tropical varieties, maps and gossip - pure - aanmelden · tropical varieties, maps and gossip /...
TRANSCRIPT
Tropical varieties, maps and gossip
Citation for published version (APA):Frenk, B. J. (2013). Tropical varieties, maps and gossip. Eindhoven: Technische Universiteit Eindhoven.https://doi.org/10.6100/IR750815
DOI:10.6100/IR750815
Document status and date:Published: 01/01/2013
Document Version:Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers)
Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can beimportant differences between the submitted version and the official published version of record. Peopleinterested in the research are advised to contact the author for the final version of the publication, or visit theDOI to the publisher's website.• The final author version and the galley proof are versions of the publication after peer review.• The final published version features the final layout of the paper including the volume, issue and pagenumbers.Link to publication
General rightsCopyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright ownersand it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.
• Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain • You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, pleasefollow below link for the End User Agreement:www.tue.nl/taverne
Take down policyIf you believe that this document breaches copyright please contact us at:[email protected] details and we will investigate your claim.
Download date: 23. May. 2020
TROPICAL VARIETIES, MAPS AND GOSSIP
Bart Frenk
This work is licensed under a Creative Commons Attribution
3.0 Unported License, which can be found via:
http://creativecommons.org/licenses/by/3.0/
Tropical varieties, maps and gossip / B.J. Frenk
Technische Universiteit Eindhoven 2013
A catalogue record is available from the Eindhoven University of Technology Library.
ISBN: 978-90-386-3343-5
NUR: 918
Cover design: Madelief Brandsma
Printed by: Koninklijke Wöhrmann, Zutphen
Printed on: 90g chlorine-free Biotop paper
Tropical varieties, maps and gossip
PROEFSCHRIFT
ter verkrijging van de graad van doctor aan deTechnische Universiteit Eindhoven, op gezag van de rector magnificus, prof.dr.ir. C.J. van Duijn, voor een
commissie aangewezen door het College voor Promoties in het openbaar te verdedigen
op woensdag 13 maart 2013 om 16.00 uur
door
Bartholomeus Johannes Frenk
geboren te Schiedam
Dit proefschrift is goedgekeurd door de promotor:
prof.dr. A.M. Cohen
Copromotor:dr.ir. J. Draisma
Preface
Tropical mathematics, or tropical geometry are loosely defined terms referring to a
subfield of mathematics in which a central role is played by an algebraic structure
called the tropical real numbers. This is the set RYt8u equipped with the operations
of addition and taking the minimum. This structure first found applications in the
algebraic formulation of optimization problems (cf. [16]). In tropical mathematics,
however, the emphasis is on geometry over the tropical semifield. For this thesis it is
convenient to make a very rough division of tropical mathematics into two (overlap-
ping) parts.
One part studies images of embedded algebraic varieties under a valuation. Such
images, called tropicalizations or tropical varieties, are (underlying sets of) polyhedral
complexes that have a characterization in terms of the algebra of the tropical real
numbers and the ideal of the variety. The founding idea is that the properties of
tropicalizations reflect properties of the algebraic variety in question. One can then
use combinatorial techniques to derive algebraic geometric results, and vice versa.
A celebrated example is G. Mikhalkin’s computation of Gromov-Witten invariants by
counting tropical curves in [45].The other part studies the algebra and geometry of the tropical real numbers with-
out reference to a valued field. For example, it includes functional analysis over the
tropical real numbers (cf. [43]), and studies abstractions of tropical varieties. A rea-
son to study such abstractions is that they are applicable to the aforementioned part
of tropical geometry, while the techniques involved do not depend on the characteris-
tic of the field. As argued in Chapter 4 the concept of a matroid falls naturally within
this part of tropical geometry.
Chapters 1 and 2 fit within the first part. The first chapter deals with reformulating
and proving the main properties of the tropicalization of an algebraic variety. None of
these results are new, but as far as we know have not been collected in a published
source. The second chapter explores the relation between unirationality of algebraic
varieties and tropical unirationality of its tropicalization.
Chapter 3, 4 and 5 fit within the second part. Chapter 4 is the largest chapter of
this thesis. Its aim is to construct a category whose objects are tropical linear spaces.
i
ii
These spaces are abstractions of tropicalizations of vector spaces and are closely re-
lated to the valuated matroids of [23]. The reason for constructing such a category
is that in general maps between tropical varieties are problematic. For example, in
the context of tropical linear spaces the naive notion of a linear map does not map
subspaces to subspaces. Chapter 4 attempts to remedy that situation.
Chapter 3 puts the tropical semifield into the broader context of idempotent semi-
fields. Its main purpose is to understand which algebraic properties of idempotent
semifields are essential for Chapter 4. The last chapter, Chapter 5 is about the monoid
generated by distance matrices under tropical matrix multiplication. This monoid is
related to the gossip problem and referred to as the gossip monoid. We make use of
techniques related to realizing a finite metric by a graph to compute polyhedral fan
structures on the gossip monoids of square matrices of size at most 4.
Acknowledgements
To me, sometimes an appropriate metaphor for being a Ph.D. candidate in mathe-
matics is tropical. It is that of deep, dark jungle. An environment without clear paths,
in which it is often difficult to distinguish the friendly from the fearful, at least at
first sight. A place easy to get lost in. That I came out in one piece is greatly due
to the efforts of my guide and copromotor, dr. ir. Jan Draisma. Of the many things
I learned from him the most important was probably to face unknown mathematics
with courage, no matter how fearsome it might appear. He also wrote the original
Tropical Algebraic Groups NWO project proposal that was the starting point of this
thesis.
During exploration one meets the people along the way that helped the project
progress in some way or another. They appear here in no particular order.
I would like to thank Prof. Dr. Eva Maria Feichtner for the invitation to do part of
my thesis work at the University of Bremen, where I spent three months. Life there
was made easier, both professionally and personally, by the people of the DiscreteStructures in Algebraic Geometry group and the capoeristas of Ginga Brasil Bremen. I
extend my thanks.
My promotor, prof. dr. Arjeh Cohen, was kind enough to accept me as his official
Ph.D. student. The three aforementioned people were also part of my reading com-
mittee, along with Prof. Dr. Thorsten Theobald and Prof. dr. Andries Brouwer, all of
which I would like to thank for their helpful corrections and comments on the first
draft of my thesis. It goes without saying that any errors and omissions left in the
thesis are entirely mine. The reading committee members are also part of my defence
committee, along with Prof. Dr. Michael Joswig en Prof. dr. Peter Butkovic. I would
like to thank all of them for their time and effort.
iii
My colleagues and fellow Ph.D. candidates at the Eindhoven University of Technol-
ogy. In particular, Jan-Willem Knopper, for his patience in dealing with the computer-
related problems I sometimes had, and Maxim Hendriks, that guided me gently
through the myriad of choices one has to make when deciding how to print one’s
thesis.
Of course, not all of my time in the last four years was spent exploring this
metaphorical jungle.
The capoeira-lessons of Mestre Tayson in Eindhoven were always a very good way
to take my mind of the mathematics. I would like to thank him and his students, my
friends, in no particular order: Danni, Damian, Tsveti, Mike, Roelof, Dirk, Imre, Qorin,
Matilde.
I would like to thank my parents for their continuous support. I consider myself
lucky that my twin sister Myrthe worked on a Ph.D. project at the University of Maas-
tricht while I worked on mine in Eindhoven. This was particularly helpful during the
periods of frustration that are bound to come up when doing research. Talking to her
about such matters was always helpful for clearing my mind. Finally, great thanks
goes to my girlfriend, Hilde, who was instrumental in the entire process for her un-
wavering support.
iv
Contents
Preface i
List of notations viii
1 Tropical geometry 1
1.1 Quick algebraic definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Tropical geometry of hypersurfaces . . . . . . . . . . . . . . . . . . . . . . 9
1.4 Tropical geometry in higher codimension . . . . . . . . . . . . . . . . . . 11
1.5 Tropical geometry and valuation theory . . . . . . . . . . . . . . . . . . . 18
2 Tropical unirational varieties 21
2.1 Some classes of tropically unirational varieties . . . . . . . . . . . . . . . 23
2.2 Combining reparameterizations . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Birational projections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Very local reparameterizations . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3 Algebra of idempotent semirings 37
3.1 Modules over semirings and homomorphisms . . . . . . . . . . . . . . . . 37
3.2 Linear functionals on idempotent modules . . . . . . . . . . . . . . . . . . 48
3.3 Ranks and vanishing conditions . . . . . . . . . . . . . . . . . . . . . . . . 52
3.4 Modules over linearly ordered semirings . . . . . . . . . . . . . . . . . . . 58
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4 A category of tropical linear spaces 61
4.1 Valuated matroids and tropical linear spaces . . . . . . . . . . . . . . . . 63
4.1.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.1.2 Basic constructions and relations . . . . . . . . . . . . . . . . . . . 71
4.1.3 Parameter spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.2 Functions and their graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
v
vi CONTENTS
4.2.1 Linear functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2.2 Extensions of matroids . . . . . . . . . . . . . . . . . . . . . . . . . 92
4.3 Morphisms between linear spaces . . . . . . . . . . . . . . . . . . . . . . . 113
4.3.1 A category of tropical linear spaces . . . . . . . . . . . . . . . . . . 113
4.3.2 The tropical linear monoid . . . . . . . . . . . . . . . . . . . . . . . 121
5 The gossip monoid 127
5.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.2 Relation to the gossip problem . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.3 Polyhedral structure of the gossip monoid . . . . . . . . . . . . . . . . . . 134
5.3.1 Graphs with detours . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.4 Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.4.1 n¤ 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.4.2 n� 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
Bibliography 145
Index 150
Summary 153
Curriculum Vitae 155
List of notations
pv : wq dual residuum of v by w . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
rns the numbers t1, . . . , nu
α� transpose of the map α . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
µ the independent set valuated matroid of µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
F extended semifield of F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
B the trivial semifield . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
BpX q Berkovich analytic space of X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Cttεuu the field of Puiseux series over C in ε . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
DrdpL q parameter space of d-dimensional tropical linear subspaces of L . . . . . . . . 80
ExtppL q elemtary extensions of the tropical linear space L . . . . . . . . . . . . . . . . . . . . . . 97
Extppµq elementary extensions of the valuated matroid µ . . . . . . . . . . . . . . . . . . . . . . . . 97
GrdpKnq the Grassmannian of d-dimensional linear subspaces of Kn (in its Plücker
embedding)
I pX q radical ideal of polynomials vanishing on X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
εR indicator of the semiring R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
invpgq initial form of the polynomial g with respect to the weight vector v . . . . . . . . 5
xv, wy the tropical inner product of v and w . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
λrX , Y s restriction of the linking system λ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
λΓ linking system induced by Γ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
¤� natural order on an idempotent commutative semigroup . . . . . . . . . . . . . . . . . 40
vii
viii CONTENTS
λK the cofactor linking system of λ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .106
Lµ tropical linear space of the valuated matroid µ . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
matpµq the underlying matroid of µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
µ{X contraction of X , i.e. µ � RzX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
µ � X contraction of µ to X . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
µY ν (valuated) matroid union . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
µ` ν (valuated) matroid direct sum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
µrX s restriction of X of µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .76
M op opposite module of M . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
φ^ dual residual of the increasing map φ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
respµq the minimal, or residual, matroid of µ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .65
TGLn the tropical linear monoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Troppgq tropicalization of the polynomial g . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
T pSq the tropical prevariety of the set of polymials S . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
R� group of units of the semiring R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
DnpFq distance matrices with entries in F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Ei jpaq elementary distance matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Frxs the semiring of polynomials over F in x� tx1, . . . , xnu. . . . . . . . . . . . . . . . . . . . . 2
Jλ the matrix all of whose entries are λ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
M� algebraic dual of the module (over a semiring) M . . . . . . . . . . . . . . . . . . . . . . . . 48
SnpFq the gossip monoid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
vK set of all w such that xv, wy vanishes tropically . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
vI generator of a tropical linear space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .68
Q8 the tropical rational numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2
R8 the tropical real numbers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2
Z8 the tropical integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2
Chapter 1
Tropical geometry
The tropicalization map is the main tool in tropical geometry. Roughly, it assigns a
polyhedral complex to an embedded algebraic variety. This chapter collects the foun-
dational results from the tropical geometry literature on the tropicalization map. It
focuses on three theorems: Bieri-Groves, the Fundamental Theorem of Tropical Ge-
ometry, and the Finite Tropical Basis Theorem, which we refer to as the main theo-rems of tropical geometry. They are respectively Proposition 1.4.9, Theorem 1.4.8 and
Corollary 1.4.7.
The route to proving the three main theorems of tropical geometry is through the
device of generic projections that was first applied in the paper by Bieri and Groves
[7]. The advantage are twofold: (i) it allows one to prove both the Fundamental
Theorem and the Bieri-Groves theorem by reducing to the hypersurface case, (ii) the
existence of a finite tropical bases is an immediate consequence of the main lemma
of the method. The proofs we give are not new. They are based on those in [52], [7]and [32].
1.1 Quick algebraic definitions
A semifield is an algebraic structure with a multiplication and an addition operator
that satisfies all the axioms of a field save the one that states that addition is invertible.
In that sense they are fields without a minus. In this thesis, and especially the first two
chapters, the semifields under consideration are of two types: (algebraically closed)
fields, or tropical semifields, which we define next.
The tropical real number semifield is the semifield R8 � pRY t8u,min,�,8, 0q.
In this semifield the "addition" operation is given by min, the "multiplication" is given
by �. It is common to write x ` y for mintx , yu and x d y for x � y , to stress
the roles of both operators. We will usually follow this convention. The tropical real
1
2 CHAPTER 1. TROPICAL GEOMETRY
numbers do not form a field, since for any a P R there is no x P RY t8u such that
a ` x � minta, xu � 8. Notice that 8 is the neutral element for tropical addition,
and 0 is the neutral element for tropical multiplication.
A tropical semifield is a subsemifield of R8. Their underlying sets are of the form
G Y t8u where G is a subgroup of pR,�, 0q. The tropical semifield with underlying
set G is denoted G8. For example, Q8 is the semifield of tropical rational numbers
and Z8 is the semifield of tropical integers. We reserve B for the tropical semifield
t0,8u and refer to it as the trivial semifield.
1.2 Foundations
Let F be a semifield and x � tx1, . . . , xnu a finite set of formal variables. A monomialin x is an element of the free commutative monoid generated by the elements of x. A
polynomial is an element of the free F -algebra generated by this monoid (having only
a finite number of coefficients different from 0 P F). The collection of all polynomials
in x over F forms an F -algebra itself and is denoted Frx1, . . . , xns, or simply Frxs.
It is common to think of polynomials as functions F n Ñ F . Let f P Frxs and
v P F n. The evaluation of f in v is the element of F obtained by substituting all
occurrences of x i in f by vi and replacing the formal operations by the operations
in the semifield. We denote it f pvq. The association of v ÞÑ f pvq to f defines a map
Frxs Ñ tF n Ñ Fu. When F is an infinite field this map is injective. However, it is
not for general semifields, even when F is infinite. In particular, when F is a tropical
semifield this map is not injective.
Example. The simplest example of the non-injectivity of the evaluation map is when
F � B and n� 1. The non-injectivity of the map is clear since the number of functions
BÑ B equals 4, while the number of polynomials in Brxs is not even finite. Moreover,
a2 � a for any a P B.
A more interesting example is the tropical univariate polynomial x2 `p1d xq` 2
over R8. It induces the function R8Ñ R8 given by v ÞÑmint2v, v�1, 2u in ordinary
notation. It is thus the minimum of three linear functions, the graphs of which are
depicted in Figure 1.1. It induces the same function as x2 ` pcd xq ` 2, when c ¥ 1.
For the remainder of this section F is a tropical semifield and K is a field equipped
with a surjective valuation ω : K Ñ F . We write R for the valuation ring ta P K |
ωpaq ¥ 0u and m for its maximal ideal ta P R | ωpaq ¡ 0u. The residue field R{m is
denoted k.
Example. A good example of such a field to keep in mind is that of the Puiseux seriesover C in ε, denoted Cttεuu. It is the union of all formal Laurent series fields Cppε1{nqq
over C in ε1{n, where n ranges over the positive integers. The order of such a formal
1.2. FOUNDATIONS 3
2
x2
1� x
Figure 1.1: The graph of the function induced by x2 ` p1d xq ` 2.
Laurent series is the minimal exponent occurring with a non-zero coefficient. For
example,
4ε�10 � ε1{2 � pi� 1qε6{8 � 20ε50,
is an element of Cppε1{2qq of order �10. The order of the trivial series 0 is defined to
be 8. The order maps on the different sets Cppε1{2qq of formal Laurent series coincide
on the intersections and hence extend to an order map on the union Cttεuu. This
extension, denoted ord, is a surjective valuation Cttεuu Ñ Q8. The field of Puiseux
series is algebraically closed (cf. [54]) and has residue field C.
Let f P Frxs and write,
f �àαPApcαd xαq,
where A is a finite subset of Nn and cα P Fzt8u. The tropical notation serves to
remind the reader of the nature of the polynomial. The tropical vanishing locus of fis the collection of points v P F n for which either the minimum in f pvq is attained for
two distinct α, or for which f pvq � 8.
Let g P Krxs. The tropicalization of g is the polynomial with coefficients in Fobtained from g by replacing the coefficients with their image under ω. With slight
abuse of notation it is denoted Troppgq and g ÞÑ Troppgq defines a surjective map
from polynomials over K to polynomials over F ,
Trop : Krx1, . . . , xns Ñ Frx1, . . . , xns , Tropp¸αPA
cαxαq �àαPApωpcαq d xαq. (1.1)
The tropical prevariety!of a polynomial associated to g is, by definition, the tropical
vanishing locus of Troppgq, as defined in the previous paragraph. It is denoted Tωpgq,or T pgq, when reference to ω is clear from the context. Since the map (1.1) is surjec-
tive the set of tropical vanishing loci of polynomials over F equals the set of tropical
prevarieties of polynomials over K .
4 CHAPTER 1. TROPICAL GEOMETRY
L1
L2
L3
x
y
Figure 1.2: The tropical prevariety of the polynomial f � pε2 � ε3qx � y � ε�1
Remark. When F � R8 the function induced by a non-zero tropical polynomial frestricts to a piecewise linear function Rn Ñ R. The vanishing locus of the polynomial
is the union of the topological closure in Rn8 of the corner locus of this restriction and
the collection of points on which f attains the value 8. In particular, it depends only
on the function induced by f .
In fact the case F � R8 is the only case we need to consider. The reason is that
any tropical semifield F embeds in R8, and hence any polynomial f over F is also
a polynomial over R8. It is an immediate consequence of the definition that the
vanishing locus of f in F n equals the vanishing locus of f as a polynomial over R8intersected with F n.
Example. Let F � Q8 and K the field Cttεuu with the order valuation ord : K Ñ F .
Consider the polynomial
f � pε2 � ε3qx � y � ε�1.
in the variables x and y . Its algebraic vanishing locus is a line in the affine plane over
Cttεuu. Moreover, Tropp f q is the tropical polynomial,
Tropp f q � p2d xq ` y ` p�1q,
whose vanishing locus is the union of three half-lines L1, L2 and L3, given by the
equalities and inequalities (specified in terms of the classical operations on Q8),
L1 : 2� x � y ¤ �1,
L2 : 2� x � �1 ¤ y,
L3 : y � �1 ¤ 2� x .
It is depicted in Figure 1.2.
Notation and terminology extend to arbitrary sets of polynomials. Let S �
Krx1, . . . , xns. The tropical prevariety associated to S is the intersection of the tropi-
1.2. FOUNDATIONS 5
cal prevarieties of the elements of S and denoted T pSq, or TωpSq when we need to
specify the valuation, i.e.,
TωpSq �£gPS
Tωpgq.
A tropical basis of S is a subset B � S such that the intersection of T pgq, with g P B,
equals T pSq. A tropical basis always exists (take B � S), but it is not a priori clear
whether a finite subset always suffices.
Tropical prevarieties are general objects and, at least in this chapter and the next,
we restrict ourselves to tropical prevarieties of ideals in Krx1, . . . , xns. Such preva-
rieties are referred to as tropical varieties over K , or over pK ,ωq, if we need to be
precise. A prevariety is simply said to be a tropical variety if it is a tropical variety
over some pair pK ,ωq. Given F � ωpKq, the distinction is irrelevant when only con-
sidering principal ideals, but to prove this we require a rather technical lemma and
some definitions.
We henceforth assume that there exists a good section of ω, i.e., a map s : F Ñ Kthat assigns to each v P F an element spvq P K such that ωpspvqq � v, spv d wq �spvqspwq andωpspvq�spwqq � v`w. For example, in the case of the Puiseux series in
ε over C a good section is spvq ÞÑ εv . The choice of a specific section is non-canonical
and not strictly necessary, but makes the proofs more intuitive, at the expense of
making them less general. We fix such a section for the remainder of this chapter and
use the suggestive notation t v for the image of v under this section. Note that t8 � 0.
In [52] the choice of a section is avoided by making use of so-called twisted coordinatesubrings of Krxs.
Let g P Krxs be a polynomial and v P F n. Let w � Troppgqpvq and suppose that
w � 8. The initial form of g with respect to v is the image of t�w gpt v1 x1, . . . , t vn xnq
under the canonical projection Rrxs Ñ krxs. It is denoted invpgq.
Example. Let f � pε2�ε3qx� y�ε�1, as in the previous example. Then, spvq � t v �
εv is a good section of ord, and
inp0,1qp f q � 1,
inp�3,0qp f q � x � 1.
Parts (ii) and (iii) of the next (technical) lemma state important properties of ini-
tial forms. They are the main technical devices for proving that tropical prevarieties of
polynomials are tropical varieties and that the tropicalization of an algebraic variety,
to be defined later, doesn’t depend on a defining ideal. Part (i) essentially states that
the Gauss norm restricted to Krxs is multiplicative (cf. [10]).
6 CHAPTER 1. TROPICAL GEOMETRY
Lemma 1.2.1 The following statements hold.(i) The map that assigns to each g P Krxs the function F n Ñ F defined by v ÞÑ
Troppgqpvq is multiplicative.(ii) Let v P F n. Denote the collection of polynomials g P Krxs that have Troppgqpvq �
8 by Dv . Then inv : Dv Ñ krxs is multiplicative.(iii) Let v P F n. Then v P T pgq if and only if either invpgq is not a monomial, or
Troppgqpvq � 8.
Proof. We prove the claims in the order stated.
(i) Let f , g P Krxs and v P F n. We need to prove that the equality Tropp f gqpvq �Tropp f qpvq d Troppgqpvq holds. Assume without loss of generality that v � p0, . . . , 0q
and that Tropp f qpvq � Troppgqpvq � 0. Then f , g P Rrxszmrxs. It is now sufficient to
show that f g P Rrxszmrxs. Since mrxs � mRrxs and the projection Rrxs Ñ Rrxs{mrxsis an R-algebra homomorphism it is suffices to prove that Rrxs{mrxs is an integral
domain. This is clear, since it equals krxs.(ii) Denote the map Kn Ñ Kn defined by pa1, . . . , anq ÞÑ pt v1 a1, . . . , t vn anq by σ
and write σ� for its pull back Krxs Ñ Krxs. Let f , g P Dv . Set w � Tropp f gqpvq,w f � Tropp f qpvq and wg � Troppgqpvq. Then w � w f � wg , by part (i). Denote the
canonical projection Rrxs Ñ krxs by π. Consider,
invp f gq � πpt�wσ�p f gqq
� πpt�wσ�p f qσ�pgqq
� πpt�w fσ�p f qt�wgσ�pgqq
� πpt�w fσ�p f qqπpt�wgσ�pgqq
� invp f q invpgq.
This proves the second claim.
(iii) Suppose v P T pgq and write g �°αPA cαxα. Let β P A be the index of
a term at which the minimum in Troppgqpvq is attained. The coefficient of xβ in
gpt v1 x1, . . . , t vn xnq equals cβ t v�β , where v � β is the ordinary inner product on Rn
extended to Rn8. The valuation of this coefficient equals Trop pgqpvq and hence the
coefficient of πpxqβ if invpgq is non-zero. Vice versa, if the coefficient of some πpxqβ
is non-zero in the initial form, then the corresponding term of g attains the minimum
in Trop pgqpvq. This proves the statement. l
Proposition 1.2.2 Let f P Krxs and write I for the ideal generated by f . Then Tωp f q �TωpIq. In particular, tropical prevarieties of polynomials are tropical varieties.
Proof. Since f P I it holds that TωpIq � Tωp f q. Let v P Tωp f q and g P I . There
exists h P Krxs such that g � hf . If either Tropp f qpvq � 8, or Tropphqpvq � 8,
1.2. FOUNDATIONS 7
then Troppgqpvq � Tropphqpvq d Tropp f qpvq � 8, by part (i) of Lemma 1.2.1. If
Tropp f qpvq � 8 and Tropphqpvq � 8 then invpgq � invphq invp f q. Since invp f q is
not a monomial invpgq is also not a monomial and hence g vanishes tropically on v.
This shows that v P T pIq. l
The next statement shows that the tropical variety of an ideal only depends on
the algebraic variety defined by this ideal. It is the essential ingredient in defining the
tropicalization of an algebraic variety, which do we after we prove the lemma.
Proposition 1.2.3 Let I � Krxs be an ideal and denote its radical by rpIq. ThenTωpIq � TωprpIqq.
Proof. Clearly, T prpIqq � T pIq, since rpIq � I . Let v P T pIq and f P rpIq. If
Tropp f nqpvq � 8 then Tropp f qpvq � 8, by Lemma 1.2.1, and hence v P T p f nq. Sup-
pose that Tropp f nqpvq � 8. Then invp f nq � invp f qn. Since invp f nq is not a monomial,
invp f q isn’t. Thus v P T f . This proves that T prpIqq � T pIq. l
Let X � Kn be an algebraic variety and let I � Krxs be an ideal whose algebraic
variety is X . Define the tropicalization of X as T pIq. By Proposition 1.2.3 it does not
depend on the specific choice of ideal. We denote it by Trop pX q. A tropical basis of
Trop pX q is defined to be a tropical basis of I pX q, i.e. a set of polynomials that vanish
on X whose tropical variety equals Trop pX q. Here we write I pX q for the radical ideal
of all polynomials vanishing on X . Note that we do not require that a tropical basis
generates I pX q, or even cut out the variety X .
Proposition 1.2.4 Let X � Kn be an algebraic variety such that X � Y Y Z, withY, Z � Kn algebraic varieties. Then Trop pX q � Trop pY q Y Trop pZq.
Proof. Let I � I pY q and J � I pZq. Then V pI Jq � X . Thus, Trop pX q � T pI Jq.Since I J � I , J it holds that Trop pX q � Trop pY q Y Trop pZq. Let v P Trop pX q and
suppose that v R Trop pY q. There exists f P I such that v is not in T p f q. By Lemma
1.2.1 the initial form invp f q exists and is a monomial. Let g P J and suppose that
Troppgqpvq � 8. Then invp f gq � invp f q invpgq is not a monomial, and hence invpgqis not a monomial, a contradiction. Thus, v P Trop pZq. l
Example. By definition every tropical variety is a tropical prevariety, but the con-
verse does not hold. As a simple example, let F � R8 and consider the tropical
prevariety T associated to the set of polynomials tx1 � x2 � 1, x1 � x2u. It is the
intersection of the tropical vanishing loci of x1 ` x2 ` 0 and x1 ` x2, which equals
tpv, vq | v P F and v ¤ 0u. Suppose that there exists an ideal I � Krx1, x2s such that
8 CHAPTER 1. TROPICAL GEOMETRY
TωpIq � T . Then I is not principal. Suppose it were and set pgq � I . Write,
g �d
i�0
mi xi2, mi P Krx1s and d ¡ 0.
Let w1 P F with w1 ¡ 0 and consider Troppgqpw1, x2q P Krx2s. This polynomial has
positive degree and hence there exists w2 P F such that w2 is in the vanishing locus
of Troppgqpw1, x2q. Thus, pw1, w2q P T pIq and pw1, w2q R T .
If I is not principal its associated variety is either empty or a finite set of points.
Since the tropical variety does not depend on the choice of ideal, T pIq � T pI pX qq.In the first case 1 P I pX q, by Hilbert’s Nullstellensatz, and hence T pIq � H. In the
second case, X is a finite union of points and hence Trop pI pX qq is a union of tropical
varieties of ideals of the form px1 � p1, x2 � p2q. In particular, it is a finite union of
points. This shows that T is not a tropical variety.
We finish the section with a lemma, which serves as the base from which to prove
the main theorem of tropical geometry for hypersurfaces. In fact, one could say the
lemma is the main theorem of tropical geometry for points on a line.
Lemma 1.2.5 (Newton-Puiseux) Let f P Krxs be a univariate polynomial. ThenωpV p f qq � T p f q.
Proof. Since K is algebraically closed there exists a P K and b1, . . . , bd P K such that
f � apx � b1q . . . px � bdq. Let c P V p f q. Then c � bi , for some i � 1, . . . , d. If bi � 0
then by Lemma 1.2.1 the image of ωp0q under Tropp f q is 8 and hence ωp0q P T p f q.If bi � 0, then inωpcqp f q is a product of d initial forms inωpcqpx � b jq, at least one of
which is not a monomial. Hence inωpcqp f q is not a monomial. Thus, ωpcq P T p f q.
Reversing the above proof yields that T p f q �ωpV p f qq. l
Remark. The above lemma implies in particular that tropical semifields that appear
as the range of a surjective valuation on some algebraically closed field satisfy that
any non-constant polynomial over them has a non-empty vanishing locus. Such a
semifield F is divisible, i.e. it has the property that for all a P F the equation xn � a has
a solution. Over such semifields it holds that any polynomial function is the product
of linear polynomial functions. The emphasis here is on the word function, since it
does not necessarily hold on the level of polynomials.
As an example, consider the polynomial x2 ` p2d xq ` 1. Suppose that it decom-
poses as px ` aq d px ` bq � x2 ` ppa` bq d xq ` pad bq. Then, by a` b � 2, either
a � 2 or b � 2. Assume without loss of generality a � 2 and hence b � �1. But then
a ` b � �1 � 2. However, x2 ` p1d xq ` 1 and x2 ` p2d xq ` 1 define the same
function, and px ` 1q d px ` 1q � x2 ` p1d xq ` 1.
1.3. TROPICAL GEOMETRY OF HYPERSURFACES 9
It is interesting to note, however, that there is a multiplicative section of the map
f ÞÑ pv ÞÑ f pvqq, assigning the induced function to a polynomial. This section assigns
to each polynomial function the polynomial with minimal coefficients. See [30] for
details.
We can now state the theorems we aim to prove in this chapter. The first, usu-
ally referred to as the Fundamental Theorem of Tropical Geometry, is the state-
ment that the tropicalization of an algebraic variety X � Kn equals ωpX q, where
ωpp1, . . . , pnq � pωpp1q, . . . ,ωppnqq, with slight abuse of notation. The second state-
ment, referred to as the Bieri-Groves theorem is the statement that the tropicalization
of X is the topological closure of (the underlying set of) a polyhedral complex in Rn,
when F � R8. The third statement is that any tropical variety has a finite tropical
basis.
The next section is dedicated to proving these three statement for the case where
X is a hypersurface.
1.3 Tropical geometry of hypersurfaces
This section deals with tropical varieties of algebraic hypersurfaces in Kn. The reason
why the distinction between hypersurfaces and general algebraic varieties is made is
essentially Proposition 1.2.2. Every tropical variety of a hypersurface is the vanish-
ing locus of a tropical polynomial and vice versa. In particular, given F there is no
dependence on the field, as long as it is algebraically closed.
We begin by stating and proving the main theorem of tropical geometry for the
case of hypersurfaces. This specific case is also referred to as Kapranov’s theorem and
it appears, with a different proof, in [24]. The proof here is essentially that in [52].For each n P N the valuation ω : K Ñ F induces a map Kn Ñ F n, given by
ωpp1, . . . , pnq � pωpp1q, . . . ,ωppnqq, that we also denote by ω. The image ωpX q is
referred to as (non-Archimedean) amoeba of X in [24].
Theorem 1.3.1 (Kapranov) Let X � Kn be an irreducible hypersurface. ThenTrop pX q � ωpX q. Moreover, for each v P Trop pX q the set ω�1pvq X X is Zariski densein X .
Proof. Let v P Trop pX q. We can assume without loss of generality that v � p0, . . . , 0q.
Let I � Krxs be the ideal I pX q of X . There exists f P Rrxszmrxs with I � p f q and by
Lemma 1.2.2 it holds that T p f q � T pIq � Trop pX q. The tropical polynomial Tropp f qvanishes at v and hence, either (i) Tropp f qpvq � 8, or, (ii) the projection f of f in
krxs � Rrxs{mrxs has at least two terms.
In the first case, let p be a point in Kn with ωppq � v. Then f ppq � 0, since
ωp f ppqq ¥ Tropp f qpvq � 8, and hence p P X .
10 CHAPTER 1. TROPICAL GEOMETRY
In the second case, there exists a variable x l which occurs with at least two distinct
exponents d and e in f . Suppose without loss of generality that l � n and d e. Write,
f � f0 � f1 xn � . . .� fd xdn � . . .� fe x e
n � . . .� fd xdn , fi P Krx1, . . . , xn�1s
Let p P pRzmqn�1 such that fdppq � 0 and feppq � 0. Such p exists since K , and hence
k is algebraically closed. Set gpyq � f pp, yq. The projection grys P krys has at least
two non-zero terms, and hence a non-zero root pn. Let pn P Rzm such that it projects
onto pn. Then, ωpgppnqq ¡ 0 and hence ωppnq P T pgq. By Lemma 1.2.5 there exists
pn P V pgq such that ωppnq � 0. Thus, pp, pnq P V p f q and ωpp, pnq � p0, . . . , 0q.
The projection of X Xω�1p0, . . . , 0q onto the first n � 1 coordinates contains a
Zariski dense set in Kn�1. Hence, X Xω�1p0, . . . , 0q it itself dense in X . l
It remains to establish the polyhedral structure of Trop pX q. Let f P Frxs and write,
f �àαPApcαd xαq, A� Nn and cα P Fzt8u.
The Newton polytope of f is the convex hull of A. The subdivision of the Newton
polytope induced by f is the regular subdivision of the Newton polytope associated to
the map that assigns the coefficient of xα to α P Nn. More precisely, it is the projection
of the lower convex hull of the set tpα, cαq | α P Au � Rn �R to Rn.
Theorem 1.3.2 Let X � Kn be the hypersurface defined by a polynomial g P Krxs. LetF � R8. The tropical variety Trop pX q is the underlying set of a pure polyhedral complexof codimension one that is dual to the subdivision of the Newton polytope associated toTroppgq.
Proof. Let f � Troppgq and write f �ÀαPA mα as a sum of monomial terms mα, i.e.,
mα is the formal product of a coefficient in F and a monomial. The function F n Ñ Fdefined by mα is an affine linear function on F , and hence v ÞÑ f pvq is the minimum
over a finite number of affine linear functions. A point v P F n lies on T pgq if and only
if the minimum is attained in at least two such affine linear functions. Thus, T pgq is
the union of the sets Cα,β � tv P F n | mαpvq � mβpvq ¤ mγpvq for all γ P Aztα,βuu,
some of which may be empty. The non-empty Cα,β ’s have dimension n� 1 and form
the maximal cells of a polyhedral complex. Moreover, if Cα,β is non-empty then tα,βu
is a face of the subdivision. The implied association defines the duality. l
We end this section with two examples of tropical hypersurfaces.
Example. Let K � Ctttuu be the field of Puiseux series over C and ω : K Ñ Q8 be
order at 0. Consider the polynomial g P Krx , ys given by g � t x y � t4 x2 � t4 y2 �
pt2� t3�3t6qx�pt2� t5qy�6. The tropicalization Troppgq of g is f � 1x y`4x2`
1.4. TROPICAL GEOMETRY IN HIGHER CODIMENSION 11
4y2`2x `2y `1. The subdivision of the Newton polytope induced by f is drawn in
Figure 1.3(a). Its vanishing locus is depicted in Figure 1.3(b).
4
4
21
1
x2x
xy
y2
y
1
(a) Polyhedral subdivision
(−1, 1)
(1,−1)
(1,−2)
(−2, 1)
xy
x2 x
1
y
y2
x
y
(b) The vanishing locus. The labels on the re-gions denote the monomials of the minimalterms.
Figure 1.3: Tropical vanishing locus of the tropical polynomial 1x y ` 4x2 ` 4y2 `2x ` 2y ` 1 (and the tropical variety of pgq).
Example. Consider SLnpKq, the algebraic variety of n� n-matrices over K of determi-
nant 1. This is a hypersurface of the n2-dimensional affine space MnpKq, the variety of
all n� n matrices with entries in K . Note that the tropicalization of MnpKq is MnpFq,which we represent by the set of F -weighted bipartite graphs on prns, rnsq. A term in
the tropical determinant of a matrix corresponds to a perfect matching in this bipar-
tite graph, and the value of this term equals the weight of the matching. The tropical
vanishing of det X �1 translates to the condition on the bipartite graph that the mini-
mal weight of a perfect matching equals 0, or that it is strictly smaller than 0, in which
case there must be two distinct perfect matchings attaining the minimum weight.
The aims of this section were to establish Theorem 1.3.1 and Theorem 1.3.2. The
existence of a finite tropical basis for a hypersurface is given by Proposition 1.2.2: any
generator of the principal ideal of the hypersurfaces suffices.
1.4 Tropical geometry in higher codimension
We start this section with a number of lemmata about the image of finite subsets of
affine spaces under generic projections. They lead up to a proof of Proposition 1.4.3
12 CHAPTER 1. TROPICAL GEOMETRY
that proves that a set is a polyhedral complex if and only if it is so under a suffi-
ciently generic projection. This last proposition is the main tool in reducing the main
theorems of tropical geometry for arbitrary tropical varieties into their hypersurface
specializations. The springboard for proving the main theorems themselves is Theo-
rem 1.4.6.
In this section we refer to affine linear surjective maps Rn Ñ Rm simply as projec-tions. In particular, a projection π is of the form v ÞÑ Av � w, with A an m� n matrix
over R and w P Rm. The kernel kerπ of π is defined to be the kernel of A, i.e. the
set of all v such that πpvq � πp0q � w. Projections are said to be equivalent if they
differ up to an affine automorphism of the range. Since we are mostly interested in
properties of the image of a projection that are invariant under affine automorphisms
(e.g., being a polyhedral complex, dimension) we focus on equivalence classes of pro-
jections. The map π ÞÑ kerπ is a bijection between equivalence classes of projections
and linear subspaces of Rn of dimension n�m.
Let A � Rn be a finite set of affine linear subspaces. It is called complete when
it is closed under intersections. The dimension of A is defined to be the maximal
dimension of its elements. We write |A | for the union of the affine subspaces inA .
A projection π : Rn Ñ Rm is said to be regular with respect toA when it respects
the dimensions of the elements ofA and detects the inclusion of subspaces inA , i.e.
when
(i) dimπpAq � dim A for all A PA ,
(ii) A� B when πpAq � πpBq for all A, B PA .
Note that m ¡ dimA when π is regular with respect to A, unless |A | is an affine
subspace itself. The next lemma shows that regular projections are abundant.
Lemma 1.4.1 LetA � Rn be a finite set of affine subspaces with dimA � d n. Thenthe collection of kernels of projections Rn Ñ Rd�1 that are regular with respect to Acontain a Zariski dense and open subset of Grn�d�1pRnq.
Proof. The proof follows that of Lemma 4.2. of [7]. Let P � |A | be a finite set such
that every affine subspace in A is the affine span of a subset of P. Write VA,p for the
linear subspace of Rn generated by A PA and p P P.
Let π : Rn Ñ Rd�1 is a projection with kerπX VA,p � t0u for all A PA and p P P.
We prove that π is regular. Firstly, π is injective on all of VA,p. Hence dim A� dimπpAqand πppq P πpAq implies p P A. It remains to prove that πpAq � πpBq implies A � B.
Let p1, . . . , pk P P be a subset whose affine span is A. Then πppiq P πpBq for all i Pt1, . . . , ku. Hence pi P B and thus A� B. This proves that π is regular. Thus, the set of
kernel of regular projections contains the set that is the intersection of all sets,
tW P Grn�d�1pRnq |W X VA,p � t0uu, with A PA and p P P.
1.4. TROPICAL GEOMETRY IN HIGHER CODIMENSION 13
Since the dimension of VA,p is at most d � 1 each of these sets is Zariski dense in
Grn�d�1pRnq and hence their intersection is. l
Lemma 1.4.2 Let A be a finite set of affine subspaces of Rn with dimA � d n andlet U � Grn�d�1pRnq be a Zariski dense subset. Then there exist d � 1 affine projectionsπ0,π1, . . . ,πm : Rn Ñ Rd�1 with kernels in U such that for every v P |A | there existsi P t0, . . . , mu such that π�1
i pπipvqq X |A | � tvu.
Moreover, for any v0 P |A | one can choose the projections such that for all i Pt0, . . . , du it holds that π�1
i pπipv0qq X |A | � tv0u.
Proof. The proof follows Lemma 4.3. of [7]. In order to make an induction we prove
the following stronger statement. Let A 1 � A , with dimA 1 � d 1. There are d 1 � 1
affine projections π0,π1, . . . ,πd1 in U such that for each v PA 1 there is i P t0, . . . , d 1uwith π�1
i pπipvqq X |A 1| � tvu.
We proceed by induction on d 1. If d 1 � �1, then A 1 � H and the statement is
vacuously true. Suppose that d 1 ¥ 0 and that the statement holds for all integers
smaller than d 1. Since the set of projections regular with respect to A contains a
dense open subset of Grn�d�1pRnq (cf. Lemma 1.4.1) and U is dense there exists a
projection π that is both in U and regular with respect to A . Let X � |A 1| be the
subset consisting of all v P |A 1| for which there exists w P |A | with w � v such that
πpvq � πpwq.
For any affine hyperplane W PA 1 it holds that,
W X X �¤V
pW Xπ�1pπpV qqq, with V PA and W � V ,
by regularity of π. Since π is regular πpW q � πpV q for all W � V and hence the
dimension of W Xπ�1pπpV qq is strictly smaller than that of W . Now set,
A 2 � tW Xπ�1pπpV qq |W PA 1, V PA and W � Vu.
By the induction hypothesis there exist m projections π0, . . . ,πd�1 such that the state-
ment holds forA 2. By construction, adding πd � π yields the statement forA 1. l
Proposition 1.4.3 Let S � Rn be an arbitrary subset. Let U � Grn�d�1pRnq a densesubset such that for every projection π : Rn Ñ Rd�1 with kernel in U the image πpSqis a polyhedral complex of dimension at most d. Then there are projections π0, . . . ,πn :
Rn Ñ Rd�1, with kernel in U, such that,
S �n£
i�0
π�1i pπipSqq. (1.2)
14 CHAPTER 1. TROPICAL GEOMETRY
In particular, S is a polyhedral complex.
Proof. The proof consists of two parts. In the first we show that S is contained in the
union of a finite collection of affine subspaces of dimension d. In the second we apply
Lemma 1.4.2 to give the statement.
We construct the collection of affine subspaces by induction. SetA0 � tRnu. Sup-
pose we have constructed a collectionAk of affine subspaces of dimension n� k ¡ dsuch that S � |Ak|. We claim there exists a projection π : Rn Ñ Rd�1 with kerπ P Usuch that kerπ and A affinely span Rn for all A PAk.
