ H1 (X , Z) forms a closed cycle, the difference
Z Z
− ∈ H1 (X , Z), (9.20)
γ γ0
and both functional determine the same element in Jac(C), which we de-
note by Z p
∈ Jac(C). (9.21)
p0
Definition 9.3.4
Given a divisor D = ai pi ∈ Divd (C) of degree d, we define
P
Z D X Z pi
µ(D) := := ai ∈ Jac(C). (9.22)

p0
240
This defines a group homomorphism
µ : Divd (C) → Jac(C), (9.23)
which is called the tropical Abel-Jacobi map.
Remark 9.3.5
Note that µ depends on the choice of p0 for d 6= 0. Indeed, choosing another
R p0
base point p00 results in an additive shift of the map by d p 0 . For d =
0
0 however, the dependence on p0 vanishes and we obtain a canonically
defined map Pic0 (C) → Jac(C).
Theorem 9.3.6 (Tropical Abel-Jacobi theorem)

The Abel-Jacobi map µ : Divd (C) → Jac(C) is constant on rational equiva-
lence classes,
D ∼ D0 =⇒ µ(D) = µ(D0 ). (9.24)
Moreover, the induced map ν : Picd (C) → Jac(C) is a bijection.
Recall that O : Divd (C) → Picd (C) is the map which associates to the
divisor D the line bundle O (D). The theorem can be summarized in the
following commutative diagram.
Divd (C)
µ
O (9.25)
∼
=
Picd (C) ν Jac(C)
Similar to the classical situation, the statement naturally splits into two
parts called Abel’s theorem (“ν is well-defined and injective”) and Jacobi
inversion theorem (“ν is surjective”). We start with tropical Abel.
Theorem 9.3.7 (Tropical Abel’s theorem)

Two divisors D, D0 are rationally equivalent if and only if µ(D) = µ(D0 ).
Proof. To prove the first direction, we show that if D is a divisor rationally

equivalent to zero, then µ(D) = 0. Let f be a rational function such that
div( f ) = D. Choose a graph structure for C which contains the support of
241
D, and fix an orientation of the edges. Then the slope of f on each inner
edge defines an (oriented) 1-chain which we denote by
d f ∈ R{inner edges}. (9.26)
By definition, µ(D) is just the class in Jac(C) of the linear functional on

Ω(C) given by integration over this 1-chain. Using again the identification
Ω(C) = H1 (C, R) ⊂ R{inner edges}, we may now exchange the roles of
integration chain and integrand. Namely, for any residue form ω let α be
the corresponding 1-cycle, we get
Z Z
ω = P(d f , α) = d f = 0, (9.27)
df α
(by slight abuse of notation, since d f is not balanced at all vertices, hence
not a residue form). This shows µ(D) = 0, which proves the first claim.
We proceed to show the second direction. We have to show that any D ∈
Div0 (C) with µ(D) = 0 is rationally equivalent to zero. Let γ an oriented
1-chain with boundary D. By definition, µ(D) is described by integration
over γ. Since µ(D) = 0, we can in fact choose γ such that
Z
= 0 ∈ Ω(C)∗ . (9.28)
γ
Again, we can essentially turn around domain and integrand and get
Z
γ=0 (9.29)
α
for all α ∈ H1 (C, R). RIt follows that we can “integrate” γ and obtain a
rational function f := γ whose divisor is equal to the boundary of γ, i.e.,
div( f ) = D.
The surjectivity of ν resp. µ can be solved in a rather explicit way, and

it is worthwhile to look at this in some detail. The key construction is to
pull back translated theta divisors from Jac(C) to C. To do so, we need
a map from C to Jac(C). In fact, we can use the Abel-Jacobi map µ via
the inclusion C ⊂ Div1 (C) which associates to each point of C the divisor
consisting of only this point.
242
Lemma 9.3.8
The Abel-Jacobi map
µ|C : C → Jac(C). (9.30)
is a tropical morphism.
Proof. The map is constant in a neighbourhood of the infinite points of C
(in fact, it contracts any end to point) since two points on the same end
are rationally equivalent. It remains to check the claim for a finite point
p ∈ C. Let us fix a chart containing p to the n + 1-valent standard tropical
line L(n + 1) ⊂ Rn with n + 1 = val(p). We may assume that the base point
p0 in the definition of µ is equal to p. Let e0 , . . . , en denote the oriented
edges of C corresponding to the rays of L(n + 1), and let ϕi ∈ Ω(C)∗ denote
the evaluations
ϕi (ω) := ω(ei ). (9.31)
It holds ϕ0 +. . . ϕn = 0. Moreover, for any point p0 on the i-th ray of L(n+1),
we find µ(p0 ) = lϕi where l is the lattice distance of p0 from p. It follows
that µ is locally given by the linear map
Rn → Ω(C)∗ ,
(9.32)
(x 1 , . . . , x n ) 7→ x 1 ϕ1 + . . . x n ϕn .
This shows that µ is a tropical morphism.
Let f : X → Y be a tropical morphism and let s = {(Ui , si )} be a Cartier
divisor on Y . We define the pull back of s along f to be the Cartier divisor
f ∗ s := {( f −1 (Ui ), si ◦ f )}. (9.33)
Exercise 9.3.9
Check that f ∗ s is a well-defined Cartier divisor on X .
Recall that the polarization P on Jac(C) provides us with a theta function
ϑ which we can either regard as a quasi-periodic function on V or a Cartier
divisor on Jac(C). For any y ∈ Jac(C), we the shift function
ψ y : Jac(C) → Jac(C),
(9.34)
x 7→ x + y,
is a tropical isomorphism. We denote the shifted theta function by
ϑ y := ψ∗− y ϑ. (9.35)
243
Definition 9.3.10
The Jacobi inverse of y ∈ Jac(C) is defined to be the divisor
D( y) := div(µ|∗C ϑ y ) ∈ Div(C). (9.36)
Note that the induced Jacobi inversion map
Jac(C) → Div(C) (9.37)
still depends on the choice of a base point p0 .
Proposition 9.3.11
For any y ∈ Jac(C) the divisor D( y) is an effective divisor of degree g.
Proof. The effectiveness follows from the fact that ϑ is locally a tropical
polynomial, i.e., a maximum of finitely many terms. To compute the degree,
let us first make the construction of D( y) more explicit. Let c1 , . . . , c g ∈ C
be a tree cut disjoint from D( y) and let e1 , . . . , e g be oriented edges with
ci ∈ ei . These choices in fact provide us with explicit bases for all vector
spaces and lattices involved. Namely we get lattice bases
ω1 , . . . , ω g ∈ ΩZ (C),
ϕ1 , . . . , ϕ g ∈ ΩZ (C)∗ ,
α1 , . . . , α g ∈ H1 (C, Z),
where
• the ωi correspond to the standard basis of R g via Proposition 9.1.5,
i.e., ωi (e j ) = δi j ,
• the ϕi form the dual basis to the ωi , i.e., ϕi (ω) = ω(ei ),
• αi is the unique simple loop in Ce ∪ {ci } which traverses ei positively

