Polytopes, rings and K-theory

Winfried Bruns
Universität Osnabrück
Joseph Gubeladze
San Francisco State University
and
Georgian Academy of Sciences, Tbilisi
Preliminary incomplete version
February 18, 2006
Contents
I Cones, monoids, and unimodular triangulations 1
1 Polytopes, cones, and complexes 3
1.A Polyhedra and their faces . . . . . . . . . . . . . . . . . . . . . . . 3
1.B Finite generation of cones . . . . . . . . . . . . . . . . . . . . . . . 12
1.C Finite generation of polyhedra . . . . . . . . . . . . . . . . . . . . 17
1.D Polyhedral Complexes . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.E Subdivisions and triangulations . . . . . . . . . . . . . . . . . . . . 29
1.F Regular subdivisions . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.G Rationality and integrality . . . . . . . . . . . . . . . . . . . . . . . 42
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2 Affine monoids and their Hilbert bases 47
2.A Affine monoids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.B Normal affine monoids . . . . . . . . . . . . . . . . . . . . . . . . 57
2.C Generating normal affine monoids . . . . . . . . . . . . . . . . . . 65
2.D Normality and unimodular covering . . . . . . . . . . . . . . . . . 74
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3 Multiples of lattice polytopes 83
3.A Knudsen-Mumford triangulations . . . . . . . . . . . . . . . . . . 83
3.B Unimodular triangulations of multiples of polytopes . . . . . . . . 88
3.C Unimodular covers of multiples of polytopes . . . . . . . . . . . . . 95
II Affine monoid algebras 109
4 Monoid algebras 111
4.A Monoid algebras and graded rings . . . . . . . . . . . . . . . . . . 111
4.B Representations of monoid algebras . . . . . . . . . . . . . . . . . 117
4.C Monomial prime and radical ideals . . . . . . . . . . . . . . . . . . 122
4.D Normality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
4.E Divisorial ideals and the class group . . . . . . . . . . . . . . . . . 129
4.F The Picard group and seminormality . . . . . . . . . . . . . . . . . 137
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
iii
iv Contents
5 Isomorphisms and Automorphisms 147
5.A Linear algebraic groups . . . . . . . . . . . . . . . . . . . . . . . . 147
5.B Invariants of diagonalizable groups . . . . . . . . . . . . . . . . . . 153
5.C The isomorphism theorem . . . . . . . . . . . . . . . . . . . . . . 156
5.D Automorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
6 Homological properties and enumerative combinatorics 179
7 Gröbner bases, triangulations, and Koszul algebras 181
III K-theory of monoid algebras 183
8 Projective modules over monoid rings 185
8.A Projective modules . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
8.B The main theorem and the plan of the proof . . . . . . . . . . . . . 187
8.C Projective modules over polynomial rings . . . . . . . . . . . . . . 189
8.D Reduction to the interior . . . . . . . . . . . . . . . . . . . . . . . 194
8.E Graded “Weierstraß Preparation” . . . . . . . . . . . . . . . . . . . 195
8.F Pyramidal descent . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
8.G How to skin a polytope . . . . . . . . . . . . . . . . . . . . . . . . 202
8.H Converse results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
8.I Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
9 Bass-Whitehead groups of monoid rings 223
9.A The functors K
1
and K
2
. . . . . . . . . . . . . . . . . . . . . . . . 223
9.B The nontriviality of SK
1
(RŒM|) . . . . . . . . . . . . . . . . . . . . 228
9.C Further results: a survey . . . . . . . . . . . . . . . . . . . . . . . . 240
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
10 Global varieties 247
10.A Toric varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
10.B Intersection theory for toric varieties . . . . . . . . . . . . . . . . 247
10.C Toric varieties with huge Grothendieck group . . . . . . . . . . . 247
10.D The equivariant Serre Problem for abelian groups . . . . . . . . . 247
Bibliography 249
Part I
Cones, monoids, and unimodular
triangulations
Chapter 1
Polytopes, cones, and complexes
1.A. Polyhedra and their faces
We suppose that the reader is familiar with the notion of affine space A over a
field K, especially the field R of real numbers, and related notions such as affine
subspace, affine map, and affine linear combination. They are introduced in almost
every textbook of linear algebra. For a subset X of Awe let aff(X) denote the affine
hull of X, i. e. the smallest affine subspace of A containing X. In the following all
affine and vector spaces are defined over the field R.
We use the symbols Z, Q, and C to denote the integral, rational, and complex
numbers respectively. The subsets of nonnegative elements in Z, Q, R will be
referred to by Z
÷
, Q
÷
, R
÷
. The symbol Ndenotes the set of positive integers.
The objects to be introduced are subsets of a finite-dimensional affine space
A, but some data depend on the choice of origin in A. For simplicity we usually
identify A with a d-dimensional vector space V and let the vector 0 represent the
origin.
In affine spaces one can introduce the notion of convexity. A subset X of V is
convex if it contains the line segment
Œx. y| = {ax ÷(1 ÷ a)y : 0 ¸ a ¸ 1]
for all points x. y ÷ X. (Open or halfopen line segments (x. y) and (x. y| are de-
fined analogously.) The intersection of convex sets is obviously convex. Therefore
one may define the convex hull conv(X) of an arbitrary subset X of V to be the
smallest convex subset of V containing X. If X = {x
1
. . . . . x
n
] is finite, then
conv(X) =
_
n

i=1
a
i
x
i
: 0 ¸ a
i
¸ 1. i = 1. . . . . n.
n

i=1
a
i
= 1
_
.
and for arbitrary X the convex hull of X is just the union of the convex hulls of the
finite subsets of X. Linear combinations as in the formula for conv(X) are called
convex combinations.