Since n� k ¤ d � 1 there exists a finite collection L of d � 1-dimensional linear
subspaces of Rn such that for each A P Ak there is a subspace L P L that is parallel
to an affine subspace of A. Moreover, since the collection of complements of each L is
open and dense in Grn�d�1pRnq and L is finite the collection of subspaces that are
complements to all L PL is open and dense as well. Thus, the intersection of U with
the latter subset is dense, thus in particular, non-empty. Let π be a projection with
kernel in this subset. Then π satisfies the stated criterion. In the second part of the
proof we refer to this specific choice of π by πn�k.
Let B be a finite set of hypersurfaces in Rd�1 such that πpSq � |B |. Set Ak�1 to
be,
Ak�1 � tπ�1pBq X A | B PB and A PAku.
By construction S � |Ak�1|. It remains to show that the elements of Ak�1 have
dimension at most n� k� 1. Suppose that AXπ�1pBq has dimension n� k for some
choice of A and B. Since A has dimension n � k it would follow that A � π�1pBq.However, the affine span of kerπ and A would be contained in π�1pBq, and hence not
equal to Rn. This contradicts the choice of π.
For the second part of the proof, first note that by construction,
|An�d | �n£
i�n�d�1
π�1i pπipSqq.
By Lemma 1.4.2 there exist d � 1 affine projections π0, . . . ,πd such that for all v P|An�d there exists i P t0, . . . , du such that tvu � |An�d | Xπ
�1i pπipvqq|. Thus,
S �An�d Xd£
i�0
π�1i pπipSqq �
n£i�0
π�1i pπipSqq.
This proves (1.2). Since all the intersectands are polyhedral complexes their intersec-
tion is. l
This ends the part on generic projections of arbitrary subsets of Rn. From here on
we apply the results to tropical varieties.
1.4. TROPICAL GEOMETRY IN HIGHER CODIMENSION 15
By abuse of terminology, a monomial map Kn Ñ Kd�1 is a map of the form
pa1xα1 , . . . , ad�1xαd�1q, where ai P K and αi P Nn. The tropicalization of a mono-
mial map restricts to an affine linear map Rn Ñ Rd�1 of the form v ÞÑ Av�w, where
A P Nd�n and w P F d�1.
Proposition 1.4.4 Let φ : Kn Ñ Kd�1 be a monomial map and X � Kn an irreduciblealgebraic variety of dimension d. Write Y � Kd�1 for the Zariski closure of φpX q. ThenTrop pφqpTrop pX qq � Trop pY q, and equality holds when Y is a hypersurface.
Proof. We use an elementary argument to show that Trop pφqpTrop pX qq � Trop pY q.Set I � I pX q and J � pφ�q�1pIq, where φ� : Kry1, . . . , yd�1s Ñ Krx1, . . . , xns is the
pullback map induced by φ. Let v P Trop pX q and consider w � Trop pφqpvq. Let f P J .
We show that w is in the tropical vanishing locus of Trop p f q. Write,
f �¸β
cβyβ , Trop pφqpvq � A � v� u.
for A P Nn�pd�1q, u P F d�1 and consider,
Trop p f qpwq �àβ
ωpcβq dwβ
� minβpωpcβq � β � Av� β � uq
� minβppωpcβq � β � uq � pβ � Avqq
� Trop pφ�p f qqpvq
Moreover, the terms occurring in the minimum of the third line and the fourth line are
the same. Since v P Trop pX q, the minimum of those terms in Trop pφ�p f qq is attained
at least twice, or Trop pφ�p f qq equals 8. This shows that w P Trop pY q.
Suppose that Y is a hypersurface and let w P Trop pY q. The set Y Xω�1pwq is
dense in Y , by Theorem 1.3.1. Hence it intersects φpX q. There is p P X such that
ωpφppqq � w. The statement follows by ωpφppqq � Trop pφqpωppqq and the fact that
ωppq P Trop pX q. l
In the next proposition it is ascertained that the projections required to prove the
main theorems are indeed dense, and hence allow us to apply Proposition 1.4.3.
Proposition 1.4.5 Let X � Kn be an irreducible algebraic variety of dimension d. WriteM for the set of monomial maps φ : Kn Ñ Kd�1 that satisfy,
(i) the map Trop pφq : Rn Ñ Rd�1 is surjective,(ii) the Zariski closure of φpX q is a hypersurface.
The set of linear spaces tkerpTrop pφqq | φ PM u is dense in Grn�d�1pRnq.
16 CHAPTER 1. TROPICAL GEOMETRY
Proof. We restrict ourselves to the case char K � 0.
Since condition (i) defines a dense subset of Grn�dpRmq it is sufficient to prove
that the maps satisfying condition (ii) define a subset of Grn�dpRnq that contains a
dense open subset.
Denote the restriction of φ to X by φX and write Y for the Zariski closure of the
image of φX . Define U � X by the following expression,
U � tp P X | dppφX q : TpX Ñ TφppqY has rank du.
If φX pUq intersect the nonsingularity locus of Y non-trivially, then Y is a hypersurface
in Kd�1. After all, if p P U and φX ppq is non-singular, then
dim Y � dimTφX ppqY ¥ dim dppφX q�TppX q
�� d.
Let q be a non-singular point in X . We claim that if dqφX has rank d, then there
exists p P U with φX ppq non-singular. Since dqφX has rank d the set U is open and
dense in X . When the characteristic of K is 0 the singularity locus of Y is contained in
a subset of Y of codimension 1. Thus, in that case φpUq intersects the non-singularity
locus of Y . This proves the statement when char K � 0, since for all φ outside of the
zero locus of detdqφX the closure of φpX q is a hypersurface. l
The next theorem is the most important statement in this section. It is a straight-
forward combination of the results on generic projections, the previous proposition
and the fundamental theorem of tropical geometry for hypersurfaces.
Theorem 1.4.6 Let X � Kn be an irreducible algebraic variety of dimension d. Thereexists monomial maps φ0, . . . ,φn : Kn Ñ Kd�1 such that,
Trop pX q �d£
i�0
Trop pφiq�1�
Trop pφipX qq�, (1.3)
where for each i P t0, . . . , nu the Zariski closure φipX q is a hypersurface and Trop pφiq isa projection. In particular, Trop pX q is a polyhedral complex of dimension at most d.
Proof. Let U denote the subset of monomial maps φ : Kn Ñ Kd�1 whose tropi-
calization is surjective and satisfy that the Zariski closure Z of φpX q is a hypersur-
face. Then the set tkerpTrop pφqq | φ P Uu is a dense subset of Grn�d�1pRnq. Then
for every π with kernel in this set there exists φ such that π � Trop pφq. The im-
age πpTrop pX qq equals the tropicalization of Z . By Theorem 1.3.2 it is a polyhedral
complex of dimension d. Hence, by Proposition 1.4.3, there exist monomial maps
φ0, . . . ,φn : Kn Ñ Kd�1 that tropicalize to projections with kernel in U such that
(1.3) holds. l
1.4. TROPICAL GEOMETRY IN HIGHER CODIMENSION 17
Corollary 1.4.7 The tropical variety Trop pX q has a tropical basis of cardinality n� 1.
Proof. Let φ0, . . . ,φn be as in the theorem and write I � I pX q. The Zariski clo-
sure Zi of φipX q is a hypersurface, and hence the zero locus of a single polynomial
hi P Kry1, . . . , yd�1s with φ�i phiq P I . We claim that tφ�i phiq P I | i � 0, . . . , nu is
a tropical basis for Trop pX q. By Proposition 1.2.2 the tropicalization Trop pZiq equals
T phiq. Then Trop pφiq�1pTrop pZiqq � T pφ�i phiqq. Applying Theorem 1.4.6 yields that
φ�i phiq form a tropical basis. l
Finally, we prove the fundamental theorem.
Theorem 1.4.8 Let X � Kn be an algebraic variety. Then ωpX q � Trop pX q.
Proof. Let v P Trop pX q. Let φ : Kn Ñ Kd�1 be a monomial map and write π : Rn Ñ
Rd�1 for its tropicalization. Suppose that the closure of φpX q is a hypersurface, that π
is surjective and that Trop pX qXπ�1pπpvqq � tvu. Such φ exist, since the combination
of the first two conditions (cf. Proposition 1.4.5), as well as the last condition, define
subsets of Grn�d�1pRnq containing an open dense set.
Write Z for the Zariski closure of φpX q. By Proposition 1.4.4 it holds that πpX q �Trop pZq. By Theorem 1.3.1 the preimage of πpvq under the valuation ω is dense in Zand hence intersects φpX q non-trivially. Thus, there exists p P X such that πpωppqq �ωpφppqq � πpvq. By the fact that ωppq P Trop pX q and the choice of φ it holds that
ωppq � v, and hence v PωpX q. l
Proposition 1.4.9 Let X � Kn be an irreducible algebraic variety of dimension d. ThenTrop pX q is the underlying set of a pure polyhedral complex of dimension d.
Proof. By Theorem 1.4.6 the set Trop pX q is the underlying set of a polyhedral complex
of maximal dimension at most d. Let P1Y. . .YPk be a non-redundant decomposition of
Trop pX q into polyhedra. Suppose that Trop pX q is not pure. There exists i P t1, . . . , kusuch that dim Pi d. Assume without loss of generality that i � 1.
Let π be (i) a projection Rn Ñ Rd�1 that is the tropicalization of a map φ :
Kn Ñ Kd�1 with the Zariski closure of φpX q a hypersurface, and (ii) regular with
respect to the collection of affine spaces spanned by the Pi . Then πpTrop pX qq is a pure
polyhedral complex of dimension d, since it is the tropicalization of a hypersurface
(cf. Proposition 1.4.4). Hence, πpP1q is contained in the union of πpPiq, with i � 1.
In fact, it is contained in the union of πpPiq with dimpπpP1q X πpPiqq � dimπpPiq
and hence πpP1q and πpP1q X πpPiq span the same affine subspace of Rd�1. Thus,
πpA1q � πpAiq where A j is the affine span of Pj . Hence A1 � Ai by regularity of π.
Since π is injective on Pj and πpP1q is contained in the union of πpPjq, as above, it
18 CHAPTER 1. TROPICAL GEOMETRY
holds that,
P1 �k¤
i�2
Pi ,
showing that P1 is redundant in the decomposition. This proves the statement. l
1.5 Tropical geometry and valuation theory
The main theorem of tropical geometry states that the non-Archimedean amoeba
ωpX q of X and the intersection of the tropical zero loci of the polynomials in the ideal
of X coincide, i.e. that the sets ωpX q and Trop pX q coincide. There is another way to
arrive at the same set, by considering the collection of (real-valued) valuations maps
on the coordinate ring of X that extend ω.
Let X � Kn be an embedded algebraic variety. The embedding gives rise to a
surjective homomorphism Krxs Ñ KrX s. We denote the image of f P Krxs under this
homomorphism by f . Write BpX q for the set of all real-valued valuations on KrX sthat extend ω, i.e.
BpX q � tν | ν : KrX s Ñ R8 is a valuation and νpaq �ωpaq for all a P Ku.
By slight abuse of terminology we refer to BpX q as the Berkovich analytic space of X(cf. [51], [31]).
The Bieri-Groves set of the embedding X � Kn is the image of BpX q un-
der the restriction of the valuations to (the monoid generated by) the coordinates
x� t x1, . . . , xnu and denoted BGxpX q, i.e.,
BGxpX q � tpνp x1q, . . . ,νp xnqq | ν PBpX qu � Rn8.
When the embedding is clear from context we write BGpX q. The main theorem of this
section is that BGxpX q X F n and Trop pX q coincide. This was proved in [4], making
use of the techniques of [5]. We prove it directly from the main theorem of tropical
geometry.
Theorem 1.5.1 Let X � Kn be an algebraic variety. Then BGxpX q X F n � Trop pX q.
Proof. Write π : Krxs Ñ KrX s for the canonical projection. We remind the reader that
we write f for πp f q.
Let w P BGxpX q. There exists a valuation ν : KrX s Ñ R8 such that νp x iq � wi .
The pullback π�pνq defined by π�pνqp f q � νp f q is a valuation on Krxs that satisfies
1.5. TROPICAL GEOMETRY AND VALUATION THEORY 19
π�pνqpx iq � wi and π�pνqp f q � 8 for all f P kerπ� I pX q. Write,
f �¸α
cαxα
and suppose the minimum in Trop p f qpwq is attained uniquely by some term α�. Then
π�pνqp f q �ωpcα�qdvα�
, by the properties of valuations, and henceωpcαqdwα �8
for all α. Thus Trop p f qpwq � 8.
The other inclusion is a direct consequence of the fundamental theorem of trop-
ical geometry (Theorem 1.4.8). Suppose w P Trop pX q. There exists p P X such
that ωppq � w. The map g ÞÑ ωpgppqq is a valuation νp : KrX s Ñ R8. It satisfies
νpp x iq �ωpx ippqq �ωppiq � wi . Thus, w P BGxpX q. l
Remark. The Berkovich analytic space is quite a large and complicated set. The main
theorem of tropical geometry essentially states that to obtain the restrictions of valu-
ations to a fixed finite subset of the coordinate ring of the variety, the point valuations
suffice. This is quite a remarkable result.
20 CHAPTER 1. TROPICAL GEOMETRY
Chapter 2
Tropical unirational varieties
Tropical geometry has proved useful for implicitization, i.e., for determining equations
for the image of a given polynomial or rational map [61, 62, 64]. The fundamental un-
derlying observation is that tropicalizing the map in a naive manner gives a piecewise
linear map whose image is contained in the tropical variety of the image of the orig-
inal map. Typically, this containment is strict, and for polynomial maps with genericcoefficients the difference between the two sets was determined in [64]. Polynomial
or rational maps arising from applications are typically highly non-generic, and yet
it would be great if those maps could be tropicalized to determine the tropical vari-
ety of their image. Rather than realizing that ambitious goal, this chapter identifies a
concrete research problem and presents several useful tools for attacking it.
The setup in this chapter differs slightly from that of Chapter 1. We use K to
refer to an algebraically closed semifield equipped with a surjective valuation ω onto
a tropical semifield F � R8. However, algebraic varieties are embedded in some
algebraic torus over K and their tropical varieties are underlying sets of polyhedral
complexes in Rn.
More precisely, let Tn denote the algebraic torus Kn� and let X � T n be a subvariety.
Write I for the ideal of X in Krx�1, . . . , x�1s � KrTns and consider the tropicalization
of a polynomial in I to be a polynomial over R8. Then the tropicalization of X is
the intersection of the tropical vanishing loci of the tropicalizations of all elements
of I X Krx1, . . . , xns, intersected with Rn. This definition coincides with the common
definition of the tropical variety of a subvariety of a torus (cf. [52]). Clearly, the main
theorem of tropical geometry does not hold as formulated in Theorem 1.4.8. Instead,
it would read Trop pX q X F n� �ωpX q.
Similarly, the tropicalization of a non-zero polynomial f P Krx1, . . . , xns is a poly-
nomial over R8 with induced map Rn Ñ R. Note that the map induced by the zero
polynomial is not defined in the setting of this chapter. By part (i) of Lemma 1.2.1,
21
22 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
(−u,−u)
(−s,−s)u
s
(−u, 0)
(s− u, 0)
(0, u− s)
(0,−s)
Figure 2.1: Outside the lines, Trop pψq is linear and of the indicated form.
we have Tropp f gqpvq � Tropp f qpvq d Troppgqpvq for non-zero polynomials f , g, and
this implies that we can extend the operator Trop to rational functions by setting
Trop p f {gq � Trop p f q � Trop pgq. We further extend this definition to rational maps
φ � p f1, . . . , fnq : Tm 99K Tn by setting Troppφq � pTropp f1q, . . . , Tropp fnqq : Rm Ñ Rn.
Now the central definition of this chapter is the following.
Definition 2.0.1 A subvariety X of Tn is called tropically unirational if there exists a
natural number m and a dominant rational map φ : T m 99K X such that the image
imTrop pφq equals Trop pX q. The map φ is then called tropically surjective.
We recall that the inclusion imTroppφq � Trop pX q always holds (cf. [21]). The
following example shows that this inclusion is typically strict, but that φ can some-
times be modified (at the expense of increasing p) so as to make the inclusion into an
equality.
Example. Let X � T2 be the line defined by y � x �1, with the well-known tripod as
its tropical variety. Then the rational map φ : T1 99K T2, t ÞÑ pt, t � 1q is dominant,
but the image of its tropicalization only contains two of the rays of the tripod, so φ is
not tropically surjective. However, the rational map,
ψ : T2 99K X � T2, ps, uq ÞÑ p1� s
u� s,1� u
u� sq
is tropically surjective. Indeed, see Figure 2: under Trop pψq, the north-west and
south-east quadrants cover the arms of the tripod in the north and east directions,
respectively, and any of the two halves of the north-east quadrant covers the arm
of the tripod in the south-west direction. So X is tropically unirational. There is no
tropically surjective rational map into X with m� 1.
The central question that interests us is the following.
2.1. SOME CLASSES OF TROPICALLY UNIRATIONAL VARIETIES 23
Question 1
Is every unirational variety tropically unirational?
This chapter is organized as follows. In Section 2.1 we review the known fact that
(affine-)linear spaces and rational curves are tropically unirational. In Section 2.2
we prove that, at least for rational varieties, our central question above is equivalent
to the apparently weaker question of whether Trop pX q is the union of finitely manyimages im Trop pφiq, i � 1, . . . , N with each φi a rational map Tpi 99K X . This involves
the concept of reparameterizations: precompositions φ�α of a dominant rational map
φ into X with other rational maps α; since tropicalization does not commute with
composition, Trop pφ �αq may hit points of Trop pX q that are not hit by Trop pφq.
In Section 2.3 we introduce a somewhat ad-hoc technique for finding suitable
(re)parameterizations. Together with tools from preceding sections this technique al-
lows us, for example, to prove that the hypersurface of singular n � n-matrices is
tropically unirational for every n. In Section 2.4 we prove that for X unirational, ev-
ery sufficiently generic point on Trop pX q has a dimpX q-dimensional neighborhood
that is covered by Trop pφq for suitable φ; here we require that K has characteristic
zero. Combining reparameterizations, we find that there exist dominant maps into
X whose tropicalization hit full-dimensional subsets of all full-dimensional polyhedra
in the polyhedral complex Trop pX q. But more sophisticated methods, possibly from
geometric tropicalization, will probably be required to give a definitive answer to our
central question.
2.1 Some classes of tropically unirational varieties
We start with some elementary constructions of tropically unirational subvarieties of
tori.
Lemma 2.1.1 If X is a tropically unirational subvariety of Tn, then so is its imageLuπpX q under any torus homomorphism π : Tn Ñ Tq followed by left multiplicationLu with u P Tq.
Proof. Ifφ : Tm 99K X is tropically surjective, then we claim that so is Lu�π�φ : Tm 99K
Y :� LupπpX qq. Indeed, since φ is a monomial map and Lu is just componentwise
multiplication with non-zero scalars, we have Trop pLu�π�φq � Trop pLuq�Trop pπq�
Trop pφq. Here the first map is a translation over the componentwise valuation ωpuqof u, and the second map is an ordinary linear map. The claim follows from the known
fact that Trop pLuq�Trop pπqpTrop pX qq � Trop pY q, which is a consequence of the main
theorem of tropical geometry (cf. Theorem 1.4.8). l
The following is a consequence of a theorem by Yu and Yuster [71].
24 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
Proposition 2.1.2 If X is the intersection with Tn of a linear subspace of Kn, then X istropically unirational.
Proof. Let V be the closure of X in Kn, by assumption a linear subspace. The support of
an element v P V is the set of i such that vi � 0. Choose non-zero vectors v1, . . . , vp P
V such that the support of each vector in V contains the support of some vi . Let
A P Kn�p be the matrix with columns v1, . . . , vp, and let vpAq P pRY t8uqn�p be the
image of A under coordinate-wise valuation. Then by Theorem 16 of [71] states that
Trop pV q � pR Y t8uqn is equal to the image of pR Y t8uqp under tropical matrix
multiplication with ωpAq. This implies that the rational map Tp 99K Tn, v ÞÑ Av is
tropically surjective. l
Another argument for the tropical unirationality of linear spaces will be given in
Section 2.3. There is a more general statement in the context of tropical linear spaces
associated to valuated matroids in Chapter 4.
Lemma 2.1.3 Let X � Tn be a closed subvariety, and write X for the the cone tpt, t pq |t P K�u over X in Tn�1. Then X is tropically unirational if and only if X is.
Proof. If X is tropically unirational, then so is T�X � Tn�1, and hence by Lemma 2.1.1
so is the image X of the latter variety under the torus homomorphism pt, pq ÞÑ pt, t pq.Conversely, if X is tropically unirational, then so is its image X under the torus homo-
morphism pt, pq ÞÑ t�1p. l
Note that Trop pX q � pt0u�Trop pX qq�Rp1, . . . , 1q; we will use this in Section 2.2.
We can now list a few classes of tropically unirational varieties.
Corollary 2.1.4 Intersections with Tn of affine subspaces of Kn are tropically unira-tional.
Proof. If X the intersection with Tn of an affine subspace of Kn, then the cone X is the
intersection with Tn�1 of a linear subspace of Kn�1. Thus the corollary follows from
Proposition 2.1.2 and Lemma 2.1.3. l
The following corollary has been known at least since Speyer’s thesis [58].
Corollary 2.1.5 Rational curves are unirational.
Proof. Let φ � p f1, . . . , fnq : T 99K Tn be a rational map, and let Y be the rational
curve parameterized by it. Let S � K be a finite set containing all roots and poles of
the fi , so that we can write
fipxq � ci
¹sPS
px � sqeis ,
2.1. SOME CLASSES OF TROPICALLY UNIRATIONAL VARIETIES 25
where the ci are non-zero elements of K and the eis are integer exponents. Let X � TS
be the image of the affine-linear linear map T 99K TS given by x ÞÑ px�sqsPS . Then X is
tropically unirational by Corollary 2.1.4. Let π : T S Ñ T n be the torus homomorphism
mapping pzsqsPS to p±
sPS zeiss qi , and let u � pciqi P T n. Then the curve Y is the image
of X under Lu �π, and the corollary follows from Lemma 2.1.1. l
Corollary 2.1.6 The variety in T m�n of m� n-matrices of rank at most 2 is tropicallyunirational.
Proof. Let φ : Tm � Tm � Tn � Tn 99K T m�n be the rational map defined by
φ : pu, x , v, yq ÞÑ diagpuqpx1tr � 1y trqdiagpvq,
where diagpuq, diagpvq are diagonal matrices with the entries of u, v along the di-
agonals; 1t , 1 are the 1 � n and the m � 1 row vectors with all ones; and x , yare interpreted as column vectors. Elementary linear algebra shows that φ is dom-
inant into the variety Y of rank-at-most-2 matrices. Moreover, φ is the composi-
tion of the linear map pu, x , y, vq ÞÑ pu, x1t � 1y t , vq with the torus homomorphism
pu, z, vq ÞÑ pdiagpuqz diagpvqq. Hence Y is tropically unirational by Proposition 2.1.2
and Lemma 2.1.1. l
Corollary 2.1.7 The affine cone over the Grassmannian of two-dimensional vector sub-spaces of an n-dimensional space (or more precisely its part in Tp
n2q with non-zero Plücker
coordinates) is tropically unirational.
Proof. The proof is identical to the proof of Corollary 2.1.6, using the rational map
Tn � Tn ÞÑ Tpn2q, pu, xq ÞÑ puiu jpx i � x jqqi j .
l
Interestingly, Grassmannians of two-spaces and varieties of rank-two matrices are
among the few infinite families of varieties for which tropical bases are known [18].It would be nice to have a direct link between this fact and the fact, used in the
preceding proofs, that they are obtained by smearing around a linear space with a
torus action.
Corollary 2.1.8 The varieties defined by A-discriminants are tropically unirational.
Proof. Like in the previous two cases, these varieties are obtained from a linear variety
by smearing around with a torus action; this is the celebrated Horn Uniformisation[28, 38]. l
26 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
In fact, this linear-by-toric description of A-discriminants was used in [20] to give
an efficient way to compute the Newton polytopes of these discriminants in the hy-
persurface case. A relatively expensive step in this computation is the computation of
the tropicalization of the kernel of A; the state of the art for this computation is [55].
2.2 Combining reparameterizations
A fundamental method for constructing tropically surjective maps into a unirational
variety X � Tn is precomposing one dominant map into X with suitable rational maps.
Definition 2.2.1 Given a rational dominant map φ : Tm 99K X � Tn, a reparameter-ization of φ is a rational map of the form φ � α where α : Tp 99K Tm is a dominant
rational map and p is some natural number.
The point is that, in general, Trop pφ�αq is not equal to Trop pφq�Trop pαq. So the
former tropical map has a chance of being surjective onto Trop pX q even if the latter
is not.
Example. In Example 2, the map ψ : ps, tq ÞÑ p 1�su�s
, 1�uu�s
q is obtained from φ : t ÞÑpt, t�1q by precomposing with the rational map α sending ps, uq to 1�s
u�s. Hence ψ is a
tropically surjective reparameterization of the non-tropically surjective rational map
φ.
This leads to the following sharpening of our Question 1.
Question 2
Does every dominant rational map φ into a unirational variety X � Tn have a tropi-
cally surjective reparameterization?
Note that if X is rational and φ : Tm 99K X � Tn is birational, then every dominant
rational mapψ : Tp Ñ X factors into the dominant rational map pφ�1�ψq : Tp 99K Tm
and the map φ. So for such pairs pX ,φq, the preceding question is equivalent to the
question whether X is tropically unirational.
We will now show how to combine reparameterizations at the expense of en-
larging the parameterizing space Tp. For this we need a variant of Lemma 2.1.3.
Let φ � pf1g
, . . . , fn
gq : Tm 99K X � Tn be a dominant rational map where
g, f1, . . . , fn P Krx1, . . . , xms. Let d ¡ 0 be a natural number greater than or equal
to maxtdegpgq, degp f1q, . . . , degp fnqu, and define the homogenizations
g :� xd0 gp
x1
x0, . . . ,
xn
x0q and fi :� xd
0 fipx1
x0, . . . ,
xn
x0q, i � 1, . . . , n.
2.2. COMBINING REPARAMETERIZATIONS 27
These are homogeneous polynomials of positive degree d in n � 1 variables
x0, . . . , xn. The map φ : Tm�1 99K Tn�1 with components p g, f1, . . . , fnq is called
a degree-d homogenization of φ. The components of one degree-d homogenization
of φ differ from those of another by a common factor, which is a rational function
with numerator and denominator homogeneous of the same degree. Any degree-
d homogenization of φ is dominant into the cone X in Tn�1 over X . Recall that
Trop pX q � t0u � Trop pX q � Rp1, . . . , 1q. The following lemma is the analogue of
this statement for imTrop pφq.
Lemma 2.2.1 Let φ be any degree-d homogenization of φ. Then the image of Trop pφq
equals t0u� pimTrop pφqq �Rp1, . . . , 1q.
Proof. For the inclusion �, let v P Rm and let u P R. Setting v :� p0, vq� udp1, . . . , 1q P
Rm�1 and using that g is homogeneous of degree d we find that
Trop p gqpvq � Trop p gqp0, vq � u� Trop pgqpvq � u;
and similarly for the fi . This proves that
Trop pφqpvq � Trop pφqpvq � up1, . . . , 1q,
from which the inclusion � follows. For the inclusion � let v � pv0, . . . , vmq P Rm�1
and set vi :� vi � v0, i � 1, . . . , m. Again by homogeneity we find
Trop p gqpvq � Trop p gqp0, vq � d v0 � Trop pgqpvq � d v0
and similarly for the fi . This implies
Trop pφqpvq � Trop pφqpvq � d v0p1, . . . , 1q,
which concludes the proof of �. l
Lemma 2.2.2 (Combination Lemma) Let φ : Tm 99K X � Tn and αi : Tpi 99K Tm
for i � 1,2 be dominant rational maps. Then there exists a dominant rational mapα : Tp1�p2�1 99K T m such that im Trop pφ � αq contains both imTrop pφ � α1q andimTrop pφ �α2q.
Proof. Consider a degree-d homogenization φ : Tm�1 99K X � Tn�1 and degree-ehomogenizations α1 : Tp1�1 99K Tm�1, α2 : Tp2�1 99K Tm�1 of α1,α2, respectively.
Define α : Tp1�1 � Tp2�1 99K Tm�1 by
αpp, qq � α1ppq � α2pqq.
28 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
We claim that
im Trop pφ � αq � imTrop pφ � αiq for i � 1, 2.
Indeed, since φ has polynomial components and since α2 is homogeneous of positive
degree e, we have
φpα1puq �α2pvqq � φpα1puqq � terms divisible by at least one variable v j .
As a consequence, for pµ, νq P Rp1�1 �Rp2�1 we have
Trop pφ � αqpµ, νq �mintTrop pφ � α1qpµq, terms containing at least one ν ju.
Hence if µ is fixed first and ν is then chosen to have sufficiently large (positive)
entries, then we find
Trop pφ � αqpµ, νq � Trop pφ � α1qpνq.
This proves that im Trop pφ � α1q � im Trop pφ � αq. Repeating the argument with the
roles of 1 and 2 reversed proves the claim.
Now we carefully de-homogenize as follows. First, a straightforward computation
shows that φ � αi is a degree-de homogenization of φ � αi for i � 1, 2, hence by
Lemma 2.2.1 we have
impTrop pφ � αiqq � t0u� imTrop pφ �αiq �Rp1, . . . , 1q.
Similarly, writing α � pa0, . . . , amq : T p1�1�p2�1 99K T m for the components of α we
define α : T p1�p2�1 99K T m as the de-homogenization of α given by
αpu, vq ��
a1p1, u, vq
a0p1, u, vq, . . . ,
amp1, u, vq
a0p1, u, vq
.
A straightforward computation shows that φ � α is a degree-de homogenization of
φ �α. Hence by Lemma 2.2.1 we have
impTrop pφ � αqq � t0u� imTrop pφ �αq �Rp1, . . . , 1q.
Now the desired containment
imTrop pφ �αq � im Trop pφ �αiq for i � 1,2
2.3. BIRATIONAL PROJECTIONS 29
follows from
t0u� im Trop pφ �αq � pt0u�Rnq X imTrop pφ � αq
� pt0u�Rnq X imTrop pφ � αiq
� t0u� imTrop pφ �αiq
l
2.3 Birational projections
In this section we show that rational subvarieties of T n that have sufficiently many
birational toric projections are tropically unirational. Here is a first observation.
Lemma 2.3.1 Let X � Tn be an algebraic variety and π : Tn Ñ Td a torus homo-morphism whose restriction to X is birational, with rational inverse φ. Then Trop pπq �
Trop pφq is the identify on Rd .
Proof. Let v P Rd be a point where Trop pφq is (affine-)linear. Such points form the
complement of a codimension-1 subset and are therefore dense in Rd . Hence it suf-
fices to prove that Trop pφqpvq maps to v under Trop pπq. Let y P T d be a point with
ωpyq � v where φ is defined and such that x :� φpyq P X satisfies πpxq � y .
Such points exist because ω�1pηq is Zariski-dense in T d (cf. Theorem 1.4.8). Now
w :� ωpxq equals Trop pφqpvq by linearity of Trop pφq at v and Trop pπqpwq � v by
linearity of Trop pπq. l
For our criterion we need the following terminology.
Definition 2.3.1 Let P � Rn be a d-dimensional polyhedron and let A : Rn Ñ Rd be
a linear map. Then P is called A-horizontal if dim ApPq � d.
Proposition 2.3.2 Let X � Tn be an algebraic variety and π : Tn Ñ Td a torus homo-morphism whose restriction to X is birational, with rational inverse φ. Using the Bieri-Groves theorem, write Trop pX q �
�i Pi where the Pi are finitely many d-dimensional
polyhedra. Then imTrop pφq is the union of all Trop pπq-horizontal polyhedra Pi .
Proof. Let Pi be a Trop pπq-horizontal polyhedron. We want to prove that Trop pφq �
Trop pπq is the identity on Pi . To this end, let v P Pi be such that Trop pφq is affine-
linear at w :� Trop pπqv. The fact that Pi is Trop pπq-horizontal implies that such v are
dense in Pi . To prove that Trop pφqpwq equals v let x P X be a point with ωpvqpxq � vsuch that φ is defined at y :� πpxq and satisfies φpyq � x . The existence of such a
point follows from birationality and the density of fibers in X of the valuation map
30 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
(cf. Theorem 1.4.8). Now w equals vpyq by linearity of Trop pπq and v � vpxq �vpφpyqq equals Trop pφqpwq by linearity of Trop pφq at w. Hence Trop pφq � Trop pπq
is the identity on Pi , as claimed. Thus imTrop pφq contains Pi . Since the projections of
Trop pπq-horizontal polyhedra Pi together form all of Rd , we also find that imTrop pφq
does not contain any points outside those polyhedra. l
Corollary 2.3.3 Let X � Tn be a birational variety and write Trop pX q ��
i Pi as inProposition 2.3.2. If for each Pi there exists a torus homomorphism π : Tn Ñ Td that isbirational on X and for which Pi is Trop pπq-horizontal, then X is tropically unirational.
Proof. In that case, there exist finitely many homomorphisms π1, . . . ,πN : Tn Ñ Td ,
birational when restricted to X , such that each Pi is Trop pπ jq-horizontal for at least
one j. Then Proposition 2.3.2 shows that the rational inverse φ j of π j satisfies Pi �
imTrop pφ jq. Now use the Combination Lemma 2.2.2. l
In particular, when all coordinate projections to tori of dimension dim X are bira-
tional the variety X is tropically unirational. This is the case in the following state-
ment.
Corollary 2.3.4 For any natural number n the variety of singular n � n-matrices istropically unirational.
Proof. A matrix entry mi j of a singular matrix can be expressed as a rational
function of all other n2 � 1 entries (with denominator equal to the corresponding
pn� 1q � pn� 1q-subdeterminant). This shows that the map Tn2Ñ Tn2�1 forgetting
mi j is birational. Any pm2 � 1q-dimensional polyhedron in Rm2is horizontal with re-
spect to some coordinate projection Rn2Ñ Rn2�1, and this holds a fortiori for the
cones of tropically singular matrices. Now apply Corollary 2.3.3. l
Corollary 2.3.3 also gives an alternative proof of Corollary 2.1.4 stating that trop-
icalizations of affine-linear spaces are tropically unirational.
Proof. [Second proof of Corollary 2.1.4] Let X be the intersection with Tn of a d-
dimensional linear space in Kn. For each polyhedron Pi of Trop pX q there exists a co-
ordinate projection π : Tn Ñ TI , with I some cardinality-d subset of the coordinates,
such that Pi is Trop pπIq-horizontal. Here we have not yet used that X is affine-linear.
Then the restriction πI : X Ñ TI is dominant, and since X is affine-linear, it is also
birational. Now apply Corollary 2.3.3. l
We continue with an example of a determinantal variety of codimension larger
than one whose unirationality is a consequence of Corollary 2.3.3.
Example. Let V � M4�5pKq be the variety of matrices of rank smaller than or equal
to 3. The ideal of V contains all maximal minors and the dimension of V equals 18.
2.3. BIRATIONAL PROJECTIONS 31
One way to see the latter statement is to write a matrix in V in the following form,
�A BC D
�, A P M3�3, B P M3�2, C P M1�3, D P M1�2.
There are no conditions on A, B and C , while D is uniquely determined by the choice
of A, B and C . The dimension thus equals 3 � 3� 3 � 2� 1 � 3� 18.
Let pmi jq denote the standard coordinate functions on M4�5. We aim to show that
the projection into any subset of X of size 18 is birational. Let z1 � mi, j and z2 � ml,k
be the indices of the coordinate functions left out of the projection. Note that if z1
appears in a maximal minor, in which z2 doesn’t, then z1 is a rational function of the
coordinate functions in the maximal minor. In particular, if z1 and z2 are in different
columns, there exist such maximal minors for z1 and z2 and hence both are rational
in the remaining 18 coordinate function.
The case that z1 and z2 are in the same column requires some calculation. Suppose
without loss of generality that z1 � m3,4 and z2 � m4,5. Then,
0 � m35 det M124,234 �m45 det M123,234 �m15 det M234,234 �m25 det M124,234,
0 � m35 det M124,134 �m45 det M123,134 �m15 det M234,134 �m25 det M134,134
by cofactor expansion of the determinants of the matrices M1234,2345 and M1234,1345.
The set of equations has a unique solution for m35 and m45 when
det M124,134 det M123,134 � det M124,234 det M123,134,
showing that the projection is birational.
We conclude this section with a beautiful example, suggested to us by Filip Cools
and Bernd Sturmfels, and worked out in collaboration with Wouter Castryck and Filip
Cools.
Example. Let Y � T5 be parameterized by ps4, s3 t, . . . , t4q, ps, tq P T 2, the affine cone
over the rational normal quartic. Write X :� Y � Y � T5, the first secant variety.
Writing z0, . . . , z4 for the coordinates on T5, we can describe X as the hyperplane
defined by
det
��� z0 z1 z2
z1 z2 z3
z2 z3 z4
�� � z0z2z4 � 2z1z2z3 � z2
1z4 � z0z23 � z3
2
� a� 2b� c� d � e.
32 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
x0
x4
x2
a
b
c
d
e
Figure 2.2: The Newton polytope of the Hankel determinant. Only the exponents ofz0, z2, z4 have been drawn.
Trop(πI)-horizontal
Trop(πJ)-horizontal
neither
e ac
bd
Figure 2.3: The tropical variety of the Hankel determinant. Two-dimensional regionscorrespond to the monomials a, b, c, d, e.
This polynomial is homogeneous both with respect to the ordinary grading of
Krz0, . . . , z4s and with respect to the grading where zi gets degree i. Hence its New-
ton polygon is three-dimensional; see Figure 2.2. Modulo its two-dimensional lineal-
ity space the tropical variety Trop pX q is two-dimensional. Intersecting with a sphere
yields Figure 2.3.
Now set I :� t1, 2,3, 4u and J :� t0,1, 2,3u. Then the coordinate projections
πI : T5 Ñ TI , πJ :� T5 Ñ TI are birational, since z0 and z4 appear only linearly in
the Hankel determinant. Let P be a full-dimensional cone in Trop pX q, and let tα,βu
be the corresponding edge in the Newton polygon. Then P is Trop pπIq-horizontal if
and only if α0 � β0 and Trop pπJq-horizontal if and only if α4 � β4. Figure 2.3 shows
that all but one of the cones are, indeed, horizontal with respect to one of these
projections. Let P denote the cone corresponding to the edge between the monomials
b � z1z2z3 and e � z32 . By Proposition 2.3.2 and the Combination Lemma 2.2.2,
there exists a rational parameterization of X whose tropicalization covers all cones
of Trop pX q except, possibly, P. We now set out to find a parameterization whose
tropicalization covers P.
Let v P P. By Lemma 2.2.1 we may assume that v is of the form pv0, v1, 2v1, 3v1, v2q.
We aim to show that there exist two reparameterizations of φ : T4 Ñ X , where
φpu0, u1, v0, v1q � pu0 � v1, u0v0 � u1v1, u0v20 � u1v2
1 , u0v30 � u1v3
1 , u0v40 � u1v4
1 q.
2.4. VERY LOCAL REPARAMETERIZATIONS 33
such that v is in the image of at least one of them. Note that from the defining in-
equalities of P it follows that v0 ¥ 0 and v2 ¥ 4v1. Let i P K be a fourth root of unity
and consider the map ψ : T3 Ñ T 4 defined by
ψpx0, x1, x2q ��1� x0,�1, i x1,�x1p1� x2 x�4
1 q�.
A short computation shows that the restriction of the tropicalization of φ � ψ to
the cone defined by w0 ¥ 0, w2 ¥ 4w1 and w2 ¤ w0 � 4w1 is the linear function
pw0, w1, w2q ÞÑ pw0, w1, 2w2, 3w1, w2q. If v satisfies v2 ¤ v0 � 4v1 then the image of
pv0, v1, v2q under this tropicalization is exactly v.
If v satisfies v2 ¥ v0�4v1 it is in the image of the tropicalization of φ�ψ�ι, where
ιpx0, x1, x2q � px�10 , x�1
1 , x�12 q.
The tropicalization is linear on the cone 0 ¥ w0, 4w1 ¥ w2 and w2 ¥ w0 � w1 and
maps �pv0, v1, v2q to v.
2.4 Very local reparameterizations
Let X � Tn be a d-dimensional rational variety that is the closure of the image of a
rational map φ : Tm 99K Tn. Suppose without loss of generality that X is defined over
a valued field pK ,ωq such that ωpK�q is a finite dimensional vector space over Q.
Write v for the image of v P R under the canonical projection RÑ R{ωpK�q. We can
now state the main result of this section and its corollary.
Theorem 2.4.1 Assume that the field K has characteristic zero. Let v P Trop pX q and setd to be the dimension of the Q-vectorspace spanned by v1, . . . , vn. There exists a rationalmap α : Td 99K Tm and an open subset U � Rd such that the restriction of Trop pφ �αq
to U is an injective affine linear map, whose image contains v.
Corollary 2.4.2 Assume that K has characteristic zero. Let tP1, . . . , Pku be a finite set ofωpK�q-rational polyhedra of dimension dim X such that
Trop pX q �k¤
i�1
Pi .
There exist a natural number p and a rational map α : Tp 99K Tm such that the image ofTrop pφ �αq intersects each Pi in a dim X -dimensional subset.
Proof. By the theorem, for each polyhedron Pi there exists a reparameterization αi
such that the tropicalization of φ �αi hits Pi in a full dimensional subset. They can be
34 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
combined using the Combination Lemma. l
The main step in the proof of the theorem is Proposition 2.4.5, which is a valuation
theoretic result.
Let v be a point of Rn. Such a point defines a valuation ωv on the field of rational
functions L � Kpy1, . . . , ynq of T n by
ωvphq � Trop phqpvq, h P L
Let Lv denote the completion of L with respect to ωv and denote its algebraic closure
by Lv . That closure is equipped with the unique valuation whose restriction to Lv
equals ωv [66, §144]. Denote by KryQ1 , . . . , yQ
n s the subring of Lv generated by all
roots of the elements y1, . . . , yn.
The next lemmata deal with the case n � 1. They allow us to prove Proposi-
tion 2.4.5 below by means of induction on the number of variables.
Lemma 2.4.3 Let v P R such that v is not in the Q-vector space spanned by vpK�q.Then Krt�1, ts is dense in Kptqv .
Proof. Let p{q P Kptq. If q is a monomial we are done, since then 1{q P Krt, t�1s.
Suppose it isn’t. Write,
q �¸
i
qi ti , with qi P K�.
Then the valuation of q equals ωvpqq � miniωpqiq � iv. Moreover, the minimum is
attained exactly once, since otherwise v would be a Q-multiple of some element of
vpK�q. Say it is attained at j. Compute,
p
q�
p
a j t j � pq� a j t jq
� pa j tjq�1 p
1� pq� a j t jq{pa j t jq
� a�1j t� j p
8
n�0
� pq� a j tjq
a j t j
n
The convergence of the power series with respect to ωv is a consequence of ωvpq�a j t
jq ¡ ωvpa j tjq. The limit is easily seen to coincide with p{q. This completes the
proof. l
Lemma 2.4.4 Suppose K is algebraically closed of characteristic 0. Then KrtQs is densein Kptqv .
2.4. VERY LOCAL REPARAMETERIZATIONS 35
Proof. Denote the residue field of Kptqv by k. Note that it is also the residue field of
Kptq under v, by the conditions on v.