oriented (where Ce = C \ {c1 , . . . , c g }).
The first and third basis actually coincide via the identification (9.12) and
we get
P(αi , ϕi ) = δi j . (9.38)
Since Ce is a tree, we can lift the map µ|Ce : Ce → Jac(C) to a map µ
e :
C → Ω(C) and can compute D( y) as the divisor of the rational function
e ∗
244
µ
e∗ ϕ y . Hence the degree of D( y) is equal to sum of the outgoing slopes
of ϕ y on the open ends of C.e Indeed, let us denote by z + , z − the ends of
i i
Ce coming from the cut point ci . Here, we use the convention that both
ends are oriented outwards and the orientation of zi+ is compatible with ei .
Denoting the slope of µ
e∗ ϕ y on zi± by si± , a straightforward generalization of
Proposition 8.2.3 shows
X
deg(D( y)) = si+ + si− . (9.39)
We now use the following two facts:
(a) The image under µ e of the primitive generator of zi± is exactly ±ϕi ∈
ΩZ (C) . This follows from the simplest case of the calculation in the
∗
proof of Lemma 9.3.8 (n = 1).
(b) The difference of the image points of the ends of zi+ minus zi− in Ω(C)∗
is equal to αi (since the difference of integration from some base point
to the endpoint of zi+ minus integration to the endpoint of zi− is the
same as integration over αi ).
We can now use the quasi-periodicity of ϑ y and compute the sum of slopes
si+ + si− simultanuously by
si+ + si− = d(ϑ y (x + αi ) − ϑ y (x))(ϕi ) by (a), (b)

= P(αi , ϕi ) by 9.2.9
=1 by (9.38) (9.40)
Since there are exactly g such pairs of ends, we obtain

X
deg(D( y)) = si+ + si− = g, (9.41)
as promised.
With just a little more effort, we are now ready to prove that the Jacobi
inversion map is essentially inverse to the Abel map ν, finishing the proof
of Theorem 9.3.6. This is in fact not quite true — we additionally have to
shift ν by a constant.
245
Theorem 9.3.12 (Jacobi inversion theorem)

There exists an element κ ∈ Jac(C) called the Jacobi inversion constant such
that for all y ∈ Jac(C) we have
µ(D( y)) + κ = y. (9.42)
In particular, the map µ : Sym g (C) → Jac(C) and hence all maps µ : Divd (C) →
Jac(C) are surjective.
Remark 9.3.13
Note that κ is constant with respect to varying y — it might however depend
the base point p0 . An investigation of this dependency and the meaning of
κ will be part of Section 9.4.
In the previous proof, we used the fact that the degree of the divisor of a
rational function can be computed in terms of the behavior of the function
on the ends of the curve. In fact, we can extract even a little more informa-
tion from this boundary behavior. This is the extra ingredient we need to
prove the Jacobi inversion theorem.
Lemma 9.3.14
Let C ⊂ Rn be a bounded tropical curve (not necessarily connected nor smooth).
We denote the boundary point of C by
∂ C = {b1 , . . . , bk }. (9.43)
Let v1 , . . . , vk be the primitive generators of corresponding open ends of C,

oriented outwards (i.e. towards the bi ), and let w1 , . . . , w k be the
P weights of
the ends. Let f be a rational function of C and let div( f ) = ai pi be its
divisor. We extend f to the points bi by continuity. Then the equation
X k
X
ai pi = w j (d f (v j )b j − f (b j )v j ) (9.44)
j=1
holds. Here, the sums on both sides refer to sums of vectors in Rn .
Proof. Both sides of the formula behave linearly with respect to the con-
nceted components of C. On the other hand, when cutting the curve into
246
more pieces by removing a two-valent vertex not contained in the support

of div( f ), both sides remain unchanged. It therefore suffices to prove the
case div( f ) = ap and p is the only vertex of C. In this case, we may write
b j = p + l j v j and then get f (b j ) = f (p) + l j d f (v j ). A straightforward com-
putation then gives
k
X k
X
w j (d f (v j )b j − f (b j )v j ) = w j d f (v j )p + w j d f (v j )l j v j
j=1 j=1