The functions considered in affine geometry are the affine forms ˛ : V ÷ R.
They are continuous with respect to the natural topology on V Š R
d
. So the open
3
4 1. Polytopes, cones, and complexes
halfspaces
H
>
˛
= {x ÷ V : ˛(x) > 0].
associated with the nonconstant forms ˛ are open sets. With such a form we asso-
ciate also the hyperplane
H
˛
= {x ÷ V : ˛(x) = 0]
and the closed halfspace
H
÷
˛
= {x ÷ V : ˛(x) ¸ 0]
For flexibility of notation we set
H
~
˛
= H
>

and H
-
˛
= H
÷

.
Each hyperplane H in V bounds exactly two open halfspaces that we denote by
H
>
and H
~
(depending on the choice of ˛ with H = H
˛
). The corresponding
closed halfspaces are denoted H
÷
and H
-
.
With respect to the standard topology we use the topological notions of the in-
terior int(X) and the boundary dX. However, in dealing with polyhedra we will
usually have to consider the relative interior and the relative boundary of P that
are taken within the affine subspace aff(X). For simplicity of notation we there-
fore use int(X) and dX to denote the relative interior and relative boundary of X.
Consequently we drop the attribute “relative” from now on. The closure of X is
denoted by X. (There is no need for a relative closure.) Note that int(X) = X if
X = {x] consists of a single point.
Note that each neighborhood of 0 contains a basis of the vector space V . It
follows that V is the affine hull of each of its open subsets.
The interior of a convex set in V is also convex: If x. y ÷ int(X), then there
exists an open neighborhood U of 0 such that x÷U. y÷U Ç X, and for z ÷ Œx. y|
the neighborhood z ÷U is contained in X.
We now introduce the main objects of this section.
Definition 1.1. A subset P Ç V is called a polyhedron if it is the intersection of
finitely many closed halfspaces. The dimension of P is given by dimaff(P). A
d-polyhedron has dimension d.
A polytope is a bounded polyhedron. A 2-polytope is called a polygon.
A morphism of polyhedra P and Q is a map ç : P ÷ Q that can be extended
to an affine map ¯ ç : aff(P) ÷aff(Q).
Figure 1.1. A polyhedron and a polytope
1.A. Polyhedra and their faces 5
Equivalently, P is a polyhedron if it is the set of solutions of a linear system of
inequalities. Since halfspaces (open or closed) are convex, polyhedra are convex
subsets of V , and the interior of a polyhedron is again convex (since the interior of
a convex set is convex, as observed above). It will be shown in Section 1.C that a
subset P of V is a polytope if and only if it is of the convex hull of a finite subset of
V .
In Proposition 1.36 we will see that ç(P) is a polyhedron if ç is a morphism of
polyhedra. For simplicity we will always consider a morphism ç : P ÷ Q to be
defined on aff(P) so that we need not distinguish ç and ¯ ç.
It is convenient to always orient the halfspaces cutting out a polyhedron P in
such a way that P = H
÷
˛
1
¨ ¨H
÷
˛
m
. Often we simply write H
÷
i
for H
÷
˛
i
, assuming
that a suitable affine form ˛
i
has been chosen to define the halfspace.
From the representation of P as an intersection of halfspaces we can easily
determine the affine hull aff(P).
Proposition 1.2. Let P = H
÷
1
¨ ¨ H
÷
m
be a polyhedron. Then aff(P) is the
intersection of those hyperplanes H
i
, i = 1. . . . . m, that contain P.
Proof. Let A be the intersection of the hyperplanes H
i
containing P. Then P is
contained in the affine space A. Replacing V by A, we must show that aff(P) = V
if P ,Ç H
i
for i = 1. . . . . m. In this case P contains a point x
i
÷ H
>
i
for i =
1. . . . . m. It follows that
1
m
(x
1
÷ ÷x
m
)
belongs to H
>
i
for i = 1. . . . . m. Therefore P ¸ H
>
1
¨ ¨ H
>
m
= 0 contains a
nonempty open subset of V , and V = aff(P).
Support hyperplanes and facets. If we visualize the notion of polytope in R
3
, then
we see a “solid body” whose boundary is composed of polygons. These facets (un-
less they are affine subspaces) are bounded by line segments, called edges, and
each edge has two endpoints. All these endpoints form the vertices of P. It is our
first goal to describe the face structure for polyhedra of arbitrary dimension.
Definition 1.3. A hyperplane H is called a support hyperplane of the polyhedron P
if P is contained in one of the two closed halfspaces bounded by H and H¨P = 0.
The intersection F = H¨P is a face of P, and H is called a support hyperplane
associated with F. A facet of P is a face of dimension dimP ÷ 1. The polyhedron
P itself and 0 are the improper faces of P.
Faces of dimension 0 are called vertices, and vert(P) is the set of vertices of P.
An edge is a face of dimension 1.
Below we will often use that the faces of P only depend on P in the following
sense: if V
t
Ç V is an affine subspace and P Ç V
t
, then the faces of P in V are its
faces in V
t
, and conversely. That a face in V is also a face in V
t
is trivial, and the
converse follows immediately from the fact that a halfspace G
÷
associated with
a hyperplane G in V
t
can be extended to a halfspace H
÷
of V such that G
÷
=
H
÷
¨ V
t
, as observed above.
6 1. Polytopes, cones, and complexes
Another easy observation: if F is a face of P, then F = aff(F) ¨ P. In fact,
aff(F) is contained in each support hyperplane associated with F.