We prove by induction on d that all zeroes in Kptqv of a polynomial of degree dover Kptqv can be approximated arbitrarily well with elements of KrtQs. For d � 1
this is the content of Lemma 2.4.3. Assume that the statement is true for all degrees
lower than d. We follow the proof of [67, §14, Satz].Let PpSq � Sd � ad�1Sd�1� . . .� a0 P KptqvrSs. After a coordinate change replac-
ing S by S � 1d
ad�1 we may assume that ad�1 � 0. Indeed, a root s of the original
polynomial can be approximated well by elements of KrtQs if and only if the root
s� 1d
ad�1 can be approximated well, since ad�1 itself can be approximated well.
If now all ai are zero, then we are done. Otherwise, let the minimum among the
numbers vpad�iq{i be ω� qτ, where ω P vpK�q and q PQ, and let c be a constant in
K with valuation ω. Setting S � c tqU transforms P into
cd tdqpUd � bd�2Ud�2 � . . .� b0q,
where each bi is an element of Kpt1{pqv of valuation at least zero, with p the denom-
inator of q. Moreover, some bi has valuation zero. Let QpUq denote the polynomial in
the brackets. The image of QpUq in the polynomial ring krUs over the residue field is
neither Ud as bi has non-zero image in L, nor a d-th power of an other linear form
as the coefficient of Ud�1 is zero. Hence the image of QpUq in krUs has at least two
distinct roots in the algebraically closed residue field k, and therefore factors over kinto two relatively prime polynomials. By Hensel’s lemma [67, §144], Q itself factors
over Kpt1{pqv into two polynomials of positive degree. By induction the roots of these
polynomials can be approximated arbitrarily well by elements of KrtQs, hence so can
the roots of Q and of P. l
Proposition 2.4.5 Let pK ,ωq be an algebraically closed field of characteristic 0 withvaluation ω and v P Rn whose entries are Q-linearly independent over R{ωpK�q. ThenKryQ
1 , . . . , yQn s is dense in Lv .
Proof. Follows from Lemma 2.4.4 by induction on the number of variables. l
We are now ready to prove the main result.
Proof. [Proof of Theorem 2.4.1] Choose τ1, . . . ,τd P R such that their projections in
R{ωpK�q form a basis of the Q-vectorspace spanned by ξ1, . . . , ξn. Let t1, . . . , tn be
variables and denote by L the field Kpt1, . . . , tnq equipped with the unique valuation
ν that extends ω and satisfies νpt iq � τi .
There exists a point x 1 P T mpLξq such that wpφpx 1qq � ξ. By Proposition 2.4.5
there exists an approximation x P T mpKrtQ1 , . . . , tQd s) of x 1 that satisfies wpφpxqq � ξ.
Choose e P N such that every coefficient of x is already in Krt�1{e1 , . . . , t�1{e
d s and set
36 CHAPTER 2. TROPICAL UNIRATIONAL VARIETIES
si � t1{ei . Thus, x P T mpKrs�1 , . . . , s�d sq, and hence defines a rational map T d Ñ T m.
Denote this map α. We show that there exists a neighborhood of σ � 1epτ1, . . . ,τdq
such that the restriction of Trop pφ �αq to U satisfies the conclusions of the theorem.
First, note that, by construction of α, Trop pφ � αqpσq � ξ. Every component
φipxq � φipαps1, . . . , sdqq of φ is a Laurent polynomial over K in the s j with a unique
term zisbi,1
1 � � � sbi,d
d of minimal valuation ξi � vpziq �°d
j�1 bi, jσ j . If we let σ vary in a
small neighborhood and change the valuations of the si accordingly, then for each ithe same term of φi has the minimal valuation. Hence Tpφ �αq is linear at σ with dif-
ferential the matrix pbi jq. Finally, as the numbers η1, . . . ,ηn span the same Q-space as
σ1, . . . ,σd modulo vpK�q, the matrix pbi jq has full rank d. This completes the proof.
l
2.5 Concluding remarks
The concept of tropical surjectivity of a rational map seems natural and concrete,
and, as far as we know, not to have been studied before. This chapter presented
some methods of determining whether a rational map is tropically surjective, and
aims to be a starting point for further study. In particular, the question whether every
unirational variety is tropically unirational is still open. It seems likely that techniques
from geometric tropicalization will prove useful in making further progress on this
question.
Chapter 3
Algebra of idempotent
semirings
In this chapter the focus is moved away from tropicalization map to the tropical semi-
fields. It gathers known results on the broader class of semirings whose addition is
idempotent, into which the tropical semifields fit naturally, and adapts them to our
specific notation and setting.
The chapter serves two purposes. First, a number of results on tropical linear
spaces, the subject of investigation in Chapter 4, hold for arbitrary modules over
idempotent semirings, and they are better stated and proved in this context. Secondly,
it serves to understand the algebraic properties that are necessary for the notion of a
matroid, and hence a tropical linear space, over a semiring to be well-behaved.
3.1 Modules over semirings and homomorphisms
The most general type of algebraic structure considered in this thesis is the semiring,
and its associated structure, the module over a semiring. We start off by defining the
former.
Definition 3.1.1 A semiring is an algebraic structure pR,�, �, 0Rq consisting of a set
R and two binary operations, addition, denoted �, and multiplication, denoted �,
satisfying the conditions
(i) pR,�, 0q is a commutative monoid with identity element 0 P R,
(ii) pR, �, 1q is a monoid with identity element 1 P R,
(iii) multiplication distributes over addition,
(iv) for all x P R it holds that 0 � x � 0 � x � 0,
(v) 0� 1.
37
38 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
A commutative semiring is a semiring in which multiplication is commutative. If each
element in Rzt0u has a multiplicative inverse the commutative semiring is said to be
a division semiring. A commutative division semiring is called a semifield. In case of
possible ambiguity we write 0R, respectively 1R, for the identity elements of R.
Note that the absorption axiom stating that 0 � x � x �0� 0 for all x P R is implied
by the other semiring axioms when additive inverses always exist. The latter property
is necessary in the implication, which is why the absorption axiom is explicitly added
here.
It is clear from the definition that rings are special cases of semirings. In fact they
are the semirings whose every element has an additive inverse. Idempotent semir-
ings are in some sense on the other end of the spectrum. Their addition operation
is idempotent and hence non-invertible. In formal terms, a semiring pR,�, �,8q is
said to be idempotent when for all x P R it holds that x � x � x , or equivalently,
when 1R � 1R � 1R. The terminology extends to commutative semirings and semi-
fields. Idempotent semirings are also known as dioids, where dioid stands for double
monoid.
Remark. We shall make no distinction between the algebraic structure and the under-
lying set, when reference to the operations is clear from the context. Thus, R denotes
the underlying set, as well as the semiring.
In this thesis all semirings are understood to be commutative, and hence we usually
omit this adjective.
Example. An important example of a semiring that is not a ring is the set of extended
real numbers RY t8u in which x ` y � mintx , yu and x d y � x � y . It is in fact
a semifield, and we refer to it as the real tropical semifield. The symbols ` and d
serve to clearly distinguish the semiring operations from the usual (partially defined)
addition and multiplication on RYt8u. We denote the real tropical semifield by R8.
The terminology and notation extends to subgroups of pR,�q. For example, the
subsemiring of R8 consisting of Q and 8 is referred to as the tropical rational semi-field, or simply the tropical rational numbers, and denoted Q8.
Definition 3.1.2 Let R be a (commutative) semiring. A module M over R is an alge-
braic structure consisting of a set M and the operations of addition M � M Ñ M ,
denoted �, and scalar multiplication R� M Ñ M , denoted �, such that the following
conditions are satisfied,
(i) pM ,�, 0M q is a commutative monoid,
(ii) x � pv�wq � x � v� x �w for each x P R and v, w P M ,
(iii) px � yq � v � x � v� y � v for each x , y P R and v P M ,
(iv) 1R � v � v for all v P M .
3.1. MODULES OVER SEMIRINGS AND HOMOMORPHISMS 39
As in the case of semirings, M is said to be an idempotent module if it has idempotent
addition, i.e. m� m � m for all m P M . Note that the axioms ensure that a module
has a zero element, denoted 0M , or simply 0, when reference to the module under
consideration is clear from the context.
Remark. Modules over semirings are often referred to as semimodules in the litera-
ture. In this thesis, we prefer the term module, for two reasons: (i) the axiom system
of a semimodule is precisely that of a module once one replaces the term ring by
semiring and hence the concept of a semimodule is an extension of that of a module
(this argument also appears in a footnote in [36]), (ii) there is no specific reference
in this thesis to modules over rings, and hence, once one keeps that in mind, there
should be no loss in clarity as a consequence of our naming convention. Moreover, it
saves writing the morpheme semi many times.
Examples. The following are examples of modules over semirings. Many more exam-
ples are to be found in [29].(i) For R a semiring and X a set denote by RX the collection of maps X Ñ R. This
is an R-module under the operations of pointwise addition and scalar multi-
plication. Moreover, such R-modules are free in the category of R-modules and
R-module homomorphisms.
(ii) Let R be the tropical rational semifield and consider the module M over R whose
underlying set is that of tropical real numbers, RY t8u, with module addition
v ` w � mintv, wu and scalar multiplication x d v � x � v. That the module
postulates are satisfied is a consequence of the tropical semifield structure on
RY t8u. It is clear that M is not finitely generated over R, since the quotient
R{Q of additive abelian groups is not finite.
(iii) Every semiring is a module over itself. Addition is the ordinary addition opera-
tion in the semiring and scalar multiplication is multiplication in the semiring.
In this sense, modules over semirings are more general than semirings.
A module over a semiring that has idempotent addition is clearly idempotent itself,
since v � v � p1R � 1Rq � v � v, but in general the implication is not an equivalence.
A simple example is a module M with idempotent addition, whose ring of scalars
acts trivially, i.e. each scalar acts as the endomorphism v ÞÑ 0M . However, if each
multiplication with a scalar defines a different endomorphism (such modules are said
to be faithful), then the implication is an equivalence.
A useful device for the analysis of modules whose addition is idempotent is the
notion of the natural preorder on the module. Since this preorder depends only on the
additive structure of the module it is most naturally defined for general semigroups.
Let S be such a semigroup. The natural preorder relation P� on S is defined by x P� ywhen there exists z P S such that x � y � z. That this actually defines a preorder is
40 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
part (i) of the following proposition. For us, part (iii) of the same proposition is the
most important part. It states that when addition on the module is idempotent it is
uniquely determined by P�. Moreover, in that case P� is a meet semi-lattice order,
and we denote it by ¤�, or ¤, when the additive structure under consideration is
clear from the context. Then it is also referred to as the natural order [29], inducedorder or canonical order [43] on the semigroup.
Proposition 3.1.1 Let S be a commutative semigroup. The following statements hold.(i) The relation P� is a preorder,
(ii) If S is a group then P� � S� S,(iii) If S is idempotent then px , yq P P� is equivalent to x � y � x and P� defines a
meet semi-lattice. Moreover, x � y � x ^ y, where ^ denotes the meet operationof P� and the unique maximum of P� is the identity element of S, if it exists. Anymeet semi-lattice is obtainable in this way.
Proof. (i) We need to show that P� is reflexive and transitive. Let x P S. Then x�0� xand hence px , xq P P�. Let x , y, z P S such that px , yq P P� and py, zq P P�. There exist
u, v P S such that x�u� y and y�v � z. Thus, x�pu�vq � z and hence px , zq P P�.
(ii) Let x , y P S. Then x � py � xq � y and hence px , yq P P�.
(iii) Suppose that x � y � x . Then clearly xP� y . Now suppose that x P� y . There
exists z P S such that x � y�z. Then x� y � y�z� y � y�z � x , which shows the
required equivalence. Since P� was already proved to be a preorder in (i) it remains
to show that it is antisymmetric. Let x , y P S and suppose that both xP� y and yP�x .
Then x � x � y � y .
The partial order P� is a lower semilattice order exactly when each set tx , yu has
a unique infimum inftx , yu. We prove that inftx , yu exists and equals x � y , in two
steps. For clarity we denote P� by ¤.
(a) The identity x � px � yq � y � px � yq � x � y implies px � yq ¤ x and
px � yq ¤ y , as required. Hence x � y is a lower bound of tx , yu(b) Suppose that z is a lower bound of tx , yu. Then z � px � yq � pz � xq � y �
z� y � z. Hence, z ¤ px� yq, showing that x� y is the largest lower bound of tx , yu.l
Remark. The correspondence of Proposition 3.1.1 restricts to a correspondence be-
tween idempotent commutative monoids and lower semilattices with a unique maxi-
mal element. The equations defining the unit element of the monoid translate directly
into the required inequalities for the maximal element. Moreover, the correspondence
is (trivially) functorial in the sense that maps that preserve the structure of the in-
duced meet semi-lattice are semigroup homomorphisms, and vice versa.
Note that commutative idempotent semigroups are also known as commutative
bands, see [16].
3.1. MODULES OVER SEMIRINGS AND HOMOMORPHISMS 41
The remainder of this section is concerned with modules that have idempotent
addition. Properties of such modules are often easier expressed in terms of the canon-
ical order. The next theorem serves to prove that the order restricted to some subset
of the semiring is a lattice order. For continuous semi-lattices this is automatic and
the subset is the entire ring, since the supremum of a set is just the infimum of its
upper bounds. Since we do not assume continuity, more work needs to be done. A
by-product of the proof is that the multiplicative inverse is an anti-automorphism of
lattices on R�, where R� denotes the group of elements of R that have a multiplicative
inverse.
Proposition 3.1.2 Let R be an idempotent semiring. If R� is closed under � then thenatural order on R restricts to a lattice order on R�.
Proof. Let x , y P R�. It is sufficient to prove that x�1 ^ y�1 � px _ yq�1. Denote
x�1 ^ y�1 by w. We show that w�1 � x _ y , in two steps.
(i) To prove that x ¤ w�1 compute,
w�1 � x � w�1 � p1� pw � xqq
� w�1 � p1� py�1 � xqq
� w�1 � px�1 � y�1q � x
� x
Hence, x ¤ w�1, and similarly, y ¤ w�1.
(ii) For the second step, let z P R satisfy x ¤ z and y ¤ z. Then,
w�1 � z � w�1 � p1� pw � zqq
� w�1 � p1� ppx�1 � zq � py�1 � zqqq
� w�1 � x�1 � px � z� px � y�1 � zqq
� w�1 � p1� py�1 � zqq
� w�1 � y�1 � py � zq
� w�1.
Hence, w�1 ¤ z, thereby proving the statement.
l
Corollary 3.1.3 Let F be an idempotent semifield. The natural order on F is a latticeorder.
Proof. It follows from the lemma that the natural order restricted to F� � Fzt0u is a
lattice order. Since x �8 � x for all x P F the element 0 is the top element of the
42 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
natural order on F . Thus, x ^ 0� x and x _ 0� 0 for all x P F . l
We continue the discussion of idempotent semirings by distinguishing the class of
tropical semirings in terms of algebraic conditions on semirings. A division semiring
F is Archimedean if for all x , y P F� the statement xZ ¤ y implies x � 1, i.e. when
the only element all of whose powers are smaller than any given invertible element
is the multiplicative identity. Recall that the tropical real numbers are the semifield
with underlying set RYt8u and operations x ` y �mintx , yu, x d y � x � y . Note
that their natural order is the common order on the extended real numbers. The next
theorem is essentially a theorem by Hölder. The statement and its proof, formulated
in the language of lattice ordered groups is found for example in [60].
Theorem 3.1.4 Let F be an idempotent semifield for which the induced order is linearthat satisfies the Archimedean property. Then F embeds into pRYt8u, min,�q.
Proof. By Theorem 2.3.10. of [60] a any lattice ordered group embeds into the lattice
ordered group pR,�q if and only if it is Archimedean and linearly ordered. It is easy
to see that pF�, �q under the natural order satisfies these properties and hence that it
embeds in pR,�q. The embedding extends to an embedding F Ñ R8. l
A tropical semiring is defined to be a subsemiring of an Archimedean idempotent
semifield. The above theorem shows that tropical semirings in our sense coincide with
the tropical semirings as defined in Chapter 1 and, for example, in [53]. In certain
types of statements the Archimedean property is not necessary and in these cases it is
more natural to focus on linearly ordered semirings.
Example. The smallest example of a tropical semiring is the two element semiring
ta, bu with addition and multiplication defined by the tables,
� a b
a a ab a b
� a b
a a bb b b
This semiring is commonly referred to as the Boolean semiring or trivial semifield and
denoted B. The embedding B ãÑ R8 that establishes it as a tropical semiring is a ÞÑ 0
and b ÞÑ 8.
A simple example of a linearly ordered semiring that is not tropical, consider the
three element semiring, R � t�8, 0,8u. It is an extension of the Boolean tropical
semiring t0,8u � R8, with addition defined �8� x � �8 and �8 � x � �8 for
all x P R.
The general notion of an idempotent module has the disadvantage that infima
of infinite sets do not exist in general. An approach often taken in the literature (e.g.
3.1. MODULES OVER SEMIRINGS AND HOMOMORPHISMS 43
[13]) is to extend the module under consideration to one that is complete in the sense
that every set has an infimum. For our purposes this condition is too strong. After all,
it is not satisfied by the tropical real numbers, since the infimum of the subset R of
R8 does not exist. Requiring that infima of bounded subsets exist will not do, since
we also wish to consider modules over the tropical rational numbers. Instead we will
require infima of very specific subsets of semirings to exist.
Let M be a module over an idempotent semiring R. Such a module is said to be
dually residuated if for each v, w P M with w � 0M the set tx P R | v ¤ x � wu has a
unique minimum. It is then denoted by v : w, or pv : wq, and called the dual residuumof v and w. In terms of the semimodule structure pv : wq is the (infinite) sum of all
elements x P R such that v � x � w � v. The notion of (dual) residual in modules
over semirings is found in [13]. It is well-studied in case the module is the ground
semiring with scalar multiplication given by semiring multiplication. See for example
[70, 65] and the more recent survey [37].
The next statement serves to single out relations between modules that preserve
the property of dual residuatedness .
Proposition 3.1.5 Let R be an idempotent semiring. The following statements hold.
(i) Let M be a module over R and N � M a submodule. If M is dually residuated overR then N is dually residuated over R.
(ii) Suppose that the induced order of R is a lattice order. If R is dually residuated thenany free module of finite rank over R is dually residuated over R.
(iii) Let A be an algebra over R and M a module over A. If A is dually residuated over Rand M is dually residuated over A then M is dually residuated over R.
Proof. We prove the claims in order stated.
(i) This is clear since both the natural order and the action of R on N are the
restrictions of those on M .
(ii) Let M be a free module over R of finite rank. Fix a finite set of free generators
G � M . Let v, w P M with w � 0M and define px gq and pygq in R such that,
v �¸gPG
x g � g, w �¸gPG
yg � g.
Set z � suptpx g : ygq | g P Gu. We claim that pv : wq � z. By a straightforward
computation,
z �w �¸gPG
pz � ygq � g ¥¸gPG
x g � g � v,
which shows that z ¥ pv : wq.
44 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
Let z1 P R satisfy v ¤ z1 � w. Then by the premise that M is free and the following
computation,
v � v� z1 �w �¸gPG
px g � z1 � ygq � g,
it holds that x g � z1 � yg � x g for all g P G. In terms of the induced order on R this
states that z1 � yg ¥ x g for each g P G. Hence, z1 ¥ z and the claim follows.
(iii) Let v, w P M with w � 0M . Subscripts A and R denote the dual residua in
the A-module M and the R-module A, respectively, e.g. pv : wqA stands for the dual
residuum of v by w in M over A. Set x � ppv : wqA : 1AqR. We claim that x � minty PR | v ¤ y �wu.
Since v ¤ pv : wqA �w ¤ px �1Aq �w � x �w, it holds that x is at least the minimum.
Suppose that y P R satisfies v ¤ y � w. We need to show that x ¤ y . Multiplying both
sides of the inequality with 1A yields that v ¤ py � 1Aq � w and thus y � 1A ¥ pv : wqA.
Hence y ¥ ppv : wqA : 1Aq � x . This proves the claim. l
Remark. In part (ii) of the preceding proposition the module is required to be of finite
rank since otherwise there is no guarantee on the existence of lowest upper bounds
of infinite subsets of R. Vice versa, if the induced order on R has the property that it
is a lattice that is complete as an upper semi-lattice then the finite rank condition is
superfluous. Since tropical linear spaces, the main geometric objects in Chapter 4, are
finitely generated, it is more natural to require finiteness, instead of completeness.
Remark. Naively one might expect that the dual residuation property of semirings
transfers to finitely generated modules over such semirings. This is not true, since
there might not be a system of generators whose mutual dual residua with respect
to R exist. A simple example is the Q8-submodule of R8 generated by t0, vu, with
v P R8zQ8. The dual residuum pv : 0q does not exist since the set tx P Q8 | v ¤ xuhas no infimum.
When the module has a finite generating set whose mutual dual residua exist it is
dually residuated over its base idempotent semiring, when the induced order of the
base ring is a lattice order. The proof is nearly as the same as for free modules of finite
rank, as in part (ii) of the Proposition 3.1.5.
Dual residuation is a device that allows one to translate order theoretic properties
of idempotent semirings to finitely generated modules over them. The key is the fact
that any element of such a module has a canonical representation as a linear combi-
nation of generators, where each coefficient is a dual residuum . More precisely, let
M be such a module over an idempotent semiring R and v P M . Fix a finite set of
3.1. MODULES OVER SEMIRINGS AND HOMOMORPHISMS 45
generators G � Mzt0Mu. Then,
v �¸gPG
pv : gq � g. (3.1)
We refer to this sum as the canonical representation of v in terms of G. Similarly, in
this setting pv : gq is referred to as the canonical coordinate of v with respect to g.
Proposition 3.1.6 Let R be an idempotent semiring having induced lattice order and Ma finitely generated module over R that is residuated over R. Then M has induced latticeorder.
Proof. Let v P M . Then v � v � p1� 1q � v � v. By part (iii) of Proposition 3.1.1 the
induced order on M is a meet semi-lattice. It remains to prove that it is also a join
semi-lattice. Here we require that M is finitely generated.
Fix a finite set of generators G � Mzt0Mu of M . Let v, w P M and define u P M by,
u�¸gPG
�pv : gq _R pw : gq
�� g.
We claim that u� v_M w. For each g P G the join pv : gq_R pw : gq is an upper bound
of both pv : gq and pw : gq. Hence u¥ v_M w.
Let u1 be an upper bound of both v and w. Since residuation is increasing in the
first argument both pu1 : gq ¥ pv : gq and pu1 : gq ¥ pw : gq. Thus pu1 : gq is at least the
smallest upper bound pv : gq_Rpw : gq and hence u1 ¥ u. This proves that u¤ v_M w.
l
Example. This example serves to direct the reader’s attention to the fact that finite
generatedness of modules over semirings is a subtle property. Even submodules of free
modules of finite rank over a linearly ordered semifield need not be finitely generated,
as the example in the proof of Proposition 2.2 in [15] shows. We repeat the relevant
part here.
Consider the free module R38 over the real tropical semiring R8 and set vn �
p� 1n, 0, 1
nq for all n P N. Denote by Vn the submodule generated by v1, . . . , vn. Clearly,
Vn � Vn�1 for all n P N. However, Vn�1zVn �H and the minimal module that contains
all vn is not finitely generated.
The last part of this section deals with homomorphisms of idempotent semimod-
ules. There is a notion of (dual) residuation of an increasing map which is an exten-
sion of that of a module (cf. [8]). Dual residuation in a module translates then to dual
residuation of the map R Ñ M defined by x ÞÑ x � w, where w P M . An increasing
map φ : M Ñ N between modules over idempotent semirings is said to be dually
46 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
residuated if there exists a map such ψ : N Ñ M such that
φ �ψ¥ idN and ψ �φ ¤ idM , (3.2)
in which the partial order on the collection of maps is the pointwise order. We then
denote the map ψ by φ^. This notation comes from the fact that ψ is the map that
assigns to each w P N the infimum of the preimage of the principal filter with minimal
element w under φ, if it exists. Since this is conceptually important we state the
equivalence in the next lemma. The proof of this lemma is to be found for example in
[8, p.10-11]. We remind the reader that a principal filter is a set of the form ty P M |
x ¤ yu, for some x P M .
Lemma 3.1.7 Suppose φ : M Ñ N is a homomorphism of modules over an idempotentsemiring. Let ψ be a map N Ñ M satisfying (3.2). Then ψpwq � mintv P M | φpvq ¥wu. In particular, the preimage of every principal filter under φ is again a principal filter.
Remark. An increasing map that has a dual residual map is referred to as dually
residuated . This is standard terminology (cf. [8]). It is inconvenient that the prime
property we are interested in is not residuatedness, but its order-theoretic dual. The
reason is essentially the fact that we have defined the natural order as we have, i.e.
v ¤ w if and only v � w � v. This convention is known as the min-convention in
tropical geometry, and its opposite is the max-convention. Had we chosen the max-
convention, the prime property would be residuatedness. Note that the min- and max-
conventions make no difference for the algebra of idempotent semirings, since order-
theoretic statements on the canonical order translate into statements on the additive
structure of the semiring.
The next statement shows that in the category of finitely generated modules that
are residuated over a fixed base semifield, residual maps always exist. The proof is
adapted to our setting from Theorem 5.2. of [8]. The statement below is fairly general
and somewhat technical.
Theorem 3.1.8 Let φ : M Ñ N be an arbitrary map between modules over an idempo-tent semiring R that respects the action of R. If (i) M is finitely generated, (ii) N is duallyresiduated and (iii) φ is a homomorphism of modules satisfying φpMzt0Muq � Nzt0Nu,then φ is dually residuated . In that case,
φ^pwq �¸gPG
pw : φpgqq � g, w P N , (3.3)
where G � Mzt0Mu is a finite set of generators of M.
3.1. MODULES OVER SEMIRINGS AND HOMOMORPHISMS 47
Vice versa, if φ is dually residuated then it is a homomorphism of modules. If, more-over, when the induced orders on both M and N are lattice orders, then φ^ is a homo-morphism of the induced join semi-lattice, i.e., φ^pv _ wq � φ^pvq _ φ^pwq for allv, w P N, that respects the action of R.
Proof. Let w P N . We need to show that the infimum of tv P M | φpvq ¥ wu exists.
We claim that it equals°
gPGpw : φpgqq � g. Denote this quantity by u. Then φpuq �°gPGpw : φpgqq � φpgq ¥ w by definition of pw : φpgqq. Thus, u is larger than the
infimum. For the other inequality, let u1 P M such that φpu1q ¥ w. Write u1 �°
gPG x g �
g, for certain x g P R. Then x g �φpgq ¥ w and hence pw : φpgqq ¤ x g . Thus, u1 ¥ u.
This shows that u is a lower bound of tv P M | φpvq ¥ wu and hence that φ is dually
residuated with residual map w ÞÑ°
gPGpw : φpgqq � g.
Suppose that φ : M Ñ N is a dually residuated map. Let v, w P M . We need to
show that φpv � wq � φpvq � φpwq, or in order theoretic terms, that φpv ^ wq �φpvq ^φpwq. Since v ^ w is a lower bound of tv, wu and φ is an increasing map it
follows that φpv^wq ¤ φpvq ^φpwq.
Let u be a lower bound of tφpvq,φpwqu. Then φ^puq ¤ φ^pφpvqq ¤ v by (3.2)
and similarly, φ^puq ¤ w. Thus, φ^puq ¤ v ^ w. Thus, u ¤ φpφ^puqq ¤ φpv ^ wq,again by (3.2). Thus, φpv^wq is the largest lower bound of tφpvq,φpwqu and hence
φpv�wq � φpvq �φpwq.
It remains to prove that φ^ respects the join semi-lattice structures on N and
M , if they exist. Let w1, w2 P N . Set R � tv P M | φpvq ¥ w1 _ w2u and Si �
tv P M | φpvq ¥ wiu, for i � 1,2. Let v P R. Then φpvq ¥ w1 _ w2 ¥ wi and
hence v P Si . Thus R � Si and inf R ¥ inf S1 _ inf S2. Let u1 � inf S1, u2 � inf S2
and suppose u1 is an upper bound of tu1, u2u. Then φpu1q ¥ φpuiq ¥ wi and hence
φpu1q ¥ w1 _ w2. Thus, u1 P R and hence inf R ¤ u1. Thus, inf R � inf S1 _ inf S2, or
equivalently, φ^pw1 _w2q � φ^pw1q _φ
^pw2q. l
Example. The condition that M is finitely generated can not be done without, unless
one imposes completeness on the induced semi-lattice. Set Q � QY t8u under the
tropical semiring structure and consider the Q-module M of functions f : N Ñ Qthat have finite support tn P N | f pnq � 8u. The map φ : M Ñ Q given by φ :
f ÞÑÀ
nPN f pnq � minnPN f pnq is well defined and a homomorphism of Q-modules.
Moreover, Q is residuated, with px : yq � x � y for y � 8. However, φ is not
residuated as a map, since t f P M | φp f q ¥ xu, for x �8 is not an order filter in Mbecause of the finiteness condition on the support.
To see that the residuatedness of N is a necessary premise of the theorem consider
the canonical injection ι : Q ãÑ R, where again R � RY t8u under tropical opera-
tions. Then the set tx P Q | ιpxq ¥ vu � tx P Q | x ¥ vu has no infimum for v P RzQ.
48 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
Example. It is not surprising, although worthwhile to point out that the dual residual
map of a homomorphism is in general not a homomorphism. We embed B � R8.
Consider the case of the submodule of B4 generated by
g1 � p0,0,8,8q, g2 � p8,8, 0, 0q, g3 � p0,8,8, 0q.
The map B3 Ñ B4 defined by px1, x2, x3q ÞÑ x1 � g1 � x2 � g2 � x3 � g3 satisfies the
conditions of Theorem 3.1.8 and hence has a dual residual map φ^. This map is not
a homomorphism of modules, since φ^pg1q � p0,8,8q, φ^pg2q � p8, 0,8q and
φ^p0,0, 0,0q � φ^pg1 � g2q � p0, 0,0q.
The final statement of this section is a distillation of the most important points
when one specializes to a certain subclass of modules over idempotent semirings with
lattice order. The proof is a direct consequence of a number of theorems in this section,
most importantly Theorem 3.1.8.
Theorem 3.1.9 Let R be a residuated semiring whose natural order is a lattice order.Consider the category ModR of modules over R that are finitely generated and embedinto a free module of finite rank. Then the following statements hold.
(i) all objects of ModR are idempotent modules whose induced order is a lattice order,(ii) the class of residuated maps and module homomorphism between objects in ModR
coincide,(iii) the dual residuation operator on morphisms maps module homomorphisms to join
semi-lattice preserving maps that respect the action of R.
Remark. The objects in the category C are essentially row and column spaces of
matrices over R. As such, examples of such modules are easy to construct. We name a
few which are of specific importance to us.
(i) (affine cones over) tropical polytopes (cf. [19]),
(ii) geometric lattices with a finite number of atoms.
3.2 Linear functionals on idempotent modules
We start by recalling some elementary definitions. Let R be a semiring. Our main
interest in this section is the map that assigns to each module M over R the collection
of homomorphisms M Ñ R. We denote this collection by M� and refer to it as the
space of linear functionals, or the (algebraic) dual space. Unlike [13] and [43] we do
not impose any continuity conditions on the linear functionals under consideration.
Clearly, M� is again a module over R under pointwise defined module operations, i.e.�φ�ψ
�pvq � φpvq �ψpvq and
�x �φ
�pvq � x �φpvq for φ,ψ P M� and x P R.
3.2. LINEAR FUNCTIONALS ON IDEMPOTENT MODULES 49
Let α : M Ñ N be a homomorphism of modules over R. Denote by α� : N�Ñ M�
the map that assigns the functional v ÞÑ ψpαpvqq to each ψ P N�. Then α� is an R-
module homomorphism, which we refer to as the transpose of α. Moreover, pα�βq� �
β� � α�, which shows that transpose and algebraic dual make up a contravariant
functor of the category of modules into itself.
The next lemma states an important property of the transpose map when we spe-
cialize to the category of modules over an idempotent semiring with induced linear
order. The first part follows by general principles, while the second part is more spe-
cific to the idempotent semirings with induced linear order. An equivalent statement
for modules over more general idempotent semirings seems not to be known (in gen-
eral injectivity of free objects in the category of modules over a fixed semiring is subtle
[69]). It is equivalent to the statement that R is injective in the category of modules
over R, and the statement that every linear functional on a submodule N � M is the
restriction of a linear functional on M . This is the form in which we require it in the
chapter on tropical linear spaces.
Lemma 3.2.1 Let α : M Ñ N be a homomorphism of modules over an idempotentsemiring R. The following statements hold.
(i) if α is surjective, then α� is injective,(ii) if R has linear canonical order and α is injective, then α� is surjective.
Proof. The first statement is an easy computation.
For the second statement, denote by IncpX , Rq the idempotent R-module of in-
creasing maps from X to R. Define the map ψ : IncpM , Rq Ñ IncpN , Rq by
�ψ f qpwq � supt f pvq | v P M and αpvq ¤ wu
We show that ψ maps M� to N�. Let f P M�. For each w P N write Sw for the
set of all u P M such that φpuq ¤ w. Then�ψ f�pwq � sup f pSwq and the linear-
ity of ψ f is equivalent to sup f pSv�wq � sup f pSvq � sup f pSwq. Clearly, Sv�w �
Sv X Sw and hence f pSv�wq � f pSvq X f pSwq. Thus, sup f pSv�wq is a lower bound
of tsupp f pSvq, supp f pSwqu and hence sup f pSv�wq ¤ sup f pSvq � sup f pSwq.
For the other inequality, let z ¤ tsup f pSvq, sup f pSwqu. It is sufficient to show that
sup f pSv�wq ¥ z. We start by showing that f pSvq is a lower set. Suppose without loss
of generality that f is surjective. Otherwise, f is the map that is identically 8 and
the image of this map under ψ is surely linear. Let x P f pSvq and suppose that y P Rsatisfies y ¤ x . There exists ux P Sv and uy P M such that x � f puxq and y � f puyq.
Then, y � x � y � f puxq� f puyq � f pux �uyq. Since ux �uy ¤ ux and Sv is a lower
set ux � uy P Sv and hence f pSvq is a lower set.
Now, since the canonical order on R is linear, there exists some x P f pSvq such
that z ¤ x . If this were not the case, then (by linearity), z would be an upper bound
50 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
of f pSvq, and hence z � sup f pSvq. Thus, there is uv P Sv and uw P Sw such that
z � f puvq � f puwq. Thus, z � f puv � uwq P f pSv�wq. Thus, sup f pSv�wq ¥ z. l
Remark. The second statement is also true in case R is an idempotent semifield.
One needs the result of [69] that the minimal injective hull of a module over an
idempotent semiring equals the maximal essential extension. Now, suppose that ι :
R ãÑ M is an essential extension, i.e. for all submodules H � M the condition ιpRq XH � t0u implies that H � t0u. Suppose for the sake of contradiction that MzιpRq �H.
Let v P MzιpRq and set H � tx � v | x P Ru. Then v P H and hence H � t0u and thus
HXιpRq � t0u. There exists some x P R such that x �v P ιpRq. However, x�1 �x �v � v PιpRq, which contradicts the choice of v. Thus, R is its own maximal extension in the
category of R-modules, and hence injective in that category. The latter is equivalent
to the statement that the pullback α� of an injection α is a surjective map.
The aim of the next part is to state a theorem that relates the algebraic dual to the
module itself, under certain conditions on the module. In passing we touch upon the
natural notion of separatedness of a module, which is naturally fulfilled for modules
over idempotent semiring that have a specific type of involution, defined in terms of
dual residues .
The map that assigns to each v P M the functional φ ÞÑ φpvq is a homomor-
phism of M to pM�q�. If this homomorphism is injective, the module M is said to
be separated by M�, or simply separated. Separatedness is equivalent to the state-
ment that for each distinct v, w P M there exists a linear functional φ P M� such
that φpvq � φpwq. In [13] the same property is also expressed by the statement that
pM , M�q is a dual pair under the canonical bracket. Our exposition here follows theirs
rather closely, save that we substitute dual residuatedness for completeness and do
not require our lattice homomorphisms to preserve arbitrary infima.
Let R be a dually residuated idempotent semiring and z0 P R. Then R is said to
be reflexive with respect to z0 if the map x ÞÑ pz0 : xq is an set-theoretic involution
R Ñ R. The notion of reflexivity as defined here is the commutative analogue of the
notion in [13]. There is a difference of convention, since their definition of natural
order is opposite to ours. Reflexivity implies in particular that the map x ÞÑ pz0 : xq is
injective and that pz0 : 0q � infty | z0 ¤ 0u � inf R is defined. Note that this infimum
does not exist in tropical semifields other than the Boolean semifield. Finally, a dually
residuated idempotent semiring is said to be reflexive if it is reflexive with respect to
at least one of its elements.
The relation between both concepts defined above is that a module M over a
reflexive semiring R is separated when that module is dually residuated over R. Fix z0
such that R is reflexive with respect to z0. The vehicle for the proof is a natural map
M Ñ M� defined by v ÞÑ inftφ | φpvq ¥ z0u. We write φv for the image of v under
3.2. LINEAR FUNCTIONALS ON IDEMPOTENT MODULES 51
this map. The following statement shows that this map exists and gives an explicit
form. Separatedness is an immediate corollary.
Lemma 3.2.2 Let R be an idempotent semiring that is reflexive with respect to z0 andlet M be a module over R that is dually residuated over R. Then the map M Ñ M� givenby v ÞÑ φv is well-defined and equals w ÞÑ pz0 : pv : wqq. In particular, M is separatedby M�.
Proof. See Theorem 36 of [13] and its corollary. l
Remark. Common operations on semirings preserve reflexiveness (cf. Proposition 31
[13]). The example that is of most interest us is the following. Let F be a semifield
whose induced order is linear. Denote the minimal lattice that has a minimal element
and contains F as a sublattice by F . The underlying set of F is equal to the underlying
set of F (if and only if F is finite), or equal to the disjoint union of the underlying set
of F and tKu, where K is the minimal element of F . The multiplication operator on Fextends to F by setting K� x � x � x �K � x for all x P F and K�K � K. We refer to Fas the extended semifield of F , by abuse of terminology, since if F is infinite K has no
multiplicative inverse and F is only a semiring, since K has no multiplicative inverse.
For modules over idempotent semirings it is not so that the module is isomor-
phic to its algebraic dual as a module, even under strong contextual conditions. The
theorem is somewhat more subtle. It is nicely stated in terms of the notion of an oppo-
site module, which we define in the setting of finitely generated modules over lattice
ordered semirings that are residuated over that semiring.
Let R be a semiring whose induced order is a lattice and let M be a finitely gener-
ated module over R. Then, by Proposition 3.1.6, it follows that M has induced lattice
order. The opposite module of M is the module with underlying set that of M , addi-
tion defined by v�1w � v_w, and scalar operation given by x �1 v � inftw | x �w ¤ vu.It is of course not automatic that this exists for every x P R and v P M . If it does the
opposite module is well-defined and we denote it by M op.
Theorem 3.2.3 Let R be an idempotent semiring that is reflexive with respect to z0 andlet M be a module over R that is both dually residuated and finitely generated over R.Then M and M� are anti-isomorphic lattices with isomorphisms given by v ÞÑ inftφ |
φpvq ¥ z0u and φ ÞÑ inftv | φpvq ¥ z0u. In other words, these maps are isomorphismsbetween the additive semigroups of M and pM�qop.
Proof. This is essentially Corollary 33 of [13]. Although completeness is required in
their statement, examination of the proof reveals that in fact only certain dual residua
are required to exist. Such dual residue exist in M by the premise, and they exist in
M� since M� embeds into a free module of finite rank by the premise that M is finitely
52 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
generated. That dual residua also exist in pM�qop is now a consequence of statement
(18b) of the article. l
Example. Certain natural submodules of the algebraic dual are infinitely generated,
even if the algebraic dual is finitely generated. The key property is the following. A
semiring is called zero-sum free when it holds that x � y � 0 implies both x � 0
and y � 0. Idempotent semirings have this property, since x � y � 0 implies that
x � x � 0 � x � x � y � x � y � 0, and similarly for y . Let R be a zero-sum free
semiring and M a module over R. Consider the space of linear functionals M� of Minto R. Let v P M . It is surprising that the set tφ P M� | φpvq � 8u Y t0M�u is a
submodule of M�. After all, φpvq �ψpvq � 8 when both φpvq � 8 and ψpvq � 8.
However, this submodule need not be finitely generated, even if M� is.
As a specific example, consider the module M � Q28 over the tropical rational
numbers and let v � p0,8q. Set M�v � tφ P M� | φpvq � 8u and write εi for the
i-th coordinate function on Q28. Then M�
v consists of the tropical linear combinations
px1 d ε1q ` px2 d ε2q having either x1 � 8 or x1 � x2 � 8. This module is not
finitely generated. The easiest way to see this is by Theorem 3.1.8, since (i) the map
φ ÞÑ φpvq is not dually residuated, since the infimum of tφ P M�v | x1 � φpvq ¥ 0u
does not exist, (ii) the image of M�v zt8u is contained in Q8zt8u and (iii) Q8 is
residuated.
3.3 Ranks and vanishing conditions
This section deals with putting the algebraic theory of idempotent semirings and rings
on equal footing. The main item is the so-called tropical vanishing property of finite
sequences, which specializes both to the algebraic definition of a vanishing sum and to
the notion of vanishing of tropical geometry. This section was inspired by the article
[17] and the different notions of dependence in tropical modules that exist in the
literature. See [1] for an overview.
We require the next definition.
Definition 3.3.1 Let R be a semiring. The indicator εR of R is the nullary operator on
R that equals �1R if it exists in R, and 1R otherwise. Note that εR ��1R if and only if
R is a ring.
The indicator essentially serves to distinguish rings from proper semirings (i.e.
those in which the additive monoid is not a group) in a way that is useful in identities
and definitions involving the algebra of the semiring.
There exist several notions of dependence in a module in the literature that are of
interest to us: (i) weak dependence, (ii) dependence in the sense of Gondran-Minoux,
3.3. RANKS AND VANISHING CONDITIONS 53
and (iii) tropical dependence. All of these notions are well-studied. For an overview,
see [1]. We postpone their definition until the end of this section.
The different notions of dependence are related to different notions of vanishing
in a module. The one corresponding to the second type of dependence is a straightfor-
ward translation. Let M be a module over a semiring R and let pv1, . . . , vnq be a finite
sequence of elements of M . This sequence is said to vanish in the sense of Gondran-Minoux if there exists a partition pP,Qq of rns such that,
¸iPP
vi � εR
¸iPQ
vi .
The partition is allowed to be trivial. If one of the sets in the partition can be chosen
a singleton, then the sequence is said to vanish weakly. Clearly, a finite sequence that
vanishes weakly also vanishes in the sense of Gondran-Minoux.
The sequence pv1, . . . , vnq is said vanish in the tropical sense when for each φ P M�
there exists a partition pPφ ,Qφq of rns such that
¸iPPφ
φpwiq � εR
¸iPQφ
φpwiq.
This notion is useful for modules over semirings with induced linear order whose
algebraic duals are finitely generated, because in that case, the potentially infinite
number of condition (one for each linear functional) reduces to a finite number (one
for each generator of the algebraic dual).
The algebraic dual of the module R consists of the maps x ÞÑ a�x , where a P R, and
hence tropical and Gondran-Minoux vanishing are equivalent in R. For sequences in Rwe simply say they vanish, or refer to them as vanishing sequences, when the vanish in
one of either equivalent manners. One could say that a sequence in a module vanishes
tropically if its image under any linear functional on the module vanishes. In general,
it is the case that any sequence that vanishes in the sense of Gondran-Minoux also
vanishes in the tropical sense. The converse, however, is false. In the case of modules
over rings, however, all notions of vanishing are equivalent. We summarize matters in
the next statement.