− w j f (p)v j − w j l j d f (v j )v j (9.45)
k
X
= (w j d f (v j ))p = ap.
j=1
Here in the second step, the second and fourth term cancel while the third
term is zero by the balancing condition.
Proof of Theorem 9.3.12. We use exactly the same setup and notation as in
the proof of Proposition 9.3.11, i.e. we fix a tree cut of C and consider the
lifted map µe : Ce → Ω(C)∗ . We pick y ∈ Ω(C)∗ from some local chart of
Jac(C) such that the Jacobi inverse D( y) has support disjoint from the tree
cut (by choosing different tree cuts, we can cover Ω(C) ∗
P completely). Since
µ is group homomorphism, the image of D( y) = ai pi under µ can be
computed as X
µ(D( y)) = [ ai µ
e(pi )]. (9.46)
Here, the sum on the right hand side is a standard sum of vectors in Ω(C)∗ ,
and the square brackets denote its equivalence class in Jac(C). By projec-
tion formula,we have X
ai µ
e(pi ) = div(ϑ
ey ), (9.47)
where ϑey = ϑ y | e is the restriction of the shifted theta function to the

µ
e(C)
image of µ
e. Hence we can apply Lemma 9.3.14 and obtain
X g
X g
X
ai µ
e(pi ) = s+j b+j − ϑ y (b+j )ϕ j + s−j b−j + ϑ y (b−j )ϕ j
j=1 j=1
247
and collecting all constant terms with respect to y on the left, we get
g
X
= κ1 − (ϑ(b−j + α j − y) − ϑ(b−j − y))ϕ j
j=1
= κ2 − P(α j , b−j − y)ϕ j by 9.2.9

= κ3 + P(α j , y)ϕ j
= κ3 + y by (9.38)
This proves the claim.
Remark 9.3.15
The Jacobi inversion theorem and the Abel-Jacobi theorem have some nice
corollaries (e.g. used in the proof of Proposition 8.4.7).
(a) Let D ∈ Div g (C) be an arbitrary divisor of degree g. Then
D0 := D(µ(D) + κ) > 0 (9.48)
is an effective divisor. Jacobi inversion implies µ(D) = µ(D0 ), hence

Abel’s theorem implies D ∼ D0 . It follows that the linear system |D| is
non-empty whenever deg(D) = g (an hence also deg(D) ≥ g).
(b) The previous statement can generalized to divisors D of arbitrary de-

gree. In the same way as before, we can then conclude
D ∼ D(µ(D) + κ) + (d − g)p0 . (9.49)
(c) By Lemma 9.3.8, the map
µ|dC : C × · · · × C → Divd (C) → Jac(C) (9.50)
is a tropical morphism. By Abel-Jacobi, its image µ(Symd (C)) corre-

sponds to non-empty linear systems of degree d (among all divisor
classes of degree d). We have
dim(C d ) = d, dim(Jac(C)) = g, (9.51)
which implies that µ|dC cannot be surjective for d < g. Hence for all
d < g, there exists a divisor D of degree d with |D| = ;.
248
The proofs we provided for the Abel-Jacobi theorem have the advantage
of being of rather explicit nature. However, there is somewhat more funda-
mental approach which exploits the full power hidden in the exact sequence
we started with, namely
0 → R → Aff ∗C → ZTC∗ → 0. (9.52)
In fact, this sequence plays a role very similar to the exponential sequence
in classical geometry, and the long exact sequence of cohomology groups
essentially contains the Abel-Jacobi theorem. To see this, let us recall from
Remark 9.3.3 that we define the jacobian of a smooth compact curve equiv-
alently as
Jac(C) = Ω(C)∗ /H1 (C, Z) ∼
= H 1 (C, R)/ΩZ (C). (9.53)
Let us now look at the non-zero terms in the long exact sequence associated
to (9.52)
0 H 0 (C, R) H 0 (C, Aff ∗C ) H 0 (C, ZTC∗ )
H 1 (C, R) H 1 (C, Aff ∗C ) H 1 (C, ZTC∗ ) 0

(9.54)
The first map H 0 (C, R) ,→ H 0 (C, Aff ∗C ) embeds the constant functions inside
the affine Z-linear functions on C. Since C is compact (and connected),
this map is fact the isormorphism H 0 (C, R) ∼ =R∼
∗
= H 0 (C, Aff C ). Hence the
sequence splits after the second term, and the latter part can be rewritten
as
0 → ΩZ (C) → H 1 (C, R) → Pic(C) → H 1 (C, ZTC∗ ) → 0 (9.55)
after using the isomorphisms
ΩZ (C) ∼
= H 0 (C, ZTC∗ ), Pic(C) ∼ ∗
= H 1 (C, Aff C ). (9.56)
We claim that this sequence encodes the Abel-Jacobi theorem, details being
left as exercise.
Exercise 9.3.16
Let C be a compact smooth curve.
249
(a) Show that H 1 (C, Aff ∗C ) ∼

= Z. Show that a unique generator is given
by the 1-cochain Ce which maps the oriented edge e to the covector
having slope 1 in the direction of e, and evaluates to zero at all other
edges.
(b) Show that with the above identification, the map Pic(C) → H 1 (C, ZTC∗ )
in (9.55) is equal to the degree map deg : Pic(C) → Z.
(c) Show that the map ΩZ (C) → H 1 (C, R) in (9.55) is equal to the em-
bedding given by integration.
(d) Let ψ : H 1 (C, R) → Pic(C) denote the map in the middle of (9.55).
Given D ∈ Div0 (C), let α be a 1-chain in C with boundary ∂ α = D.
Let P(α, .) ∈ H1 (C, R)∗ ∼
= H 1 (C, R) be the linear functional. Show that
ψ(P(α, .)) = O (D). (9.57)
These statements imply the Abel-Jacobi theorem Jac(C) ∼