Proposition 1.4. Let P Ç V be a polyhedron, and H a hyperplane such that H ¨
P = 0. Then H is a support hyperplane associated with a proper face of P if and
only if H ¨ P Ç dP.
Proof. We may assume that V = aff(P). Suppose first that H is a support hyper-
plane. If H¨int(P) = 0, then both inclusions P Ç H
÷
or P Ç H
-
are impossible
since none of the closed halfspaces contains a neighborhood of x ÷ H ¨ int(P).
Thus H ¨ P Ç dP.
As to the converse implication, if H intersects P, then one of the open halfspa-
ces, say H
>
, must contain an interior point x of P since int(P) Ç H is impos-
sible. If H is not a support hyperplane, then the other open halfspace H
~
must
also contain a point y ÷ P. But all points on the half open line segment Œx. y)
belong to int(P), as is easily checked. Exactly one of these points is in H so that
H ¨ int(P) = 0.
In the representation
P = H
÷
1
¨ ¨ H
÷
n
the halfspace H
÷
i
can be omitted if it contains the intersection of the remaining
halfspaces. Therefore P has an irredundant representation as an intersection of
halfspaces. It is evident from two-dimensional polyhedra in R
3
that we can achieve
the uniqueness of these halfspaces only if dimP = dimV . But this is not an essen-
tial restriction, since we can always consider P as a polyhedron in aff(P). First a
lemma:
Lemma 1.5. Suppose P = H
÷
1
¨ ¨ H
÷
n
is a polyhedron. Then a convex set
X Ç dP is contained in H
i
for some i .
Proof. On the contrary we assume that X ,Ç H
i
for all i and choose x
i
÷ X \ H
i
.
Then
x =
1
n
(x
1
÷ ÷x
n
) ÷ X Ç P.
but x ÷ H
>
i
for all i (as one sees immediately upon choosing an affine form defin-
ing H
÷
i
). The intersection H
>
1
¨ ¨ H
>
n
is open and contained in P. Thus
x ÷ int(P), a contradiction.
Theorem1.6. Let P Ç V be a polyhedron such that d = dimP = dimV . Then the
halfspaces H
÷
1
. . . . . H
÷
n
in an irredundant representation P = H
÷
1
¨ ¨H
÷
n
are
uniquely determined. In fact, the sets F
i
= P ¨ H
i
, i = 1. . . . . n, are the facets of
P.
Proof. We choose an irredundant representation P = H
÷
1
¨ ¨ H
÷
n
.
By Proposition 1.4 and Lemma 1.5, every facet F of P is contained in one of
the sets F
i
. Let H be the support hyperplane defining F. Then H = aff(P ¨
H) = aff(F) since F is a facet. But H is contained in H
i
, and so H = H
i
and
F = P ¨ H = P ¨ H
i
= F
i
.
1.A. Polyhedra and their faces 7
It remains to show that each F
i
is a facet of P. It is enough to consider F
n
. Let
P
t
= H
÷
1
¨ ¨H
÷
n-1
and set H = H
n
. Then H ¨P
t
= 0, but H is not a support
hyperplane of P
t
. By Proposition 1.4 H intersects the interior of P
t
. Therefore
H ¨ P
t
= H ¨ P contains a nonempty subset X that is open in H. But then
aff(P ¨ H) ¸ aff(X) = H, and so dimP ¨ H = dimH = dimP ÷ 1.
The face structure of a polyhedron. We note an immediate corollary of Theorem
1.6:
Corollary 1.7. Let P be a polyhedron.
(a) Then dP is the union of the facets of P.
(b) Each proper face of P is contained in a facet.
Part (b) of the corollary is very useful for inductive arguments, provided the
relation “F is a face of P” is transitive, and this is indeed true.
Proposition 1.8. Let F be a face of the polyhedron P and G Ç F. Then G is a face
of P if and only if it is a face of F.
Proof. Evidently a face G of P is a face of every face F ¸ G.
For the converse implication there is nothing to show if F = P, and this in-
cludes the case in which dimP = 0. Suppose dimP > 0. Then the proper face F
is contained in a facet F
t
, as we have seen in the corollary above, and, by induction,
G is a face of F
t
. Therefore we may assume that F itself is a facet.
For simplicity we may further assume that V = aff(P). Let H
t
be a support
hyperplane in aff(F) for the face G of F, and H the (uniquely determined) sup-
port hyperplane of P through F. There exists a point x ÷ H
~
that is contained
in the interior of the polyhedron P
t
defined as the intersection of the halfspaces
associated with the remaining facets of P. Set
¯
H = aff(H
t
L {x]). Then
¯
H is a
hyperplane in V . We claim that it intersects P exactly in G.
P
F
P
t
x
G
¨
H
Figure 1.2. The construction of
¨
H
Assume that y ÷ P ¨
¯
H, y … G. Then y … F, since
¯
H ¨ F = G. The line
segment Œy. x| is not contained in P, and it can leave P only through F. Since
Œy. x| Ç
¯
H, it intersects F in a point z ÷ G. On the other hand, all points on (y. x|
belong to the interior of P
t
, since y ÷ P
t
and x ÷ int(P
t
). Thus z ÷ int(F) =
H ¨ int(P
t
). This is a contradiction.
As we have seen, the maximal elements (with respect to inclusion) among the
proper faces are the facets. Now we turn to the minimal faces:
8 1. Polytopes, cones, and complexes
Proposition 1.9. Let P be a polyhedron and let U be the vector subspace given by
the intersection of the hyperplanes through 0 parallel to the facets of P.