Proposition 3.3.1 Let M be a module over a semiring R.
(i) Any finite sequence in M that vanishes weakly, vanishes in the sense of Gondran-Minoux.
(ii) Any finite sequence in M that vanishes in the sense of Gondran-Minoux vanishestropically.
When M is the module R, then the implication of (ii) is an equivalence. If R is a ringand M is separated by M�, then both implications (i) and (ii) are equivalences, and a
54 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
sequence pv1, . . . , vnq vanishes if and only if°n
i�1 vi � 0.
Proof. Statement (i) is immediate. Statement (ii) is a consequence of the linearity of
the elements of M�.
Suppose R is a ring and M is separated by M�. If a finite sequence indexed by
rns vanishes in the sense of Gondran-Minoux, then any partition of rns is a vanishing
partition. Thus, it vanishes weakly. If a sequence v1, . . . , vn vanishes tropically, then
for any partition pP,Qq of rns and φ P M�,
¸iPP
φpviq � �¸jPQ
φpv jq.
Hence φp°n
i�1 viq � 0. Thus, by separatedness,°n
i�1 vi � 0. l
Examples of modules over semiring in which tropical vanishing is not equivalent
to Gondran-Minoux vanishing are well-known, and we give one later. The notions of
vanishing lead to two distinct notions of structure preserving maps, as defined next.
Let M be an R-module and N an S-module for semirings R and S. A map α : M Ñ
N is said to be a Gondran-Minoux pseudomorphism of modules when for all v, w P Mthe following conditions are satisfied,
(i) αp0M q � 0N ,
(ii) αpεR � vq � εS �αpvq,(iii) the sequence pαpvq,αpwq,εS � αpv � wqq vanishes in the sense of Gondran-
Minoux.
If instead of (iii) the sequence pαpvq,αpwq,εS �αpv�wqq vanishes in the tropical sense,
then the map α is said to be a tropical pseudomorphism of modules. The terminology
extends to maps between semirings and algebras over semirings. Such maps are said
to be pseudomorphisms of semirings, or pseudomorphisms of algebras, if they preserve
the multiplicative structure, in addition to the above axioms.
Remark. Denote the ring of scalars of M by R and the ring of scalars of N by S.
Clearly, if S � R, and α is a homomorphism of R-modules it is a pseudomorphism.
Moreover, if N � S, then by Proposition 3.3.1, a map α : M Ñ N is a Gondran-
Minoux pseudomorphism if and only if it is a tropical pseudomorphism, and we omit
the classifier.
Example. Let K be a field equipped with a non-Archimedean valuation ω : K Ñ Fthat is surjective onto a tropical semifield F � R8. Note that we use the term valua-
tion in the sense of [5]. The more descriptive term would be additive ring valuation.
We denote the operators on F by ` and d, as is customary for tropical semifields.
Then, ω is a pseudomorphism of semirings. The valuation ω respects the multiplica-
tive structure, since ωpx yq � ωpxq �ωpyq � ωpxq dωpyq, where � denotes the
3.3. RANKS AND VANISHING CONDITIONS 55
(extension of the) ordinary addition operator on R. Pseudomorphism axioms (i) and
(ii) state that ωp0q � 8 and ωp�1q � 0, which are fulfilled. The vanishing axiom
(iii) states that either,ωpx� yq �ωpxqdωpyq �mintωpxq,ωpyqu, orωpxq �ωpyqand ωpx � yq ¥ωpxq. This property is satisfied by ω.
The importance of Gondran-Minoux pseudomorphisms lies in the following propo-
sition. Note that we refer to axiom (iii) and its tropical counterpart as the vanishingaxiom.
Proposition 3.3.2 Let M be a module over R and N a module over S. Consider anarbitrary map α : M Ñ N that satisfies αp0M q � 0N and αpεR � vq � εS � αpvq forall v P M. Then α is a Gondran-Minoux pseudomorphism if and only if for each se-quence pv1, v2, . . . , vnq P M n that vanishes in the sense of Gondran-Minoux, the imagepαpv1q,αpv2q, . . . ,αpvnqq vanishes in the sense of Gondran-Minoux as well.
Proof. We start by proving the reverse implication. Let v, w P M . The se-
quence pv, w,εRpv � wqq clearly vanishes. Thus, by the premise, the sequence
αpvq,αpwq,αpεRpv�wqq � pαpvq,αpwq,εSαpv�wqq vanishes.
Let pP,Qq be a vanishing partition for v1, . . . , vn. In particular, there exists mutually
disjoint sets P1, . . . , Pm and Q1, . . . ,Qn that together partition rns such that
m
j�1
α�¸
iPPj
vi
�� εS
n
j�1
α� ¸
iPQ j
vi
�.
Choose these sets such that maxt|Pi |, . . . , |Pm|, |Q1|, . . . , |Qn|u is minimal and suppose
that the maximum is larger than one and that the number of sets of maximal size
is minimal. Without loss of generality, since ε2S � 1, let |Q1| attain the maximum and
write w1, . . . , wk for the elements indexed by Q1. By the vanishing condition εSαpw1�
. . .�wkq,αpw1q,αpw2 � . . .�wkq vanishes. Thus, at least one of the following holds,
(i) αpw1 � . . .�wkq � αpw2 � . . .�wkq �αpw1q,
(ii) εSαpw2, . . . , wkq � εSαpw1 � . . .�wkq �αpw1q,
(iii) εSαpw1q � εSαpw1 � . . .�wkq �αpw2 � . . .�wkq.
We show that all cases lead to a contradiction.
(i) Immediate.
(ii) Adding αpw1q to both sides gives a contradiction.
(iii) Adding αpw2 � . . .�wkq to both sides gives a contradiction.
Hence, the maximal size is 1 and the expression αpv1q � . . .�αpvnq vanishes. l
Example. A potentially confusing issue is that the axioms defining a Gondran-Minoux
pseudomorphism are not preserved under common operations on maps. For example,
let K be a field and ω : K Ñ B a surjective valuation onto the trivial semifield B �
56 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
t0,8u. Then ω is a Gondran-Minoux pseudomorphism between K and B. However,
the coefficient-wise application ofω is not a Gondran-Minoux pseudomorphism K3 Ñ
B3. Set v � pa, b, 0q and w � p�a, 0, cq, with a, b, c P K�. The images of v, w and
v � w in B3 are p0, 0,8q, p0,8, 0q and p8, 0, 0q, whose sum does not vanish in the
sense of Gondran-Minoux. Note however that it does vanish tropically and that the
coefficient-wise application of ω is a tropical pseudomorphism.
According to Proposition 3.3.2 above the Gondran-Minoux pseudomorphisms
have the desirable property that they form the largest class of maps that respect the
notion of Gondran-Minoux vanishing in the entire category of modules over a semi-
ring. However, by the example, such maps are not suited for the purpose of tropical
geometry, since the coefficient-wise application of an additive valuation is not in the
class. The solution to this dilemma is to restrict to a full subcategory of modules over
a ring, in which the strictly larger class of tropical pseudomorphisms shares some of
the desirable properties of that of the Gondran-Minoux pseudomorphisms. The next
lemma is essential in proving that this works.
Lemma 3.3.3 Let α : M Ñ N be a tropical pseudomorphism and suppose N is separatedby N�. If any of the following conditions hold,
(i) N is a module over a ring,(ii) both M and N are modules over semirings with idempotent addition,
then α is an additive map. Moreover, if (i) holds and 2 � imα � t0Nu, then M is also amodule over a ring.
Proof. We prove the claims in the order stated. Let v, w P M .
(i) For each φ P N� the sum �φpαpv � wqq �φpαpvqq �φpαpwqq vanishes. This
implies that φpαpv � wqq � φpαpvq � αpwqq, by linearity of φ. Since the canonical
morphism N Ñ N�� is injective it follows that αpv�wq � αpvq �αpwq.(ii) Let φ P N�. Denote the composition φ � α by ψ. Then ψ : M Ñ S is a
pseudomorphism, where S is the semiring of scalars of N . By the arguments in the
previous item, it is sufficient to prove that ψpv � wq � ψpvq �ψpwq. By the pseudo-
morphism axioms, the sumψpv�wq�ψpvq�ψpwq vanishes. This implies that either
ψpv � wq � ψpvq �ψpwq, ψpvq � ψpv � wq �ψpwq, or ψpwq � ψpv � wq �ψpvq.By symmetry, it is sufficient to consider the case ψpvq � ψpv � wq �ψpwq. By idem-
potency of the addition in S the equality ψpvq�ψpwq �ψpv�wq�ψpwq and hence
ψpv�wq ¥ψpvq �ψpwq.For the other inequality, consider the image of v � w � w under ψ. By the pseu-
domorphism axioms, the sum ψpv � w � wq � ψpv � wq � ψpwq vanishes. Thus,
ψpwq � ψpv � wq, or ψpv � wq � ψpwq � ψpv � wq. In terms of the natural or-
der, ψpv � wq ¤ ψpwq. The inequality, ψpv � wq ¤ ψpvq follows by symmetry. Thus,
ψpv�wq ¤ψpvq �ψpwq, and the statement follows.
3.3. RANKS AND VANISHING CONDITIONS 57
Denote the semiring of scalars of M by R. Let v P M such that αpvq � αpvq � 0.
Suppose that R is not a ring. Then εR � 1 and hence αpvq � �αpvq. This contradicts
the choice of v, and hence R is a ring. l
The different notions of vanishing lead to different notions of rank as well. Con-
sider a finite sequence v1, . . . , vn of elements of M . Such a sequence is said to be
Gondran-Minoux, tropically, or weakly dependent if there exist scalars x1, . . . , xn P R,
not all equal to 0, such that the sequence x1 � v1, . . . , xn � vn vanishes in the sense of
Gondran-Minoux, tropically, or weakly, respectively. It is said to be independent if it is
not dependent. If S � M is some subset, then the Gondran-Minoux, tropical, or weak
rank of this set is the maximal size of a subset of V that is independent in the sense of
Gondran-Minoux, tropically, or weakly. For a more complete overview of the different
notions of rank we refer to reader to [1]. An interesting paper on the notion of weak
rank in specific classes of modules over idempotent semirings is [68]. Note that what
we refer to as weak dependence is referred to as dependence in that paper.
For us the notion of tropical rank is of most importance, since it is most strongly
connected to tropical linear spaces. Although the tropical rank of a matrix was already
defined and studied in the paper [18] it was first connected to a notion of dependence
in [35], albeit in the context of supertropical algebra. Tropical rank is most useful for
modules whose semiring of scalars has linear induced order and we postpone further
discussion of tropical rank to the end of the next section.
The weak rank of a set of vectors is most important when restricted to subsets
of modules over semifields. In that case it is strongly connected to the generating di-
mension of the module, i.e. the minimal cardinality of a spanning subset. The theorem
relating generating dimension to weak dependence it best stated in terms of a concept
we borrow from lattice theory. Let M be a module over an idempotent semifield F and
u P N � M . Then u is said to be reducible in M if there exist distinct v, w P Mztuusuch that u� v�w, and it is said to be irreducible otherwise. The following statement
and their proofs can be found in [68] and are quite well known for finitely generated
modules over the tropical real numbers (cf. [19, 27]). We reformulate the relevant
statements in [68].
Theorem 3.3.4 Let S � M be a finite set with M a module over an Archimedean idem-potent semifield. Then S is weakly independent if and only if every u P S is irreducible inthe module generated by S.
Proof. This is Proposition 2.5.3 of [68]. l
A basis of a module M is a weakly independent generating set for M . The previous
theorem has the next statement as an immediate corollary.
Corollary 3.3.5 Let M be a finitely generated module over an Archimedean idempotent
58 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
semifield. Then it has a unique basis up to scalar multiples.
Proof. This is Theorem 5 of [68]. l
3.4 Modules over linearly ordered semirings
The next lemma is particular for the case of modules with a linear order, and states
an essential non-trivial property that holds for such modules.
Lemma 3.4.1 Let M be a module with linear canonical order. Let k and n be naturalnumbers. For each i P t1, . . . , ku, let σi be a Gondran-Minoux vanishing sequence oflength n with elements in M. Then, the sequence σ1 � . . .� σk is a Gondran-Minouxvanishing sequence as well.
Proof. Set τ� σ1�. . .�σk and let j P t1, . . . , nu be the index of a minimal element of
τ. There exists i such that j is the index of a minimal element of σi . By the character-
ization of vanishing sequences in modules with induced linear order there is another
index k P t1, . . . , nu such that pσiq j � pσiqk. Thus, τ j ¤ τk ¤ pσiqk � pσiq j � τ j and
τ j � τk are both minimal entries of τ. l
The above statement is the main reason to restrict to modules over semirings Rhaving linear induced order. It states equivalently that the collection of vanishing
sequences of fixed length n with elements in R is a module over R. The implication
is that the analogue of linear tropical prevarieties over a semiring are modules over
this semiring and hence that they behave much like actual tropical prevarieties and
tropical convex sets (cf. [19]). In particular, the statement is an essential ingredient
in the proof that the tropical linear spaces of Chapter 4 are both finitely generated
modules and tropical prevarieties.
Remark. The condition that M has linear canonical order seems to be necessary in the
above statement. The essential obstruction to proving the statement for semiring with
non-linear canonical order seems to be that the vanishing partitions of the different
sequences pσiq may differ. If this were not the case, then the sequence τ in M k of
length n that is defined by τ j � ppσ1q j , . . . , pσkq jq is a vanishing sequence. Since the
map M k Ñ M given by pv1, . . . , vkq ÞÑ v1 � . . .� vk is a Gondran-Minoux pseudomor-
phism the entry-wise image of τ under this map would vanish as well, by Proposition
3.3.2. This image is precisely the sum σ1 � . . .�σk.
Example. This an example of Gondran-Minoux vanishing sequences whose sum is not
a Gondran-Minoux vanishing sequence, when the induced order on M is not linear.
3.4. MODULES OVER LINEARLY ORDERED SEMIRINGS 59
Let M be the module B4. Define the Gondran-Minoux vanishing sequences σ and
τ by,
σ ��p8,8, 0, 0q, p8,8,8, 0q, p8,8, 0,8q
�τ �
�p0,8,8,8q, p0, 0,8,8q, p8, 0,8,8q
�.
The sum σ�τ is the sequence,
σ�τ��p0,8, 0, 0q, p0, 0,8, 0q, p8, 0, 0,8q
�.
Thus, σ�τ is not a Gondran-Minoux vanishing sequence. This shows that the condi-
tion that the canonical order on M is linear is not superfluous.
Remark. There remains the question whether there is a natural class of modules on
which the implication of Lemma 3.4.1 is an equivalence. A natural such condition
might be stated in the following way. Let M be a module with idempotent addition
and denote by k the minimal number such that for every v1, . . . , vn P M there exists
a subset S � rns of size k such that°n
i�1 vi �°
jPS v j . For example, M has linear
induced order precisely when k � 1.
The above example shows that any module into which B4 embeds has vanishing
sequences whose sum is not a vanishing sequence. Modules over idempotent semir-
ings which have k ¥ 2 always contain a B-submodule isomorphic to B2. Hence, to
prove equivalence in Lemma 3.4.1 it would be sufficient to find two vanishing se-
quence of equal length, whose elements are in B2 and whose sum does not vanish.
The next corollary is important since it shows that for a sequence to tropically
vanish, one needs only to check vanishing of the image sequence under a finite num-
ber of linear functionals, if the algebraic dual is finitely generated. In particular, any
module over a ring whose induced order is linear that is a submodule of a finitely
generated free module has a finitely generated algebraic dual, by Lemma 3.2.1. In
particular, tropical convex subsets of Rn8 and tropical linear spaces are of this type.
Corollary 3.4.2 Let R be a semiring with linear order and M a module over R whosealgebraic dual is generated by some finite set Γ. A finite sequence pv1, v2, . . . , vkq vanishestropically if and only if for each γ P Γ the sequence pγpv1q,γpv2q, . . . ,γpvkqq vanishes inR.
Proof. Let φ P M�. We show that pφpviqq vanishes. Write φ �°γPΓµγ � γ with
µγ P R. The sequence pφpviqq is the sum of the sequences µγ � pγpviqq, all of which are
vanishing sequences in R. By the lemma, pφpviqq is a vanishing sequence R. l
60 CHAPTER 3. ALGEBRA OF IDEMPOTENT SEMIRINGS
The corollary states in particular that a finite subset tv1, . . . , vku � Rn is tropically
dependent if there exist α1, . . . ,αk P R such that the minimum in the expression α1 �
v1i � . . .� αk � vki is attained at least twice for i P rns, where vi j is the j-th entry of
the vector vi . Thus, for free modules over linearly ordered semirings tropical rank is
the same as the notion in [35]. Moreover, our notion of vanishing corresponds to the
notion of summing to a ghost element. Hence, we have the following theorem.
Theorem 3.4.3 Let V � Rn be a finite set of vectors. The tropical rank of V is the sizeof the largest non-vanishing minor of the matrix whose rows are the elements of V .
Proof. This is Corollary 3.14. of [35]. l
3.5 Conclusion
Tropical linear spaces over R are in their most general linear prevarieties in a free
module over R of finite rank. The aim of this chapter was to restrict the choice of Rsuch as to have a comfortable setting in which to think about tropical linear spaces. A
number of properties are convenient to have.
(i) Duality: tropical linear spaces are finitely generated modules.
(ii) Tropical rank: there is a meaningful notion of tropical rank.
(iii) Residuation: homomorphisms between tropical linear spaces are residuated
mappings.
(iv) Lattice: the induced order on a tropical linear spaces is a lattice order.
(v) Unique basis: tropical linear spaces (as modules) have an essentially unique
basis.
All of these conditions are fulfilled if R is a tropical semifield. If R is a residuated idem-
potent semifield whose induced order is linear (or if R is an idempotent semifield with
induced linear order) then all conditions except possibly the last one are fulfilled.
Chapter 4
A category of tropical linear
spaces
Inspired by Figure 4.1(a) from talks by Mikhalkin (see [39, Paragraph 12]), in which
three points on a tropical projective line are brought into special position by a se-
quence of modifications and projections (except the last map, which is an invertible
translation corresponding to an invertible diagonal tropical matrix), we aim to define
and study tropical combinatorial analogues of affine algebraic groups. A first attempt
would be to define these as certain groups of invertible tropical matrices. However, it
is easy to see that all invertible tropical matrices are in fact monomial, and the process
in Figure 4.1(a) cannot be captured by a single monomial matrix (though the last step
can).
(a) Moving three points to standard position onthe projective line.
(b) A morphism from thetropical projective line toitself.
Figure 4.1: Moving three points, stepwise or at once.
The composition of the morphisms in Figure 4.1(a) is depicted in Figure 4.1(b):
it is a metric tree with two isometric embeddings of the tropical projective line
61
62 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
1 23
a b ∞c d ∞∞ ∞ a� d
a� d < b� c
56
a b ∞c d ∞∞ ∞ b� c
a� d > b� c
4
89
a b ∞c d ∞∞ ∞ e
a� d = c� b < e
a ∞ ∞∞ d ∞∞ ∞ a� d
∞ b ∞c ∞ ∞∞ ∞ b� c
7 10
a b ∞c d ∞∞ ∞ e
a� d = b� c = e
a ∞ ∞b d ∞∞ ∞ a� d
a b ∞∞ d ∞∞ ∞ a� d
∞ b ∞c d ∞∞ ∞ b� c
a b ∞c ∞ ∞∞ ∞ b� c
12
3
4
5
6
78
9
10
Figure 4.2: TPGL2: polyhedral complex and matrices. The dashed lines and the dashed(white) vertices are not really there, so that, for instance, face 3 has the topology of ahalf-plane.
r�8,8s—one for the source (solid) and one for the target (dashed). Moreover, the
tree does not have ends that are superfluous in the sense that they are covered by
neither the source nor the target. Such pairs of embeddings form a pure polyhedral
complex of dimension three (see Figure 4) which is a monoid with a natural composi-
tion: embed everything into a larger tree, and delete superfluous ends. Therefore, it is
reasonable to call this complex TPGL2, the tropical projective general linear group of
the line. Incidentally, it is isomorphic, as a polyhedral complex with monoid structure,
to the tropicalization of the variety GL2 as sitting inside 3� 3-matrices in the usual
way (modulo scalars, to be precise). In higher dimensions, we still have a monoid
homomorphism from TGLn to the tropicalization of GLn, but it is not injective.
Our aim for the remainder of this section is to give an appropriate construction
that mimics Figures 4.1 and 4.2 for tropical linear spaces other than R28. A clue on
how this is to be done is given by the observation that generic morphisms from the
tropical projective line to itself consist of a pair of embeddings into a three-valent trees
on four leaves. The latter are common objects in the tropical geometry literature. They
are projectivizations of generic 2-dimensional tropical subspaces of R48.
Our approach is to define a category, whose objects include Rn8, and define the
tropical linear group of rank n to be some subgroup, or submonoid, of the endomor-
phisms of Rn8 in this specific category. This is the program that we attempt to carry
out in this chapter.
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 63
The category is constructed in terms of general tropical linear spaces, of which Rn8
is a specific instance. Tropical linear spaces are modules with a combinatorial struc-
ture that abstract the properties of tropicalizations of vector spaces over valued fields,
in the same way that matroids abstract properties of linear independence. They were
implicitly introduced in [23] and put into a tropical context in [59]. It is inconvenient
that tropical linear spaces have no intrinsic definition. The definition of a tropical lin-
ear space is through an intermediate object called a valuated matroid. Section 4.1 is
devoted to gathering and reformulating known results on tropical linear spaces and
valuated matroids. A difference between our approach and the more common one in
the tropical geometry literature is that we tend to understand tropical linear spaces as
finitely generated submodules, instead of as tropical prevarieties, and that we develop
the theory over arbitrary tropical semifields. The latter does not yield new results, but
makes the connection to the theory of matroids cleaner.
A morphism between tropical linear spaces L andM is defined to be a tropical
linear subspace of L �M that is obtainable from L by iteratively taking graphs oflinear functions. In Figure 4.1(b) the morphism is the tree with four leaves in the
middle. It is a twofold linear extension of the tropical projective line, with each linear
extension glueing a branch onto the projective line. The concept of a general linear
extension is developed in Section 4.2. The actual category and TGLn are defined in
Section 4.3.
4.1 Valuated matroids and tropical linear spaces
This chapter serves to gather results on tropical linear spaces and valuated matroids
from various sources, and to restate and prove them in a unified language. All of the
results in the chapter are known in some way.
In Section 4.1.1 we start by defining valuated matroids. These are best seen as
functions that assign a finite element of a tropical semifield to each basis of a matroid.
They include the concept of an ordinary matroid by taking t0,8u as the tropical semi-
field. To each valuated matroid we associate a finitely generated semimodule called a
tropical linear space. This is not the ordinary definition of a tropical linear space seen
for example in [59], but is more convenient for our purposes. It is also more fitting
with the algebraic approach to tropical linear spaces we adopt here. The remainder of
Section 4.1.1 is devoted to stating and proving some important properties of tropical
linear spaces and valuated matroids.
In Section 4.1.2 we give constructions to make valuated matroids from other ones.
The main purpose is to understand which operations on tropical linear spaces yield
tropical linear spaces, and which operations are induced on the valuated matroids to
which the tropical linear spaces are associated. For example, the Cartesian product of
64 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
two tropical linear is a tropical linear space, as is the coordinate projection.
In the last short Section 4.1.3 we deal with parameter spaces of tropical linear
spaces.
4.1.1 Foundations
Valuated matroids
The main objects under consideration are introduced in [23] under the name valuatedmatroids. They were put into a tropical context in [59]. To emphasize the connection
with matroids and the dependence on the tropical semifield F we use the names
matroid over F , F-valuated matroid, or even F-matroid, for shortness.
Definition 4.1.1 Let R be a finite set, F a tropical semifield and d P N. A valuatedmatroid on R over F of rank d is a function µ : PdpRq Ñ F that satisfies the valuatedexchange condition for all X , Y P PdpRq, i.e. for all x P X zY there exists y P Y zX such
that
µpX q dµpY q ¥ µpX � x � yq dµpY � x � yq. (4.1)
Valuated matroid µ and ν on R over F are called proportional if they are scalar mul-
tiples of each other, i.e. if there exists λ P F� such that µpX q � λ d νpX q for all
X � R.
Proposition 4.1.1 (Immediate properties) Let µ be a valuated matroid on R over F.(i) For h : F Ñ G a homomorphism of semifields, the composition h �µ is a valuated
matroid on R over G.(ii) The set tA� R | µpAq � 8u forms the collection of bases of an ordinary matroid.
(iii) The set tA � R | µpAq is minimalu forms the collection of bases of an ordinarymatroid.
Proof. We prove the claims in the order in which they are stated. All proofs are
elementary, which we give in full for the sake of completeness.
(i) Let A and B be subsets of R of size d. Let a P AzB. There exists b P BzA such
that µpAq d µpBq ¥ µpA� a � bq d µpB � a � bq. Since h is increasing and a
multiplicative homomorphism, hpµpAqqdhpµpBqq ¥ hpµpA� a� bqqdhpµpB�a� bqq.
(ii) Let A and B such that µpAq � 8 and µpBq � 8. Let a P AzB. There exists
b P BzA such that µpA� a� bq d µpB � a� bq � 8. Thus, µpA� a� bq � 8
and µpB� a� bq � 8, since F has no zero divisors.
(iii) Let A and B be such that both µpAq and µpBq are minimal. Then for each a P AzBthere exists b P BzA such that µpA� a� bqdµpB� a� bq ¤ µpAqdµpBq. Thus,
µpA� a� bq and µpB� a� bq are minimal.
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 65
l
The second and third item in the above proposition give natural ways of asso-
ciating ordinary matroids to valuated matroids. These are made explicit in the next
definition.
Definition 4.1.2 Let µ be a valuated matroid on R over F . The underlying ma-troid of µ is the matroid on R that has tX � R | µpX q � 8u as its collection
of bases. The residual matroid or minimal matroid is the matroid on R that has
tX � R | µ is minimal at Xu as its collection of bases. We write matpµq for the un-
derlying matroid, and respµq for the residual matroid of µ.
Proposition 4.1.2 The map µ ÞÑ matpµq establishes a bijection between valuated ma-troids over B and ordinary matroids.
Proof. This is a trivial consequence of part (ii) of Proposition 4.1.1 and the fact that
µpAq � 8 implies µpAq � 0. l
Remark. If we reinterpret matroids as valuated matroids over B, which is valid by
the above proposition, then the map µ ÞÑ matpµq is induced in the sense of part (i)
of Proposition 4.1.1 by the homomorphism F ÞÑ B that maps 8 to 8 and everything
else to 0.
The underlying matroid is used to transfer matroid terminology to valuated ma-
troids. We shall for example speak of the rank of the valuated matroid to mean the
rank of the underlying matroid. Similarly, the bases of a valuated matroid are the
bases of the underlying matroid, i.e. the collection of sets on which the valuated ma-
troid is finite.
An important class of valuated matroids is formed by the realizable valuated ma-troids. The class depends on a field K with a valuation ω : K Ñ F . Let A be a d � n-
matrix where d ¤ n and suppose that A has maximal rank. For each subset J of
t1, . . . , nu we write ArJs for the submatrix of A consisting of the columns indexed by
J . Then the function J ÞÑ ωpdetpArJsqq is a valuated matroid; the valuated matroid
associated to A. A valuated matroid realizable over K , or pK ,ωq if we wish to be pre-
cise, is a valuated matroid that arises in this way. This example hints at the relation
between tropicalized linear spaces and valuated matroids, which we discuss later.
Before moving on to tropical linear spaces, we state and prove a number of alter-
native characterizations of valuated matroids, that put them more clearly in a tropical
context, and relate them to subdivisions of matroid basis polytopes. Note that part
(iv) is essentially the definition found in [59].For brevity, we write Si j for the set SYti, ju andBpµq denotes the set of bases of
the valuated matroid, i.e. the collection of size d sets whose image under µ is finite.
66 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
Theorem 4.1.3 (Speyer, Murota) Let R be a finite set and F a tropical semifield, andµ :P pRq Ñ F a function. Then the following statements are equivalent,
(i) µ is a valuated matroid of rank d,(ii) µ satisfies the exchange condition (4.1) for all subsets A, B � R of size d such that
|AzB| � 2,(iii) the minimum of the expression,
àiPY zX
µpY � iq dµpX � iq (4.2)
is attained at least twice, for all X , Y � R satisfying |Y | � d � 1 and |X | � d � 1,(iv) the minimum of (4.2) is attained at least twice, for all X , Y � R satisfying |Y | �
d � 1, |X | � d � 1 and |Y zX | � 3,(v) for each S � R with |S| � d � 2 and i, j, k, l P RzS the minimum of�
µpSi jq dµpSklq`�µpSikq dµpS jlq
`�µpSilq dµpSk jq
is attained at least twice.(vi) the convex hull of tχB | µpBq � 8u is a matroid basis polytope in the sense of [9]
and the assignment χB ÞÑ µpBq P F� � R of values to the vertices of this polytopeinduces a subdivision all of whose faces are again matroid polytopes.
Proof. We establish this theorem by first proving the equivalence of (i) and (ii) and
the equivalence of (i) and (iii). The equivalence of (ii) and (iv) is then a byproduct of
the latter proof. Part (v) is simply a restatement of (iv), with S � X X Y . Finally, we
prove the equivalence of (iv) and (vi).
We start with the equivalence of (i) and (ii). It is trivial that (i) implies (ii). That
(ii) also implies (i) is more complicated. We follow the proof by contradiction in [48].Assume without loss of generality that F contains at least three elements. Otherwise,
extend F .
Suppose that µ satisfies the basis exchange condition for all pairs pA, Bq satisfying
|A| � |B| � d and |A∆ B| � 4 and that there exist pairs which do not satisfy the basis
exchange condition. Let pA�, B�q be such a pair, chosen such that |A� ∆ B�| ¡ 4 is
minimal. The aim is to arrive at a contradiction by constructing a pair not satisfying
the basis exchange condition whose sets have a symmetric difference that has strictly
smaller size. In fact, it is sufficient to show that this pair does not satisfy the basis
exchange condition for τ �µ, for some τ P FR�, with,
pτ �µqpBq ��ä
iPBτi
dµpBq, B PPdpRq,
since τ �µ is a valuated matroid if µ is.
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 67
Let a� P A�zB� be such that for all b P B�zA� the strict inequality µpA�qdµpB�q µpA�� a�� bqdµpB�� a�� bq holds. We choose τ P FR
� and set µ1 � τ �µ such that
µ1pA�q � µ1pA��a��iqwhen i P B�zA� and A��a��i PBpµq, and µ1pB��a��iq ¡µ1pB�q if i P B�zA�. For example, we might choose,
τi �
$'&'%µpA�q �µpA�� a�� iq, i P B�zA� and A�� a�� i PBpµq,µpB�� a�� iq �µpB�q � ε, i P B�zA� and A�� a�� i RBpµq,0 otherwise.
,
for some ε ¡ 0.
Let a0 P A�zB� such that a0 � a�. Choose b0 P B�zA� such that µ1pB� � a0 � bqis minimal for b � b0 and set B0 � B� � a0 � b0. It is sufficient to show that µ1pB0 �
a�� bq ¡ µpB0q for all b P B�zA�, since then
µ1pA�q dµ1pB0q � µ1pA�� a�q dµ1pB0q
µ1pA�� a�� bq dµ1pB0 � a�� bq,
for all b P B0zA� � B�zA� and |A� ∆ B�| ¡ |A� ∆ B0|.
Let b P B�zA� and consider µ1pB0 � a�� bq �µ1pB0q,
µ1pB0 � a�� bq �µ1pB0q � µ1pB�� a�� a0 � b� b0q �µ1pB�� a0 � b0q
¥�µ1pB�� a0 � b0q dµ
1pB�� a�� bq
`�µ1pB�� a0 � bq dµ1pB�� a�� b0q
�µ1pB�q
�µ1pB�� a0 � b0q
¡ 0, (4.3)
where the first inequality follows by applying the exchange condition to the pair pB��a�� a0 � b� b0, B�q and a� R B�. This yields that,
µ1pB�� a�� a0 � b� b0q dµ1pB�q
¥�µ1pB�� a0 � b0q dµ
1pB�� a�� bq�`�
µ1pB�� a0 � bq dµ1pB�� a�� b0q�
The positivity of expression (4.3) is a consequence of the choice of b0 and the choice
of τ.
The proof of the equivalence of (i) and (iii) is elementary. To see that (i) implies
(iii), let X and Y be subsets of R satisfying |X | � d�1 and Y � d�1. Suppose that the
minimum in (4.2) in attained at i�. Then, by the valuated exchange condition stated
68 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
in (4.1) applied to X � i�, Y � i� and i�, there exists j� P Y zX with j� � i� such that
µpX � i�q d µpY � i�q ¥ µpX � j�q d µpY � j�q. By choice of i� right and left hand
side of the inequality are equal, showing that the minimum is also attained at j�.
For the converse, (iii) implies (i), let A, B � R such that |A| � |B| � d, and a P AzB.
Consider (4.2) with X � A� a and Y � B � a. Let b P BzA such that µpA� a� bq dµpB� a� bq is minimal among the terms in (4.2). Such b exists by the premise that
the minimum is attained twice. Then µpAq dµpBq ¥ µpA� a� bq dµpB� a� bq.
We obtain the equivalence of (ii) and (iv) by restricting this proof to the collection
of pairs of subsets described parts (ii) and (iv) of the statement.
l
All characterizations above are analogous to various axiomatizations of matroid
in terms of their bases. Items (i)-(iv) differ in the scope and representation of the
exchange axiom. Characterization (vi) is of a different nature. It is perhaps best seen
in light of Gale’s description of matroid basis sets in terms of unique maxima [26]. A
number of different axiomatizations of valuated matroids will appear later in this text,
the first of which is strongly connected to certain submodules associated to valuated
matroids called tropical linear spaces. It appears in the end of the following section.
The characterization (vi) associates combinatorial data to a valuated matroid in
terms of a polyhedral subdivision. This polyhedral subdivision is referred to as the
combinatorial type of the valuated matroid.
Tropical linear spaces
The main reason for studying valuated matroids is that they parameterize nice tropical
prevarieties, called tropical linear spaces. Let µ be a valuated matroid on R over F .
The tropical linear space associated to µ will be defined as a certain finitely generated
subsemimodule of FR. The reason for introducing it this way instead of as a tropical
prevariety is twofold. First, it makes the relation between tropical linear spaces (for
valuated matroids) and lattices of flats (for matroids) clearer from the outset, and
secondly, it allows us to define the prevariety structure using the dual matroid. This
will be made more clear later. It is also more in line with the algebraically flavored
approach to tropical linear spaces that is preferred in this thesis. See also [49].
Let I � R be an independent set of µ of corank 1 and define vµI P FR to be the
point with i-th coordinate µpI � iq, when i R I , or 8, when i P I . When the reference
to the valuated matroid is clear from context we simply write vI . The next lemma is
about minimal sets of generators for the F -module generated by the vectors vI .
Lemma 4.1.4 Let I be a corank 1 independent set of matpµq and let J be a collectionof such sets. The following statements hold:
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 69
(i) the point vI is in the tropical linear hull of tvJ | J P J u if and only if the hyperplaneof µ spanned by I equals the hyperplane of µ spanned by J, for some J P J ,
(ii) the support of vI is RzH, where H is the hyperplane spanned by I in matpµq.
Proof. Part (ii) of the statement is immediate: Let I be an independent corank 1 set in
matpµq. Then i P clpIq is equivalent to I� i is not a basis of matpµq. Thus, µpI� iq � 8
if and only if i P clpIq.
For part (i), suppose that there exists J P J such that clpIq � clpJq. By part (i),
vI and vJ have equal support. If the size of the support is at most 1 the implication is
trivially true. Suppose that it is at least two. Let i, j P supppvIq � supppvJq. Consider
(4.2) with Y � I Yti, ju and X � J . If k P I , then k P clpIq and hence k P clpJq. Thus,
the expression reduces to, µpI� iqdµpJ� jq`µpI� jqdµpJ� iq. Since the minimum
is attained twice, µpI � iq dµpJ � jq � µpI � jq dµpJ � iq, which shows that vI and
vJ are proportional.
Suppose that for each J P J there exists λJ P F such that,
vI �àJPJλJ d vJ .
Let J be such that clpJq � clpIq. Since any two hyperplanes are incomparable in �-
order, there must exist i P clpIqz clpJq. Thus, J� i is a basis, and I� i isn’t and λJ �8.
Thus, vI is already in the module generated by all vJ with clpJq � clpIq and J P J . In
particular, there exists J P J such that clpIq � clpJq. l
By the above lemma, any minimal set of generators consists of a set of vectors
tvI | I P J u subject to the condition that every hyperplane of µ is the closure of
exactly one I P J . To avoid a choice for a specific set of generators we associate to
every hyperplane H the point vH defined as the sum of the vI , where I runs over all
corank 1 subsets that span H. By the lemma, it holds that vH � vI for some I . This
allows for a clean definition of the tropical linear space as a tropical linear hull.
Definition 4.1.3 Let µ be a valuated matroid on R over F . The tropical linear space of
µ is the underlying set of the F -module generated by tvH | H is a hyperplane of µu. It
is denoted Lµ.
Example. Let n P N and µ a valuated matroid of rank 2 over R8, such that µpBq 8
for each B P P2rns. Then the underlying matroid of µ is the uniform matroid of rank
2 on n. The hyperplanes of the latter matroid are exactly the singletons tiu and the
corresponding generator is
vtiu � pµpti, 1uq, . . . ,µti� 1, iu,8,µti, i� 1u, . . . ,µti, nuq.
70 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
The tropical linear spaceLµ has topological dimension two (we will prove this later),
and hence its projectivization has topological dimension 1. This projectivization is a
tree on n leaves, with the i-th leave being the projectivization of vtiu.
The next proposition makes clear that the tropical linear space associated to the
valuated matroid is an analogue of the lattice of flats of the matroid. It shows in fact
that if F � B, or equivalently, when µ is an ordinary matroid the tropical linear space
is isomorphic the lattice of flats of µ.
Proposition 4.1.5 (Immediate properties) (i) The tropical linear space Lµ has thestructure of a lattice under the partial order restricted from FR. (ii) The function thatassigns Rz supppvq to each point v P FR is a lattice morphism that maps Lµ surjectivelyonto the lattice of flats of the underlying matroid of µ.
Proof. Part (i) is an immediate consequence of the fact that a finitely generated sub-
module of FR is a lattice (cf. Proposition 3.1.6). Part (ii) follows from Lemma 4.1.4
and the fact that supppv`wq � supppvq X supppwq. l
The following is a characterization of valuated matroids in terms of tropical de-
pendence (cf. Section 3.3).
Lemma 4.1.6 Let µ : PdpRq Ñ F be a function such that the set tA P PdpRq | µpAq �8u forms the set of bases of a matroid M. Then µ is a valuated matroid if and only if foreach triple I , J , K of independent (in M) sets of size d�1 in such that I X J XK has sized � 2 the set tvI , vJ , vKu is tropically dependent.
Proof. Suppose µ is a valuated matroid and set S � IXJXK . Then |S| � d�2, where
d � rkµ. Write IzS � tiu, JzS � t ju and KzS � tku. We need to show that there
exists x I , xJ , xK P F such that
px I dµpSilqq ` pxJ dµpS jlqq ` pxK dµpSklqq, (4.4)
vanishes tropically for every l R S. Set x I � µpS jkq, xJ � µpSikq and xK � µpSi jq.The expression vanishes by characterization (v) of Theorem 4.1.3.
Conversely, let S � R with |S| � d � 2 and i, j, k, l P RzS distinct. Set I � Si,J � S j and K � Sk. By the condition on the tropical rank there exist x I , xJ , xK P Fsuch that the expression px I d vIq ` pxJ d vJq ` pxK d vKq vanishes tropically, or in
other words, that the minimum of (4.4) is attained at least twice. Without loss of
generality, suppose that x I � µpS jkq. Then, by the expression for l � k it follows that
xJ � µpSikq. Setting l � j gives xK � µpSi jq. Hence, µ satisfies (v) of Theorem 4.1.3.
l
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 71
Examples. We end this section with a number of typical examples of tropical linear
spaces.
(i) Let R be a finite set. Then FR is a tropical linear space. The proportional equiva-
lence class of valuated matroids on R is that of the free valuated matroids on R,
i.e. those valuated matroids on R of rank |R|. The hyperplanes of this matroid
are all sets of the form Rztiu, and vRztiu equals ei , i.e. the vector all of whose
entries are 8, except the i-th, which is 0.
All tropical linear spaces are by definition contained in some free F -module.
Since they are also invariant under tropical scalar multiplication no information
is lost by drawing the projectivizations, instead of the actual tropical linear
spaces. The picture one should have in mind of FR is the |R| � 1 dimensional
simplex (at least if F � R8, a subset thereof if F is strictly contained in R8).
The reason is that PpRR8q is homeomorphic to the |R| � 1-dimensional simplex.
Distances, however, are of course distorted. To clarify matters we have drawn
R48 and B3 in Figures 4.3(a) and 4.3(b).
(∞,∞, 0,∞)
(∞, 0,∞,∞)
(0,∞,∞,∞)
(∞,∞,∞.0
(a) (Projectivization of) four dimensionaltropical affine space over R
8.
(0, 0, 0)
(0, 0,∞)
(0,∞, 0)(∞, 0, 0)
(∞,∞, 0)
(0,∞,∞)(∞, 0,∞)
(b) (Projectivization of) three dimen-sional tropical affine space over B.
Figure 4.3: Tropical affine spaces.
(ii) Any geometric lattice embedded as a meet-subsemilattice in the full power set
lattice of a finite set is a tropical linear space. The full power set lattice is the
ambient tropical affine space. Equivalently, the underlying set of the Bergman
fan of a matroid is the intersection of a tropical linear space in Rn8 with Rn [2].
(iii) More specific examples of (projectivizations of) tropical linear spaces are given
after the prevariety structure on a tropical linear space is dealt with, in Section
4.1.2
4.1.2 Basic constructions and relations
This section deals with various ways of constructing new valuated matroids from a
set of given ones: orthogonal dual, restriction, contraction, direct sum, matroid union
72 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
and matroid intersection. A number of these constructions have definitions which are
more easily stated in terms of an equivalent, or cryptomorphic characterization of val-
uated matroids in terms of a function on independent sets of the underlying matroid.
For valuated matroid this characterization is stated and proved to be cryptomorphic
by Murota in [48]. We follow his exposition.
Let R be a finite set and ν : P pRq Ñ F a function. Then ν is said to be a valuatedmatroid on independent sets if the following conditions are satisfied.
(i) ν is not identically 8,
(ii) ν is increasing, i.e. νpIq ¤ νpJq if I � J ,
(iii) if I is a proper subset of J and νpJq � 8 then there exists i P JzI such that
νpJq � νpI � vq,(iv) if |I | � |J | � 1, there exists v P JzI such that νpIqd νpJq ¥ νpI � vqd νpJ � vq.
The collection tI � R | νpIq � 8u form the collection of independent sets of an (or-
dinary) matroid on R and hence the maximal sets of this collection are equicardinal,
of, say, cardinality d. The translation from valuated matroids on independent sets
to the valuated matroids of Definition 4.1.1 is by restricting to PdpRq. The inverse
translation is µ ÞÑ µ, where,
µpIq �à
BPBpµqI�B
µpBq,
and Bpµq is the set of basis of matpµq. Definition 4.1.1 and the independent set
characterization are cryptomorphic in the sense that the restricting to PdpRq subsets
with finite image and µ ÞÑ µ are inverse bijections between the valuated matroids on
independent sets and the valuated matroids of Definition 4.1.1.