= Pic0 (C).
9.4 Riemann’s theorem

Recall that the appearance of the Jacobi inversion constant κ ∈ Jac(C) in
Theorem 9.3.12 (in general, κ depends on the choice of base point p0 ). Rie-
mann’s theorem gives a more explicit reinterpretation of this constant. As
before, we denote by Symd (C) ⊂ Divd (C) the subset of all effective divisors
in Divd (C). We set
Wd := µ(Symd (C)) ⊂ Jac(C). (9.58)
We regard the Wd only as sets here, even though they carry the structure
of closed tropical subspaces. We already know Wg = Jac(C) by the Jacobi
inversion theorem. Riemann’s theorem identifies Wg−1 .
Theorem 9.4.1 (Tropical Riemann’s theorem)

Let C be a compact smooth tropical curve (with some fixed base point p0 ).
Then
Wg−1 + κ = Θ, (9.59)
where Θ ⊂ Jac(C) is the Theta divisor and κ is the Jacobi inversion constant.
250
The proof of this statement requires a couple of constructions which are

of minor importance for the rest of this text. We will hence move some
of the details to the exercises. Our first aim will be to contruct special
fundamental domains of the quotient map V → A for a principally polarized
abelian variety A based on the theta divisor.
Exercise 9.4.2
Let A be a principally polarized abelian variety. Let Θ
e ⊂ V be the lift of the
Theta divisor to V . Let ∆ be a connected component of V \ Θ
◦ e.
(a) Show that the restriction of the quotient map to
∆◦ → Jac(C) (9.60)
is injective and has dense image A \ Θ.
(b) Show that for generic vectors w ∈ V , the set
∆w := {x ∈ V : x − εw ∈ ∆◦ for sufficiently small ε} (9.61)
is a fundamental domain for V → A consisting of unions of relative

interiors of the faces of the closure of ∆◦ . Moreover, this union is
closed with respect to inclusions, i.e. for two faces τ, σ we have
τ ⊂ σ, τ ⊂ ∆w ⇒ σ ⊂ ∆w . (9.62)
Next, it will be helpful to slightly extend our notion of tree cut. Let Γ be
a compact metric graph (where we allow finite closed ends in this section).
A tree map of Γ is a tuple (Γ 0 , ϕ) where
• Γ 0 is a metric tree (with possibly open and closed finite ends),
• ϕ : Γ 0 → Γ is a bijective isometry.
This generalizes the notion of tree cut in two ways. First, we now also allow
cutting at higher-valent vertices. Second, we cut infinitesimally near the cut
point, such that e.g. for a two-valent cut point the new ends are open and
closed, one each.
Let c : Ce → C be the universal cover of C. We can identify Ce with the
set of all homotopy classes of paths with fixed endpoints, starting in p0 .
251
Integration along the path then provides us with a map µ e : Ce → Ω(C)∗ ,

a canonical lift of µ. For w ∈ Ω(C) , let Tw be a connected component of
∗
µ
e−1 (∆w ). Let
cw := c| Tw : Tw → C (9.63)
denote the restriction of the covering map. (The choices of the open cell ∆◦
and the connected component Tw are not essential in the following, since
different choices are isomorphic up to deck transformations of Ce → C). The
following diagram summarizes the setup.
cw
c
Tw Ce C
µ
e µ (9.64)
∆w Ω(C)∗ Jac(C)
Lemma 9.4.3
For generic w ∈ Ω(C)∗ , the map cw : Tw → C is a tree map.
Proof. By definition, Tw is a tree and cw is an isometry. Since ∆w is a fun-

damental domain for generic w, surjectivity follows as well. It remains to
check injectivity. Indeed, let p 6= p0 ∈ Tw with cw (p) = cw (p0 ). The path
connecting p to p0 in Tw projects down to a closed loop γ ∈ π1 (C). Since
γ is non-zero, it contains a segment which is a simple loop, and hence by
relabeling p and p0 accordingly, we can assume that γ is a simple loop. In
particular, we can assume that its homology class [γ] ∈ H1 (C, Z) is non-
zero. This implies µe(p0 ) = µ
e(p) + [γ] 6= µ
e(p0 ). But again, we may assume
that ∆w is a fundamental domain and therefore µ e(p) = µ e(p0 ), which is a
contradiction.
Remark 9.4.4
Recall that for any y ∈ Jac(C) (or, by abuse of notation, y ∈ Ω(C)∗ ) we
denote the shifted theta function resp. divisor by ϑ y resp. Θ y (and Θ e y ).
Note that the statement and proof of the previous lemma (and exercises)
continue to hold without any changes for ∆◦ being an open cell of Θ e y (for
any y). Hence, for any y ∈ Jac(C) and generic w ∈ Ω(C) we obtain a
∗
252
associated tree map denoted by
cw ( y) : Tw ( y) → C. (9.65)
We now want to use the construction of tree maps in order to compute

the Jacobi inverse D( y) more explicitly. Given a tree map ϕ : Γ 0 → Γ and a
point q ∈ Γ , we can define the cut divisor by
X
Eϕ := (valΓ (ϕ(p)) − valΓ 0 (p)) ϕ(p) Div(Γ ). (9.66)
p∈Γ 0
Note that valΓ (ϕ(p)) − valΓ 0 (p) can be interpreted as the number of open
ends e in Γ 0 for which ϕ(p) provides the “missing” endpoint of ϕ(e). For the
special tree maps cw ( y) : Tw ( y) → C, we use the notation Ew ( y) := Ecw ( y) .
Proposition 9.4.5
For any y ∈ Jac(C) and generic w ∈ Ω(C)∗ , we have
D( y) = Ew ( y). (9.67)
The proof uses reduction to the case of transverse intersection of W1 and