(a) The sets x ÷U, x ÷ P, are the maximal affine spaces contained in P.
(b) Every minimal face F of P is a maximal affine subspace of P.
(c) In particular, if P has a bounded face, then its minimal nonempty faces are
its vertices.
Proof. Clearly x÷U Ç P since the affine forms defining P are constant on x÷U.
On the other hand, if Ais an affine subspace contained in P, then these forms must
be constant on A, and so A Ç x ÷U for x ÷ A. This proves (a).
For (b) there is nothing to prove if P has no proper face. Otherwise every
minimal face of P is contained in a facet, and it follows by induction on dimP that
the minimal faces F are affine subspaces of V . So F Ç x ÷ U Ç P for every
x ÷ F. If H is a support hyperplane associated with F, then U ÷ x Ç F, as is
easily seen (for an arbitrary face F).
(c) follows immediately from (b).
We can now prove several important properties of the set of faces of P:
Theorem 1.10. Let P be a polyhedron and Aa maximal affine subspace of P. Then
the following hold:
(a) The intersection F ¨ G of faces F. G is a face of F and G.
(b) Every face F is the intersection of facets. In particular, the number of faces
is finite.
(c) Let F
0
Ç Ç F
m-1
Ç F
m
= P be a strictly ascending maximal chain of
nonempty faces of P. Then dimF
0
= dimA and dimF
i÷1
= dimF
i
÷1
for all i = 0. . . . . m÷ 1.
(d) For each x ÷ P there exists exactly one face F such that x ÷ int(F). It is
the unique minimal element in the set of faces of P containing x.
Proof. (a) If F ¨G = 0, then F ¨G is a face by definition. Otherwise, a support
hyperplane of P associated with the face F is also a support hyperplane of the
polyhedron G that intersects G exactly in F ¨ G. Thus F ¨ G is a face of G, and
therefore of P.
(b) If F is a facet, then it is certainly an intersection of facets. Otherwise F
is strictly contained in a facet F
t
, and, by induction, the intersection of facets of
F
t
. But the facets of F
t
arise as intersections of facets of P with F
t
, as follows
immediately from Theorem 1.6.
(c) Follows immediately by induction from Corollary 1.7 and Proposition 1.9.
(d) By (a) and (b) the intersection of all faces of P containing x is a face F of
P. By Corollary 1.7 x ÷ int(F). Every other face G of P with x ÷ G has F as a
proper face, and so x ÷ dG.
It follows from Theorem 1.10 that for all faces F. G of P there are a unique
maximal face contained in F ¨G (namely F ¨G itself) and a unique minimal face
containing F LG. Therefore the set of faces, partially ordered by inclusion, forms
a lattice, the face lattice of P. (A partially ordered set M is called a lattice if each
1.A. Polyhedra and their faces 9
two elements x. y ÷ M have a unique infimum and a unique supremum. Most
often we will use the term “lattice” in a completely different meaning; see Section
1.G.)
In general a polyhedron need not have a bounded face, and as we see from
Proposition 1.9, it has a bounded face if and only if its minimal nonempty faces are
vertices. As the next proposition shows, the structure of polyhedra in general is
essentially determined by those with vertices.
For each affine subspace Aof V we can find a complementary subspace A
t
char-
acterized by the conditions that A ¨ A
t
consists of a single point x
0
and V =
A÷(A
t
÷x
0
) = (A÷x
0
) ÷A
t
. The map A× A
t
÷V , (x. x
t
) ÷x ÷x
t
÷x
0
, is
then a bijection, and we may write V = A× A
t
.
Proposition 1.11. Let P Ç V be a nonempty polyhedron, A a maximal affine sub-
space of P, A
t
a complementary subspace, and P
t
= A
t
¨ P. Then P = A × P
t
under the identification V = A× A
t
, and P
t
has vertices.
Proof. Let x
0
be the intersection point of A and A
t
. We can assume that x
0
= 0.
For a point y = x ÷x
t
÷ P, x ÷ A, x
t
÷ A
t
, one has y ÷x ÷ P, since y ÷A Ç P
for all y ÷ P. But y ÷ x = x
t
÷ P
t
. Conversely, if x
t
÷ P
t
Ç P, then x
t
÷x ÷ P
for all x ÷ A. This shows P = A × P
t
. Since A × B Ç P for each affine subspace
B of P
t
, the largest affine subspaces contained in P
t
are points.
There meet exactly two edges in a vertex of a polyhedron in R
2
and two facets
in an edge of a 3-dimensional polyhedron in R
3
. This fact can be generalized to all
dimensions:
Proposition 1.12. Let F be a nonempty face of the polyhedron P. Then dimF =
dimP ÷ 2 if and only if there exist exactly two facets F
t
and G
t
containing F. In
this case one has F = F
t
¨ G
t
.
Proof. If there exist exactly two facets F
t
and G
t
containing F, then dimF =
dimP ÷ 2 as follows immediately from Theorem 1.10, and clearly F = F
t
¨ G
t
.
For the converse implication we can assume that V = aff(P). Let dimF =
dimP ÷2, and let F
t
be a facet containing F. Then F is a facet of F
t
. By Theorem
1.10 the facets of F
t
are of the form F
t
¨ G
t
where G
t
is another facet of P. It is
enough to show that G
t
is uniquely determined. Suppose that F = F
t
¨ G
tt
for a
third facet G
tt
of P.