Truncation
The first construction of this section is strongly related to the independent set axiom.
One can wonder whether a valuated matroid on independent sets, restricted to sets
of a certain cardinality is a valuated matroid, if it is not identically 8. The following
lemma, combined with an easy induction show that this is indeed the case.
Lemma 4.1.7 Let µ : PdpRq Ñ F be a valuated matroid. The map Pd�1pRq Ñ Fdefined by,
A ÞÑàA�B
BPBpµq
µpBq, A PPd�1pRq,
is a valuated matroid.
Proof. For a proof the reader is referred to [48]. l
The matroid in the lemma is referred to as the truncation of µ and denoted Tpµq.It is the valuated analogue of the truncation of an ordinary matroids (cf.[11, Section
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 73
7.4]).
Orthogonal dual
The characterization of the tropical linear space as a tropical prevariety, i.e. an inter-
section of tropical hyperplanes is conveniently stated in terms of the orthogonal dual
of the valuated matroid. The definition is straightforward.
Definition 4.1.4 The orthogonal dual of µ is the function µK : P pRq Ñ F defined
by µKpX q � µpRzX q. When confusion is unlikely to occur, we will omit the adjective
orthogonal and refer to it simply as the dual of µ.
Lemma 4.1.8 The orthogonal dual is a valuated matroid of rank |R| � rkµ.
Proof. Let A, B � R satisfy |A| � |B| � |R| � rkµ and a P AzB. Then a P pRzBqzpRzAqand hence there exists b P pRzAqzpRzBq such that,
µKpAq dµKpBq � µpRzAq dµpRzBq
¥ µpRzA� a� bq dµpRzB� a� bq
� µKpA� a� bq dµKpB� a� bq.
l
Clearly, pµKqK � µ. For matroids over B the dual is precisely the orthogonal dual
of an ordinary matroid. The relation between the tropical linear space of a matroid
and its dual is given in the next proposition. To state it we define the map x�, �y :
FR � FR Ñ F by,
xv, wy �àiPRpvi dwiq, v, w,P FR.
This is a tropical bilinear form in the sense that both v ÞÑ xv, by and w ÞÑ xa, wy are
linear functions on FR. The set of all w P FR for which the expression xv, wy tropically
vanishes (i.e. it equals 8, or the minimum of the terms is attained at least twice) is
denoted vK. Similarly, for any set V � FR we set,
VK � tw P FR |À
iPR vi dwi vanishes tropically for all v P Vu
Proposition 4.1.9 Let µ be a valuated matroid on R. Then LKµ �LµK .
Proof. We start by proving the inclusion LµK �LKµ . The rank of µK is |R| � d, where
d is the rank of µ. It is sufficient to prove that for any set I of size |R| � d � 1 the
74 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
vector αI � pµKpI � iqqi is tropically perpendicular to Lµ. Let J � R be a set of size
d � 1. Consider pαI , vJq,
pαI , vJq �àiPRµKpI � iq dµpJ � iq
�àiPRµpRzI � iq dµpJ � iq
�àiPRzI
µpRzI � iq dµpJ � iq.
The minimum of this expression is attained twice by (4.2).
To prove the converse the following lemma is required. For v P FR define µJ to be
the function P psupp vq Ñ F ,
Z ÞÑ vZXsupp v dµpZ Y Jq,
where vX denote the tropical product of the coordinates of v indexed by X and J is
an independent set of matpµq such that supp v Y J is spanning in matpµq. This is a
valuated matroid again, and hence the collection of arguments on which it is minimal
(its residual matroid) is the collection of bases of a matroid. We prove that if v KLµthen respµJq has no isthmuses, where an isthmus is a point i P R that is in every basis
of µ. Suppose that it has an isthmus j P R. Let A P respµJq be a basis. Consider the
expression,
vIXsupp v dàiRIpvi dµppA� jq Y J � iqq
By definition, the minimum of this expression is attained at j. Suppose that it were
also attained at i. Then vi �8. But i R supp v, since j is an isthmus of µJ .
Let v PLKµ . We need to show that,
v �à
Spv : wSq dwS ,
where |S| � d � 1 and wS has j-th coordinate µpS � jq. These are the canonical
generators of LµK . Let j P R and suppose that v j � 8. In case v j � 8 the j-thcoordinate of the expression on the right hand side is also8, and we are done. Choose
A and J such that respµJpAqq is minimal among all such choices, with j R A. This is
possible, since no µJ has isthmuses. Set S� � AY J � j. We claim that,
v j � pv : wS�q dwS� � pv : wS�q dµpS�� jq.
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 75
This translates to,
v j � suptvi �µpS�� iq | i P S�,µpS�� iq � 8udµpS�� jq.
It is thus sufficient to prove that the maximum over all vi �µpS�� iq, with i P S� and
µpS� � iq � 8, is attained at j. This is a consequence of the construction of S�. This
proves the statement. l
The proof is a bit technical, but this is due to the limited mathematical machinery
we have at our disposal at this point. Later, in Section 4.2.1 more geometric insight is
given.
Corollary 4.1.10 The tropical linear space Lµ is the intersection of the hyperplanesassociated to αRzC , where C runs over the circuits of µ.
Proof. By the previous proposition, Lµ �LKµK
. Thus, Lµ is the collection of points in
FR that are tropically perpendicular to the hyperplane vectors αH of LµK , where Hruns over the hyperplanes of µK. The statement follows by the fact that the setwise
complements of the hyperplanes of the dual are exactly the circuits of µ. l
The above corollary is the usual definition of a tropical linear space, as found in
[59]. We give some more specific examples of tropical linear spaces.
Example.
(i) Tropical affine space F n is the tropical linear space associated to the matroid µ
on R that assigns 0 to R. Its dual is the matroid that assigns 0 to H, which has
no hyperplanes. Hence F n is the tropical prevariety of H.
(ii) A more specific example is the tropical linear space L associated to µ on r4s
over R8, where µ has underlying matroid the uniform matroid of rank 3 on
r4s. It consists of 6 cones that intersect in the principal submodule generated
by p�µp234q,�µp134q,�µp124q,�µp123qq, where � denotes the the ordinary
minus on R. Its projectivization is depicted in Figure 4.4(a). The white dots on
the boundary of the simplex represent the canonical generators. Equivalently,
it is the subset of R48 perpendicular to LµK , which is the module generated by
vµK
H � pµp234q,µp134q,µp124q,µp123qq.
(iii) A function µ :P2r4s Ñ R8 is a valuated matroid when the expression,
pµp12q dµp34qq ` pµp13q dµp24qq ` pµp14q dµp23qq, (4.5)
vanishes tropically. If matpµq � U2r4s, the hyperplanes of µ are the singletons,
and Lµ � xvt1u, vt2u, vt3u, vt4uy. Equivalently, L is the locus of points perpen-
76 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
(a) Projectivization of a tropicallinear space associated to a valu-ated matroid with underlying ma-troid U3,4.
v{2}
v{1}
v{4}
v{3}
(b) Projectivization of the tropi-cal linear space associated to avaluated matroid µ with underly-ing matroid U2,4.
Figure 4.4: Projectivizations of tropical linear subspaces of R48.
dicular to LµK . The latter is generated by the elements,
wt1u � p8,µp34q,µp24q,µp23qq
wt2u � pµp34q,8,µp14q,µp13qq
wt3u � pµp24q,µp14q,8,µp12qq
wt4u � pµp23q,µp13q,µp12q,8q,
and hence L is the intersection of the four hyperplanes wKtiu. It is depicted in
Figure 4.4(b) in case µp14q dµp23q is the unique maximum in (4.5).
Restriction and contraction
The operations of restriction, and its dual, contraction, have analogues for valuated
matroids. They were already given in [23] for equivalence classes of valuated ma-
troids, but with minor modifications their constructions lift to the individual valuated
matroids themselves. In this subsection µ is a valuated matroid on R over a tropical
semifield F .
Let X � R. The restriction of µ to X is the function on subsets of X that assigns to
each J � X the minimum of µpZq, where J � Z � R. It is denoted µrX s. Contractionis the dual of this operation. i.e., the contraction of µ to X is pµKrX sqK. It is denoted
µ � X , or, when more convenient, by µ{pRzX q. The fact that both operations indeed
yield valuated matroids and the characterization of their tropical linear spaces is given
by the next lemma.
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 77
Lemma 4.1.11 Let µ be a valuated matroid on R and X � R. Then both µrX s and µ � Xare valuated matroids on X . Moreover,
(i) the tropical linear space associated to the restriction of µ to X is the image of Lµunder the natural projection FR Ñ F X ,
(ii) the tropical linear space associated to the contraction to X is the image of thetropical linear space tv P Lµ | vi �8 for all i R Xu under the natural projectionFR Ñ F X .
Proof. We prove the statements in the numbered order.
(i) We should start by showing that the restriction is indeed a valuated matroid.
Choose I a set of size rkµ R � rkµ X such that I Y X is spanning. Denote the map
P X Ñ F defined by A ÞÑ µpAY Iq by µI for the purpose of this proof. That µI is a
valuated matroid is a direct consequence of the exchange condition. Moreover, µrX sis the pointwise tropical sum (minimum) of the µI . To show that µrX s is a valuated
matroid it is sufficient to show that the µI for varying I are proportional. For this,
assume without loss of generality that rkµ R� rkµ X � 1. Let i, j P RzX such that both
X � i and X � j are spanning sets of µ and let A, B � X be sets of size rkµ R� 1. We
need to show that µpA� iq dµpB� jq � µpA� jq dµpB� iq.Suppose that one of the terms is 8. Assume without loss of generality that µpA�
iq � 8. Then A� i is not a basis of matpµq and hence A does not span X . Thus, A� jis not a basis of matpµq, or, equivalently, µpA� jq � 8.
Suppose that none of the terms is is 8. Consider (4.2) with Y � AY ti, ju and
X � B. If k P A, then B � k � X and hence µpB � kq � 8. The required equality
µpA� iqdµpB� jq � µpA� jqdµpB� iq follows since (4.2) now has only two terms,
and the minimum must be attained at least twice. Thus, we have shown that the µI
are proportional.
We now show that the projection contains the tropical linear space associated to
µrX s. Let H be a matroidal hyperplane of µrX s. Then H � X X G for some matroidal
hyperplane G of µ (cf. [11, Proposition 7.3.1. on page 131]). There exists a spanning
set I for G such that IXH spans H. Thus, the rank of IzH is |I |�rkµ H � rkµ R�rkµ X .
Set J � IzH. Then, πpvµI q � vµJJ , where the superscript denote the matroid with
respect to which the generating vector is defined. Thus, πpvµHq is proportional to vµrX sG ,
proving the inclusion.
To prove the other inclusion, we make use of Corollary 4.1.10. Let v P πX pLµq and
C a circuit of µrX s. Then C is also a circuit of µ (cf. [11, Proposition 7.3.1. on page
131]). Let S � C be a minimal dependent spanning set of µrX s such that αC � αS .
Let I � R such that µrX spAq � µpAY Iq. Then S Y I is a minimal dependent spanning
set of µ containing C , and hence the minimum,
àiPCµpSY I � iq d vi �
àiPCµrX spS� iq d vi
78 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
is attained twice, proving that v PLµrX s.(ii) That the contraction is a valuated matroid is a direct consequence of part (i)
of this Lemma and Lemma 4.1.8. For the purpose of this proof, write L8 for the set
tv | vi �8 for all i R Xu and denote the canonical coordinate projection FR Ñ F X by
π. We prove that πpL8q �Lµ�X .
Note that Lµ�X � πpLµKqK by Proposition 4.1.9 and part (i) of this lemma. Thus,
it is sufficient to show the statement that v K w is equivalent to πpvq K πpwq, for
any v P L8 and any w P FR. Let v K w. Then the minimum of xv, wy is attained at
least twice. If xv, wy � 8, then the minimum xπpvq,πpwqy is attained at least twice;
if xv, wy � 8, then xπpvq,πpwqy � 8. Proof of the converse is similar. l
This section ends with a useful corollary of the proof.
Corollary 4.1.12 There exists I�, J� � R such that I�Y X is spanning in µ and J� is abasis of µzX such that µrX spZq � µpZ Y I�q and µ{X pZq � µpZ Y J�q, for all Z � R ofthe appropriate size.
Direct sum
The next construction is very simple and natural from the point of view of tropical
linear spaces.
Proposition 4.1.13 Let µ and ν be valuated matroids on disjoint sets R and S over F.Then the function,
P pR\ Sq Ñ F, X ÞÑ µpX X Rq d νpX X Sq
defines a valuated matroid on R\S over F. Its associated tropical linear space isLµ�Lν .
Proof. The valuated exchange condition is easily seen to follow from either the valu-
ated exchange condition for µ or that of ν .
The corank 1 independent sets of µ` ν are of the form A\ J or B \ I , where Iis a corank 1 independent set of µ, J is a corank 1 independent set of ν , A is a basis
of µ and B is a basis of ν . The fundamental generators associated to these sets are
p8, . . . ,8q�µpAq d vνJ and νpBq d vµI � p8, . . . ,8q. They also generate Lµ�Lν . l
The valuated matroid defined in the above proposition is called the direct sum of µ
and ν and denoted µ`ν . It generalizes the direct sum operation on matroids (cf. [11,
Section 7.6.]) in the sense that matpµ`νq �matpµq`matpνq, where the `-operator
on the left hand side is the direct sum as defined above, and the `-operator on the
right hand side is the direct sum of ordinary matroids.
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 79
Union and intersection
A construction that is important in the proof of Proposition 4.2.21 is that of valuated
matroid union. This construction is dual to the matroid intersection construction on
valuated matroids that appears in [59] and is the valuated equivalent of an earlier
matroid construction [11]. The construction assigns to valuated matroids µ and ν on
a common ground set R another valuated matroid, µY ν on R by setting,
µY νpZq �à
XYY�ZµpX q d νpY q,
in the independent set characterization. When at least one basis µ is disjoint from a
least one basis of ν , then the definition in terms of the bases cryptomorphism simpli-
fies to the following,
pµY νqpZq �à
X\Y�ZµpX q d νpY q.
That both these definitions in fact define the same valuated matroid is stated and
proved in the next proposition.
Proposition 4.1.14 The union µ Y ν is a valuated matroid on R that is dual to thevaluated matroid intersection of µ and ν .
Proof. We simplify the situation somewhat to be able to make use of the results in
[59]. First of all, note that it is sufficient to prove that µ Y ν defines a valuated
matroid in case that some basis of µ is disjoint from some basis of ν . Since, suppose
that this is not the case, then there exist truncations µ1 of µ and ν 1 of ν such that
µY ν � µ1Y ν 1 and µ1 and ν 1 have disjoint bases.
We proceed by proving that µY ν equals pµKX νKqK. Let Z P R. Then,
pµKX νKqKpZq �à
XXY�RzZ
µKpX q d νKpY q
�à
XXY�RzZ
µpRzX q d νpRzY q
�à
RzXYRzY�Z
µpRzX q d νpRzY q
�à
X 1YY 1�Z
µpX 1q dµpY 1q.
The statement follows by Proposition 3.1. of [59]. l
Note that direct sum is a special case of the union. Let µ and ν be valuated ma-
troids on R and S over F , with R and S disjoint sets. We write υ0pX q for the valuated
matroid on X of rank 0 that assigns 0 to H. Then µ` ν � pµ`υ0pSqqY pν `υ0pRqq.
80 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
Extensions and quotients
The inclusion relation on tropical linear spaces contained in the same ambient space
is reflected in the quotient relation between valuated matroids on the same ground
set. Let R be a finite set. Let µ and ν valuated matroids on R over F . Then ν is said
to be quotient of µ when Lν � Lµ. We write ν � µ when ν is a quotient of µ. By
Proposition 4.1.9 it is immediate that ν � µ if and only if µK � νK.
Concluding remarks
In sections 4.2 and 4.3 of this chapter the focus is on the tropical linear space as-
sociated to a valuated matroid, rather than on the valuated matroid itself. To avoid
having to define the appropriate maps every time we speak of the tropical linear space
associated to a construction on a valuated matroid, we apply the notations of certain
constructions to the tropical linear space as well. For example, instead of writing the
projection of Lµ to FS , where µ is a valuated matroid on a finite set R and S � R, we
write L rSs. Where there already exists good notation for the linear spaces associated
to constructions (e.g. in the case of the direct sum), we always prefer the old notation.
The convention is mostly useful for restrictions and contractions, and even in these
case, we will not always use it. Note that by Proposition 4.1.9 this notational con-
vention agrees with the orthogonal dual of the tropical linear space, that we defined
earlier.
4.1.3 Parameter spaces
Let L be a tropical linear space in FR. A tropical linear spaceM in FR is said to be
a subspace of L ifM � L . The collection of tropical linear subspaces of dimension
d of L is denoted DrdpL q, and the collection of all linear subspaces of L is DrpL q.Following [33], we refer to these sets as Dressians, after Andreas Dress. By definition
of a tropical linear space, the collection DrdpL q can be embedded as a subset of the
collection of proportional equivalence classes of valuated matroids on R over F . This
is the content of Proposition 4.1.16, which we prove after the next lemma.
Lemma 4.1.15 Let µ and ν be valuated matroids on R over F. Then Lµ � Lν if andonly if µ and ν are equivalent.
Proof. If µ and ν are equivalent, there exists λ P F� such that µ� λdν . In particular,
they have the same underlying matroid and rank, say d. Let I be a set of cardinality
d � 1. Then vµI � λd vνI � Lν and vice versa. The statement follows since the vµIgenerate Lµ as an F -module.
For the other implication, suppose that Lµ �Lν . By part (ii) of Proposition 4.1.5
the lattices of flats of matpµq and matpνq are the same and hence matpµq and matpνq
4.1. VALUATED MATROIDS AND TROPICAL LINEAR SPACES 81
are the same matroid M . By considering the supports one sees that vµH and vνH are
proportional for all hyperplanes H of M . By Proposition 4.1.4 for all I of cardinality
d�1 there exists λI P F such that vµI � λI d vνI . We show that µpAq�νpAq is constant
for all bases A of M , by induction on the symmetric distance between bases.
Let A and B be bases of M such that |A ∆ B| � 2. Write I � AX B, A� I � i and
B � I � j. Since vµI � λI d vνI it follows that µpAq � λI d νpAq and µpBq � λI d νpBq.Let A and B be bases of M and suppose that for all bases C and D such that
|A∆ B| ¡ |C ∆ D| it holds that µpCq�νpCq and µpDq�νpDq are equal, to say λ P F�.
Since, by the basis exchange axiom for ordinary matroids, there exists a basis C� such
that |A∆ C�|, |B ∆ C�| |A∆ B| the statement follows. l
This lemma has the interesting consequence that the space of all linear subspaces
of a fixed dimension d of some ambient affine space is essentially a projective tropical
prevariety in the projectivization of the space of maps PdpRq Ñ F . As a matter of
fact, in this thesis the Dressian shall be identified with this embedding whenever it is
convenient to do so.
Proposition 4.1.16 Let L � FR be a tropical linear space. The Dressian DrdpL q em-beds as a tropical prevariety in DrdpF
Rq. In particular, it embeds as a tropical prevarietyin the projectivization of the space of maps PdpRq Ñ F.
Proof. Let µ be a valuated matroid associated to L . The Dressian DrdpL q is the
space of equivalence classes of valuated matroids µ on R over F of rank d satisfying
Lν � L . Thus, it is clearly a subset of DrpFRq. We need to show that the condition
Lν � L is given by a the vanishing of a finite number of tropical polynomials. By
Proposition 4.1.9 the condition is equivalent to Lν K LK � LµK . By linearity this
condition is equivalent to the perpendicularity of a set of generators of Lν and a set
of generators of LµK , both of which can be taken finite. Thus, the statement follows.
l
A question related to parameter spaces which deserves mention is the following.
Let K denote a field with surjective valuation v : K Ñ F and consider a linear space
L � KR with tropicalization L . Tropicalization maps GrdpLq to DrdpL q, but is in
general not surjective. In fact, the image depends on the specific choice of field Kand L such that vpLq � L . Determining the image is a difficult question, unless
d � 2 and K is algebraically closed. The general question reduces to determining
when a matroid is realizable over a K , in case v is the trivial valuation and F � B. To
show that in general the Dressian can be expected to have a much bigger dimension
that the image of the Grassmannian under the tropicalization map, we state the next
proposition, first stated and proved in [33] (as Theorem 3.6).
Proposition 4.1.17 (Herrman, Jensen, Joswig, Sturmfels) Let K be an alge-
82 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
braically closed field with surjective valuation ω : K Ñ R8. The dimension of Dr3pRn8q
is of order Θpn2q, while Gr3pKnq, and hence ωpGr3pK
nqq, has dimension 2n� 9.
Remark. It is worthwhile to mention here the well-known fact that ωpGr2pKnqq �
Dr2pRn8q, i.e. that all projectivized tropical linear are realizable. However, we later
give the explicit example Dr2pL q, with L the tropical linear space of µ of rank 3
on r4s that assigns 0 to each of its bases. To do so, it is necessary to study functions
on tropical linear spaces. An interesting feature of the example is that this particular
Dressian is actually a tropical linear space of dimension 3 and hence, that not all lines
in L are realizable as a line in a fixed realization of L . Or in other words, Dr2pL q is
not the tropicalization of (an embedding of) the Grassmannian Gr2pK3q.
4.2 Functions and their graphs
This section is devoted to studying functions from a tropical linear space to the semi-
field the space is defined over. The collection of all such functions that are linear is the
algebraic dual of the tropical linear space, as defined in Section 3.2. We use Lemma
3.2.1 to derive that linear functions are the restrictions of homomogeneous polyno-
mial functions of degree 1. This is not specific to tropical linear spaces, but holds
for any submodule of FR. However, the tropical linear space, through its valuated
matroid, defines a polyhedral structure on the homogeneous polynomial functions of
degree 1, which is intimately related to restrictions of these functions to the tropical
linear space. The main statement is Proposition 4.2.2. It has as a consequence that dterms are sufficient to specify the restriction to a tropical linear space of rank d. In
Theorem 4.2.3 a subcomplex of the polyhedral structure is shown to have underlying
set equal to the dual tropical linear space. This statement is essentially the character-
ization of a tropical linear space stated in Proposition 2.3 of [59].In the second part of this chapter we classify the elementary extension functions
on a fixed tropical linear space, i.e. the functions whose graph, defined in a suitable
manner (we refer to it as the extended graph), is again a tropical linear space. The
theory developed in the first of part of Section 4.2.2 is an extension of that devel-
oped by H. Crapo on elementary extensions of matroids in [14]. Our main results are
that, although such functions need not be linear, they are defined by their image on
the canonical generators of the tropical linear space. This is the content of Theorem
4.2.11. Moreover, the collection of all such functions forms a module, that is also a
tropical prevariety in a natural way. The exact result is stated in Theorem 4.2.9 and
Corollary 4.2.10. These are the main result. The remainder of the second part of this
chapter is devoted to examining elementary extension functions in more detail. In
particular, we prove that the linear functions are always elementary extension func-
4.2. FUNCTIONS AND THEIR GRAPHS 83
tions and that they are the best-behaved, both with respect to realizability over a field
(Proposition 4.2.15), and with respect to linear subspaces (Proposition 4.2.17).
The third part is the second section of Section 4.2.2, and aims to understand the
situation in which we iteratively take the extended graph of a linear function. More
precisely, if L over F is a tropical linear space, and α : L Ñ F is a linear function,
then the extended graph L1 is a tropical linear space. The extended graph of a linear
function α1 :L1 Ñ F is again a tropical linear space L2. Repeating the process yields
a sequence of tropical linear space that we represent as,
Lα1 // L1
α2 // L2α3 // . . . αk // Lk (4.6)
This situation can be usefully represented by a weighted directed graph on R Ytp1, . . . , pku nodes, with R such that L � FR. The valuated matroid of Lk is de-
termined by that of L and the weighted directed graph, or rather by its valuatedlinking system, i.e. the weights of a certain collection of vertex disjoint paths. A gen-
eral valuated linking system is an abstraction of this concept, and they are used as a
tool in the second section of Section 4.2.2 to understand the situation represented by
(4.6). The linking systems that come from weighted directed networks are then used
to construct the morphisms in the category of tropical linear spaces.
4.2.1 Linear functions
This section deals with the algebraic dual of a tropical linear space, as defined in
Section 3.2. We remind the reader that the map that assigns the module of linear
functionals to a tropical linear space, and the pullback to a homomorphism between
tropical linear spaces, is a contravariant functor on the subcategory of tropical linear
spaces and homomorphisms. The following statement is essential for our analysis,
and is an immediate consequence of Lemma 3.2.1.
Lemma 4.2.1 Let L � FR be a tropical linear space. Any linear functional in L � is therestriction of a linear functional FR Ñ F. Or more precise, for each φ P L �. there existpλiq P FR such that for all v PL ,
φpvq �¸iPR
λi d vi . (4.7)
Proof. By Lemma 3.2.1 there exists a linear functional ψ : FR Ñ F such that ι�pψq �
φ, where ι :L ãÑ FR is the canonical injection. In other words, φpvq �ψpvq for each
v PL . Denote the i-th canonical free generator of FR by ei . Thenψpvq �°
iPRψpeiqd
vi , which proves the second statement. l
84 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
(∞, 0,∞)
(∞,∞, 0)
CM
(0,∞,∞)
{{2, 3}}
{{1, 2}}
{{1, 3}}
Figure 4.5: The polyhedral subdivision of Q38ÑQ8 induced by the matroid U2pr3sq.
The matroids associated to the maximal cones are specified by their collection ofbases.
Let µ be a valuated matroid on R over F with associated tropical linear space
Lµ � FR. For each linear functional φ : FR Ñ F define the support of φ as the set of
i P R for which φpeiq � 8 and denote it by suppφ. In the case that suppφ � R the
vector�φpeiq
�lies in FR
� and hence the twisted matroid�φpeiq
��µ, with
��φpeiq
��µqpBq �
�àiPBφpeiq
dµpBq (4.8)
is well-defined. We denote it simply by φ �µ.
Write S for the support of φ and denote the injection FS Ñ FR that maps the
i-th canonical generator of FS to the i-th canonical of generator of FR by ι. Then the
support of ι�pφq is S and hence ι�pφq �µrSs is well-defined. Now define the ordinary
matroid µφ as the direct sum of matpµ{Sq and respι�pφq �µrSsq, i.e.,
µφ �matpµ{ suppφq ` respι�pφq �µrSsq. (4.9)
The map φ ÞÑ µφ induces a subdivision of the space of linear functionals FR Ñ F . We
give an example below.
Example. Consider the valuated matroid µ on r3s over Q8 with µpBq � 0 for all
B P P2pr3sq. The association φ Ñ µφ defines a polyhedral subdivision of the space
of linear functionals Q38 Ñ Q8. We identify this space of linear functionals with F3
through the natural map φ ÞÑ pφpe1q,φpe2q,φpe3qq. The polyhedral subdivision of
PpQ8q induced by the above association is drawn in Figure 4.5.
As a first example, consider the linear map φpz1, z2, z3q � z3, identified with the
point p8,8, 0q PQ38. Its has support t3u and the associated matroid is the direct sum
4.2. FUNCTIONS AND THEIR GRAPHS 85
of matpµq{t3u, which has bases tt1u, t2uu, and resp0 �µrt3usq, which has bases tt3uu.
Thus, M � µφ has collection of bases tt1, 3u, t2, 3uu. We set out to determine which
other functions induce this matroid as well. Note that M is a direct product in two
ways,
M � tt3uu` tt1u, t2uu, M �H`M ,
where we have specified the matroids by their collection of bases. The first decom-
position comes either from points p8,8,φ1q (all of which indeed induce M), or
from points pφ1,φ2,8q, where φ1, φ2 and φ3 are finite. The latter induce a matroid
µ{t1, 2u ` resppφ1,φ2q � µq � H` tt1, 2uu. Remains, the case resppφ1,φ2,φ3q � µqq.
This equals M exactly when,
φ1 �φ3 � φ2 �φ3 φ1 �φ2
in non-tropical notation. This is the cone CM � tpv, v, wq P Q | w vu. It is invari-
ant under classical addition of p1, 1,1q, or, stated differently, under tropical scalar
multiplication and hence has a well defined image in PpQ38q.
There is a strong relation between the matroid µφ and the restriction of φ to Lµ,
which is stated in the next proposition.
Proposition 4.2.2 Let µ be a valuated matroid on R over F and φ : FR Ñ F a linearfunctional. Write S for the support of φ and define the linear map ι : FS
ãÑ FR byιpeiq � ei . Then, for a subset X � R, we have,
φpvq �àiPXφpeiq d vi , for all v PLµ, (4.10)
if and only if X X S is a spanning set of respι�pφq � µrSsq. Moreover, if for each v PLµ � p8, . . . ,8q there is i P X such that φpvq � φpeiq d vi with vi � 8 then X is aspanning set of µφ .
Proof. Since (4.10) is equivalent to φpιpvqq � `tφpeiqdvi | i P XXSu for all v PLµrSsit is sufficient to prove the first statement in the case that S � R, which we assume in
the remainder of its proof.
Let X be a spanning set of respφ �µq, containing a basis B. It is sufficient to prove
(4.10) for the fundamental generators vI , with I an independent set of µ of corank 1.
Consider,
φpvIq �àiPRφpeiq dµpI � iq
and let i� be an index at which the minimum is attained. Suppose that i� R B, since
otherwise the statement is trivial. Then φpeiqI dφpei�qdµpI� i�q is minimal among
the terms of φpeiqI d φpvIq, where we have written φpeiqI to denote the tropical
86 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
product dtφpeiq | i P Iu. Note that φpeiqI dφpei�qdµpI � i�q � pφ �µqpI � i�q. The
exchange condition for φ �µ states that there exists j� P BzI such that
pφ �µqpBq d pφ �µqpI � i�q ¥ pφ �µqpB� i�� j�q d pφ �µqpI � j�q.
Since pφ �µqpBq ¤ pφ �µqpB� i�� j�q, by choice of B, the inequality pφ �µqpI� i�q ¥pφ �µqpI � j�q follows. Thus, the minimum in φpvIq is also attained at j� P B.
To prove the reverse implication, let X � R such that φpvq � `tφpeiqd vi | i P Xufor all v PL and let B� be a basis of respφ �µq for which |B�zX | is minimal. Assume,
for the sake of contradiction, that the difference is not empty and let j P B�zX . Clearly,
the minimum of φpvB�� jq is attained at index j. It is also attained at some i P X , by
choice of X . Then i R B� and B� � j � i is a basis of respφ � µq. Moreover, |X zB�| ¡|X zB� � j � i|, contradicting the assumption that B�zX �H. Thus B� � X , showing
that X is a spanning set of respφ �µq.
For the second statement we prove a characterization of the spanning sets of
matpµ{Sq in terms of φ. The claim is that a set X � RzS is spanning in matpµ{Sqif and only for each v P L satisfying vi � 8 for all i P S there exists j P X such that
v j � 8. It is sufficient to prove this statement for v a fundamental generator. Such
fundamental generators correspond to hyperplanes that contain S and the statement
translates to the following: X � RzS is a spanning set of matpµ{Sq if and only if X � Hfor each hyperplane H � S. Since the hyperplanes of matpµ{Sq are of the form HzS,
where H is a hyperplane of matpµq containing S, it is sufficient to show that the span-
ning sets of a matroid are exactly those sets not contained in any hyperplane. But
this equivalence is clear by the fact that both matroid closure and rank are increasing
functions.
A set X � R is spanning in µφ if and only if X X S is spanning in respι�pφq �µrSsqand X zS is spanning in matpµ{Sq. Let v P Lµzp8, . . . ,8q. If supp v X S � H, then
φpvq � `tφpeiqd vi | i P X XSu �8, since X XS is spanning in respι�pφq �µrSsq, and
hence there exists some index at which the minimum is attained. If supp v X S � H,
then φpvq � 8. Since X XRzS is spanning in matpµ{Sq it contains j such that v j � 0.
The above argument in reverse is a proof of the other implication. l
Example. We continue the example just before Proposition 4.2.2. Let φ be the linear
map Q38 Ñ Q8 defined by the point pφpe1q,φpe2q,φpe3qq � p3, 3,1q. This map is in
the cone CM , and hence µφ � tt1,3u, t2,3uu. Thus,
φpvq � p3d v1q ` p1d v3q � p2d v2q ` p3d v3q, when v PLµ, (4.11)
according to Proposition 4.10 above. Write Hi for the set of v PQ38 for which φpvq �
φpeiqdvi . ThenLµ � H1YH3 andLµ � H2YH3, by (4.11). ThusLµ � pH1XH2qYH3.
4.2. FUNCTIONS AND THEIR GRAPHS 87
(∞, 0,∞)
(∞,∞, 0)
(0,∞,∞)
H3
H2 H1
Lµ
Figure 4.6: The sectors Hi of Q38 on which the tropical linear function φ attains its
minimum at coordinate i overlayed on the (dashed) tropical linear space Lµ.
The situation is depicted in Figure 4.6.
Proposition 4.2.2 is an essential step in proving the theorem below. The second
part of this theorem gives an alternative characterization of the tropical linear space
associated to the orthogonal dual of µ, by Proposition 4.1.9. Of course, one needs to
consider this tropical linear space to be a submodule of the space of linear functionals
on FR, through the map v ÞÑ φv , where
φvpwq � xv, wy �àiPRpvi dwiq, w P FR.
The characterization thus obtained is essentially that of Proposition 2.3 of [59].
For the first part we need a definition. The restriction map ρ : pFRq� Ñ L �
is a homomorphism of semimodules, and since pFRq� is finitely generated and L �
is residuated, the dual residual of ρ exists when ψ P pFRq� and ψpvq � 8 for all
v P L implies that ψpwq � 8 for all w P FR. This is exactly the case when µ has no
loops. Then for all φ P L � the image ρ^pφq exists and is referred to as the minimalrepresentative of φ. Note that,
ρ^pφq � inftψ P pFRq� | φpvq ¤ψpvq for all v PL u. (4.12)
Note that since ρ is surjective, by abstract residuation theory one gets that the com-
position ρ^ �ρ : pFRq�Ñ pFRq� is a dual closure mapping (cf. [8, Theorem 2.7]), i.e.
it is increasing, idempotent and at most the identity.
Theorem 4.2.3 (i) the functional φ is the minimal representative of its restriction toLµ if and only µφ has no loops, (ii) the functional φ vanishes tropically on Lµ if and
88 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
only if µφ has no isthmuses, other than those of µ and φpe jq � 8 for all isthmuses j ofµ.
Proof. (i) For the first statement it is sufficient to prove that j P R is a non-loop
(i.e. t ju is independent) in µφ if and only if there is v P Lµ � p8, . . . ,8q such that
φpvq � φpe jq d v j and v j � 8. Suppose that j is independent. There exists some
basis B that contains j. Since B � j is not a spanning set, there must exist some
v P Lµ � p8, . . . ,8q such that for all i P B � j either φpvq φpeiq d vi or vi � 8.
Since B is a spanning set, for this v both φpvq � φpe jq d v j and v j �8.
Let v P Lµ � p8, . . . ,8q such that φpvq � φpe jq d v j and v j � 8. Denote the
support of φ by S. We distinguish two cases:
If j P S then the statement holds with v replaced by πpvq, where π : FR Ñ FS is
the canonical projection. Hence we can assume without loss of generality that R� S.
There exists a fundamental generator vI for which φpvIq � φpe jq d µpI � jq and
j R clpIq, by Lemma 4.1.4. Let B be a basis of µφ . By the assumption that R � S the
latter matroid equals respφ �µq. If j P B then j is independent and there is nothing to
prove, so suppose that j R B. By the exchange condition for φ �µ there exists i� P BzIsuch that,
pφ �µqpI � jq d pφ �µqpBq ¥ pφ �µqpI � i�q d pφ �µqpB� i�� jq.
Since pφ �µqpI � jq ¤ pφ �µqpI � i�q and neither equals 8 it follows that pφ �µqpB�i�� jq ¤ pφ �µqpBq. Thus B� i�� j is a basis of µφ and hence j is independent in µφ .
If j R S then φpvq � 8 and by linearity there exists some fundamental generator
vI whose j-th coordinate is not equal to 8 satisfying φpvIq � 8. Thus, clpIq � S and
j R clpIq, by Lemma 4.1.4. Thus j R clpSq and hence j is independent in matpµ{Sq.
(ii) Suppose that there exists v PLµ�p8, . . . ,8q such that φpvq � φpe jqd v j and
v j �8, and for all i P R� j either φpvq φpeiq d vi or vi �8. Then by Proposition
4.2.2 this is equivalent to the statement that all spanning sets of µφ contain j. This
characterizes j as an isthmus of µφ . Ifφ vanishes on v thenφpvq � 8. Hence,φpe jq �
8 and vi �8 for all i � j. Thus, R� j is a flat and j is an isthmus of µ.
(iii) Suppose that j is an isthmus of µφ that is not also an isthmus of µ. Then
R is a spanning set of µφ , while R � j is not and thus, by Proposition 4.2.2, there
exists v P Lµ such that φpvq � φpe jq d v j and v j � 8. Moreover, for all i � j either
φpe jq d v j φpeiq d vi or vi � 8. Clearly, φ doesn’t vanish tropically at v, unless
φpvq � 8. Suppose, for the sake of contradiction, that φpvq � 8. Then for all i � j it
holds that vi �8. Thus, R� j is a flat of µ and hence j is an isthmus of µ, contradicting
the assumption. Hence φ does not vanish at v.
Suppose that φ does not vanish tropically on Lµ and that every isthmus of µφ is
also an isthmus of µ. We show that there exists an isthmus j of µ for which φpe jq � 8.
4.2. FUNCTIONS AND THEIR GRAPHS 89
There exists v P Lµ such that φpvq � φpe jq d v j � 8 and φpvq φpeiq d vi for all
i � j. By Proposition 4.2.2 this means that j is an isthmus of µφ and thus, by the
premise, that j is an isthmus of µ. But clearly, φpe jq � 8. l
Remark. In the case that F � B, that is, in the case of ordinary matroids the above
statements specialize to known facts. Let µ be an ordinary matroid on R. Recall that
Lµ � BR is isomorphic, as a lattice, to the lattice of flats of µ (cf. Proposition 4.1.5).
A linear function φ : BR Ñ B is defined by its support S. The matroid µφ equal µ{S`µrSs, where the operations are the well-known operations of restriction, contraction
and direct sum on ordinary matroids. Since µφ has no loops precisely when S is a flat
of µ it is a consequence of Theorem 4.2.3 that the linear functionsLµÑ B correspond
bijectively to the flats of µ.
By the last remark the lattices L and L � are anti-isomorphic in case F � B.
We cannot expect this result to be true for more general semifields, since when F is
infinite L has no minimum, whereas it has a unique maximum. However, a similar
result holds. To prove it we make use of some of the results in Section 4 of [13] that
we restated in Section 3.2 of this thesis.
Theorem 4.2.4 Let L be a tropical linear subspace of FR and denote the canonicalinclusion by ι,
(i) if F � B then the composition of the map BR Ñ pBRq� that assigns the linearfunctional defined by ei ÞÑ 0 if vi � 8 and ei ÞÑ 8 if vi � 0 to v P BR with ι� isan anti-isomorphism of lattices when restricted to L .
(ii) if F is infinite and L X pF�qR � H, then the composition of the map pF�qR ÑHomppF�qR, F�q that assigns the linear functional defined by ei ÞÑ v�1
i to v PpF�qR with ι� is an anti-isomorphism of lattices when restricted to L X pF�qR.
Proof. Let µ be a valuated matroid that has L as its tropical linear space. Denote the
matrix whose rows are the vectors vH , where H runs over set of hyperplanesH of µ,
by M . Then the row space of M is L , while the column space is t�φpvHq
�HPH | φ P
L �u. The latter is isomorphic toL � as an F -module. Part (i) now follows immediately
from Proposition 3.2.3. For part (ii) it is sufficient to prove that the isomorphism
between FH dM and Md FR restricts to an isomorphism between pFH dMqXpF�qR
and pM d FRq X pF�qH and is of the form as stated in the theorem.
Let v P pFH dMq X pF�qR. The image of v under the isomorphism of Proposition
3.2.3 is M dw, where w � inftw1 | xw1, vy ¥ 0u � v�1. Since L XpF�qR is not empty
µ has no loops. Thus, for each i P R there is H P H such that i R H and all of the
rows of M are different from p8, . . . ,8q. Hence, M d w P pF�qH . This proves that
the isomorphism indeed restricts as stated. Note that the inverse satisfies the same
property by symmetry. l
90 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
Remark. The result above is a consequence of Theorem 23 of [19] which gives a
particular map that establishes that the tropical convex sets generated by the rows
and column of a matrix over the finite tropical real numbers are isomorphic polyhedral
complexes. This result is very much related to the that of [13], which we refer to in
this thesis. Interpreted in the spirit of [19] a tropical linear space is the row space
of the matrix A whose rows are the canonical generators vH . The space of linear
functionals, defined by their images on vH , is the tropical linear subspace of FH that
is the column space of A.
The theorem gives a simpler way of recognizing which linear functionals are min-
imal representatives of their restrictions to a tropical linear subspace, at least when
the functional is finite on the finite part of FR.
Corollary 4.2.5 Suppose F is infinite and φ : pF�qR Ñ F� is linear. Then φ is aminimal representative of its restriction to L if and only if φpeiq
�1iPR PL .
Proof. This follows immediately from the proof of the theorem. However, it is also
a consequence of Theorem 4.2.3 once one realizes that µφ � pµKqKz , where z �
�pφpeiqqiPR. l
Remark. It follows from Theorem 4.2.4 that L � is not a tropical linear space in
general. In fact, for the case that F � B, the statement thatL � BR is a tropical linear
space is equivalent to the statement that it is a geometric lattice. Since it is known
that the opposite lattice of a geometric lattice is not necessarily geometric L � is not
in general a tropical linear space (cf. [12]).
The next statement is quite remarkable. Since it holds for both a valuated matroid
and its underlying matroid it states that the tropical rank of the matrix of generating
vectors equals the rank of the same matrix with every finite entry replaced by 0.
Proposition 4.2.6 LetL � FR be a tropical linear space. The space of linear functionalsL � is finitely generated by the restrictions of αi : FR Ñ F for i P R, where αipe jq � 0
if i � j and 8 otherwise. Let H denote the collection of hyperplanes of µ. The tropicalrank of the set of vectors t
�αipvHq
�HPH | i P Su equals the rank of S in µ.
Proof. The first statement follows immediately from the surjectivity of the restriction
map. Let S � R and write M for the matrix whose rows are the vectors in the state-
ment. We need to show that the largest tropically non-singular square submatrix of
M has size rkµpSq. Clearly, the row space of M is the image of L under the canonical
projection π : FR Ñ FS and the rank of S is the rank of the matroid µrSs. By Lemma
4.1.11 and Lemma 4.1.15 the row space of M generates LµrSs. It is thus sufficient to
prove the statement for S � R.
4.2. FUNCTIONS AND THEIR GRAPHS 91
Let B be a basis of µ. Then M has |B| rows such that each is proportional to
precisely one of vB�i , for i P B. The submatrix of M with these rows and columns
indexed by the elements of B is 8 everywhere, except on the diagonal, where the
entries are elements of F�. This submatrix is clearly tropically non-singular. Thus, the
tropical rank of M ¥ |B| � rkµ � rkµpRq. Let M 1 be a square, tropically non-singular
submatrix of M . Suppose its columns are indexed by some set S � R. Then the tropical
rank of the full submatrix of M with columns indexed by S is |S|. In particular M has
at least |S| distinct rows that form a minimal set of generators of its row space. Since
the row space is precisely LµrSs it must be that these generators are of the form vH ,
for H a hyperplane of µrSs, by Lemma 4.1.4. Thus, matpµrSsq is a matroid on S with
|S| distinct hyperplanes satisfying the additional condition that there exists exactly
one bijection σ : H pµrSsq Ñ R such that σpHq R H, for all H P H pµrSsq. We show
that this implies that matpµrSsq is the unique matroid of rank |S| on S.