Θ( y). More precisely, we say that µ : C → Jac(C) is transverse to Θ y if
µ−1 (Θ y ) consists of finitely many 2-valent vertices of C.
Exercise 9.4.6
Show that for generic y, µ and Θ y intersect transversely.
Proof of Proposition 9.4.5. First assume that µ and Θ y are transverse. In
this case, cw ( y) in fact corresponds to the ordinary tree cut along the points
in µ−1 (Θ y ) (each point p ∈ µ−1 (Θ y ) corresponds bijectively to a closed finite
end of Tw ( y)) and hence
X
Ew ( y) = p. (9.68)
p∈µ−1 (Θ y )
(The choice of w only affects which of two associated ends is closed). Let
us compute D( y). We first show supp(D( y)) = µ−1 (Θ y ). The inclusion
supp(D( y)) ⊂ µ−1 (Θ y ) is clear by definition. For the inverse, if p ∈ µ−1 (Θ y )
is not in supp(D( y)), this implies that ϑ y ◦ µ is affine linear in a neigh-
bourhood of p and hence the whole neighbourhood lies in µ−1 (Θ y ). This
253
is a contradiction to transversality, hence the equality of set follows. But

µ−1 (Θ y ) consists of g points by the previous argument, and by Proposition
9.3.11 we know that D( y) is effective and of degree g, hence each point
must occur with multiplicity one and the claim follows.
Let us now prove the general case. We may choose generic w such that
µ and Θ y−εw intersect transversely for all sufficiently small ε > 0. By conti-
nuity of the divisor and pull-back constructions, it holds
lim D( y − εw) = D( y). (9.69)

ε→0
It remains to prove continuity for Ew ( y). In terms tree maps, one can check
that Tw ( y −εw) is obtained from Tw ( y) by shortening all open ends a certain
bit, while growing a corresponding number of small closed segments at the
“opposite” point p ∈ Tw ( y). In fact, we argued before that this number is
just valC (ϕ(p)) − val Tw ( y) (p). Since all points in Ew ( y − εw) appear with
multiplicity one, this shows
lim Ew ( y − εw) = Ew ( y). (9.70)

ε→0
This proves the claim.
Exercise 9.4.7
Give the details of the proof of Equation (9.69).
The last important ingredient we need in order to prove Riemann’s theo-

rem is a link to the notion of moderator that appeared in Section 8.4. Given
a tree map ϕ : Γ 0 → Γ and a point q ∈ Γ , we orient all edges in Γ 0 towards
the unique lift of q and using ϕ get an induced orientation Oϕ on γ (for a
sufficiently fine graph strucuture). Repeating Definition 8.4.12, we define
a divisor X
K+ := (val+ (x) − 1)x (9.71)
x∈Γ
where val+ (x) denotes the number of outgoing edges at x. (In 8.4.12, we
only considered the case when O is acyclic).
Exercise 9.4.8
Show that Eϕ = K+ + q.
254
Lemma 9.4.9
For any y ∈ Jac(C), generic w ∈ Ω(C)∗ , and q ∈ Jac(C) such that µ(q) ∈
/ Θy,
the orientation induced by cw (λ) and q is acyclic.
Proof. Let e q ∈ Tw ( y) be the unique lift of q and let q0 := m Ýu(e q) ∈ ∆w be

its image under the lifted Abel-Jacobi map. Since µ(q) ∈ / Θ y , the point q0
sits in fact in the interior, i.e. q ∈ ∆ . Note that the orientation on Tw ( y) is
0 ◦
obviously acyclic, since all edges are oriented towards e q. Hence an oriented
cycle in C must lift to an oriented path γ1 inTw ( y), with “open” starting point
p1 and “closed” endpoint p2 . Let γ2 the unique (oriented) path from p2 to
q. Let us now change w to a generic vector u close to −w. It follows that the
e
new fundamental domain ∆u contains p1 , but not p2 . We choose as Tu ( y)
−1
the unique connected component of m Ýu (∆u ) which contains e q. Let γ3 be
the unique path in Tu ( y) connecting e q to the unique point in the same fiber
as p1 , say p3 . We denote by γ0i projected paths in C. Since µ e(p1 ) = µe(p3 ),
the homology class of γ3 ◦ γ2 ◦ γ1 is zero, and hence
0 0 0
[γ03 ◦ γ02 ] = −[γ01 ] ∈ H1 (C, Z). (9.72)
We now check on the chain level that this is impossible. Let e be an edge
used by γ01 such µe maps its interior to ∆w ∩ ∆u , but its endpoint to ∆w \ ∆u .
Since γ01 is an oriented and hence simple path, this edge occurs with non-
zero coefficient in the chain representation of [γ1 ]. On the other hand, this
edge can neither be used by γ02 (since γ02 ◦ γ01 is still oriented and hence sim-
ple) not γ0 3 (since e corresponds to an open end in Tv ( y)). Hence the chain
representations of [γ03 ◦γ02 ] and [γ01 ] are different and the claim follows.
We will now formulate the main statement of the section, of which Rie-
mann’s theorem is an easy corollary. Here, the support of a linear system is
defined to be [
supp |D| := supp D0 . (9.73)
D0 ∈|D|
Theorem 9.4.10
Let D be a divisor of degree g. Set y = µ(D) ∈ Jac(C). Then
supp |D| = µ|−1