We choose affine-linear forms ˛. ˇ. ; with nonnegative values on P such that
F
t
= P ¨H
˛
, G
t
= P ¨H
ˇ
, G
tt
= P ¨H
;
. The affine linear forms vanishing on
F, and therefore on aff(F), form a two-dimensional vector subspace of the space
of all affine-linear forms. Thus there is a linear relation
a˛ ÷bˇ ÷c; = 0.
Since ˛. ˇ. ; define pairwise different facets, none of a. b. c can be zero. We can
assume that two of a. b. c are positive, say a and b. If c < 0, then H
÷
;
contains
H
÷
˛
¨ H
÷
ˇ
– impossible, since the facets define an irreducible representation of
P. But c > 0 is also impossible, since ˛. ˇ. ; are simultaneously positive on an
interior point of P.
10 1. Polytopes, cones, and complexes
Another “evident” property of polyhedra P is the connectedness of their
boundaries – unless P is a “sandwich” with exactly two parallel facets. In fact,
more is true: dP is connected in codimension 1, a property explained by
Theorem 1.13. Let P be a polyhedron with at least 2 nonparallel facets. Then for
each pair F. G of facets of P there exists a chain
F = F
0
. F
1
. . . . . F
n
= G
of facets of P such that F
i
¨ F
i÷1
is a facet of F
i
and F
i÷1
for i = 0. . . . . n ÷ 1
(so the intersection F
i
¨ F
i÷1
has codimension 1 with respect to F
i
and F
i÷1
). In
particular, dP is connected.
Proof. We may assume that V = aff(P), and use induction on the number of
facets of P. Let H
÷
1
. . . . . H
÷
m
be the halfspaces for which F
i
= H
i
¨ P are the
facets of P, and set P
t
= H
÷
1
¨ ¨ H
÷
m-1
. There is nothing to prove if m = 1.
For m > 1 we distinguish three cases:
(1) m = 2. If H
1
[ H
2
, then P is excluded by the hypothesis of the theo-
rem. Otherwise H
1
and H
2
meet in an affine space of codimension 2 and they are
“connected” by it.
(2) m = 3, and H
1
[ H
2
. Then H
3
is parallel neither to H
1
nor to H
2
, and
H
1
¨ H
3
and H
3
¨ H
2
“connect” the three facets of P.
(3) m ¸ 3 and dP
t
is connected in codimension 1. We have to show that this
property does not get lost when P
t
is intersected with H
÷
m
. First observe that H
m

P
t
. Otherwise all facets of P were parallel to H
m
, and this is impossible if m ¸ 3.
Pick x ÷ H
m
\ P
t
.
Obviously H
m
intersects int(P
t
). (Otherwise P
t
were contained in one of the
halfspaces of H
m
, and so P = P
t
.) We choose y ÷ int(P
t
) ¨H
-
m
. The line segment
Œy. x| leaves P
t
through a facet F
t
of P
t
, and the intersection point z is in H
~
m
. We
cannot have F
t
Ç H
-
m
, because H
÷
m
¨ F
t
is a facet of P (different from H
m
¨ P).
Thus F
t
¨ H
>
m
= 0 as well. This is only possible if H
m
¨ int(F
t
) = 0, so that
H
m
¨ F
t
has codimension 2 in V .
Let F
t
1
. . . . F
t
m-1
be the facets of P
t
. In general, a common facet F
t
i
¨ F
t
j
of
F
t
i
and F
t
j
may intersect P in a face of codimension > 2. This is only possible if
F
t
i
¨ F
t
j
¨ H
~
m
= 0. However, in this case H
m
meets both int(F
t
i
) and int(F
t
j
), as
seen above, and we are done.
That dP is connected, is now immediate. In fact, the facets, being convex, are
connected, and the facets along a chain as in the theorem intersect in nonempty
sets.
Theorem 1.13 can be generalized to faces of arbitrary dimension. Its hypoth-
esis is equivalent to the existence of faces of P of dimension < dimF, and in the
corollary we use the hypothesis in this (formally) weakest possible version.
Corollary 1.14. Let F and G be faces of P such that dimF = dimG, and sup-
pose P has a face F
t
with dimF
t
< dimF. Then there exists a chain F =
F
0
. F
1
. . . . . F
n
= G of faces of P such that F
i
¨ F
i÷1
is a facet of F
i
and F
i÷1
for
i = 0. . . . . n ÷ 1. In particular dimF
i
= dimF for all i .
1.A. Polyhedra and their faces 11
Proof. The faces F and G are contained in facets
¯
F and
¯
G to which we can apply
the theorem. Using induction on the chain of facets connecting them, we conclude
that we may assume that F and G are properly contained in the same facet of P.
But then we are done by induction on dimP ÷dimF: if P has a face of dimension
< dimF, then this holds for all facets of P, too, as follows from Theorem 1.10.
Finally we want to give another, “internal” characterization of faces. We call a
subset X of P extreme, if it is convex and if it contains every line segment Œx. y| for
which x. y ÷ P and (x. y) ¨ X = 0.
Proposition 1.15. Let P be a polyhedron. Then the faces of P are exactly its extreme
subsets.
Proof. Every face is an extreme subset for trivial reasons. Conversely, let X be an
extreme subset. If X ¨ int(P) = 0, then X is contained in a facet by Lemma 1.5,
and we are done by induction on dimP. Otherwise X contains an interior point x
of P. Let y ÷ P, x = y. Then the line through x and y contains a point z ÷ P
such that x is strictly between y and z. Hence Œy. z| Ç X, and in particular y ÷ X
which shows X = P.
The notion of face and extreme subset make sense for arbitrary convex sets.