First, it is more convenient to prove the dual statement: if there exists precisely
one bijection σ between the circuits C and the ground set R of a matroid M such
that σpCq P C , then M is the rank 0 matroid on R. Suppose that this is not the case.
There exists p0 P R such that p0 is not a loop in M . Let C0 be the preimage of p0,
i.e. σpC0q � p0. Then C0 � p0 � H. Let p1 P C0 � p0 and let C1 be the preimage of
p1 under σ. By minimality of C0 among dependent sets, p1 is not a loop either. Since
p1 P C0 and σpp1q � σpp0q it follows that p1 is also not a loop. Repeating this process,
one obtains sequences pp0, p1, p2, . . .q and pC0, C1, C2, . . .q such that
σpCiq � pi , pi � pi�1 and pi�1 P Ci X Ci�1, for all i � 0, 1,2, . . ..
Since R is finite, there exist minimal indices k l such that Ck � Cl . Now construct a
new bijection σ1 : C Ñ R that equals σ on C ztCk, Ck�1, . . . , Cl�1u and has σ1pCiq �
pi�1 for all i � k, . . . , l � 1. Moreover, σ1pCq P C for all C P C . Thus, we have found
another bijection, contradicting the premise.
Since the circuits of the dual are the complements of the hyperplanes of the primal
matroid, we have proved the statement and hence matpµrSsq is the matroid of rank
|S| on S. Thus, |S| ¤ rkµ and hence the tropical rank of M is rkµ. l
Remark. The last part of the proof of the theorem also follows by the characterization
of the rank of a t0,8u-matrix given by Proposition 4.3. of [18].
Corollary 4.2.7 The tropical rank of the set of vectors tvH | H PH u equals the rank ofµ.
The above corollary has an important consequence for tropical linear spaces over
RYt8u. By Theorem 4.2 of [19] the dimension of a maximal face of the polyhedral
complex structure on the tropical linear space induced by the type function on points
92 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
equals the rank of an associated matroid. Note that although the results of [19] were
proved for tropical convex sets in Rn they hold for finitely generated submodules of
Rn8.
4.2.2 Extensions of matroids
Elementary extensions of matroids
This section deals with functions on tropical linear spaces that are restriction of trop-
ical polynomial functions. Let α : L Ñ F be such a function. In case that F � R8 a
polyhedral structure onL defines a polyhedral structure on the graph of α. However,
the resulting complex is in general not balanced. It is a result, proved for example in
[57] that one can always add maximal dimensional weighted faces to the complex
to make the result balanced. Moreover, this extension is unique and we refer to it
as the extended graph of α. Note that it is not so clear how to define the extended
graph intrinsically for the finite tropical semifield and therefore, in the first half of
this section we restrict ourselves to the case F � R8. We state a specialization of the
result exactly as we need it, but take some liberty with the terminology of polyhedral
complexes. For purposes of this section, a polyhedral complex in Rn8 is the collection
of topological closures of polyhedra in Rn that together form a polyhedral complex.
Lemma 4.2.8 Let L be a tropical linear subspace over RR8 and α : L Ñ R8 the
restriction of a tropical rational function to L . There is a unique balanced polyhedralcomplex whose underlying set is of the form
tpv,αpvqq | v PL uY tpv,λq | v PK and λ¥ αpvqu � RR8�R8, (4.13)
for some K � L , and whose polyhedral structure is induced by the refinement of thedomains of linearity of α on L and the polyhedral structure on L . Moreover, K doesnot depend on the polyhedral structure on L .
Since underlying sets of extended graphs do not depend on the polyhedral struc-
ture on L the notion of the extended graph of a function on a tropical linear space
makes sense. The extended graph then no longer comes with a polyhedral structure.
For our purposes we are mostly interested in extended graphs that are tropical
linear spaces. In particular, the following two questions are answered in the next
theorem: (i) for which functions α : L Ñ F is the extended graph a tropical linear
space, (ii) which tropical linear spaces are extended graphs of tropical polynomial
functions on L . We start by describing the setup on the matroid side.
Let µ be a matroid on R over F . A matroid ν on S over F is called an extension of
µ if: (i) R � S, (ii) rkµ� rkν , and (iii) νrRs � µ. If |SzR| � 1, then ν is said to be an
4.2. FUNCTIONS AND THEIR GRAPHS 93
elementary extension of µ. Note that in the case that F � B these terms coincides with
standard matroid terminology (cf. Section 7.2. of [50]).
Let ν be an elementary extension of µ. Write S � R\ tpu and set d � rkµ. By
the second and third condition the values of ν on d-size subsets of S are fixed by µ.
Thus, the elementary extensions of µ coincide with a certain subset of all functions
Pd�1pSq Ñ F . The map establishing the coincidence assigns to any elementary ex-
tension ν the function I ÞÑ νpI \ tpuq. We refer to the functions in the image of this
map as elementary matroidal extension functions of µ, or simply, matroidal extensionfunctions.
Before we state the next theorem we remind the reader that pv : wq � inftλ P F |v ¤ λdwu, for v, w P FR and w � p8, . . . ,8q.
Theorem 4.2.9 Let µ be be a matroid on R over F of rank d. Then α :Pd�1pRq Ñ F isan elementary extension function of µ if and only if,
(i) αpIq � 8 if rk I d � 1, and(ii) αpIq � pvI : vJqdαpJq for all size d � 1 independent sets I and J that have equal
closure in µ, and(iii) for all S � R and i, j, k P R satisfying |S| � d � 2 and i, j, k R S the expression,
�µpSi jq dαpSkq
�`�µpSikq dαpS jq
�`�µpSk jq dαpSiq
�(4.14)
vanishes tropically
Proof. Let ν be an elementary extension of µ and denote the associated elementary
extension function by α. Note the difference between the Greek letter ν , which stands
for the elementary extension, and the symbols vI , written with a minuscule Latin v,
that stand for the canonical generators of Lµ.
Let I � R be a dependent subset of size d�1. Then rkνpI�pq ¤ rkµpIq�1¤ d�1.
Thus, I � p is not a basis of ν , or equivalently, νpI � pq � 8.
Let I , J � R be independent sets of size d � 1 such that clµpIq � clµpJq. Let i PRz clµpIq. We prove that νpI�iqdνpJ�pq � νpJ�iqdνpI�pq. By part (iii) of Theorem
4.1.3 the sum of µpJ�i�p� jqdµpI� jq for j running over J�i�p vanishes tropically.
The terms with j R ti, pu are 8 and hence νpI � iq d νpJ � pq � νpJ � iq d νpI � pq.Since vI and vJ are proportional it holds that pvI : vJq � pνpI � iq : νpJ � iqq. The
statement follows, since pνpI � iq : νpJ � iqq � νpI � iq d νpJ � iq�1.
Let S � R and i, j, k P R as in part (iii) of the theorem. Setting X � SYti, j, ku and
Y � S� p and applying part (iii) of Theorem 4.1.3 again yields the statement.
Now let α be a function satisfying (i), (ii) and (iii) of the theorem. We need to
show that the map ν : PdpR� pq Ñ F defined by νpAq � µpAq if A � R and νpAq �αpA� pq if p P A is a valuated matroid. We use characterization (iv) of Theorem 4.1.3.
94 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
Let X , Y � R� p of satisfying |Y | � d � 1, |X | � d � 1 and |Y zX | � 3. Set S � X X Yand write ti, j, k, lu � X zY \ Y zX . We distinguish three cases.
(1) Suppose X , Y � R. Then the condition (4.2) is satisfied since ν restricts to µ
and µ is a valuated matroid.
(2) Suppose p P ti, j, k, lu. Then the vanishing condition (iii) of Theorem 4.1.3 for
ν translates immediately to (4.14).
(3) Suppose p P S and denote S�p by T . If T is dependent in µ, then by condition
(i) all terms of expression (4.2) equal8, and hence the expression vanishes tropically.
Suppose that T is independent. For ease of notation write αpT Y Zq by x 1Z . Then for
each m R T the expression,
µpT mi jq d x 1kl `µpT mikq d x 1jl `µpT mk jq d x 1i j ,
vanishes tropically, by the second item in this proof. Since T is independent, µpT YZq � ξpZq, where ξ is the contraction of µ by T (cf. Corollary 4.1.12). For ease of
notation we assume without loss of generality that ti, j, k, lu � t1,2, 3,4u and omit
the multiplication sign from our computations. It is now sufficient to prove that the
expression,
x 112 x 134 ` x 113 x 124 ` x 114 x 123
vanishes tropically, if the expressions,
ξp j34qx 112 ` ξp j24qx 113 ` ξp j23qx 114, ξp j34qx 112 ` ξp j14qx 123 ` ξp j34qx 124
ξp j24qx 113 ` ξp j14qx 123 ` ξp j12qx 134, ξp j23qx 114 ` ξp j13qx 124 ` ξp j12qx 134
vanish tropically for any j R r4s, where ξ is a valuated matroid of rank 3. As a final
substitution, set xZ � x 1Z d ξp jr4szZq for some fixed j. Now it is sufficient to show
that if all ofx12 ` x13 ` x14, x12 ` x23 ` x24
x13 ` x23 ` x34, x14 ` x24 ` x34
vanish tropically, then x12 x34 ` x13 x24 ` x14 x23 vanishes tropically.
Take two 2-sets having empty intersection, say 12 and 34, without loss of gener-
ality, and suppose the minimum is attained at x12 x34. Now we need to show that,
x12 x34 ¥ x13 x24 or x12 x34 ¥ x14 x23, (4.15)
given (leaving out the equality conditions coming from the tropical equations)
px12 ¥ x13 or x12 ¥ x14q and px12 ¥ x24 or x12 ¥ x23q and
px34 ¥ x13 or x34 ¥ x32q and px34 ¥ x24 or x34 ¥ x14q
4.2. FUNCTIONS AND THEIR GRAPHS 95
Suppose that (4.15) is not true to arrive at a contradiction. Without loss of generality
we may assume that x12 ¥ x13 (otherwise apply the permutation p3, 4q). Then x34
x24 and hence x34 � x14. But then x12 x23 and thus x12 � x24. Then x34 x13 and
thus x34 � x23. Here we used that the assumption that the quadratic equation is not
satisfied. But then x13 ¡ x34 � x14 � x23 ¡ x12 � x24 ¥ x13. l
Remark. Consider the uniform matroid µ � U3r4s over F , defined by µpAq � 0 for
all size 3 subsets A of r4s. The proof of the statement that functions Pd�1pRq Ñ Fthat satisfy conditions (i), (ii) and particularly (iii) of the theorem reduced to proving
that the solutions of the tropical equations describing quotients of µ are valuated
matroid. Closer examination of the proof in fact shows that the conditions describing
elementary quotients, if non-trivial, imply the valuated exchange conditions.
Corollary 4.2.10 The elementary extension functions of µ form a module under thepointwise operations, that embeds into FH , where H is the collection of (matroidal)hyperplanes of µ.
Remark. It would be nice to have the statement that the induced order is a lattice
order. This is true for F � B, since any submodule of a free B-module of finite rank is
finitely generated. One can then use Proposition 3.1.6. However, finite generatedness
is not immediate for F infinite. Note that it is known that for F � B the induced order
is a lattice, by [44] (Weak Maps of Combinatorial Geometries).
Example. Consider the matroid U3p4q over F and denote its set of hyperplanes byH .
Then H � t12, 13,14, 23,24, 34u, where we take i j to mean ti, ju. Note that H is
also the collection of subsets of size 2 of r4s. By the theorem, a function α :H Ñ F is
a matroidal extension function of U3p4q if the expressions,
αp12q `αp13q `αp14q, αp12q `αp23q `αp24q
αp13q `αp23q `αp34q, αp14q `αp24q `αp34q. (4.16)
vanish tropically. The solution set is the tropical linear space associated to the cycle
matroid of the complete graph K4, which has affine diagram:
12
13
143424
23
96 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
This is not immediately obvious. It is clear that the linear space of MpK4q is contained
in the space of extensions, since the linear forms (4.16) come from the induced cyclesof K4, where induced cycles are those cycles that are also induced subgraphs. The
picture of K4 we have in mind is given below.
13 12
14
34
23
24
The linear space is cut out by all cycles of K4. However, by a theorem of Yuster and Yu
([71], Theorem 8) the linear forms associated to induced cycles form a tropical basis,
and hence already cut out the tropical linear space.
The tropical linear spaceLMpK4qis generated as a subsemimodule of FH by vectors
corresponding to its hyperplanes
H1 � t12,13, 14u, H2 � t12,23, 24u, H2 � t13, 23,34u, H4 � t14, 24,34u
and
H12 � t12,34u, H13 � t13,24u, H14 � t14, 23u.
The corresponding generating vectors are
v1 � p8,8,8, 0, 0, 0q, v12 � p8, 0, 0, 0, 0,8q,
v2 � p8, 0, 0,8,8, 0q, v13 � p0,8, 0, 0,8, 0q,
v3 � p0,8, 0,8, 0,8q, v14 � p0,0,8,8, 0, 0q,
v4 � p0, 0,8, 0,8,8q,
where the coordinates are in the order p12,13, 14,23, 24,34q.
Let L � FR be a tropical linear space and µ a valuated matroid associated to L .
Denote the canonical projection FR � F Ñ FR by πR. Then, each tropical linear space
K � FR � F of rank equal to d having πRpK q � L has an associated valuated ma-
troid ν that is an elementary extension of µ. After all, by Lemma 4.1.11 and Lemma
4.1.15, it follows that the restriction of any valuated matroid associated to K is pro-
portional to µ. Moreover, any elementary extension of µ gives rise to a tropical linear
space in FR � F of equal rank whose image under π equals L . This shows that there
is a bijective correspondence between such tropical linear spaces and elementary ma-
troidal extension functions. The next theorem describes how to obtain the extension
function from the tropical linear space, without constructing the elementary extension
matroid.
4.2. FUNCTIONS AND THEIR GRAPHS 97
We shall write ExtppL q for the tropical linear subspaces of FR�p of rank equal
to that of L whose image under πR equals L . Similarly, we write Extppµq for the
module of elementary extension of µ, whose ground set is R� p.
Theorem 4.2.11 Let L � FR be a tropical linear space of rank d and µ a valuatedmatroid associated to L . The map that assigns to each K P ExtppL q the functionεK :Pd�1pRq Ñ F given by
εK pIq �
#πp
�π^K pvIq
�, if rkµ I � d � 1,
8, if rkµ I d � 1
where πK is the restriction of πR : FR�p Ñ FR to K , establishes a bijection betweenExtppL q and the elementary matroidal extension functions of µ. Moreover,
πppπ^K pvqq �
àIpv : vµI q d εK pIq,
for all v PL , where I runs over all size d � 1 subsets of R.
Proof. Let ν be an elementary extension of µ such that Lν �K and set εL :L Ñ Fto be the elementary extension function corresponding to the extension ν of µ. Since
εK pIq � νpI � pq � πppvνI q, it is sufficient to prove that
π^K pvq �à
Ipv : vµI q d vνI , v PL ,
where I runs over all size d � 1 subsets of R. Let v P L . By Theorem 3.1.8 it holds
that,
π^K pvq �à
Jpv : πK pv
νJ qq d vνJ , v PL ,
where J runs over all size d � 1 subsets of R� p. Since πKpvνJ q � vνJ if J � R and that
fact that v PL it holds that
v � πK�à
Ipv : vµI q d vνI
�, v PL ,
where I runs over all size d � 1 subsets of R. Now the claim follows from πKppv :
vJq d vJq ¥ v and πpppv : vJq d vJq � 8 for when p P J and |J | � d � 1. l
Remark. For general F the notion of an elementary matroidal extension function is
analogous to the notion of a linear subclass for ordinary matroids (cf. [50, Section
7.2]). More precisely, if F � B and ε is an elementary matroidal extension function,
the set tclpIq | εpIq � 8u is a linear subclass. The inverse construction associates the
elementary matroidal extension function I ÞÑ 8 if and only if clpIq P X to a linear
98 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
subclass X �H .
In a similar vein, the function πp � π^K : L Ñ F of the above theorem is the
analogue of a modular cut.
Example. We continue the example of MpK4q with F � Q8. Consider the extension
matroid associated to the vector v12 � p8, 0, 0, 0, 0,8q, or equivalently, to the exten-
sion function ε defined by,
εpt1, 2uq � εpt3,4uq � 8 and εpt1, 3uq � εpt1,4uq � εpt2,3u � εpt2, 4uq � 0.
We construct the function πp �π^K . By Theorem 4.2.11 it equals,
w ÞÑ pw : v13q ` pw : v14q ` pw : v23q ` pw : v24q
� pw2 _w4q ` pw2 _w3q ` pw1 _w4q ` pw1 _w3q,
where wi _ w j � maxtwi , w ju. Note that wi _ w j equals pwi d w jq � pwi ` w jq on
tw | wi �8 or w j �8u. The latter expression is the tropical quotient of two tropical
polynomial functions.
By Theorem 4.2.11 the elementary matroidal extension functions and functions of
the form πp �π^K for K P ExtppL q are in bijective correspondence. We shall use the
name elementary matroidal extension function for the latter class of functions as well.
This is unlikely to lead to confusion, since the domain of a function is included in its
definition, or, at least, clear from the context.
Proposition 4.2.12 Let L � FR be a tropical linear space and Lν P ExtppL q. Denotethe associated matroidal extension function L Ñ F by ε. Then,
Lν � tpv,εpvqq | v PL uY tpv,λq | v PLν{p and λ¥ εpvqu.
When defined (e.g. when F � R8,Q8) Lν is the extended graph of ε.
Proof. We start by proving the easier inclusion �. By Theorem 4.2.11 any point in
FR�p of the form pv,εpvqq, with v P L is in K . Moreover, by Lemma 4.1.11 the
set tpv,8q | v P Lν{pu � K . For v P Lν{p and λ P F such λ ¥ εpvq the equality
pv,λq � pλ : εpvqqdpv,εpvqq`pv,8q holds. SinceK is closed under tropical addition
and scalar multiplication, the inclusion follows.
For every circuit C of ν choose a size d � 1-spanning set S of ν that contains Csuch that αCpeiq � νpS � iq. For the purpose of the proof we say such S represents
C . Note that by Corollary 4.1.10 we have that K is the intersection of the tropical
vanishing loci of functions of the form αC .
4.2. FUNCTIONS AND THEIR GRAPHS 99
Let v P K . Then πK pvq P L , since K is an extension of L . We write π for πK .
By definition of ε the inequality vp ¥ εpvq holds. Suppose πpvq R Lν{p. Then there
exists a circuit C of ν{p such that πpvq is not in the vanishing locus of αν{pC . Let S � Rrepresent C and consider the expression,
àiPS�p
νpS� p� iq d vi ��à
iPSpν{pqpS� iq d vi
`�νpSq d vp
�,
where we have used that tpu is independent in ν . If this were not the case, then
K �L�t8u and the theorem holds. The first sum of the right hand side of the above
expression equals αν{pC pπpvqq, and since the entire expression vanishes tropically it
must hold that νpSq d vp � αν{pC pπpvqq. Thus, the fiber of π above πpvq is singleton.
Thus, v � pπpvq,εpvqq since pπpvq,εpvqq is contained in the fiber, by Theorem 4.2.11.
That K is the extended graph of ε is a consequence of the form of K as estab-
lished by the theorem and the fact that extended graphs are unique by Lemma 4.2.8.
l
The space Lν{p is a tropical linear subspace of L . It equals L if and only if
ε :L Ñ F is identically 8, but otherwise it is of codimension 1. One should see Lν{pas the locus of vanishing of ε. In the setting of tropical linear spaces there is no harm
in referring to it as the cycle of ε. See also [57] and [25].
Proposition 4.2.13 Linear functions L Ñ F are elementary extension functions. Anelementary extension function L Ñ F is the pointwise maximum of a finite number oflinear functions.
Proof. Let φ : L Ñ F be a linear extension. Then φ extends to a linear function
FR Ñ F . We denote the extension by φ as well. Set εpIq � φpvIq, for I a corank 1
independent subset of µ and εpIq � 8 otherwise. We first need to show that ε satisfies
the conditions of Theorem 4.14. Condition (i) is fulfilled by definition, and condition
(ii) is true by proportionality of canonical generators vI and vJ , when clpIq � clpJq,by Lemma 4.1.4. The third condition is a priori non-trivial. We prove an equivalent
statement, that is trivially stronger.
Let X , Y � R with |Y | � d�1 and |X | � d�2. We aim to show that the expression,
àiPY zX
µpY � iq d εpX � iq �à
iPY zX
µpY � iq dφpvX�iq (4.17)
�à
iPY zX
àjPRzpX�iq
φpe jq dµpY � iq dµpX � i� jq
100 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
vanishes tropically. Let i� P Y zX and j� P RzpX � i�q such that the minimum in the
above expression is attained for j � j� and i � i�. Then the above expression equals,
àiPY zX
φpe j�q dµpY � iq dµpX � j�� iq.
By the valuated exchange condition for µ the minimum is attained at i � i� and at
least one other index i. This shows that the minimum of (4.18) is attained at least
twice.
Let ε : L Ñ F be the extension function of the extension K of L and fix an
elementary extension ν of µ such that Lν � K . Let Sν denote the set of dependent
S � R such that S� p is a basis of ν . For each S P Sν denote the function,
w ÞÑ νpS� pq�1 d¸
iPS�p
νpS� iq dwi , w PL ,
by φS . We claim that ε � maxtφS | S P Sνu. For each v P L the values vp P F such
that pv, vpq PK are defined by the vanishing of the expressions,
� ¸iPS�p
νpS� pq�1 d νpS� iq d vi
` vp,
where S runs over the elements of Sν . Thus, εpvq �maxtφSpvq | S P Sνu. l
Corollary 4.2.14 Any elementary extension function of a valuated matroid of rank 2
induces a linear function on the associated tropical linear space, in the sense of Theorem4.2.11.
Proof. Let L be a tropical linear subspace of FR of rank 2 and let µ be a matroid
associated to L . Then a function ε :P1pRq Ñ F is an elementary matroidal extension
function if and only if the expressions,
µpi jq dαptku`µpikq dαpt juq`µp jkq dαptiuq
vanish tropically, for all distinct i, j, k P R. Comparing these expression with the ex-
pressions definingL as a prevariety yields that α is an elementary extension function
if and only if the vector�αptiuq
�P FR is an element of L .
Let Mµ denote the matrix whose pi, jq-th coefficient is the µpti, juq. The rows of
Mµ are the canonical generators of L and hence the row space of Mµ is precisely L .
Since a linear function is determined by its image on the canonical generators, and
every linear function onL is the restriction of a linear function FR Ñ F it follows that
the column space of Mµ is isomorphic to the module of linear functions on L , under
4.2. FUNCTIONS AND THEIR GRAPHS 101
the isomorphism that maps such a function α : L Ñ F to the vector�αpvtiuq
�P FR.
Since Mµ is symmetric (this is the feature that distinguished the rank 2 case from
cases of higher rank) the image under this isomorphism is precisely L itself.
We have shown that both the space of linear functions and the space of elementary
extension functions are isomorphic to L , under the same isomorphism, when one
takes into account Theorem 4.2.11 and hence coincide. l
Proposition 4.2.15 Let µ be a valuated matroid on R over F which is realizable overthe valued field pK ,ωq, with ω : K Ñ F surjective. Then µ�α p is realizable over analgebraic extension of pK ,ωq when α is linear.
Proof. Denote the rank of µ by d and assume without loss of generality that R � rns.There exists a matrix M of size d� n with entries in K such that µpAq �ωpdet MrAsq,where MrAs denotes the full submatrix of M whose columns are indexed by the
elements in A. Let x1, . . . , xn be indeterminates and consider the linear function
f : pKdqn Ñ Krx1, . . . , xnsd defined by f pa1, . . . , anq � x1a1� . . .� xnan, with ai P Kd .
Denote the columns of M by mi , set m� f pm1, . . . , mnq and write M 1 for the d�n�1
matrix whose first n columns coincide with those of M and whose n� 1-th column is
m. For each A � rns of size d clearly ωpdet M 1rAsq � µpAq. Let A � rn� 1s of size dsuch that n� 1 P A. Then
detpMrA1sq �¸
iPrns
p�1qn�i x i detpMrA1� pn� 1q � isq
Assume without loss of generality that α : FR Ñ F . There exists an algebraic extension
of pL,ω1q of pK ,ωq and elements ai P L with ω1paiq � αpeiq such that for all A1 �rn� 1s of size d such that n� 1 P A,
ω1pdetpMrA1sq �àiPrns
ω1paiq dω1pdetpMrA1� pn� 1q � isqq
�àiPrns
αpeiq dµpA1� pn� 1q � iq
� αpvµA1�pn�1qq,
which proves the statement. l
Corollary 4.2.16 Let K be an algebraically closed field with surjective valuation ω :
K Ñ F and R a finite set. Then the image Gr2pKRq in its Plücker embedding under the
entry-wise application of ω equals Dr2pFRq in its realization of the space of proportional
equivalence classes of valuated matroids of rank 2.
102 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
The behavior of linear subspaces under arbitrary elementary extensions of the
containing linear space is complicated. Matters are somewhat simpler once restricted
to linear extensions.
Proposition 4.2.17 Let ι :Lµ �Lν be tropical linear spaces and φ :Lν Ñ F a linearfunctional. Then Lµ�ι�pφq p �Lν �φ p,
Proof. Denote the ground set of µ by R. Write d be the rank of µ and e for the rank of
ν . Denote µ�ι�pφq�p by µ1 and ν �φ p by ν 1. We need to show that for all I � R� pand S � R� p with |I | � d �1 and |S| � e�1 the function φS : FR�p Ñ F defined by
φSpeiq � ν1pS� iq if i P S and 8 otherwise, vanishes on vν
1
I .
We distinguish a number of cases. Write π for the canonical projection FR�p Ñ FR.
(i) Suppose p R S. Then π�pφSq vanishes on Lν . Since πpvIq P Lµ � Lν and
φSpepq � 8 it shows that φS vanishes on vI .
(ii) Suppose p P S and p R I . Then φS vanishes on Lν 1 and vI P Γφ1 � Γφ � Lν 1and hence φS vanishes on vI .
(iii) The most complicated case is when p P S and p P I . In effect we show that
πpvIq PLν 1{p. Here the linearity of φ is crucial. Consider the expression, where
we drop the superscripted matroids, and write wJ , respectively vJ for the fun-
damental generators of Lν , respectively Lµ,
φSpvIq �àiPSzI
ν 1pS� iq dµ1pI � iq
�àiPSzI
φpwS�i�pq dφpvI�i�pq
�àiPSzI
�àjφpe jqνpS� i� j� pq
d�à
k
φpekqµpI � i� k� pq
�àiPSzI
�àj,k
φpe jqφpekqνpS� i� j� pqµpI � i� k� pq
(4.18)
Suppose that the minimum in the above expression is attained in some triple
pi�, j�, k�q. Set S� � S � j�� p and I� � I � k�� p. There exists some i1 � i�
such that νpS� i�qµpI � i�q ¥ νpS� i1qµpI � i1q, since Lµ �Lν and hence the
minimum in 4.18 is also attained at pi1, j�, k�q. This shows that φSpvIq vanishes
tropically.
l
Example. As already stated, the behavior of general elementary extension functions
under restriction is much more complicated. In general such a restriction is not an
elementary extension function for the linear subspace and examples are not difficult
to construct.
4.2. FUNCTIONS AND THEIR GRAPHS 103
Consider the so-called rigid line L in the tropical linear spaceM over B associ-
ated to the matroid U3pt1,2, 3,4uq. This is example [25]. Then L is the cycle ofMcut out by the extension function ε defined by ε : vi j ÞÑ 8 if and only if ti, ju � t1,2u
or t3, 4u, and 0 otherwise. The restriction of ε itself to L is not an extension func-
tion. After all, L � tv12, v34u � tp8,8, 0, 0q, p0, 0,0, 0q, p0,0,8,8q, p8,8,8,8qu
and εp0, 0,0, 0q � 0. In canonical coordinates,
p0, 0,0, 0q � 0d p8,8, 0, 0q ` 0d p0, 0,8,8q,
and thus ε does not satisfy Theorem 4.2.11.
Let H denote the collection of hyperplanes of µ and suppose that the injection
ι : Extpµq ãÑ FH of Corollary 4.2.10 is dually residuated (for example when F � Bor when Extpµq is finitely generated). The map γ� ι � ι^ is a closure operator whose
image is the module of extension functions Extpµq � FH . A possible construction that
presents itself naturally now is the following.
Let ν be a valuated matroid on R such that j : Lµ � Lν and let ε P Extpνq. Then
j�ε is not necessarily a extension function of µ, however γp j�εq is. Then construct
µ1 � µ�γp j�εq p. However, by the following example Lµ1 �Lν�εp.
This following is an example from Chapter 7 of [50] of two incompatible single
element extensions.
1
2
3
4
5
6
(a) the base matroid ν
1
2
3
4
5
6p
(b) the extension of the codi-mension one subspace of Lν
1
2
3
4
5
6q
(c) the matroid of the tropicallinear space Lν �φ q
Figure 4.7: Two non-linear extensions of a matroid µ
The contraction of Figure 4.7(b) by p is the rank two matroid with atoms (and
hyperplanes) t3u, t6u, t2,5u, t1,4u. We denote it by µ. The restriction j�φ of φ toLµmaps vµ3 ÞÑ 0, vµ6 ÞÑ 0, vµ25 ÞÑ 8 and vµ14 ÞÑ 8. This is not an extension function. The
closure γp j�φq is the function that is identically 8. This follows since all extension
functions on a rank 2 matroid are linear. And hence the extension µ �γp j�φq q is
µ`U0pqq, i.e. it adds q as a loop. The corresponding linear space is not a subspace of
Lν �φ q.
104 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
Linear extensions, directed networks and linking systems
This section aims to give a different interpretation of linear extensions of a matroid in
terms of matroids induced by graphs. Such an interpretation makes it easier to handle
a large of number of extensions, and in particular, leads the way to a better picture of
the tropical linear monoid, to be defined later.
The central concept in this section is that of a linking system, or rather of the
valuated analogue thereof. Linking systems were introduced in [56] to generalize
induction of a matroid through a directed network or a bipartite graph. Independently,
they came up in [42], in essence to abstract the patterns of singular minors of a
matrix. In this context there is an interesting link with the Bruhat decomposition of
the general linear group, which is expanded upon in the articles [46]. Finally, valuated
linking systems were introduced by Murota. We mention the articles [47] and [40].
The next definition is from [47]. The definition of a linking system is derived
from that of a matroid, and hence they are really nothing new. However, they form a
useful device to put certain results on valuated matroids that are induced from other
valuated matroids on directed networks in a nice context. Another reason to define
them here is that they are a useful vehicle to prove statements on the induction of
valuated matroids through directed networks and bipartite graphs, as we shall see
later.
Definition 4.2.1 (Linking system) Let R and S be finite sets. A function λ : P pRq �P pSq Ñ F is called a linking system on pR, Sq over F when the map µλ :P pR>Sq Ñ Rdefined by,
µλpX q � λpRzX , SX X q
is a valuated matroid on R > S over F satisfying µλpRq � 0. The map µλ is referred to
as the graph, or the representation matroid, of λ.
The name representation matroid comes from [42]. The next example of a class
of valuated matroids is the key to [42]. Over B linking systems abstract the patterns
of singularity and non-singularity of a matrix, in the same way that matroids abstract
the patterns of singularity of full submatrices (i.e. linear dependence).
Example. There is a notion of realizability for valuated linking systems, that reduces
to realizability of the representation matroid. Let K be a field equipped with some
valuation ω : K Ñ F and let A be a matrix over K whose rows and columns are
indexed by R, respectively S. Then the function that assigns to each equicardinal pair
pX , Y q of subsets X � R and Y � S the value ωpdetpArX , Y sq, where ArX , Y s denotes
the submatrix with rows indexed by X and columns indexed by Y , is a linking system.
Its representation matroid is realizable as well. The matrix that realizes it is pIR Aq,
4.2. FUNCTIONS AND THEIR GRAPHS 105
where IR is the identity matrix, with entries in K , whose rows and columns are indexed
by R.
The exchange condition for valuated matroids translates easily to similar defining
conditions for valuated linking systems. For a proof of the equivalence we refer the
reader to [47]. We refer to R as the domain set or row set of and to S as the range setor column set of λ.
Proposition 4.2.18 A map λ :P pRq�P pSq Ñ F is a linking system if and only if bothof the following symmetrical vanishing conditions are satisfied.
(i) For each pX1, Y1q, pX2, Y2q PP pRq�P pSq and i1 P X1zX2 there is either j1 P Y1zY2
such that,
λpX1, Y1q dλpX2, Y2q ¥ λpX1 � i1, Y1 � j1q dλpX2 � i1, Y2 � j1q
or there is i2 P X2zX1 such that,
λpX1, Y1q dλpX2, Y2q ¥ λpX1 � i1 � i2, Y1q dλpX2 � i1 � i1, Y2q
(ii) For each pX1, Y1q, pX2, Y2q PP pRq�P pSq and j1 P Y1zY2 there is either i1 P X1zX2
such that,
λpX1, Y1q dλpX2, Y2q ¥ λpX1 � i1, Y1 � j1q dλpX2 � i1, Y2 � j1q
or there is j2 P Y2zY1 such that,
λpX1, Y1q dλpX2, Y2q ¥ λpX1, Y1 � j1 � j2q dλpX2, Y2 � j1 � j2q
Proof. See [47]. l
Constructions that produce linking systems from linking systems induce construc-
tions on the associated representation matroids. We define the cofactor linking sys-
tem and the restricted linking system, and prove, through the vehicle of the repre-
sentation matroid, that they indeed yield linking systems. These are the construc-
tions we need for what follows. We start by defining them on an arbitrary function
λ :P pRq �P pSq Ñ F .
(i) The cofactor system of λ, denoted λK, is the function P pRq�P pSq Ñ F defined
by λKpX , Y q � λpRzX , SzY q.(ii) Let pX , Y q P P pRq � P pSq. The restriction!of a linking system of λ to pA, Bq,
denoted λrA, Bs, is the function P pAq �P pBq Ñ F defined by λrA, BspX , Y q �λpX , Y q.
The next statement proves that these indeed define valuated linking systems and gives
106 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
the induced construction on the representation matroids.
Proposition 4.2.19 Let λ :P pRq �P pSq Ñ F be a function.
(i) The cofactor system λK is a linking system if and only if λ is a linking system andλpR, Sq � 8. Its representation matroid is the orthogonal dual of µλ.
(ii) If λ is a linking system and pA, Bq PP pRq�P pSq then λrA, Bs is a linking system.Its representation matroid is proportional to the representation matroid
�µλ � pAY
Sq�rAY Bs.
Proof. We prove the claim in order stated.
(i) Let λ be a linking system. Since the map λ ÞÑ λK is an involution, it is sufficient
to prove that λK is a linking system with representation matroid as stated. Let pX , Y q PP pRq �P pSq. Then,
λKpX , Y q � λpRzX , SzY q
� µλpX > pSzY qq
� µλppR > SqzpRzX > Y qq
� µKλ pRzX > Y q
Moreover, µKλ pRq � µλpSq � 8.
(ii) That λrA, Bs is a linking system follows immediately from Proposition 4.2.18.
The representation matroid of λrA, Bs is the matroid on A> B defined Z ÞÑ µλpRzZ >SX Zq, for Z � A> B. Consider the image of Z under µ1 �
�µ � pAY Sq
�rAY Bs,
µ1pZq � pµ � pAY SqqpZq
� µpZ Y pRzAqq
� µppRzZq Y pSX Zqq,
where the proportionality, denoted �, is a consequence of Lemma 4.1.11. l
The significance of linking systems is that they induce a matroid on the range set
from a matroid on the domain set. The construction is as follows. Let λ be a linking
system on pR, Sq over F . Let µ be a matroid over F on R. The image of µ under λ is the
map P pSq Ñ R given by, in its independent set characterization (cf. Section 4.1.2),
Y ÞÑàX�R
µpX q dλpX , Y q, Y � R. (4.19)
We denote this image by λpµq, by slight abuse of notation. The content of the next
statement is that λpµq is indeed a matroid.
4.2. FUNCTIONS AND THEIR GRAPHS 107
Proposition 4.2.20 Let λ be a linking system over F on pR, Sq and µ a matroid over Fon R. Then λpµq is a matroid over F on S.
Proof. First note that without loss of generality, that there exists a basis A of µ and a
set B � S of equal cardinality such that λpA, Bq � 8. One can always truncate µ such
that this is the case.
We prove that the representation matroid of λpµq is pµλ Y µq � S. This is a fairly
simple computation. First note that the matroid µλYµ has R as a basis, and hence it
follows by Lemma 4.1.11 that the contraction to S is the map that assigns to Y � Sthe value of pµλYµqpY Y Rq. It is thus sufficient to show that
µλYµpY Y Rq �à
XλpX , Y q dµpX q,
where X runs over the subsets of R of cardinality |Y |. This is an easy computation
once on realizes that there exists a basis B of µ such that RzBY Y is a basis of µλ. l
By the above proposition linking systems function as maps on the set of all ma-
troids with ground set the domain set of the system. We pursue this analogy further.
Let κ and λ be linking systems on pR, Sq and pS, Tq. The composition of κ and λ is the
map P pRq �P pTq Ñ F defined by,
pX , Zq ÞÑàY�S
κpX , Y q dλpY, Zq.
We denote it by λ � κ. The next proposition states that this indeed defines a linking
system and that composition behaves naturally with respect to the action of a linking
system on sets of matroids.
Proposition 4.2.21 Let κ and λ be linking systems on pR, Sq and pS, Tq. Then the com-position λ � κ is a linking system on pR, Tq. Moreover, pλ � κqpµq � λpκpµqq for anymatroid µ on R.
Proof. The proof is similar to that of Proposition 4.2.20. We compute the graph of
λ �κ, as in Definition 4.2.1. Set γ� λ �κ and let A� R\ T . Then,
µγpAq � γpRzA, AX Tq
�àY�S
κpRzA, Y q dλpY, AX Tq
�àY�S
µκppAX Rq Y pSX Y qq dµλppSzY q Y pAX Tqq.
This is precisely the value of A in the matroid pµκ `µ0pTqq Y pµ0pRq `µλq{S, where
µ0pX q, for X � R, T denotes the valuated matroid of rank 0 on X that assigns 0 to H.
108 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
The proof of the second statement follows immediately by expanding both expressions
according to their definitions. l
In matroid theory linking systems as presented in [56] simultaneously generalize
a number of known constructions of matroids from other matroids and a bipartite
graph or network. In the valuated case the same results holds.
Let F be a semiring and R an arbitrary set. An F-weighted network on R is a directed
graph on R equipped with a weight function that assigns an element of F� to every
arc of the network. A node is a source node when it has no incoming arcs, and it is a
terminal node, or sink node when it has no outgoing edges.
A weighted directed network Γ induces a function on pairs of equicardinal subsets
of the nodes. Let pX , Y q be such a pair of subsets. A linking between X and Y is a
collection of node disjoint paths from a node in X to a node in Y . The weight of alinking is the tropical product (ordinary sum) of the weights of all paths contained in
it, where the weight of the path is the tropical product of the weights of the edges that
make up the path. The function induced from Γ, denoted λΓ, assigns to each pX , Y qthe weight of a linking between X and Y of minimal weight. In other words, it assigns
the tropical sum over all linkings between X and Y to pX , Y q.
The construction is similar for bipartite graphs. Let ∆ be a bipartite graph with
color classes R and S with edge weights in F . To each pair pX , Y q with X � R and
Y � S and of equal cardinality, the function λ∆ assigns the minimal weight of a
matching between X and Y .
Both constructions are intimately related. The precise statement is that a linking
systems induced by a directed network is the cofactor linking system of one induced
by a bipartite graph.
Proposition 4.2.22 Let λ : P pRq2 Ñ F be a function. Then λ is the weight functionof an F-weighted directed network on R if and only if λK is the weight function of anF-weighted bipartite graph on pR, Rq.
Proof. Follows the proof of Lemma 2.4.3 of [50], which establishes a bijection be-
tween vertex disjoint path systems in matchings and networks in weighted graphs.
l
The above statement is also the means to prove that the functions induced by
directed networks are indeed linking systems.
Proposition 4.2.23 Let F be a tropical semifield. The following statements hold.
(i) The weight function of an F-weighted bipartite graph is a linking system.(ii) The weight function of an F-weighted directed network is a linking system.
4.2. FUNCTIONS AND THEIR GRAPHS 109
Proof. By Lemma 4.2.19 and Proposition 4.2.22 it is sufficient to prove the first state-
ment. Let ∆ be a bipartite graph and let pX1, Y1q and pX2, Y2q be pairs whose image
under the induced function λ∆ are not 8. If they were, the exchange condition holds
trivially. Let M1 and M2 be matchings between pX1, Y1q and pX2, Y2q that realize the
minimal weight. Let i P X1zX2. There is a path in M1 YM2 that starts in i and ends in
either Y1 or in X2zX1.
Suppose the former and denote the end point by j. Then pX1 � i, Y1 � jq and
pX2 � i, Y2 � iq are matchings having the same combined edges as M1 and M2. Thus,
the sum of their weights are equal and hence λ∆pX1� i, Y1� iqdλ∆pX2� i, Y2� iq is
at most the combined weight of M1 and M2.
Now, suppose that the path p ends in j P X2zX1. Create two new matchings M 11
and M 12 between pX1 � i� j, Y1q and pX2 � i� j, Y2q by
M 11 � M1zpYte | e P pXM2u
M 12 � M2zpYte | e P pXM1u.
In effect, M 1i is obtained from Mi by flipping the edges in p. Clearly, the combined
weights of M 11 and M 1
2 is that of M1 and M2, and hence λ∆pX1 � i � j, Y1q dλ∆pX2 �
i� j, Y2q is smaller. l
Given linking systems κ and λ with the same domain set R, but disjoint range
sets S and T , over the same tropical semifield one constructs a new linking system on
pR, S > Tq by,
pX , Y q ÞÑàκpX1, Y X Sq dλpX2, Y X Tq,
where the sum ranges over the partitions pX1, X2q of X . Elementary linear extensions
are the result of applying a linking system of a very specific form to the base matroid.
This is the content of the next proposition.
To state it we require a simple construction. The identity linking system on R,
denoted idR is the linking system induced by the bipartite graph with color classes Rand a disjoint copy of R, and edges pi, iq with i P R all of weight 0. In other words,
idRpX , Y q � 0 if X � Y and 8 otherwise.
Proposition 4.2.24 Let λ be a linking system on pR, tpuq over R and µ a matroid on Rover R. Let idR be the identity linking system on R and set ν �
�idR �λ
�pµq. Then ν is a
linear extension of µ. Vice versa, any linear extension is obtained in this way.Moreover, idR �λ is induced by the weighted bipartite graph on pR, R� pq with edge
weight function,
pi, jq ÞÑ
$'&'%
0 if i, j P R and i � j8 if i, j P R and i � jλptiu, tpuq if j � p.