C
(Θ y+κ ). (9.74)
255
Let us start the proof with the following lemma. Let C be a smooth tropi-
cal curve. Let Γ ⊂ C by a proper subgraph. By abuse of notation, we denote
by ∂ Γ the divisor containing each boundary point of Γ with multiplicity 1.
Lemma 9.4.11
Let D = a p p be a divisor on C. Let ϕ : Γ 0 → Γ be a tree map such that
P
D − ∂ Γ − Eϕ is effective. Then
Γ ⊂ supp |D|. (9.75)
Proof. We proceed by induction. We must distinguish two cases. First, let

us assume there exists a vertex p in ∂ Γ such
valΓ (p) > a p ≥ Eϕ + 1. (9.76)
This implies that the lift p0 of p in Γ is connected to at least two edges. Hence
p0 is not an end in Γ 0 and Γ 0 \ {p0 } is disconnected. Restricting ϕ to the clo-
sures of the connected components (and Γ to the corresponding image),
the assumptions of the lemma still hold. Hence we can apply the induc-
tion hypothesis and the claim follows. The case occurs when all boundary
points p satisfy a p ≥ valΓ (p). In this case, we can deform D to an rationally
equivalent divisor by moving one point along all the edges adjacent to all
boundary points. The corresponding rational function is of the form
f (x) = max{− dist(C \ Γ , x), b} (9.77)
for some (not too) negative constant b. Let Γ b be the subgraph of Γ where
f (x) = b. If we choose −b equal to the minimal distance of an interior
vertex of Γ to the boundary, Γ b is a “smaller” graph then Γ and D + div( f )
still satisfies the assumption of the lemma (just use the shrinked tree map).
Again, by induction the claim follows.
Proof of Theorem 9.4.10. We set y := µ(D). Let us start with the inclusion
supp |D| ⊂ µ|−1
C
(Θ y+κ ). Choose q ∈
/ µ|−1
C
(Θ y+κ ). We have to show q 6 supp |D|
or equivalently |D − q| = ;. By the Jacobi inversion theorem, we have
D ∼ D( y + κ). Let K+ the divisor associated to a tree map cw ( y + κ) and
the point q ∈ C. By Proposition 9.4.5 and Exercise 9.4.8 we conclude
|D − q| = |D( y + κ) − q| = |Ew ( y + κ) − q| = |K+ |. (9.78)
256
Since µ(q) ∈ / Θ y+κ , we can apply Lemma 9.4.9 which says that the corre-
sponding orientation on C is acyclic and hence K+ is a moderator in the
sense of Definition 8.4.12. Then Lemma 8.4.14 implies |K+ | = ;, and we
are done.
Let us now prove the second inclusion supp |D| ⊃ µ|−1 C
(Θ y+κ ). Let q ∈
µ|C (Θ y+κ ). Choose a tree map c := cw ( y + κ). Since µ(q) ∈ Θ y+κ , the
−1
corresponding image point in ∆w sits in the interior of some boundary face

of ∆w , say σ. Let Γ 0 be the connected component of c −1 (σ) containing q.
Let Γ = c(Γ 0 ) ⊂ C. Obviously, c|Γ 0 : Γ 0 → Γ is again a tree map. If Γ is not
yet closed in C, we can add boundary points to Γ and Γ 0 (just compactify
one of the open ends mapping to the open end on Γ ), ending up with a
compactified tree map ϕ : Γ 0 → Γ . By Lemma 9.4.11 we can finish the
proof by showing that Ec − ∂ Γ − Eϕ is effective. Since ϕ is (essentially) just
a restriction of c to a smaller domain, Ec − Eϕ is obviously effective. For a
boundary point p ∈ ∂ Γ , let x be the image point in σ. By the balancing
condition, there must be edges of µ e(C)e adjacent to x but outside of ∆w .
Such edges provide the necessary extra contribution of p in Ec .
Proof of Riemann’s Theorem 9.4.1. Let D be a divisor of degree g − 1. Set

y = µ(D) = µ(D+ p0 ). The sequence of equivalences (with Theorem 9.4.10
at third place)
y ∈ Wg−1 ⇔ |D| 6= ; ⇔ p0 ∈ supp |D + p0 | ⇔ 0 ∈ Θ y+κ ⇔ y + κ ∈ Θ

(9.79)
proves the claim. In the last step, we use Θ y = Θ− y since ϑ( y) = ϑ(− y).
257
Bibliography
[AR10] Lars Allermann and Johannes Rau. First steps in tropical intersec-
tion theory. Math. Z. 264.3 (2010), pp. 633–670. arXiv: 0709.
3705.
[Ber76] David N. Bernstein. The number of roots of a system of equations.
Functional Analysis and its Applications 9 (1976), pp. 183–185.
[BM07] Erwan Brugallé and Grigory Mikhalkin. Enumeration of curves
via floor diagrams. Comptes Rendus. Mathématique. Académie
des Sciences, Paris 345.6 (2007), pp. 329–334.
[BM09] Erwan Brugallé and Grigory Mikhalkin. Floor decompositions of
tropical curves: the planar case. Proceedings of the 15th Gökova
geometry-topology conference, Gökova, Turkey, May 26–31, 2008.
Cambridge, MA: International Press, 2009, pp. 64–90. arXiv:
0812.3354.
[BS11] Erwan Brugallé and Kristin M. Shaw. Obstructions to approxi-
mating tropical curves in surfaces via intersection theory. Preprint
(2011). arXiv: 1110.0533.
[BG09] Winfried Bruns and Joseph Gubeladze. Polytopes, rings, and K-
theory. New York, NY: Springer, 2009, pp. xiii + 461. DOI: 10.
1007/b105283.
[Cho92] Christian Choffrut. Rational relations and rational series. Theor.
Comput. Sci. 98.1 (1992), pp. 5–13.
[CLS11] David A. Cox, John B. Little, and Henry K. Schenck. Toric va-
rieties. Vol. 124. Graduate Studies in Mathematics. American
Mathematical Society, Providence, RI, 2011, pp. xxiv+841.
[Dan78] V. I. Danilov. Geometry of toric varieties. Russ. Math. Surv. 33.2
(1978), pp. 97–154.
258
Bibliography
[DM69] Pierre Deligne and D. Mumford. The irreducibility of the space of