However, while every face of a convex set is an extreme subset, the converse is false
in general (the reader should find an example).
Convex and strictly convex functions. Let A be an affine space and X Ç A a
convex set. Then a function ¡ : X ÷ R is convex if ¡ (tx ÷(1 ÷ t )y) ¸ t¡ (x) ÷
(1 ÷t )¡ (y) for all x. y ÷ X and t ÷ Œ0. 1|. In other words: ¡ is convex if the graph
of ¡ [Œx. y| is below the line segment
_
(x. ¡ (x)). (y. ¡ (y))
_
Ç A × R. (We use [ to
denote the restriction of functions or mappings to subsets.)
One can immediately prove Jensen’s inequality: ¡ (t
1
x
1
÷ ÷ t
n
x
n
) ¸
t
1
¡ (x
1
) ÷ ÷ t
n
¡ (x
n
) for all convex combinations of points in X. The reader
should check that convexity is a local property: ¡ is convex if and only if for each
x ÷ X there exists a convex neighborhood U of x such that ¡ [U ¨ X is convex.
The function ¡ is strictly convex if ¡ (tx ÷ (1 ÷ t )y) < t¡ (x) ÷ (1 ÷ t )¡ (y)
for all x. y ÷ X, x = y, and t ÷ (0. 1). Like convexity, strict convexity is a local
property.
Remark 1.16. If X is open in A, then for (strict) convexity it is enough that ¡ is
(strictly) convex on all line segments Œx. y| that are parallel to one of d = dimA
linearly independent vectors :
1
. . . . . :
d
. Since X is open, and (strict) convexity
is a local property, we can assume that X contains the smallest parallelotope that
has the points x and y as vertices and edges parallel to :
1
. . . . . :
d
. After an affine
transformation one can identify the parallelotope with a unit cube, and for the case
in which X is the unit cube, the claim is an easy exercise.
Concave and strictly concave functions are defined in the same manner, the
inequalities being replaced by the opposite ones.
12 1. Polytopes, cones, and complexes
1.B. Finite generation of cones
We have introduced polyhedra as the set of solutions of finite linear systems of
inequalities. In this subsection we study the special case in which the inequalities
are homogeneous, or, equivalently, the affine forms are linear forms on the finite-
dimensional vector space V (over the field R, just as in Section 1.A). The main
advantage of the linear case is the duality between a vector space V and the space
V
+
= Hom
R
(V. R), for which there is no affine analogue.
Definition 1.17. A cone in V is the intersection of finitely many linear closed half-
spaces. (We say that a halfspace is linear if it is defined by a linear form on V .)
A morphism of cones C and C
t
is a map ç : C ÷ C
t
that extends to a linear
map ¯ ç : RC ÷RC
t
.
Asubset X of V is conical if it is closed under nonnegative linear combinations:
a
1
x
1
÷ ÷a
n
x
n
÷ C for all x
1
. . . . . x
n
÷ C and all a
1
. . . . . a
n
÷ R
÷
, n ÷ Z
÷
.
A cone C is evidently conical. Often this property is used to define the notion
of cone. However, all our conical sets are polyhedra, and therefore the term “cone”
will always include the attribute “polyhedral”.
The faces of a cone are themselves cones since each support hyperplane H of
a cone C must contain 0: the ray R
÷
x, x ÷ C ¨ H, x = 0, cannot leave C, and
0 ÷ H
>
is impossible.
The intersection of conical sets is certainly conical. Therefore we may define
the conical hull R
÷
X of a subset X of V as the smallest conical set containing X.
In analogy with the convex hull, one has
R
÷
X = {a
1
x
1
÷ ÷a
n
x
n
: n ÷ Z
÷
. x
1
. . . . . x
n
÷ X. a
i
÷ R
÷
. i = 1. . . . . n]
if X = {x
1
. . . . . x
n
] is a finite set, and if X is infinite, then R
÷
X is the union of the
sets R
÷
X
t
, where X
t
Ç X is finite. We say that a conical set is finitely generated if
it is the conical hull of a finite set.
The central theorem in this subsection shows that finitely generated conical
sets are just cones, and conversely. In addition to the “external” representation of
cones as intersections of halfspaces one has an “internal” description by finitely
many elements generating C as a conical set.
Theorem 1.18. Let C Ç V be a conical set. Then the following are equivalent:
(a) C is finitely generated;
(b) C is a cone.
Proof. Suppose that C = R
÷
x
1
÷ ÷ R
÷
x
m
. For (a) == (b) we have to find
linear forms z
1
. . . . . z
n
such that C = H
÷
z
1
¨ ¨ H
÷
z
n
. We use induction on m,
starting with the trivial case m = 0 for which C = {0] is certainly polyhedral.
Suppose that
C
t
= R
÷
x
1
÷ ÷R
÷
x
m-1
= H
÷
×
1
¨ ¨ H
÷
×
r
for linear forms ×
1
. . . . . ×
r
. The process by which we now construct z
1
. . . . . z
n
is
known as (the dual version of) Fourier–Motzkin elimination. (Instead of solving a
system of inequalities we construct such a system for a given set of solutions.)
1.B. Finite generation of cones 13
Set x = x
m
. We may assume that
×
i
(x)
_
ˆ
_
ˆ
_
= 0 i = 1. . . . . p.
> 0 i = p ÷1. . . . . q.
< 0 i = q ÷1. . . . . r.
Set
j
ij
= ×
i
(x)×
j
÷ ×
j
(x)×
i
. i = p ÷1. . . . . q. j = q ÷1. . . . . r.