110 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
Proof. Let X � R� p be a set of size rkµ. If X � R then the image of X under λpµq is
the image of X under idRpµq which equals µpX q. On the other hand, if p P X , then,
pidR �λqpµqpX q �àiPX
idRpX � p, X � pq dλptiu, tpuqdµpX � p� iq
� φpvµX�pq,
where φ is the restriction to Lµ of the map FR Ñ F that is determined by ei ÞÑ
λptiu, tpuq.The second statement is a consequence of the same computation backwards,
where we set φ to be the restriction to Lµ of the map ei ÞÑ λptiu, tpuq. The state-
ment on the bipartite graph follows immediately by writing down the linking system
associated to the weighted bipartite graph given. l
Another way to describe the linking system in the proposition is by means of
weighted directed networks. This is more convenient for iterated linear extensions.
Consider pidR � λq as in the proposition. Let Γ denote the graph on R� p with arc
weights pi, jq ÞÑ 8 if j � p and pi, pq � λptiu, tpuq. It is obtained from the bipartite
graph in the lemma by contracting the edges pi, iq, with i P R, and orienting the
remaining edges towards p. It hence carries the same information as this bipartite
graph. Now a simple computation shows that λΓrR, R�ps � idR�λ. We have depicted
an example of both the bipartite graph and the directed network associated to a linear
function φ : F4 Ñ F in Figure 4.8.
φ1 φ2 φ3 φ4
1 2 3 4
1 2 3 4 p
(a) Bipartite graph through which a val-uated matroid induces its elementary ex-tension associated to φ.
φ1 φ2 φ3 φ4
1 2 3 4
p
(b) The simpler directed network associ-ated to the bipartite graph. The edges aredirected towards p.
Figure 4.8: Bipartite graph and network associated to the elementary extension asso-ciated to the linear function φ : F4 Ñ F .
Consider the sequence of elementary linear extensions,
µ� µ0φ1 // µ1
φ2 // . . . φm // µm . (4.20)
Repeating the procedure above allows one to represent µm as the image of µ under
a linking system that comes from a directed network. Denote the ground set of µi
4.2. FUNCTIONS AND THEIR GRAPHS 111
by Ri and write tpiu � RizRi�1. The function φi is a linear map FRi Ñ F and hence
different choices of φi might restrict to the same function of Lµi�1. Build up the
directed network inductively in m steps and denote the weighted network obtained
after the i-th step by Γi .
(i) Γ0 is the network with node set R and no arcs,
(ii) Γi�1 is obtained from Γi by adding pi�1 to the node set and arcs p j, pi�1q with
weights φi�1pe jq to the arc set, for j in the node set of Γi .
Note that the node set of Γi is exactly Ri . The next theorem proves that the iterated
linear extension µm of µ is represented by a restriction of the linking systems induced
by Γm.
Theorem 4.2.25 Consider the sequence (4.20) of linear extensions and set Γ � Γm.Then µm � λΓrR, R Y tp1, . . . , pmuspµq. Moreover, λΓrR, R Y tp1, . . . , pmus � idR �
λΓrR, tp1, . . . , pmus.
Proof. The first statement is a consequence of Proposition 4.2.24. The last statement
follows since any vertex disjoint linking between pX , Y q must link YXR to YXR. Such
a linking only exists when Y X R� X . l
Example. This is an example for F � B, i.e. when the valuated matroids under consid-
eration are ordinary matroids. We remind the reader of the fact that a minimal linear
(functional) extension of a linear function Lν Ñ B is of the form φF for some flat Fof ν , where φF peiq � 0 if and only if i P F . Moreover, distinct flats are associated with
distinct linear extensions and the restriction of some φX to Lν equals the restriction
of φX to Lν , where X is the closure of X in ν .
1
2
3
4
5
6
7 8
(a) Iterated linear (principal) extension ofthe free matroid on t1,2, 3,4u
1 2 3 4
5
6
7
8
(b) A graph representing the iter-ated linear extension of (a)
Figure 4.9: Four-fold linear extension of a free matroid and graph Γ4
112 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
For an example, denote the matroid of the point configuration restricted to points
t1,2, . . . , iu in Figure 4.9 by µi , e.g. µ4 is the free matroid on 4 points. Then the
elementary extension µ6 of µ5 has the restriction of φt1,2,3,5u to Lµ5as extension
function. This restriction equals the restriction of φt1,2,3u, φt1,2,5u, φt1,3,5u, or φt2,3,5u
to Lν . Different choices of functional extension give rise to different linking systems
as in Proposition 4.2.24.
The graph associated to the canonical order of extension, i.e. µ4 Ñ µ5 Ñ µ6 Ñ
µ7 Ñ µ8 and minimal functional extensions of the matroidal extension functions is
depicted in Figure 4.9(b). Note that since we do not draw arcs of weight 8 the drawn
arcs necessarily have weight 0.
Proposition 4.2.26 Let λ be a linking system on pR, Sq. The image of the (trivial) ma-troid that assigns 0 to R under idR �λ is the representation matroid of λ.
Proof. The proof is a small computation. Denote the trivial matroid referred to in the
statement by µ and let Y � S be of size |R|. Then,
pidR �λqpµqpZq � idRpZ X R, Z X Rq dλpRzZ , Z X Sq dµpRq
� λpRzZ , Z X Sq
� µλ
��RzpRzZq
�\�Z X S
�� µλpZq,
which proves the statement. l
It is convenient to extend the action of linking systems on valuated matroids on the
appropriate ground set to tropical linear subspaces of an appropriate ambient affine
space. Formally, let λ be a linking system over F on pR, Sq and L � FR a tropical
linear space. Then the image of L under λ is defined to be the tropical linear space
associated to λpµq, where µ satisfies Lµ �L . Since the map on matroids induced by
λ is homogeneous, the image λpK q does not depend on the specific choice of µ.
Linking systems on pR, Sq over F fulfill the role of linear maps from FR to FS in
the sense that they map linear subspaces of FR (matroids on R) to linear subspaces on
FS (matroids on S). In the light of the previous theorem, the name graph for what is
called the representation matroid in [42] is appropriate in this setting. The situation
is more complex when we wish to restrict linking systems to proper subspaces of FR,
since in this case distinct linking systems on pR, Sq might represent the same map.
However, as in ordinary linear algebra, linear maps are represented by their graphs.
In the broader context of a linking system λ on pR, Sq restricted to a matroid µ on Rthe graph of the restriction of λ to µ is the matroid pidR � λqpµq. Linking systems λ
and κ, both on pR, Sq are said to define the same restriction to µ when their graphs
4.3. MORPHISMS BETWEEN LINEAR SPACES 113
pidR � λqpµq and pidR � κqpµq are equal. Note that when µ has rank |R| the graphs
are only equal when the linking systems are equal, by the above proposition and the
fact implicit in the definition that linking systems are uniquely determined by their
representation matroids.
4.3 Morphisms between linear spaces
In this section we use the results of Section 4.2.1 to construct the morphisms in a
category whose objects are tropical linear spaces. We let F denote a fixed semifield.
A morphism from L to M is a tropical linear subspace of L �M that is of the
form pidR �λqpL q, where λ is a linking system induced by iterated linear extensions
and a projection. The linking systems are fairly easy to handle, while the morphism
itself is not. Most of this section is devoted to proving that certain construction are
independent of the specific linking system that we chose to represent the morphism.
The central statement in the first part of this section (Proposition 4.3.3) is an example.
It states that restriction is well-defined. Unfortunately, we have not been able to prove
it. Its validity, and thus the validity of most of the theorems following it (including
Proposition 4.3.6, that states that composition is well defined) depends on Conjecture
4.3.1.
However, since a morphism with domain FR is uniquely defined by a linking sys-
tem (cf. Corollary 4.3.2), none of these problems occur when we consider only such
morphisms. In particular, HompFR, FRq has the structure of a monoid under com-
position. In Section 4.3.2 we define the tropical linear monoid as a submonoid of
HompFR, FRq and prove its most important properties in Theorem 4.3.7
4.3.1 A category of tropical linear spaces
The objects of the categoryB are the tropical linear spaces. By definition, a morphismα in B from a tropical linear space L to a tropical linear space M , denoted α :
L ÑB M , is a tropical linear subspace of L �M of the same rank as L , that is
the projection of an iterated linear extension of L . In other words, α P DrdpFR� FSq,
where d � rkL , L � FR andM � FS , is a morphism L ÑB M when there exists a
set T � R > S and a tropical linear space K � F T such that,
(i) there exist a filtration R � T0 � T1 � T2 � . . . � Tk � T , with |Ti�1zTi | � 1 and
a sequence of tropical linear spaces L �K0 �K1 �K2 � . . . �Kk �K such
that Ki � F Ti and Ki�1 is a linear extension of Ki ,
(ii) the image of K under the canonical projection to FR � FS equals α.
When confusion is unlikely to occur we drop the subscript B and write α : L ÑMinstead. In this section HompL ,M q denotes the set of all morphisms with domain L
114 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
and rangeM .
The next theorem is concerned with obtaining a representation of morphisms in
terms of linking systems induced by certain weighted acyclic networks. It has as an
important corollary that morphisms between affine spaces can be identified with such
linking systems.
Theorem 4.3.1 Let Γ be an F-weighted acyclic directed network with source vertices Rand sink vertices S having R X S � H. Let K � FR be a tropical linear space. ThenpidR �λΓrR, SsqpK q is a morphism K Ñ FS . Moreover, all morphisms in HompK , FSq
are of this form.
Proof. Denote the edge weight function of Γ by w. Consider the topological orderon the nodes N of Γ, i.e. p ¤1 q if there is a directed path in Γ from p to q. Fix an
arbitrary linear refinement ¤ of ¤1. Let p1 be the ¤-minimal element of NzR. Then
all arcs with end point p start in R. Let φ1 be the linear function FR Ñ F defined by
φpeiq � wpi, p1q. Add the rest of the nodes inductively. Finally one obtains a tropical
linear space in F N , which equals pidR � λrR, NzRsqpK q by Theorem 4.2.25, and is
an iterated linear extension of K . Finally, the projection of this space to R \ S is
pidR �λrR, SsqpK q P HompK , FSq.
If α P HompK , FSq. Then there exist T � R\ S and α1 P DrpF T q such that α1rR\Ss � α and α1 is an iterated linear extension of K , i.e.,
Kφ1 // K1
φ2 // K2φ3 // . . . φm // α1
Choose a matroid associated to K and construct the weighted acyclic network Γm
as in the paragraph preceding Theorem 4.2.25. By the preceding paragraph pidR �
λΓmrR, Ssq � α. l
Let L � FR andM � FS be tropical linear spaces. A linking system λ on pR, Sqover F is said to represent a morphism α P HompL ,M q when α � pidR � λqpL q. In
this terminology the previous theorem states that every morphism is represented by
the restriction of a linking system induced by a weighted directed acyclic network. It
is an immediate consequence of Proposition 4.2.26 that a linking system representing
a morphism with domain FR is unique. This is the content of the next statement.
Corollary 4.3.2 The map λΓrR, Ss ÞÑ pidR � λΓrR, SsqpFRq implicit in the theorem isbijective.
In essence, the corollary combined with the theorem state that any morphism
K Ñ L is a restriction of a morphism FR Ñ FS , where K � FR and L � FS . This
statement has a more general form, in the sense that a morphism on K restricts to a
4.3. MORPHISMS BETWEEN LINEAR SPACES 115
morphism on any subspace. The proof is by representing a morphism as a linking sys-
tem. The main difficulty lies in proving that different representations yield the same
restriction. As a matter of fact, the proof of the statement depends on the following
conjecture, which we have not been able to prove.
Conjecture 4.3.1 Let R� S be finite sets and denote the canonical projection π : FR Ñ
FS by π. Consider K �L � FR and suppose that there exists an iterated extension,
σ : πpK q // . . . // K ,
such that the iterated extension τ : Lztpu Ñ L associated to the order on RzS corre-sponding to σ is the restriction. Then for any p P RzS the extension function LztpuÑLis the restriction of the extension function KztpuÑK .
Proposition 4.3.3 Suppose Conjecture 4.3.1 is true. Let α P HompK ,L q be a mor-phism, for K � FR and L � FS . Let ξ be a matroid such that Lξ � K . For λ and κlinking systems on pR, Sq representing the same morphism in HompK ,L q it holds thatpidR �λqpξq � pidR �κqpξq.
Proof. Let α be a morphism in HompK ,L q. Let FR and FS be the ambient spaces of
K and L . Assume without loss of generality that RX S �H. There exists a tropical
linear spaceM � F T , with R\ S � T , whose image under the canonical projection
π : F T Ñ FR � F T equals α. Moreover,M is an iterated linear extension of K �K0,
as follows,
K0φ1 // K1
φ2 // K2φ3 // . . . φm //M (4.21)
Write T � R \ tp1, . . . , pmu, where R \ tp1, . . . , piu is the index set of the ambient
affine space containing Ki . Fix a linear order on RzT by pi ¤ p j if and only i ¤ j.Write S � tq1, . . . , qnu � TzR, with qi ¤ q j for i ¤ j. The situation is summarized in
the next diagram.
K0φ1; p1 //
ε1; q1
!!BBB
BBBB
BK1
φ2; p2 // K2φ3; p3 // K3
φ4; p4 // K4φ5; p5 // . . . φm; pm //M
π
��L1
ε2; q2 // L2ε3; q3 // L3
ε4; q4 // L4ε5; q5 // . . . εn; qn // α,
(4.22)
in whichLi is the projection ofM to R\tq1, . . . , qiu and εi is defined by the diagram,
i.e. it is the unique extension function associated to the extension Li of Li�1.
Now letK 1 �K andK 1i �Ki , whereK 1
i is defined iteratively to be the extension
of K 1i�1 associated to φ1i , with φ1i the restriction of φi to Ki . Again, set L 1
i to be
the projection of K 1m to FR � Ftq1,...,qiu. Thus, L 1
i � Li and we have the diagram of
116 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
subspaces,
K 10
φ11; p1//
ε11; q1
AA
AAAA
AAK 1
1
φ12; p2// K 1
2
φ13; p3// K 1
3
φ14; p4// K 1
4
φ15; p5// . . .
φ1m; pm// K 1
m
π
��
L 11
ε12; q2// L 1
2
ε13; q3// L 1
3
ε14; q4// L 1
4
ε15; q5// . . .
ε1n; qn// L 1
n,
(4.23)
The aim is to prove that L 1n does not depend on the specific choice of iterated se-
quence, but only on α. Since α is uniquely determined by K and ε1, . . . ,εn it suffices
to prove that ε1i is the restriction of εi to L 1i�1. We then set L 1
n to be the restricted
morphism.
By induction on n we prove a slightly more general statement. Consider the dia-
grams (4.22) and (4.23) in which the φi are general matroidal extension functions
that restrict to matroidal extension functions φ1i on the K 1i�1 (as opposed to linear
functions). Then ε1i is the restriction of εi .
If n � 1, then without loss of generality we may assume that q1 � pm. We start
by examining (4.23). Write ρ1i for the dual residual of the canonical projectionK 1i Ñ
K 1i�1 and denote the dual residual of the canonical projection L 1
1 ÑK1
0 by ρ1. Note
that ρ1ipvq � pv,φ1ipvqq and ρ1pvq � pv,ε1pvqq, by Theorem 4.2.11. Write πi for the
canonical projection FR� Ftp1,...,piu� FtpmuÑ FR� Ftp1,...,pi�1u� Ftpmu. Then repeated
application of Lemma 4.3.1 yields that
ρ1 � π1 �π2 � . . . �πm�1 �ρ1m �ρ
1m�1 � . . . �ρ11.
The same procedure, applied to the containing sequence (4.22), gives that,
ρ � π1 �π2 � . . . �πm �ρm �ρm�1 � . . . �ρ1.
By the above expressions for ρ and ρ1, and the fact that ρ1i is the restriction of ρi it
follows that ρ1 is the restriction of ρ to K 10 .
Suppose that the statement is true for all integers smaller than n. By the previous
paragraph ε11 is the restriction of ε1 to K 10 . Let j � 1, . . . , m such that q1 � p j . The
restriction of M to R \ tp1, . . . , p ju. For each i ¥ j, denote by π� the canonical
projection of FR � Ftp1,...,piu to FR � Ftp j ,...,piu. Note that we use the same symbol to
denote different projections. Consider the diagram of elementary extensions,
π�pK jqψ j�1
//
ε2
%%LLLLLLLLLLLπ�pK j�1q
ψ j�2// . . .
ψm�1// π�pKm�1q
ψm�1// π�pM q
��L2
ε3 // . . .εn�1
// Ln�1εn // α,
4.3. MORPHISMS BETWEEN LINEAR SPACES 117
obtained from (4.22) by applying π� to the upper row. Note that π�pK jq �L1, which
has become the base in the diagram to extend from.
Denote the extension from π�pK 1i�1q to π�pK 1
i q by ψ1i . If it is true that ψ1
i is the
restriction of ψi to π�pK 1i�1q we can use induction and conclude the truth of the
statement. Here we make use of Conjecture 4.3.1. l
Let L � FR andM � FS . Let α P HompL ,M q be a morphism and K a tropical
linear subspace of L . The restriction of α toK is the morphism pidR�λqpK q, where
λ is a linking system representing α. The previous proposition shows that restriction
is well-defined, if Conjecture 4.3.1 is true. Moreover, since linear extensions respect
subspaces, it is a morphism K Ñ M (as opposed to a morphism K Ñ FS). We
denote the restriction of α to K by α|K .
The proposition above also allows us to generalize the action of linking systems
on tropical linear spaces, defined at the end of Section 4.2.2. A natural way is to
represent the morphism by a linking system and have the linking system act on the
tropical linear space. It is a consequence of the previous proposition and the next
lemma that this is well-defined.
Lemma 4.3.4 Let L � FR be a tropical linear space and λ a linking system on pR, Sqover F. Then λpL q � pidR �λqpL qrSs.
Proof. Let µ be a valuated matroid on R over F associated to L . For Y � S,
pidR �λqpµqpY q �à
X�YXRλpX zY, Y X Sq dµpX q �
àXλpX , Y q dµpX q,
where X � R. This proves the statement. l
Let L � FR be a tropical linear space. Denote the collection of all valuated ma-
troids on R over F whose associated linear space is a subspace ofL by Matpµq, where
µ is a valuated matroid satisfying L � Lµ. By the characterization of valuated ma-
troids associated to a tropical linear space DrpL q is the projectivization of Drpµq. In
particular, the latter only depends on the proportional equivalence class of µ.
Given a tropical linear spaceM � Lν � FS and a morphism α : L ÑM define
a map Matpµq Ñ MatpSq by Matpαqpξq � λpξq, where λ is a linking system on pR, Sqrepresenting α. By Proposition 4.3.3 the image λpξq is independent of the choice of
representing linking system and hence Matpαq is well-defined.
Its most important properties are stated in the next proposition.
Proposition 4.3.5 Suppose Conjecture 4.3.1 is true. The map α ÞÑ Matpαq is injective.Moreover, the image of Matpαq is contained in Matpνq.
Proof. Let α P HompLµ,Mνq and ξ P Matpµq. Then by Proposition 4.3.3 the matroid
pidR � αqpµq is well defined. Since the tropical linear space associated to αpµq is
118 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
contained in α, by Proposition 4.1.11 the tropical linear space associated to αpµq �
pidR �αqpµqrSs is contained in Lν . Thus, αpµq PMatpνq.
We need to show that this map is injective. Let Y � R > S with |Y | � rkµ. Then,
�pidR �λqpµq
�pY q �
àX
pidR �λqpX , Y q dµpX q
�à
X�YXRλpX zY, Y X Sq dµpX q
�à
X�YXRλpX zY, Y X Sq dµppX zY q Y pY X Rqq,
where X runs over the basis of µ. In particular, if this expression is not equal to 8,
there exists a basis of µ containing Y XR, and hence Y XR is independent in µ. Thus,
the contraction�µ{pRX Y qqpZq � µpZ Y pRX Y qq (cf. Lemma 4.1.11). Thus,
�pidR �λqpµqpY qq �
àYXRPI pµq
λpµ{pRX Y q `υ0pRX Y qq,
with υ0pR X Y q the matroid of rank 0 on R X Y that maps H to 0. It remains to
show that ξ � µ{pR X Y q ` υ0pR X Y q is in Matpµq. This is a consequence of the
characterization of contraction and direct sum. The tropical linear space associated to
ξ is tv PLµ | vi �8 for i P clµpRX Y qu. l
By homogeneity of the map Matpαq it induces a map DrpL q Ñ DrpM q and hence
morphisms determine natural maps on subspaces, if Conjecture 4.3.1 is true. It is
clearly not the case that a morphism is determined by its map on subspaces and it
is not even evident that in general the morphism is determined up to tropical scalar
multiplication by its map on subspaces.
Example. A simple example to see that the morphism α is not determined by its
map on subspaces of L is given by HompFR, Fq. Such a morphism induces a map
DrpFRq Ñ DrpFq. The latter consists only of two points F and 8. Any morphism
HompFR, Fq is given by a single linear elementary extension of a free matroid on
R and hence by a linear function FR Ñ F . Two linear function that have the same
support, induce the same map DrpFRq Ñ DrpFq.
Example. We give a number of examples of the above map. Among other things it
serves to convince the reader that the map on points of a projectivized tropical linear
space (i.e. the map on rank 1 quotient matroids) is the projectivization of a linear
map, and that any such map can be obtained in this way.
(i) The first example is of the case F � B. Consider the free matroid µ on R� r3s,
i.e, µpRq � 0. The associated lattice of flats (or tropical linear space) is the power set
lattice of R. The projectivization of this lattice as a module is the poset RztHu, which
4.3. MORPHISMS BETWEEN LINEAR SPACES 119
we represent as the points in the barycentric subdivision of the triangle, as drawn in
Figure 4.10.
(∞,∞, 0)
(0,∞,∞)(∞, 0,∞)
(0, 0, 0)
(∞, 0, 0) (0,∞, 0)
(0, 0,∞)
Figure 4.10: Poset isomorphic to the projectivization of the tropical linear spaceLµ �B3 under the map v ÞÑ Rz supp v.
Note that the image of p8,8,8q under any morphism with domain B3 is always
the point p8, . . . ,8q and hence no information on the morphism is truly lost by pro-
jectivizing (this is peculiar to the case F � B). Let |S| � 3. A morphism α : BR Ñ BS
(or, with slight abuse of notation, a morphism B3 Ñ B3) defines a map on the points
of PpBRq (of which there are 7), and a map on the lines of PpB3q (of which there are 7
as well). Consider for example the morphism represented by network in Figure 4.11
and its induced map on subspaces of B3 of projective dimension 0 and 1.
7→
7→
7→
7→
7→
7→
7→
7→
7→
7→
7→
7→
7→
7→
λ ∼
Figure 4.11: The map induced by λ on the tropical linear subspaces of B3 of projectivedimension (rank) 0 and 1. It maps B3 to B3 (of projective dimension 2). The linesconsist of the points they contain.
The homogeneous map on points PpB3q Ñ PpB3q induced by λ equals pv1 : v2 :
v3q ÞÑ pv1 ` v3 : v2 : v1q. This is easily seen by realizing that such a projective point
has associated matroid µ : t1u ÞÑ v1, t2u ÞÑ v2, t3u ÞÑ v3 and thus that the action of λ
120 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
on such a point is
t1u ÞÑ λpt1u, t1uqdµpt1uq`λpt3u, t1uqdµpt3uq � v1 ` v3,
t2u ÞÑ λpt2u, t2uqdµpt2uq � v2,
t3u ÞÑ λpt1u, t3uqdµpt1uq � v1.
(ii) Example of the projective line over Q8. We aim to clarify figures 4.1(a) and
4.1(b). Morphisms with domain a tropical linear space of rank 2 are much easier to
understand than general morphisms, for the same reason that valuated matroids of
rank 2 are easier to understand. The essential difference between the rank 2 case and
the higher rank cases is that any elementary extension of a rank 2 valuated matroid
is linear. In particular, the set of morphisms HompL ,M q, with L of rank 2 are all
tropical linear subspaces of L �M of rank 2 that project surjectively onto L . We
give an example with L � F2.
The elementary extensions of F2 correspond bijectively to linear functions
px , yq ÞÑ pad xq ` pbd yq.
It remains to define composition of morphisms inB . To do so the next proposition
is required. First we state a construction that is required in the proof. Let R1, R2, S1, S2
be finite sets, which, without loss of generality, we can assume to be disjoint. For
linking systems κ on pR1, S1q and λ on pR2, S2q define the linking system pκ,λq on
pR1 > R2, S1 > S2q by the expression,
pκ,λqpX , Y q � κpX X R1, Y X S1q dλpX X R2, Y X S2q, .
for all X � R1 > R2 and Y � S1 > S2. This is easily seen to define a linking system,
either by using the definition (the representation matroid is the direct sum of the rep-
resentation matroids), or the exchange conditions for linking system (cf. Proposition
4.2.18). We denote it by pκ,λq.
Proposition 4.3.6 Suppose that Conjecture 4.3.1 is true. Let α P HompK ,L q and β PHompL ,M q. For any choice of representing linking systems λ for α and κ for β , thecomposition κ �λ represents the same morphism in HompK ,M q.
Proof. The proof relies on a decomposition lemma for the linking system pidR�pλ�κqq.
We prove that pidR�λ�κq � pidR,λq�pidR�κq. Let X � R and Z � R>T and consider
4.3. MORPHISMS BETWEEN LINEAR SPACES 121
the image of pX , Zq under the composition of idR �κ and pidR,λq,
pX , Zq ÞÑà
Y�R>Spid�κqpX , Y q d pid,λqpY, Zq
�à
Y�R>SYXR�X
κpX zY, Y X Sq d idRpY X R, Z X Rq dλpY X S, Z X Tq
�à
Y 1�S
κpX zY, Y 1q dλpY 1, Z X Tq
� pλ �κqpX zZ , Z X Tq
� idR � pλ �κqpX , Zq.
This proves the decomposition.
It remains to prove that the tropical linear space pidR � pλ �κqqpK q � pidR,λqpαq
depends only on α � pidR � κqpK q and β � pidS �λqpL q. Let X � R and Z � T and
compute,
pidR,λqpαqpX > Zq �àY�S
αpX Y Y q dλpY, Zq
Now Y 1 ÞÑ αpX YpY 1XSqq is a valuated matroid (it is α{X ) if X is independent in the
matroid associated to α (or equivalently, in the matroid associated to K ). Its tropical
linear space is contained inL , since it is the projection of α{X`υ0pRq, whose tropical
linear space is contained in α (by the same argument as used in Proposition 4.3.5).
Now,
pidR,λqpαqpX > Zq �à
XPI pµq
�λpα{X q
�pZq.
By Proposition 4.3.5 the summands are determined uniquely by the morphism β , and
hence pidR,λqpαq is determined uniquely by β . This concludes the proof. l
The composition of α and β in the proposition is the morphism in HompK ,M q
represented by κ�λ. Note that in fact the proposition states that the restriction of the
composition of appropriate extensions of morphisms does not in fact depend on the
specific extensions. This is a natural and somewhat expected condition.
4.3.2 The tropical linear monoid
Definition and basic results
This section specializes the result of the previous section to morphisms HompFR, FRq.
It ends with a conjecture on the structure of HompFR, FRq that depends on the graphs
representing the linking system (and hence the morphism, since morphisms are rep-
resented uniquely by linking systems, cf. Corollary 4.3.2.
The set HompFR, FRq carries a composition which makes it into a monoid with
122 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
identity the morphism represented by idR. Moreover, the results which depend on
Conjecture 4.3.1, hold without this dependence in HompFR, FRq (or in fact for any set
of morphisms whose domain is a tropical affine space), since in that case morphisms
are uniquely represented by linking systems and there is no need to prove the inde-
pendence of maps on quotients from the representing linking system. We will exploit
the latter fact by mostly working with the linking systems themselves, as opposed to
the morphisms.
Let F be a tropical semifield and n P N. The tropical linear monoid of rank n over Fis the submonoid of HompF n, F nq consisting of the morphisms (linking systems) that
map F n to F n. It is denoted TGLnpFq. By Proposition 4.2.21 it is indeed a submonoid
of HompF n, F nq.
Theorem 4.3.7 Let F be a tropical semifield and n P N. The following statements aboutTGLnpFq hold.
(i) There is a dually residuated surjective homomorphism from TGLnpFq to the subsetMnpFq of matrices with finite tropical determinant. The dual residual maps thematrix A to the linking system induced by the weighted bipartite graph on prns, rnsqwith edge weights ti, ju ÞÑ Ai j .
(ii) Let K be an algebraically closed field equipped with a surjective valuation ω :
K Ñ F. For each λ P TGLnpFq there exists a matrix A P MnpKq such that λ isrealized by A, i.e. λpX , Y q � ωpdet ArX , Y sq. Equivalently, the actual morphismpidrns�λqpF
nq is an element of ωpGrnpK2nqq.
(iii) Every element λ P TGLnpFq induces a map Matprnsq Ñ Matprnsq that preservesmatroid rank and the quotient relation. Moreover, λ is uniquely determined by thismap. In particular, λ determines a rank preserving map between tropical linearsubspaces of F n.
Proof. We prove the claim in order stated.
(i) The map π : TGLnpFq Ñ MnpFq defined by λ ÞÑ pλptiu, t juqi j is well-defined.
To show that its dual residual is as stated, it is sufficient to show that for any linking
system κ with πpκq � πpλq and all X , Y � rns of equal cardinality κpX , Y q ¥ λpX , Y q.By an iterated application of the exchange condition for linking systems, with X2 �H
and Y2 �H the inequality,
κpX , Y q ¥äiPXκptiu, tσpiquq �
äiPXλptiu, tσpiquq ¥ λpX , Y q,
holds for some σ : X Ñ Y . This proves the statement.
(ii) This is an immediate consequence of Proposition 4.2.15 and the fact that the
restriction of a valuated matroid realizable over some valued field K is again realizable
4.3. MORPHISMS BETWEEN LINEAR SPACES 123
over K .
(iii) The first statement is a specialization of Proposition 4.3.5. For the second
statement, let Γ be a directed weighted acyclic network in Γ that represents the mor-
phism λ. There is a vertex disjoint linking between the n source vertices and the nterminal vertices, that restricts to a matching of any subset of the source vertices to a
subset of equal size of the terminal vertices. Hence, rkλpµq � rkµ, for any matroid µ
on the source vertices. l
Example. This example serves to explain Figure 4.2 in detail. For convenience, denote
the domain set by R and the range set by S, say R� t1, 2u and S � t3,4u. A morphism
F2 Ñ F2 is a linking system on pR, Sq that assigns a finite value to the pair pR, Sq.Identifying such linking systems with their graphs (or representation matroids) yield
that TGL2pFq is the subset of Dr2pFR � FSq consisting of tropical linear spaces whose
image under the projection FR�FS Ñ FR is FR and whose image under the projection
FR � FS Ñ FS is FS . The condition that such a tropical linear space needs to be the
projection of an iterated linear extension is superfluous, since by Corollary 4.2.14,
any elementary extension is linear. In other words, there is a single directed network
Γ that yields all morphisms if we very the arc weights in F . It is depicted in Figure
4.12.
1 2
3
4
α1 α2
β3 β1 β2
Figure 4.12: Directed network inducing the morphisms in TGL2.
For example, to obtain the morphisms in the cone labeled 1 in Figure 4.2 choose
α1 � a, α2 � c, β1 � b, β2 � d and β3 � 8. To obtain the morphisms in the cone
labeled 3 one chooses α1 � a, α2 � c, β1 �8, β2 � e� a and β3 � b� a � d � c.
Any such choice of arc weights gives a linking system λΓrR, Ss. The matrix depicted
below the combinatorial type of the cone is,
��� λpt1u, t3uq λpt1u, t4uq 8
λpt2u, t3uq λpt2u, t4uq 8
8 8 λpt1,2u, t3,4uq
��
124 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
and hence TGL2 is a four-dimensional complex. Dividing out the actions of multiplica-
tion of the columns with a constant and the multiplication of the rows with a constant
gives the two dimensional complex depicted in Figure 4.2.
Complete intersections, transversal matroids and morphisms
An n�d dimensional tropical linear subspaceL of F n is called a complete intersectionif there exist d linear forms α1, . . . ,αd on F n such that L is the intersection of the
hypersurfaces αKi . The aim is to characterize the tropical linear subspaces that are
complete intersections, and use these results to study generic elements of TGLnpFq.
An essential ingredient in the characterization is Theorem 5.3. of First steps intropical geometry, which we will rephrase for convenience. Fix integers d ¤ n and let
A be a d � n matrix over F . Such a matrix determines a matroid ξA of rank d on rns,by the following expression,
ξApJq � detpAJq, J � rns and |J | � d.
Here AJ denotes the d�d submatrix of A consisting of the columns indexed by J . The
matroids that arise in such a way are the valuated analogues of transversal matroids,
but one can also think of them as matroids realizable over F .
Interpret the rows of A as linear forms α1, . . . ,αd on F n. Their intersection is a
tropical prevariety that contains the stable intersection of the hypersurfaces αKi . In
particular, it contains a tropical linear space of dimension n� d.
Lemma 4.3.8 Write LA for the linear space associated to ξA. The stable intersection ofthe hypersurfaces αKi equals LK
A .
Proof. We do the proof on the level of matroids. Let µi be the valuated matroid αi , i.e.
µip jq � αipe jq. By definition ξA is the matroid union of the matroids µi . This extends
the fact that a transversal matroid is a union of rank 1 matroids. The statement follows
by the facts that matroid intersection is dual to matroid union and that the linear
space of the intersection of matroids is the stable intersection of the associated linear
spaces. For the last statement see Tropical linear spaces. l
Now Theorem 5.3 establishes a necessary and sufficient condition for when LKA
equals the intersection of hyperplanes αKi .
Theorem 4.3.9 (Theorem 5.3) The tropical linear space LKA is contained in the inter-
section of hypersurfaces αKi . They are equal if and only if AJ is tropically non-singularfor all J � rns of size d.
4.3. MORPHISMS BETWEEN LINEAR SPACES 125
Proof. The first statement is an immediate consequence of the previous lemma. By
another method, it is easy to see that αi is contained inLA, since it is the image of the
point ei under the linking system coming from the bipartite graph on prds, rnsq with
weight function wpi, jq � Ai j .
Proof of the second statement is much more complicated and we refer to reader to
First steps in tropical geometry. It would be nice to have a purely combinatorial proof
though. l
We are now ready to prove our main claim.
Theorem 4.3.10 An n� d dimensional tropical linear subspace L of F n is a completeintersection if and only if its dual is of the form LA, where A is a d � n matrix over Fsatisfying the property that AJ is tropically nonsingular for every J � rns of size d.
Proof. Suppose thatL �LKA . By Theorem 4.3.9 the intersection of hypersurfaces αKi
equals LKA , where α1, . . . ,αd are the linear forms coming from the rows of A. Thus,
L is the complete intersection of the hypersurfaces αKi .
Suppose that L is the complete intersection of hypersurfaces αKi , i � 1, . . . , d. Let
A be the d�n matrix whose rows are the αi and consider the tropical linear spaceLA.
Since αi PLA it follows that L �LKA . Now it is sufficient to show that dimLA � d.
The dimension of LA equals the rank of the underlying matroid MA of ξA. But MA
is the transversal matroid of the set system S � pS1, . . . , Sdq, where j P Si whenever
the j-th coefficient of αi is finite. We need to show that S has a transversal. Suppose
that it does not. By Hall’s Marriage theorem (cf. Theorem 12.2.1 of [50]) there exists
an X � rds satisfying |X | ¡ |SX |, where
SX �¤iPX
Si .
The intersection of all αKi with i P X equals its full preimage under the projection
onto the coordinates SX , since there are no conditions on the coordinates outside of
SX . Hence it is a prevariety of dimension larger than n� |X |, contradicting the fact
that L is a complete intersection.
Thus, L � LKA and, by Theorem 4.3.9, the matrix A satisfies the required condi-
tion. l
Remark. The theorem above generalizes Theorem 6.3 of The tropical Grassmannianby providing both a converse and extending to arbitrary dimension. Note that in the
rank 2 case the projectivization of a linear space of the formLA is almost a caterpillar
tree, save for the fact that different legs might be attached to the same point of the
body. The condition that A is tropically non-singular precludes this. This is most easily
seen by considering ξA as the matroid obtained by extending the free matroid on
126 CHAPTER 4. A CATEGORY OF TROPICAL LINEAR SPACES
two elements by linear functions represented by the columns of A, say βi : F2 Ñ F ,
i � 1, . . . , n and deleting the two original coordinates. The tropical non-singularity
condition states that for two such functions βi1dβ j2 � βi2dβi1. This is equivalent to
the condition that βKi � βKj .
The application to the HompF n, F nq is via Proposition 4.2.22.
Theorem 4.3.11 Let λ P HompF n, F nq be a generic morphism in the sense that thematrix pλprnsztiu, rnszt juqi j satisfies the genericity condition of Theorem 4.3.10. Thenthere exists m¥ n and a tropical linear space L � F m such that,
(i) L is an iterated linear extension of F n,(ii) the image of L under the canonical projection F m Ñ F n equals pidn �λqpF
nq,(iii) the orthogonal dual of L has ξA as associated matroid, with A a matrix over F.
Proof. By definition of a morphism there exists an iterated linear extension L of F n.
This iterated linear extension is of the form pidB � λΓrB, RzBsqpF nq for some linking
system λΓ induced by the weighted directed network Γ on a set of nodes R, where Bis the set of n source nodes of Γ. Thus, L � λΓpF
Bq and a matroid µ associated to Lis Y ÞÑ λΓpB, Y q. The dual satisfies,
µKpZq � µpRzZq � λΓpB, RzZq � λΓKpRzB, Zq � λ∆pRzB, Zq,
for some weighted bipartite graph ∆ on pR, Rq. However, since the only nodes in a
basis on the domain nodes of ∆ are RzB it holds that λ∆pRzB, Zq � λ∆1pRzB, Zq,with ∆1 the induced subgraph of ∆ on pRzB, Rq. Thus, µK equals µA, where A is the
pRzBq � R-matrix over F with pi, jq-th entry λ∆ptiu, t juq � λΓpRztiu, Rzt juq. l
Chapter 5
The gossip monoid
Suppose somebody wishes to travel between three locations: the center of Eindhoven
(E), the center of Amsterdam (A) and the boundary of Amsterdam (B), and has two
types of transportation at its disposal: a bicycle, and a car. The time (in minutes) it
takes to travel between the different sites is given in the tables below.
bicycle E B A
E 0 630 640
B 630 0 20
A 640 20 0
car E B A
E 0 90 140
B 90 0 60
A 140 60 0
The car has a bike rack. This is convenient when travelling from E to A, since one
can take the car (and bicycle) to B and proceed on bicycle to A. The total travelling
time is 110 minutes. However, it doesn’t help when travelling from A to E. In that case
it is fastest to leave the bike and go by car, for a total travelling time of 140 minutes.
The time it takes to travel between A, B and E in this setup is given in the next table.
car and bicycle E B A
E 0 90 110
B 90 0 20
A 140 20 0
This is not a metric, since the travelling times are not symmetric and do not satisfy
the triangle inequality.
In this chapter we study the monoid of finite tropical products of distance matri-
ces. In this setting the situation above is represented by taking the product of two
distance matrices, one corresponding to the bike metric, the other to the car metric.The example shows that the tropical product of distance matrices is not necessarily a
127
128 CHAPTER 5. THE GOSSIP MONOID
distance matrix.
Section 5.1 gives the basic definitions, and establishes some easy results. Section
5.2 gives the relation with the gossip problem, discussed for example in [3]. It ex-
plains the name of this chapter. Section 5.3 deals with a possible way of giving a
polyhedral fan structure to the monoid. And finally, Section 5.4 explicitly constructs
the monoid generated by the distance matrices of size 3� 3 and that of size 4� 4.
5.1 Foundations
Let F be a tropical semifield and n a natural number. In this chapter it is more nat-
ural to write the semifield operations in the classical manner, and hence we have an
explicit inclusion F � R Y t8u such that x d y � x � y and x ` y � mintx , yu.Consider the semiring MnpFq of square matrices of size n with entries in F under the
operations of tropical matrix multiplication d and tropical matrix addition `, defined
by the expressions,
�X d Y
�pi, jq � mintX pi, kq � Y pk, jq | k � 1, . . . , nu, (5.1)�
X ` Y�pi, jq � mintX pi, jq, Y pi, jqu (5.2)
where X , Y P MnpFq. Moreover, when Fzt8u is a ring under the ordinary operations
restricted from R, the Fzt8u-module structure of MnpFzt8uq � MnpRq extends natu-
rally to MnpFq. We denote the extended scalar multiplication by λ�X and the extended
matrix addition by X�Y , for λ P F . We refer to them as the ordinary or classical matrix
operations.
Remark. It is convenient to interpret the entry at position pi, jq of an element of thesemiring MnpFq as the cost of moving from site i to site j. In this interpretation the
product X d Y of two such elements X and Y has as entries the minimal cost of
moving from one state to another in precisely two steps, where the costs of the first
step are determined by X , and the cost of the second step are determined by Y . The
interpretation of an m-fold (finite) product is similar. The entries of the sum X ` Yare the minimal costs of moving from site to site where one can choose to calculate
costs either according to X , or to Y .
A symmetric matrix X P MnpFq is said to be a distance matrix if its entries are
non-negative, satisfy the triangle inequality,
X pi, kq ¤ X pi, jq � X p j, kq, i, j, k P rns (5.3)
and are equal to 0 on the diagonal. Note that the non-negativity criterion is a con-
sequence of the other two by taking i � j � k in (5.3). Denote the collection of
5.1. FOUNDATIONS 129
distance matrices by DnpFq. Since the number of inequalities in (5.3) is finite, DnpFqis a polyhedral complex. It is also a cone in MnpFq in the sense that (i) it is closed
under ordinary scalar multiplication by non-negative elements of F , (ii) it is closed
under ordinary matrix addition. However, by the following example, DnpFq is neither
closed under tropical matrix addition, nor under tropical matrix multiplication.
Example. Let X , Y P DnpBq, where B is the trivial semifield t0,8u, be given as,
X �
��� 0 0 8
0 0 8
8 8 0
�� , Y �
��� 0 8 8
8 0 0
8 0 0
��
Then Z � X ` Y is not a distance matrix, since 8 � Zp1,3q ¦ Zp1,2q � Zp2, 3q � 0.
Moreover, X d Y is the non-symmetric matrix,
X d Y �
��� 0 0 0
0 0 0
8 0 0
��
Remark. Equip MnpFq with the order relation ¤ that is defined by X ¤ Y when
X i j ¤ Yi j , for all i, j P rns. We write I for the multiplicative neutral element, i.e. the
tropical identity matrix, with entries
Ii j �
#0 if i � j,8 if i � j
,
and Jλ for the matrix all of whose entries are λ, for some λ P F . The set of distance
matrices is equivalently characterized as the subset of the symmetric matrices in the
closed interval rJ0, Is that are idempotent under tropical matrix multiplication.
The submonoid of pMnpFq,dq generated by DnpFq is denoted SnpFq and referred
to as the gossip monoid or the gossip monoid on n points, for reasons to be explained
later. By the previous example, SnpFq is in general strictly bigger than DnpFq. The next
example shows that there exist matrices all of whose entries are non-negative that are
not in SnpFq.
Example. The following illustrates that not every element of MnpFq is in the monoid
SnpFq generated by the elements of DnpFq. The simplest such example occurs for
n� 2. The matrix,
X �
�0 8
0 0
�
130 CHAPTER 5. THE GOSSIP MONOID
is not a product of distance matrices. Suppose that it were. Write X � X1 d . . .d Xm.