curves of a given genus. Publications Mathématiques 36 (1969),
pp. 75–109.
[EKL06] Manfred Einsiedler, Mikhail Kapranov, and Douglas Lind. Non-
archimedean amoebas and tropical varieties. J. Reine Angew. Math.
601 (2006), pp. 139–157. arXiv: math/0408311.
[Ewa96] Günter Ewald. Combinatorial convexity and algebraic geometry.
Vol. 168. New York, NY: Springer, 1996, pp. xiv + 372.
[EI06] Günter Ewald and Masa-Nori Ishida. Completion of real fans and
Zariski-Riemann spaces. Tohoku Math. J. (2) 58.2 (2006), pp. 189–
218.
[Ful93] William Fulton. Introduction to toric varieties. The 1989 William
H. Roever lectures in geometry. Princeton, NJ: Princeton Univer-
sity Press, 1993.
[Ful98] William Fulton. Intersection theory. 2nd ed. Berlin: Springer, 1998.
[FS97] William Fulton and Bernd Sturmfels. Intersection theory on toric
varieties. Topology 36.2 (1997), pp. 335–353. arXiv: alg-geom/
9403002.
[GKZ08] I. M. Gel’fand, M. M. Kapranov, and A. V. Zelevinsky. Discrimi-
nants, resultants, and multidimensional determinants. Reprint of
the 1994 edition. Boston, MA: Birkhäuser, 2008, pp. x + 523.
[IM12] I. Itenberg and G. Mikhalkin. Geometry in the tropical limit. Math-
ematische Semesterberichte 59.1 (2012), pp. 57–73.
[Ite93] Ilia Itenberg. Contre-exemples à la conjecture de Ragsdale. French.
Comptes Rendus de l’Académie des Sciences. Série I 317.3 (1993),
pp. 277–282.
[IV96] Ilia Itenberg and Oleg Viro. Patchworking algebraic curves dis-
proves the Ragsdale conjecture. The Mathematical Intelligencer
18.4 (1996), pp. 19–28.
[Kat09] Eric Katz. A tropical toolkit. Expo. Math. 27.1 (2009), pp. 1–36.
arXiv: math/0610878.
[Kat12] Eric Katz. Tropical intersection theory from toric varieties. Collect.
Math. 63.1 (2012), pp. 29–44. arXiv: 0907.2488.
259
Bibliography
[KT01] Allen Knutson and Terence Tao. Honeycombs and sums of Her-
mitian matrices. Notices of the American Mathematical Society
48.2 (2001), pp. 175–186.
[Kou76] Anatoli G. Kouchnirenko. Polyèdres de Newton et nombres de Mil-
nor. French. Invent. Math. 32 (1976), pp. 1–31.
[Mik01] Grigory Mikhalkin. Amoebas of algebraic varieties. Preprint (2001).
[Mik04] Grigory Mikhalkin. Decomposition into pairs-of-pants for complex
algebraic hypersurfaces. Topology 43.5 (2004), pp. 1035–1065.
[Mik05] Grigory Mikhalkin. Enumerative tropical algebraic geometry in
R2 . J. Am. Math. Soc. 18.2 (2005), pp. 313–377. arXiv: math/
0312530.
[Mik06] Grigory Mikhalkin. Tropical geometry and its applications. Pro-
ceedings of the international congress of mathematicians (ICM),
Madrid, Spain, August 22–30, 2006. Volume II: Invited lectures.
Zürich: European Mathematical Society (EMS), 2006, pp. 827–
852. arXiv: math/0601041.
[Pas08]
P
Mikael Passare. How to compute 1/n2 by solving triangles. Amer-
ican Mathematical Monthly 115.8 (2008), pp. 745–752.
[Pin98] Jean-Eric Pin. Tropical semirings. Idempotency. Based on a work-
shop, Bristol, UK, October 3–7, 1994. Cambridge: Cambridge
University Press, 1998, pp. 50–69.
[Rag06] V. Ragsdale. On the arrangement of the real branches of plane
algebraic curves. American Journal of Mathematics 28 (1906),
pp. 377–404.
[Rau09] Johannes Rau. Tropical intersection theory and gravitational de-
scendants. PhD-Thesis. Technische Universität Kaiserslautern, 2009.
URL : http://kluedo.ub.uni-kl.de/volltexte/2009/
2370/.
[Rau16] Johannes Rau. Intersections on tropical moduli spaces. Rocky Mt.
J. Math. 46.2 (2016), pp. 581–662. arXiv: 0812.3678.
260
Bibliography
[RST05] Jürgen Richter-Gebert, Bernd Sturmfels, and Thorsten Theobald.