Then j
ij
(x
u
) ¸ 0 for u = 1. . . . . m÷ 1 and j
ij
(x) = 0. Let
{z
1
. . . . . z
n
] = {×
1
. . . . . ×
q
] L {j
ij
: i = p ÷1. . . . . q. j = q ÷1. . . . . r].
Evidently
C Ç D = H
÷
z
1
¨ ¨ H
÷
z
n
.
In fact, x
1
. . . . . x
m
are contained in D, and so this holds for each of their nonneg-
ative linear combinations.
It remains to prove the converse inclusion D Ç C. Choose y ÷ D. We must
find t ÷ R
÷
and z ÷ C
t
for which y = tx ÷ z. So we consider the ray y ÷ tx,
t ÷ R
÷
, and have to show that it meets C
t
.
z
x
y
Figure 1.3. Construction of z
We have to find the values of t ÷ R for which y ÷ tx ÷ C
t
. Equivalently, we
have to find out when y ÷ tx ÷ H
÷
×
i
for i = 1. . . . . r. We distinguish three cases:
(i) i ¸ p. Then ×
i
(x) = 0, ×
i
÷ {z
1
. . . . . z
n
], and so ×
i
(y) ¸ 0. Clearly
y ÷ tx ÷ H
÷
×
i
for all t ÷ R.
(ii) p ÷ 1 ¸ i ¸ q. Then ×
i
(x) > 0, ×
i
÷ {z
1
. . . . . z
n
], and ×
i
(y) ¸ 0.
Evidently y ÷ tx ÷ H
÷
×
i
if and only if t ¸ ×
i
(y),×
i
(x).
(iii) q ÷ 1 ¸ i ¸ r. Then ×
i
(x) < 0, and y ÷ tx ÷ H
÷
×
i
if and only if t ¸
×
i
(y),×
i
(x).
Thus the range of values of t ÷ R for which y ÷ tx ÷ C
t
is an intersection of
intervals of type (÷o. a
i
|, i = p ÷ 1. . . . . q, and Œb
j
. o), j = q ÷ 1. . . . . r. The
intersection of all these intervals and R
÷
is nonempty if and only if a
i
¸ 0 and
a
i
¸ b
j
for all values of i and j . But a
i
= ×
i
(y),×
i
(x) ¸ 0 since ×
i
(y) ¸ 0 and
×
i
(x) > 0, and a
i
¸ b
j
, as follows immediately from the inequality j
ij
(y) ¸ 0
(observe that ×
i
(x)×
j
(x) < 0).
The implication (a) == (b) has been proved, and we will see below that
(b) ==(a) follows from it (and is in fact equivalent to it).
We are nowjustified in speaking of the cone C generated by x
1
. . . . . x
m
. Clearly,
each face F of C is generated by the intersection F ¨ {x
1
. . . . . x
m
].
For the efficiency of Fourier–Motzkin elimination one should always produce
a shortest possible list z
1
. . . . . z
n
of linear forms in each step. This task can be
14 1. Polytopes, cones, and complexes
split into two parts: (i) Find a minimal set of linear forms cutting out the vector
subspace U generated by x
1
. . . . . x
m
, and (ii) find linear forms in U
+
that represent
the facets of P. Part (i) (if at all necessary) is done by Gaussian elimination. Once
one has a basis of U (chosen from x
1
. . . . . x
m
), one introduces new coordinates,
and considers C as a full-dimensional cone in U. Thus we may assume that U =
V . Part (a) of the next proposition tells us how to get started, and part (b) says how
to preserve irredundance when we add the next generator.
Proposition 1.19. Let V be a d-dimensional vector space.
(a) Suppose that the cone C is generated by a basis x
1
. . . . . x
d
of V , let H
i
be
the linear hyperplane generated by the elements x
j
, j = i , and choose H
÷
i
such that x
i
÷ H
÷
i
. Then C = H
÷
1
¨ ¨ H
÷
d
is the unique irredundant
representation of C as an intersection of halfspaces.
(b) Suppose that C
t
= H
÷
×
1
¨ ¨ H
÷
×
r
is a cone of dimension d, given as an
irredundant intersection of halfspaces. Let F
i
, i = 1. . . . . r be the facet
C
t
¨ H
÷
×
i
. Then,with the notation of the proof of Theorem 1.18, C = C
t
÷
R
÷
x is the intersection of the halfspaces H
÷
×
i
, i = 1. . . . . q, and those H
÷
µ
ij
for which there exists no k = i. j such that F
i
¨ F
j
Ç F
k
.
Proof. (a) By construction the intersections C ¨ H
i
are the only facets of C. The
rest follows from Theorem 1.18 and Theorem 1.6.
(b) Since H
×
i
, i = 1. . . . . q, intersects the d-cone C
t
in a facet and is a support
hyperplane of C, it must intersect C, too, in a facet. It remains to find those j
ij
=
×
i
(x)×
j
÷ ×
j
(x)×
i
that belong to facets of C.
Suppose first that H
µ
ij
intersects C in a facet F and let U be the vector space
generated by F. Similarly let F
t
= C
t
¨H
µ
ij
and U
t
be the vector space generated
by F
t
. Note that F
t
= F
i
¨ F
j
.
Then x … U
t
since U
t
Ç H
×
i
¨ H
×
j
and x … H
×
i
. On the other hand, U =
U
t
÷ Rx. We conclude that dimU
t
= dimU ÷ 1 = dimC
t
÷ 2. According to
Proposition 1.12 there exist exactly two facets of C
t
containing F
t
, and these are F
i
and F
j
.
For the converse one argues similarly, using the converse implication of Propo-
sition 1.12.