Since n � 2 there is some matrix X j for which the p2,1q-st entry equals 0. But then
the p1, 2q-st entry of X must be equal to zero, by the fact that both right and left
multiplication by matrices with diagonal equal to 0 are decreasing maps.
The gossip monoid is generated by a finite number of one parameter submonoids,
which we define now. Let i, j P rns with i � j and a P F . The matrix Ei jpaq whose
pk, lq-th entry is, $'&'%
0 if k � l,8 if k � l and tk, lu � ti, jua if tk, lu � ti, ju
In other words, it is the matrix that differs from the tropical identity matrix only at
entries pi, jq and p j, iq, where it is a. Since the triangle inequalities (5.3) are trivially
satisfied, Ei jpaq is a distance matrix. A matrix of the form Ei jpaq is referred to as
an elementary distance matrix. The next statement shows that they indeed suffice to
generate SnpFq.
Lemma 5.1.1 The gossip monoid SnpFq is generated by the elementary distance matri-ces.
Proof. It is sufficient to show that any element of DnpFq can be written as a product
of elementary distance matrices. Let X P DnpFq. We claim that,
X �äi j
Ei jpX i jq, (5.4)
where the product is taken in any order. Let k, l P rns and consider the entry at position
pk, lq of the right-hand side of the above expression. By elementary matrix algebra it
equals,nà
i1�1. . .
nàir�1
�pEI1
qki1 � pEI2qi1 i2 . . .� pEIr�1
qir�1 ir� pEIr
qir l
, (5.5)
where I1, . . . , Ir are the size 2 subsets of rns in the order of (5.4). Consider a finite
summand of the above expression corresponding to the tuple I1, . . . , Ir . Then k � i1,
or I1 � tk, i1u, and i j�1 � i j or I j � ti j�1, i ju, and ir � l, or Ir � tir , lu. Let pk �j0, j1, . . . , jm � lq be a maximal, non-staggering subsequence of pk, i1, . . . , ir , lq, i.e.
js � js�1 for all s, and maximal as such. Then the summand under consideration
equals,
X j0, j1 � X j1, j2 � . . .� X jm�1, jm ,
which is at least Xkl by the triangle inequality. Since Xkl also appears in (5.5) the
statement follows. l
5.1. FOUNDATIONS 131
Let A P SnpFq. The length `pAq of A is the minimal length of a product of elementary
distance matrices equal to A, or more precisely,
`pAq �mintm | there exist elementary distance matrices X1, . . . , Xm
such that A� X1 d . . .d Xmu (5.6)
There is an easy bound on the length of an element of SnpFq.
Proposition 5.1.2 The length of an element of SnpFq is at most n3 � n2.
Proof. Let X be an arbitrary finite product of elementary distance matrices, i.e.,
X �mä
k�1
EIkpakq,
for sequences I1, . . . , Im of pairs and a1, . . . , am P F . The entry at position pi, jq of X is
ak1� . . .� aks
, where pIk1, . . . , Iks
q is a path from i to j in the complete graph on rns.Since revisiting a site is never cheaper, without loss of generality s ¤ n. The bound
follows since X has at most n2 � n non-zero entries pi, jq. l
To make use of the techniques available for representing a distance matrix by a
graph it is useful to have the following construction. To a matrix A we associate the
matrix A� of minimal costs of moving from one site to another in any number of steps.
Algebraically, this translates to,
A� � I ` A` A2 ` . . .� I ` A` A2 ` . . .` An, (5.7)
where exponentiation is with respect to d. The latter equality holds since it is never
cheaper to revisit a site, by the positivity of the entries of A. Incidentally, the opera-
tor A ÞÑ A� makes MnpPq into a structure known as a Kleene algebra. The relevant
properties of the operator are given in the next lemma.
Lemma 5.1.3 Let A P MnpFq. The matrix A� is idempotent with diagonal entries 0.If A has nonnegative entries and 0 on the diagonal then A� � An. If A P SnpPq thenA� P DnpPq.
Proof. Let A P MnpFq. The idempotency follows by expanding the product A� d A�,
making use of (5.7). Let X P MnpFq have non-negative off-diagonal entries and di-
agonal entries 0 and let Y P MnpFq have non-negative entries. Consider the entry at
position pk, l) of the product X d Y . It equals,
nài�1
�Xki � Yilq � Ykl `
� nài�1i�k
Xki � Yil
.
132 CHAPTER 5. THE GOSSIP MONOID
Thus, X d Y ¤ Y . It follows that An ¤ Am, with m¤ n and hence A� � An, when A has
diagonal entries 0.
By the algebraic characterization of distance matrices given earlier it is sufficient to
show that A� is symmetric. By Lemma 5.1.1 there exist elementary distance matrices
X I1, . . . , X Ir
such that,
A�rä
j�1X I j
.
The entry at position pk, lq of A� � An equals the minimum of all AJ1� . . . � AJm
,
where the Ji P tI1, . . . , Iru and pJ1, . . . , Jmq form an edge path between k and l in the
complete (loopless) graph on rns. Then AJm� . . .�AJ1
is at least A�lk, since pJm, . . . , J1q
is a path from l to k. Thus, A�lk ¤ A�kl and the statement follows by symmetry. l
5.2 Relation to the gossip problem
We refer to SnpFq as the gossip monoid, as the case F � B is strongly related to the
gossip problem. This mathematical problem appears to be first published in [3] and is
stated there as follows.
There are n ladies, and each of them knows some item of gossip not
known to the others. They communicate by telephone, and whenever one
lady calls another, they tell each other all that they know at that time.
How many calls are required before each gossip knows everything?
In our setting, it is the problem of determining the length of the all-zero matrix
J0. A scheme of calls solving the gossip problem translates to an expression for J0
as a product of a minimal number of elementary distance matrices Ei j , where in this
section Ei j is shorthand for Ei jp0q. Each such elementary distance matrix represents a
phone call between person i and j. The following example serves to clarify the link.
Example. Let n � 4 and consider the tropical product E12 d E23 d E34. The product
represents a sequence of three calls starting with a call between 1 and 2, then a call
between 2 and 3, and ending with a call between 3 and 4. The pieces of gossip 1
knows after this sequence are his own, and that of 2, the pieces of gossip 2 knows
after the sequence is that of 1, that of 2 and that of 3, while both 3 and 4 know all
the gossip. This is reflected in the columns of the product E12 d E23 d E34, i.e.
E12 d E23 d E34 �
������
0 0 0 0
0 0 0 0
8 0 0 0
8 8 0 0
�����
5.2. RELATION TO THE GOSSIP PROBLEM 133
In the article [3] a simple solution to the gossip problem is given. The authors
show that any product of less than 2n� 4 fundamental generators is strictly bigger
than J0, for n ¥ 4. Their solution depends on four chief gossips, an initial collector
i and a final spreader j, both of whom are chief gossips. Assume without loss of
generality that the chief gossips have labels 1, 2, 3 and 4. Then,
J0 �� nä
k�5
Eik
d XC d
� näk�5
E jk
where XC represents a calling scheme among the chief gossips to distribute the in-
formation among themselves. It has length 4 and can be taken equal to the product
E12 d E34 d E13 d E24. The length of J0 for n� 1,2, 3 is 0,1, 3.
There are some questions related to the gossip problem that are very natural from
the point of view of the gossip monoid, that to our knowledge have not even been
asked, let alone answered, in published sources.
(i) What is the maximal length of an element of SnpBq?(ii) What is the size of SnpBq?
(iii) Is there an efficient way to determine whether an n � n-matrix over B is an
element of SnpBq?Table 5.1 presents some computational results on questions (i) and (ii) obtained by
Jochem Berndsen in [6]. Notice that for n � 6,7, 8 the all-zero matrix is not an
element of maximal length in SnpBq (in those cases `pJ0q � 8, 12,14). Note that there
is rather large gap between the values in the table and the simple upper bound of
n3 � n of Proposition 5.1.2.
n |SnpBq| max. length1 1 02 2 13 11 34 189 45 9152 66 1,092,473 107 293,656,554 138 166,244,338,221 16
Table 5.1: Sizes and maximal lengths of SnpBq, for n� 1, . . . , 8.
Remark. The monoid SnpR8q can be seen as modelling a related real world situation
of sharing n gossips between n gossipers. The setup is the same is for the B case, with
the added complication that gossipers now know a certain fraction of each gossip.
134 CHAPTER 5. THE GOSSIP MONOID
Moreover, phone calls are over an imperfect channel, in the sense that they only
transmit a fixed fraction of the information between both participants. More precise,
an elementary distance matrix Ei jpaq corresponds to a phone call between i and j in
which a fraction e�a of the information is interchanged between gossip i and j.
5.3 Polyhedral structure of the gossip monoid
In this section we restrict our attention to the case F � R8. The following theorem
exploits the fact that SnpR8q is contained in a tropicalization of the orthogonal group.
Theorem 5.3.1 The elements SnpR8q X Rn2are contained in the underlying set of a
polyhedral fan of dimension pn2q.
Proof. Let K � Ctttuu be the field of Puiseux series over C in t with order valuation
ord : K ÑQ8. We prove that the matrices in SnpR8qwith entries in Q8 are contained
the tropicalization of OnpKq. The statement then follows by the Bieri-Groves theorem
(Proposition 1.4.9) and Proposition 1.2.4. The polyhedral fan in the statement is com-
mon refinement of the unions of the polyhedral fan structures on the tropical varieties
of the irreducible components of OnpKq.
Let x P K and i, j P rns with i � j. Define the matrix Gi jpxq by,
Gi jpxqpk, lq �
$''''''&''''''%
cospxq if k � l � i, or k � l � j,� sinpxq if k � i and l � j,sinpxq if k � j and l � i,0 if k � l and tk, lu � ti, ju,1 if k � l and k, l R ti, ju.
Then Gi jpxq P OnpKq and ordpGi jpxqq � Ei jpordpxqq.
Let I1 � pi1, j1q, . . . , Ir � pir , jrq be a sequence of pairs and a1, . . . , ar P K . Consider
the product,
X �rä
k�1
EIkpakq.
Since the residue field C of K is infinite, there exist t1, . . . , t r P K such that ordptkq �
ak and
ord� r¹
k�1
GIkptkq��
räk�1
ord�GIkptkq�� X .
Thus, X P Trop pOnq and SnpR8q is contained in the topological closure of Trop pOnq.
l
5.3. POLYHEDRAL STRUCTURE OF THE GOSSIP MONOID 135
The theorem does not give a natural polyhedral structure on SnpR8q. We attempt
to give SnpR8q a natural polyhedral structure in the next section.
5.3.1 Graphs with detours
The next part is concerned with the problem of realizing an element of MnpR8q by a
graph with a finite number of prescribed paths between the nodes. When restricted to
elements of DnpR8q this realization problem is that of embedding a metric on a finite
number of points in a weighted graph (cf. [34]), which we explain first.
Let Γ� pV, Eq be an undirected graph and w : E Ñ F a function assigning nonneg-
ative weights to the edges. We do not assume Γ to be simple. The weight of a path in
pΓ, wq is the sum of the weights of the individual edges in the path. It is thought of as
the cost of traversing the path. A map ` : rns Ñ V is called a labeling, or rns-labeling,
if we need to be precise, and the pair pΓ,`q is referred to as a labeled graph, or an
rns-labeled graph.
A weighted rns-labeled graph gives rise to a matrix ApΓ, w,`q in DnpFqwhose entry
at position pi, jq is the minimal weight of a path between `piq and `p jq. We say that
the weighted labeled graph pΓ, w,`q realizes a matrix X if X � ApΓ, w,`q. Moreover,
such a realization is said to be optimal if the total sum of the edge weights of pΓ, wq is
minimal among all realizations. Optimal realizations always exist, as proven in [34].To exclude superfluous edges we require that an optimal realization of X is minimalin the sense that the result of a contraction or deletion of any edge does not realize
X .
The terminology extends to labeled graphs without weights. Such a graph is said
to realize a certain matrix when there exist a weight function on its edges that realizes
the matrix in the previously defined sense. The qualifiers optimal and minimal extend
in the same manner.
In what follows we restrict our attention to F � R8. The first step in describing
the cones of the gossip monoid is to find a minimal set of labeled graphs that realize
the elements of DnpR8q optimally (and thus minimally, by definition). We do so in
the next example. See also [22].
We remind the reader that J0 stands for the matrix of the appropriate size with all
entries 0.
Examples. We give optimal realizations of the elements of DnpR8q, for n � 2,3, 4.
For the cases n� 5,6 see [41] and [63].
(i) Optimal realizations of the elements of D2pR8q are easily determined. An ele-
ment of D2pR8qztJ0u is minimally realized by the graph on two vertices having
one edge. The choice of labeling is inconsequential as long as it is injective. The
matrix J0 is optimally realized by the graph on one vertex.
136 CHAPTER 5. THE GOSSIP MONOID
1 323 32 21 1
1, 2 3 1, 3 2 2, 3 1
1, 2, 3
1 3
2
Figure 5.1: The Hasse diagram of the poset of minimal graphs that optimally realizethe elements of D3pR8q.
(ii) We consider the case n� 3. Any matrix in D3pR8q is realized by the top labeled
graph of the poset depicted in Figure 5.1, although only the matrices in the
interior of the cone are optimally realized by it. However, any matrix of D3pR8qis minimally realized by some graph in the poset. The graph depends on the
facet of D3pR8q the matrix is contained in. In effect, the order complex of the
poset of Figure 5.1 is isomorphic to the polyhedral cone D3pR8q.
1 2
3 4
1 2
34
1
2
3
4
Figure 5.2: The three distinct labeled graphs that realize all matrices in D4pR8q Theparallel sides of the middle rectangle have equal weight.
(iii) The case n � 4 is similar to that of n � 3 in the sense that there exists a single
graph Γ such that for any element X of D4pR8q there exists a labeling ` of Γsuch that X is realized by pΓ,`q. However, three distinct labelings are required,
unlike in the case n� 3. The labeled graphs are depicted in Figure 5.2.
We wish to extend the representation of DnpR8q in terms of realizing labeled
graphs to the entire monoid SnpR8q. For this we need an extension of the concept
5.3. POLYHEDRAL STRUCTURE OF THE GOSSIP MONOID 137
of a labeled weighted graph. Let i and j be distinct elements of rns. A detour from ito j in an rns-labeled weighted graph is simply a path p starting at `piq and ending
at `p jq that has larger total weight than the path of minimal weight between `piqand `p jq. Note that a path is allowed to traverse the same edge more than once. The
data specifying the detour is the triple pi, j, pq. A labeled weighted graph with detoursis a tuple consisting of a labeled weighted graph and a finite set of detours between
distinct pairs of vertices.
Let pΓ, w,`,Dq be an rns-labeled weighted graph with set of detours D. It gives rise
to a matrix ApΓ, w,`,Dq whose entry at position pi, jq equals the weight of the detour
from i to j, if there is any, or the weight of a path of minimal weight between i and j,if there is no detour between i and j in D. In particular, ApΓ, w,`,Dq need not be sym-
metric, while its diagonal entries must be 0. Again, if X P MnpFq and X � ApΓ, w,`,Dq,then pΓ, w,`,Dq is said to realize X and similar terminological considerations as for
the detourless case apply. In particular, a matrix X is said to be realized by a triple
pΓ,`,Dq if there exists an assignment of edge weights w to Γ such that pΓ, w,`,Dqis a weighted labeled graph with detours and X � ApΓ, w,`,Dq. By slight abuse of
terminology we refer to the triple pΓ,`,Dq as a labeled graph with detours.
Examples. We give three examples of labeled (weighted) graphs with detours.
1 2va b
(a) Linear graph with a sin-gle detour from 1 to 2.
1
2
3 4
(b) Graph with a single detour from 1to 4.
1, 2, 3
3 → 1
(c) A r3s-labeled graph onone vertex and one edge(a loop) with a single de-tour from 3 to 1.
Figure 5.3: Examples of labeled weighted graphs with detours.
(i) The first example is very simple. Consider the linear graph pΓ,`q on three ver-
tices one of whose end vertices is labeled 1 and the other 2. For ease of notation
we label the middle vertex v and define p to be path pt1, vu, t1, vu, t1, vu, tv, 2uq
from 1 to 2. Set the weight wpt1, vuq � a and wptv, 2uq � b. It is represented
138 CHAPTER 5. THE GOSSIP MONOID
in Figure 5.3(a). The matrix it gives rise to is,
ApΓ, w,`, tpuq �
�0 3a� b
a� b 0
�
Unless a � 0, or a � 8 this matrix is not in S2pR8q � D2pR8q. The matrices
that can be realized by pΓ,`, tpuq are precisely those in the cone tX P M2pR8q |X11 � X22 � 0, X12 ¥ X21, and X21 8uY I .
(ii) An example of a labeled graph with detours is displayed in 5.3(b). In Section
5.4.2 we will see that it realizes an element of S4pR8q. Note that the matrix Xrealized by this graph satisfies X41 � X13 � X34 ¤ X14 ¤ X12 � X24.
(iii) The third example is displayed in Figure 5.3(c). The graph Γ has a single point,
say v, and a single edge e connecting v to itself. The data specifying the detour
is the triple p3, 1, eq, in which 3 is the start and 1 is the end of the detour.
This detour graph appears in the classification of the cones of S3pR8q. A matrix
realized by this detour graph has 0 at all positions except at position p3, 1q,
where it may have any value in R8.
The following easy statement allows one to realize the transpose of a matrix real-
ized by a labeled weighted graph with detours.
Lemma 5.3.2 Let pΓ, w,`,Dq be a weighted rns-labeled graph with detours. Write D 1
for the set of detours,
D 1 � tp j, i, pq | i, j P rns and pi, j, pq P Du.
Then ApΓ, w,`,D 1q is the transpose of ApΓ, w,`,Dq.
Proof. Set X � ApΓ, w,`,Dq and Y � ApΓ, w,`,Dq. If there are no detours from i to
j, or from j to i then X i j � X ji � Yji � Yi j . If there is a detour from i to j in D, then
there is a detour from j to i in D 1, with the same path. Thus, X i j � Yji . This proves
the statement. l
The matrices realized by a labeled (unweighted) graph with detours G � pΓ,`,Dqform a polyhedral cone that is a linear image of pR�8q
E , where E is the set of edges of
Γ. This cone is referred to as the cone of G and denoted C pGq � MnpR8q. We aim to
find a finite set of labeled graphs with detours tG1, . . . , Gku such that,
SnpR8q �k¤
i�1
C pGiq. (5.8)
As the examples in the next section show, for n � 2,3, 4 the Gi can be chosen such
that C pGiq are the maximal cones of a fan structure on SnpR8q. We do not know if
5.4. CASES 139
such a choice is possible in general.
5.4 Cases
In this section we determine the gossip monoids on 3 and 4 points explicitly. Up to
symmetry there are only 2 distinct maximal cones in the fan S3pR8q, while there seem
to be 16 in S4pR8q.
5.4.1 n¤ 3
Since Ei jpaq d Ei jpbq � Ei jpa � bq and S2pR8q � D2pR8q it is clear that S2pR8q is
isomorphic to the monoid of positive tropical real numbers under addition (tropical
multiplication). The first non-trivial case is that of S3pR8q, the structure of which we
determine in this section.
Since S3pR8q is a pointed fan no information is lost by intersecting S3pR8q with
a sphere centered around the point of the fan (which is J0). We denote the resulting
spherical polyhedral complex by S13pR8q. It is depicted in Figure 5.4. The maximal
cones are those consisting of the matrices that are realized by the labeling detour
graphs.
1
23
1 3 2
3 2 1
3 2 1
1 3 2
1 3 2
1 3 2
Figure 5.4: Representation of the spherical complex S13pR8q. The labeled graphs withdetours corresponding to the maximal cells are indicated. The middle triangle repre-sents the cone of distance matrices.
The computations to show that Figure 5.4 is a represention of S13pR8q are elemen-
tary and can be done by hand. To prove that the matrices realized by the graphs with
140 CHAPTER 5. THE GOSSIP MONOID
a single detour in Figure 5.4 are indeed in S3pR8q it suffices to check that,
Api j ka b
c
q � E jkpbq d Api j k
a c− aq � Ap
i j k
c− b bq d Ei jpaq, (5.9)
for any valid choice of a, b and c. Notice that,
Api j ka b
c
q d Ei jpdq, and E jkpdq d Api j ka b
c
q
are contained in the complex of Figure 5.4 for all choices of a, b, c and d by (5.9).
Thus, to see that these cones are all it is sufficient to check that,
Ei jpdq d Api j ka b
c
q, �
$''''''''&''''''''%
Api j ka b
c
q, c� b ¤ d,
Api j ka b
b + d
q, a ¤ d ¤ c� b,
Api j kb
a + b
dq, 0¤ d ¤ a
,
Eikpdq d Api j ka b
c
q �
$''''''''''''''''&''''''''''''''''%
Api j ka b
c
q, c ¤ d,
Api j kb
d
aq, a� b ¤ d ¤ c,
Api j
k
a+b−d2
b+d−a2
a+d−b2
q, b� a ¤ d ¤ a� b,
Apij k
b
daq, 0¤ d ¤ b� a.
(when a ¤ b),
5.4. CASES 141
Eikpdq d Api j ka b
c
q �
$'''''''''''''''&'''''''''''''''%
Api j ka b
c
q, c ¤ d,
Api j kb
d
aq, a� b ¤ d ¤ c,
Api j
k
a+b−d2
b+d−a2
a+d−b2
q, a� b ¤ d ¤ a� b,
Apij k bd
a
q, 0¤ d ¤ a� b.
(when b ¤ a).
That the products,
Api j ka b
c
q d Eikpdq, and Api j ka b
c
q d E jkpdq
are also contained in one of the cones of Figure 5.4 follows by transposing the product
matrices and applying Lemma 5.3.2.
5.4.2 n� 4
The case n � 4 is sufficiently large to make the computations intractable by hand.
Instead we used Mathematica to compute a fan structure on SnpR8q. Figure 5.5 gives
the realizing graphs with detours of all the cones of S4pR8q, up to transposition and
the action of Σ4, the symmetric group on t1, 2,3, 4u. The surplus length of a detour
from i to j is defined as the difference between the length of the detour and the
minimal distance between i and j in the graph. Two detours between i to j and k to
l have the same color if their surplus lengths are equal.
These graphs were obtained by generating all 66 possible piecewise linear affine
maps R68Ñ M4pR8q of the form
px1, . . . , x6q Ñ EI1px1q d EI2
px2q d . . .d EI6px6q,
up to Σ4 and transposition, and determining the �-maximal cones in their images. It
then suffices to check that the collection of cones obtained in such manner is closed
under left multiplication by an elementary distance matrix. Connectivity in codimen-
sion 1 is proved by Figure 5.6. It shows that any maximal cone is connected in codi-
mension 1 to at least on cone of distance matrices. Note that each labeling depicted
in Figure 5.2 corresponds to a different maximal cone. Since the distance matrices
are themselves connected in codimension one it follows that S4pR8q is connected in
codimension 1.
Most intersections in Figure 5.6 are of a special type. Recall that a cone determined
142 CHAPTER 5. THE GOSSIP MONOID
by a graph with detours is the image of a linear map pR�8qE Ñ MnpR8q, where E is
the set of edges of the graph. According to Figure 5.5 every cone C in SpR8q is the
image of such a map αC with domain pR�8q6. Moreover, it is simple to check that all
cones are six dimensional. In particular, the intersection of such a cone with a sphere
around the origin is a simplex. Most intersections in Figure 5.6 are of the image of the
restriction of αC to a five dimensional boundary hyperplane of pR�8q6, in other word,
they are obtained by setting one coordinate equal to 0. In terms of the graphs this is
expressed by contracting an edge. The only exception is the connection of C7 to C12.
Although these cones intersect in a five dimensional boundary cone, the boundary
cone is obtained from the parameterizations specified by the graphs with detours by
restricting them to a hyperplane of the form x i � x j .
This leads to the following theorem.
Theorem 5.4.1 The cones realized by the graphs of Figure 5.5 give a polyhedral fanstructure on S4pR8q. This polyhedral fan is pure of dimension 6 and connected in codi-mension 1. Its intersection with a sphere around the origin is a simplicial spherical com-plex. Moreover, every element of SnpR8q is the product of (at most) 6 elementary distancematrices.
We end this chapter with a conjecture. The results of this section give evidence for
the following statement.
Conjecture 5.4.1 There exists a polyhedral fan structure on SnpR8q such that this fanis pure of dimension pn
2q and connected in codimension 1. Every element of SnpR8q is the
product of (at most) pn2q elementary distance matrices.
5.4. CASES 143
C1 C2 C3
C4 C5 C6
C7 C8 C9
C10 C11 C12
Figure 5.5: The graphs realizing a polyhedral fan structure on S4pR8.
144 CHAPTER 5. THE GOSSIP MONOID
2
3 4
1 ∗
C2C10
1 2
34
∗
1
2
34
1
2
3 4∗
C9
1
234
∗
1
234
∗
4 3 2
1
C1 C2
12
3
4
1 2
3
4
1
23
4
∗∗
C8 C11
1
2
3
4
1
2
34
1
2
34
∗∗
C5 C4
1 2
34
12
34
∗
4
∗1
23
C12C11
1
2
34
1
2
343
1 4
2
∗ ∗
C4 C8
12
34
∗
1
3 4
2
1 2
3
4∗
C6 C3
1
2
3 4
C2
1 2
34∗
C4
1
2
34
∗
1 2
3
4
∗
C3
C7 C12
1
2
34
12
34
12
34
Figure 5.6: Intersection of cones realized by labeled graphs that are themselves re-alized by a contraction of the labeled graph (except in case C7-C12. The edge to becontracted is indicated by an asterix �. This shows that the cones in the grey boxesintersect in a cone of dimension 5. The intersection between C7 and C12 is obtainedby setting equal surplus lengths in the graphs representing C7 and C12.
Bibliography
[1] Marianne Akian, Stephane Gaubert, and Alexander Gutermann. Linear inde-
pendence over tropical semirings and beyond. Preprint, 2008.
[2] Federico Ardila and Caroline J. Klivans. The Bergman complex of a matroid and
phylogenetic trees. J. Combin. Theory Ser. B, 96(1):38–49, 2006.
[3] Brenda Baker and Robert Shostak. Gossips and telephones. Discrete Math.,2(3):191–193, 1972.
[4] George M. Bergman. The logarithmic limit-set of an algebraic variety. Trans.Amer. Math. Soc., 157:459–469, 1971.
[5] George M. Bergman. A weak Nullstellensatz for valuations. Proc. Amer. Math.Soc., 28:32–38, 1971.
[6] Jochem Berndsen. Three problems in algebraic combinatorics. Master’s thesis,
Eindhoven University of Technology, 2012.
[7] Robert Bieri and J. R. J. Groves. The geometry of the set of characters induced
by valuations. J. Reine Angew. Math., 347:168–195, 1984.
[8] T.S. Blyth and M.F. Janowitz. Residuation Theory. Pergamon Press, 1972.
[9] Alexandre V. Borovik, I. M. Gelfand, and Neil White. Coxeter matroids, volume
216 of Progress in Mathematics. Birkhäuser Boston Inc., Boston, MA, 2003.
[10] S. Bosch, U. Güntzer, and R. Remmert. Non-Archimedean analysis, volume 261
of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles ofMathematical Sciences]. Springer-Verlag, Berlin, 1984. A systematic approach to
rigid analytic geometry.
[11] Thomas Brylawski. Constructions. In Theory of matroids, volume 26 of Encyclo-pedia Math. Appl., pages 127–223. Cambridge Univ. Press, Cambridge, 1986.
145
146 BIBLIOGRAPHY
[12] Alan L. C. Cheung. Adjoints of a geometry. Canad. Math. Bull., 17(3):363–365;
correction, ibid. 17 (1974), no. 4, 623, 1974.
[13] Guy Cohen, Stéphane Gaubert, and Jean-Pierre Quadrat. Duality and separa-
tion theorems in idempotent semimodules. Linear Algebra Appl., 379:395–422,
2004. Tenth Conference of the International Linear Algebra Society.
[14] H.H. Crapo. Single element extensions of matroids. J. Res. Natl. Bureau Stan-dards B, pages 55–65, 1965.
[15] R.A. Cuninghame-Green and P. Butkovic. Bases in max-algebra. Linear Algebraand its Applications, 389:107–120, 2004.
[16] Raymond Cuninghame-Green. Minimax Algebra. Springer-Verlag, 1979.
[17] Vladimir I. Danilov, Alexander V. Karzanov, and Gleb A. Koshevoy. Planar flows
and quadratic relations over semirings. Journal of Algebraic Combinatorics,36(3):441–474, 2012.
[18] Mike Develin, Francisco Santos, and Bernd Sturmfels. On the rank of a tropical
matrix. In Combinatorial and computational geometry, volume 52 of Math. Sci.Res. Inst. Publ., pages 213–242. Cambridge Univ. Press, Cambridge, 2005.
[19] Mike Develin and Bernd Sturmfels. Tropical convexity. Documenta Mathematica,
9:1–27, 2004.
[20] Alicia Dickenstein, Eva Maria Feichtner, and Bernd Sturmfels. Tropical discrim-
inants. J. Am. Math. Soc., 20(4):1111–1133, 2007.
[21] Jan Draisma. A tropical approach to secant dimensions. J. Pure Appl. Algebra,
212(2):349–363, 2008.
[22] Andreas Dress, Katharina T. Huber, and Vincent Moulton. Hereditarily opti-
mal realizations: why are they relevant in phylogenetic analysis, and how does
one compute them? In Algebraic combinatorics and applications (Gößweinstein,1999), pages 110–117. Springer, Berlin, 2001.
[23] Andreas W. M. Dress and Walter Wenzel. Valuated matroids. Adv. Math.,93(2):214–250, 1992.
[24] Manfred Einsiedler, Mikhail Kapranov, and Douglas Lind. Non-Archimedean
amoebas and tropical varieties. J. Reine Angew. Math., 601:139–157, 2006.
[25] Georges François and Johannes Rau. The diagonal of tropical matroid varieties
and cycle intersections. Collectanea Mathematica, pages 1–26, 2012.
BIBLIOGRAPHY 147
[26] David Gale. Optimal assignments in an ordered set: An application of matroid
theory. J. Combinatorial Theory, 4:176–180, 1968.
[27] S. Gaubert and R. Katz. The Minkowski theorem for max-plus convex sets.
Linear Algebra and Applications, 421:356–369, 2007.
[28] Israel M. Gelfand, Mikhail M. Kapranov, and Andrei V. Zelevinsky. Discriminants,resultants, and multidimensional determinants. Mathematics: Theory & Applica-
tions. Birkhäuser, Boston, MA, 1994.
[29] Jonathan S. Golan. Power algebras over semirings, volume 488 of Mathemat-ics and its Applications. Kluwer Academic Publishers, Dordrecht, 1999. With
applications in mathematics and computer science.
[30] Nathan B. Grigg. Factorization of tropical polynomials in one and several vari-
ables. Brigham Young University Honors Thesis, 2007. Undergraduate Thesis.
[31] Walter Gubler. A guide to tropicalizations. Preprint, 2011.
[32] Kerstin Hept and Thorsten Theobald. Tropical bases by regular projections. Proc.Amer. Math. Soc., 137(7):2233–2241, 2009.
[33] Sven Herrmann, Anders Jensen, Michael Joswig, and Bernd Sturmfels. How to
draw tropical planes. Electron. J. Combin., 16(2, Special volume in honor of
Anders Bjorner):Research Paper 6, 26, 2009.
[34] Wilfried Imrich, J. M. S. Simões-Pereira, and Christina M. Zamfirescu. On op-
timal embeddings of metrics in graphs. J. Combin. Theory Ser. B, 36(1):1–15,
1984.
[35] Zur Izhakian. On the tropical rank of a tropical matrix. Communications inAlgebra, 37(11):3912âAS–3927, 2009.
[36] Zur Izhakian, Marianne Johnson, and Mark Kambites. Pure dimension and pro-
jectivity of tropical polytopes. Preprint, 2011.
[37] P. Jipsen and C. Tsinakis. A survey of residuated lattices. In Ordered algebraicstructures, volume 7 of Dev. Math., pages 19–56. Kluwer Acad. Publ., Dordrecht,
2002.
[38] M.M. Kapranov. A characterization of A-discriminantal hypersurfaces in terms
of logarithmic Gauss map. Math. Ann., 290(2):277–285, 1991.
[39] Eric Katz. Tropical geometry. Notes by Eric Katz of a course by Grigory Mikhalkin
at the University of Texas, February and March 2008.
148 BIBLIOGRAPHY
[40] Yusuke Kobayashi and Kazuo Murota. Induction of M-convex functions by link-
ing systems. Discrete Appl. Math., 155(11):1471–1480, 2007.
[41] Jack Koolen, Alice Lesser, and Vincent Moulton. Optimal realizations of generic
five-point metrics. European J. Combin., 30(5):1164–1171, 2009.
[42] Joseph P. S. Kung. Bimatroids and invariants. Adv. in Math., 30(3):238–249,
1978.
[43] G. L. Litvinov, V. P. Maslov, and G. B. Shpiz. Idempotent functional analysis. An
algebraic approach. Mat. Notes, 69, 2001.
[44] Dean Lucas. Weak maps of combinatorial geometries. Trans. Amer. Math. Soc.,206:247–279, 1975.
[45] Grigory Mikhalkin. Enumerative tropical algebraic geometry in R2. J. Amer.Math. Soc., 18(2):313–377, 2005.
[46] Mikhail Mitrofanov. Bimatroids and Gauss decomposition. European J. Combin.,28(4):1180–1195, 2007.
[47] K. Murota. Finding optimal minors of valuated bimatroids. Appl. Math. Lett.,8(4):37–41, 1995.
[48] Kazuo Murota. Convexity and Steinitz’s exchange property. Adv. Math.,124(2):272–311, 1996.
[49] Kazuo Murota and Akihisa Tamura. On circuit valuation of matroids. Adv. inAppl. Math., 26(3):192–225, 2001.
[50] James G. Oxley. Matroid theory. Oxford Science Publications. The Clarendon
Press Oxford University Press, New York, 1992.
[51] Sam Payne. Analytification is the limit of all tropicalizations. Math. Res. Lett.,16(3):543–556, 2009.
[52] Sam Payne. Fibers of tropicalization. Math. Z., 262(2):301–311, 2009.
[53] Jean-Eric Pin. Tropical semirings. In J. Gunawardena, editor, Idempotency. Cam-
bridge University Press, 1998.
[54] V. Puiseux. Recherches sur les fonctions algébriques. J. Math. Pures Appl.,15:365–480, 1850.
[55] Felipe Rincón. Computing tropical linear spaces, 2011. Preprint available from
http://arxiv.org/abs/1109.4130.
BIBLIOGRAPHY 149
[56] A. Schrijver. Matroids and linking systems. J. Combin. Theory Ser. B, 26(3):349–
369, 1979.
[57] Kristin M. Shaw. A tropical intersection product in matroidal fans. Preprint,
2011.
[58] David E. Speyer. Tropical Geometry. PhD thesis, University of California, Berke-
ley, 2005.
[59] David E. Speyer. Tropical linear spaces. SIAM J. Discrete Math., 22(4):1527–
1558, 2008.
[60] Stuart A. Steinberg. Lattice-ordered Rings and Modules. Springer, 2009.
[61] Bernd Sturmfels and Jenia Tevelev. Elimination theory for tropical varieties.
Mathematical Research Letters, pages 543–562, 2008.
[62] Bernd Sturmfels, Jenia Tevelev, and Josephine Yu. The Newton polytope of the
implicit equation. Mosc. Math. J., 7(2):327–346, 351, 2007.
[63] Bernd Sturmfels and Josephine Yu. Classification of six-point metrics. Electron.J. Combin., 11(1):Research Paper 44, 16 pp. (electronic), 2004.
[64] Bernd Sturmfels and Josephine Yu. Tropical implicitization and mixed fiber
polytopes. In Software for algebraic geometry, volume 148 of IMA Vol. Math.Appl., pages 111–131. Springer, New York, 2008.
[65] K. L. N. Swamy. Dually residuated lattice ordered semigroups. Math. Ann.,159:105–114, 1965.
[66] Bartel L. van der Waerden. Algebra. Zweiter Teil, volume 23 of HeidelbergerTaschenbücher. Springer Verlag, Berlin Heidelberg New York, fifth edition, 1967.
[67] Bartel L. van der Waerden. Einfürung in die algebraische Geometrie, volume 51
of Die Grundlehren der mathematischen Wissenschaften. Springer Verlag, Berlin
Heidelberg New York, second edition, 1973.
[68] Edouard Wagneur. Moduloïds and pseudomodules 1. dimension theory. DiscreteMathematics, 1998:57–73, 1991.
[69] Huaxiong Wang. Injective hulls of semimodules over additively-idempotent
semirings. Semigroup Forum, 48:377–379, 1994.
[70] Morgan Ward and R. P. Dilworth. Residuated lattices. Trans. Amer. Math. Soc.,45(3):335–354, 1939.
[71] Josephine Yu and Debbie Yuster. Representing tropical linear spaces by circuits.
In Formal Power Series and Algebraic Combinatorics, Proceedings, 2007.
Index
F -matroid, 64
F -valuated matroid, 64
F -weighted network, 108
rns-labeled graph, 135
absorption axiom, 38
additive ring valuation, 54
algebraic dual space, 48
algebraic torus, 21
Archimedean, 42
basis of a module, 57
Berkovich analytic space, 18
Boolean semiring, 42
canonical order, 40
classical matrix operations, 128
cofactor system, 105
column set, 105
combinatorial type, 68
commutative semiring, 38
complete, 12
complete intersection, 124
composition
of linking systems, 107
of morphisms, 121
contraction, 76
cycle, 99
degree-d homogenization, 27
dependent, 57
detour, 137
dimension, 12
dioids, 38
direct sum, 78
distance matrix, 128
divisible semifield, 8
division semiring, 38
domain set, 105
Dressian, 80
dual
algebraic, 48
orthogonal, 73
dual residuum, 43
dual space, 48
dually residuated
increasing map, 46
module, 43
elementary distance matrix, 130
elementary extension, 93
elementary matroidal extension function,
93, 98
equivalent projections, 12
extended graph, 92
extension, 92
elementary, 93
faithful, 39
free, 39
Gondran-Minoux dependent, 57
Gondran-Minoux pseudomorphism, 54
good section, 5
150
INDEX 151
gossip monoid, 129
gossip problem, 132
graph of a linking system, 104
idempotent, 38
idempotent module, 39
implicitization, 21
independent, 57
indicator, 52
induced order, 40
initial form, 5
irreducible, 57
kernel of a projection, 12
labeled graph, 135
labeled graph with detours, 137
weighted, 137
length, 131
linear functional, 48
linking, 108
linking system, 104
matroid over F , 64
matroid union, 79
matroidal extension function, 93
minimal, 135
minimal matroid, 65
minimal representative, 87
module, 38
monomial, 2
morphism, 113
natural order, 40
natural preorder, 39
Newton polytope, 10
optimal, 135
ordinary matrix operations, 128
orthogonal dual, 73
polynomial, 2
principal filter, 46
projection, 12
proportional, 64
pseudomorphisms of algebras, 54
pseudomorphisms of semirings, 54
Puiseux series, 2
range set, 105
rank, 57
real tropical semifield, 38
realizable valuated matroids, 65
realized by a graph, 135
reducible, 57
reflexive, 50
regular projection, 12
regular subdivision, 10
reparameterization, 26
represent a morphism, 114
representation matroid, 104
residual matroid, 65
residue field, 2
restriction
of a linking system, 105
of a matroid, 76
of a morphism, 117
row set, 105
semifield, 1, 38
semiring, 37
separated, 50
sink node, 108
source node, 108
support, 84
surplus length, 141
terminal node, 108
transpose, 49
trivial semifield, 2, 42
tropical basis
of a set, 5
of a variety, 7
152 INDEX
tropical identity matrix, 129
tropical linear monoid, 122
tropical linear space, 69
tropical prevariety
of a polynomial, 3
of a set, 4
tropical pseudomorphism, 54
tropical rational numbers, 38
tropical rational semifield, 38
tropical real number semifield, 1
tropical semifield, 2
tropical semiring, 42
tropical vanishing locus, 3
tropical variety, 5
tropicalization
of a variety, 7
of a polynomial, 3
of a rational function, 22
tropically dependent, 57
tropically unirational, 22
truncation, 72
underlying matroid, 65
valuated exchange condition, 64
valuated matroid, 64
on independent sets, 72
valuation ring, 2
vanish, 53
in the sense of Gondran-Minoux, 53
in the tropical sense, 53
weakly, 53
vanishing axiom, 55
vanishing sequence, 53
weakly dependent, 57
weight of a linking, 108
zero-sum free, 52
Summary
Tropical varieties, maps and
gossip
Tropical geometry is a relatively new field of mathematics that studies the tropicaliza-
tion map: a map that assigns a certain type of polyhedral complex, called a tropical
variety, to an embedded algebraic variety. In a sense, it translates algebraic geomet-
ric statements into combinatorial ones. An interesting feature of tropical geometry
is that there does not exist a good notion of morphism, or map, between tropical
varieties that makes the tropicalization map functorial. The main part of this thesis
studies maps between different classes of tropical varieties: tropical linear spaces and
tropicalizations of embedded unirational varieties.
The first chapter is a concise introduction to tropical geometry. It collects and
proves the main theorems. None of these results are new.
The second chapter deals with tropicalizations of embedded unirational varieties.
We give sufficient conditions on such varieties for there to exist a (not necessarily
injective) parametrization whose naive tropicalization is surjective onto the associated
tropical variety.
The third chapter gives an overview of the algebra related to tropical linear spaces.
Where fields and vector spaces are the central objects in linear algebra, so are semi-
fields and modules over semifields central to tropical linear algebra and the study
of tropical linear spaces. Most results in this chapter are known in some form, but
scattered among the available literature. The main purpose of this chapter is to col-
lect these results and to determine the algebraic conditions that suffice to give linear
algebra over the semifield a familiar feel. For example, under which conditions are
varieties cut out by linear polynomials closed under addition and scalar multiplica-
tion?
The fourth chapter comprises the biggest part of the thesis. The techniques used
are a combination of tropical linear algebra and matroid theory. Central objects are
153
154 INDEX
the valuated matroids introduced by Andreas Dress and Walter Wenzl. Among other
things the chapter contains a classification of functions on a tropical linear space
whose cycles are tropical linear subspaces, extending an old result on elementary
extensions of matroids by Henry Crapo. It uses Mikhalkin’s concept of a tropical mod-
ification to define the morphisms in a category whose objects are all tropical linear
spaces. Finally, we determine the structure of an open submonoid of the morphisms
from affine 2-space to itself as a polyhedral complex.
Finally, the fifth and last chapter is only indirectly related to maps. It studies a cer-
tain monoid contained in the tropicalization of the orthogonal group: the monoid that
is generated by the distance matrices under tropical matrix multiplication (i.e. where
addition is replaced by minimum, and multiplication by addition). This monoid gen-
eralizes a monoid that underlies the well-known gossip problem, to a setting where
information is transmitted only with a certain degree accuracy. We determine this
so-called gossip monoid for matrices up to size 4, and prove that in general it is a
polyhedral monoid of dimension equal to that of the orthogonal group.
Curriculum Vitae
Bartholomeus Johannes Frenk was born on February 3, 1981 in Schiedam, the Nether-
lands. After finishing VWO in 1993 at the Bisschop Bekkers College in Eindhoven, he
studied Applied Mathematics at the Eindhoven University of Technology. In 2006 he
graduated within the Discrete Algebra and Geometry group on algebras related to
knot theory. After working for nearly a year as a risk modelling analyst he started a
Ph.D. project in 2008 at the Eindhoven University of Technology, of which the results
are presented in this dissertation.
155