First steps in tropical geometry. Idempotent mathematics and math-
ematical physics, Proceedings of the international workshop, Vi-
enna, Austria, February 3–10, 2003. Providence, RI: American
Mathematical Society, 2005, pp. 289–317.
[Sim88] Imre Simon. Recognizable sets with multiplicities in the tropical
semiring. Mathematical Foundations of Computer Science. Springer
Berlin Heidelberg, 1988, pp. 107–120.
[SS04] David Speyer and Bernd Sturmfels. The tropical Grassmannian.
Adv. Geom. 4.3 (2004), pp. 389–411. arXiv: math/0304218.
[Stu02] Bernd Sturmfels. Solving systems of polynomial equations. CBMS
Regional Conferences Series 97. Providence, RI: American Math-
ematical Society, 2002.
[Vig09] Magnus Dehli Vigeland. The group law on a tropical elliptic curve.
Mathematica Scandinavica 104.2 (2009), pp. 188–204.
[Vir80] O. Ja. Viro. Curves of degree 7, curves of degree 8, and the Ragsdale
conjecture. Soviet Mathematics. Doklady 22 (1980), pp. 566–
570.
[Vir79] O. Ya. Viro. Construction of M-surfaces. Russian. Funktsional’nyui
Analiz i Ego Prilozheniya 13.3 (1979), pp. 71–72.
[Vir06] Oleg Viro. Patchworking real algebraic varieties. Preprint (2006).
[Vir10] Oleg Viro. Hyperfields for Tropical Geometry I: Hyperfields and
dequantization. Preprint (2010). arXiv: 1006.3034.
[Zie95] Günter M. Ziegler. Lectures on polytopes. Vol. 152. Berlin: Springer-
Verlag, 1995, pp. ix + 370.
261

Tropical Geometry

Uploaded by

Tropical Geometry

Uploaded by

Tropical Geometry

Grigory Mikhalkin, Johannes Rau

November 16, 2018

3 Projective space and other tropical toric varieties 76

4 Tropical cycles in Rn 100

4.6 The projection formula . . . . . . . . . . . . . . . . . . . . . . . . 131

5 Projective tropical geometry 145

6 Tropical cycles and the Chow group 172

7 Tropical manifolds 179

8 Tropical curves 210

9 Tropical jacobians 229

We benefited from numerous valuable conversations, in particular with Vic-

Figure 1.1: A complex elliptic curve.

Figure 1.2: Two real elliptic curves.

RP2 can be obtained from a disc D2 by identifying the antipodal points on

Figure 1.3: Real projective plane.

Figure 1.4: Elliptic curves in the real projective plane.

Figure 1.5: An elliptic curve in another presentation of RP2 by a disk.

Furthermore, the inclusion R ⊂ C gives us an inclusion RP2 ⊂ CP2 as well

oval non-trivial component

Figure 1.6: Complex elliptic curve with its real locus.

commutativity in the case of H or even associativity in the case of O.

1.2 Tropical Numbers

T = [−∞, +∞) = R ∪ {−∞}

enhanced with the arithmetic operations

“x + y” = max{x, y}, (1.1)

where we set “(−∞)+ x” = “x +(−∞)” = x and “(−∞)x” = “x(−∞)” =

The set T is a semigroup with respect to tropical addition. It is commu-

Axiom 1.2.4 (Archimedes)

Clearly, the conventional linear order > on T = R ∪ {−∞} makes per-

of x = (x 1 , . . . , x n ) ∈ Tn is a tropical polynomial P in n variables. Here, a j ∈

An infinite tropical Laurent series does not have to be well-defined even

1.3 Tropical monomials and integer affine

as Tn minus all points x = (x 1 , . . . , x n ) with x l = −∞ whenever jl < 0.

Now we do everything tropically borrowing the notions from the main

Figure 1.7: Tropical projective plane TP2 .

A curve in TP2 might be represented as a picture in R2 which has to be

Figure 1.8: A tropical cubic curve in R2 (before compactification).

Figure 1.9: A tropical cubic curve embedded in TP2 .

attached after the compactification in TP2 . In the next section we look at

Figure 1.10: Collapse of a complex cubic curve to a tropical cubic curve.

1.4 Forgetting the phase leads to tropical

The exponent α is called the argument, or the phase of z. It gets ignored

x = log t a = log t |z|,

where the base t of the logarithm is a large real number.

We should stress here that Log−1 t

i.e. x ¾ t y is an interval in T instead of a specific number.

“x + t y” = max{x ¾ t y} = log t (t x + t y ). (1.4)

This expression has a well-defined limit when t → +∞. We get

“x + y” = lim “x + t y” = lim log t (t x + t y ) = max{x, y},

thus recovering tropical addition as a certain limit of addition of complex

“x + y” = lim min{x ¾ t y} = lim log t |t x − t y | = max{x, y}.

The operations ¾ t and + are examples of hypergroup additions, see [Vir10]

Multiplication in C induces a well-defined single-valued operation

1.5 Amoebas of affine algebraic varieties

Figure 1.11: The amoeba A t and three rays inside.

show that the coamoeba consists of two congruent triangles in (R/2πZ)2

lim Log t (Vt ) = { y = “Ax” + C},

Figure 1.13: Amoeba of a parabola.

that if V ⊂ Cn is irreducible and is not contained in a coordinate hyperplane

A¯t = A t ∪ {(−∞, 0)} ∪ {(0, −∞)}

at the far left and far lower apex of A t at Figure 1.11.

F t = A0 ¾ t “A1 x” ¾ t · · · ¾ t “Ad x d ” (1.7)

Figure 1.14: A tropical parabola as the limit of amoebas when t → +∞.

if a j (t) = α j t A j . The case of d = 2, i.e. when G t is a parabola is depicted on

1.6 Patchworking and tropical geometry

to three out of four quadrants of R2 . Which arc corresponds to which

(−, +) R2>0 (−, +) R2

Figure 1.16: The logarithmic limit for each quadrant separately.

(−, −), (+, −)

As indicated, we denote the resulting set in RP2 by RΓ . Note that RΓ is

Figure 1.19: A real tropical line.

Figure 1.20: A tropical curve of degree 10.

b1 (RX ) ≤ h1,1 (X ), (1.8)

Figure 1.21: The quasitropical limit of the counterexample to the Ragsdale

where X is a smooth complex algebraic surface defined over R and RX is its