Polarity. The concept by which we will complete the proof of Theorem 1.18 is that
of polarity (or duality) of cones. For a subset X of V we set
X
+
= {z ÷ V
+
: z(x) ¸ 0 for all x ÷ X].
and call (the evidently conical set) X
+
the polar set of X. Since dimV < o, the
natural map h : V ÷V
++
, defined by
(h(:))(z) = z(:).
is an isomorphismof vector spaces. This shows that X
+
is not only conical, but also
an intersection of linear halfspaces. (The symbol
+
has two different meanings;
applied to V it always denotes the space of linear forms.)
1.B. Finite generation of cones 15
Furthermore, the identification V Š V
++
allows us to consider the bipolar set
X
++
as a subset of V , and, by definition,
X Ç X
++
.
Since X
++
is always conical and an intersection of halfspaces, the inclusion X Ç
X
++
is strict in general.
Using the implication (a) == (b) of Theorem 1.18 we now derive the duality
theorem for cones:
Theorem 1.20. Let C Ç V . Then the following hold:
(a) C is the intersection of (possibly infinitely many) halfspaces if and only if
C = C
++
.
(b) In particular, C = C
++
if C is a cone.
(c) Suppose C is a cone, and let z
1
. . . . . z
n
÷ V
+
. Then C = H
÷
z
1
¨ ¨ H
÷
z
n
if and only if C
+
= R
÷
z
1
÷ ÷R
÷
z
n
.
Proof. (a) By definition, C
++
is the intersection of linear halfspaces, namely of the
halfspaces H
÷
z
, z ÷ C
+
. On the other hand, if C =
_
z÷/
H
÷
z
for a subset ¹ of
V
+
, then ¹ Ç C
+
, and
C
++
=
_
z÷C

H
÷
z
Ç
_
z÷/
H
÷
z
= C.
Together with the inclusion C Ç C
++
, this shows that C = C
++
.
(b) follows immediately from (a).
(c) Suppose that C = H
÷
z
1
¨ ¨H
÷
z
n
Ç V
+
and set C
t
= R
÷
z
1
÷ ÷R
÷
z
n
.
Then C
t
is a finitely generated conical set, C
t
Ç C
+
, and C = (C
t
)
+
. By Theorem
1.18(a) ==(b) C
t
is a cone, and so (b) implies
C
t
= (C
t
)
++
= ((C
t
)
+
)
+
= C
+
.
Conversely, if C
+
= R
÷
z
1
÷ ÷R
÷
z
n
, then (C
+
)
+
= {z
1
. . . . . z
n
]
+
, and so
C = C
++
= {z
1
. . . . . z
n
]
+
, as desired.
As a first application we complete the proof of Theorem 1.18. Let C be a cone.
We have to showthat C is finitely generated as a conical set. By Theorem1.20(c) the
polar conical set C
+
is finitely generated. So we can apply Theorem1.18(a) ==(b)
and obtain that C
+
is a cone. It follows that C
++
is finitely generated, and since
C = C
++
, we are done.
Remark 1.21. The equality C = C
++
has many interpretations, for example as a
separation theorem: If C is a cone and x a vector not in C, then there exists a linear
form ç such that ç(y) ¸ 0 for all y ÷ C and ç(x) < 0. In other words: C and x
are separated by the hyperplane H
ç
.
It is a version of the Farkas lemma (see [103]) and has far reaching gener-
alizations in the Hahn-Banach separation theorem. The most general version of
Minkowski’s theorem is the theorem of Krein-Milman (see [52]).
Let F be a face of the cone C. Then
F
+
C
= {z ÷ C
+
: z(F) = 0]
16 1. Polytopes, cones, and complexes
is a conical set. Moreover, if az ÷ bj, z. j ÷ C
+
, a. b > 0, vanishes on F, then
z. j ÷ F
+
C
. So F
+
C
is an extreme subset of C
+
, and by Proposition 1.15 it is a face
of C
+
, the face polar to F (the subscript C is necessary since F
+
C
depends on C in
an essential way).
Theorem1.22. Let C be a cone in the vector space V . Then the assignment F ÷F
+
C
defines an order reversing bijection of the face lattices of C and C
+
. Moreover, one
has dimF ÷dimF
+
C
= dimV .
Proof. Let us first show that the assignment is injective. Choose a support hy-
perplane H of C associated with F and a linear form z such that H = H
z
and
C Ç H
÷
z
. Then z is contained in G
+
C
for a face G of C if and only if G Ç F. Thus
F
+
C
= G
+
C
implies F = G.
The assignment is evidently order reversing, and if we apply it to C
+
and C
++
=
C, then the composition of the two maps is an injective order preserving map of
the face lattice of C to itself. But the face lattice is a finite set, and so the composi-
tion must be the identity.
Since we can exchange the roles of C and C
+
, the first statement of the theorem
has been proved. By Theorem 1.10, which describes the chain structure of the face
lattice, we must have dimF
+
C
= dimG
+
C
÷1 if dimG = dimF ÷1. Therefore it is
enough for the second statement to show dimC ÷dimC
+
C
= dimV . This however
is evident: all linear forms vanishing on C belong to C
+
, and a linear formvanishes
on C if and only if it vanishes on the vector subspace generated by C.
Pointed cones. The most important class of cones are the pointed ones: we say
that C is a pointed cone if 0 is a face of C. Equivalently we can require that
x ÷ C. ÷x ÷ C ==x = 0.
and every cone decomposes into the direct sum of a pointed one and a vector sub-
space. This follows immediately from
Proposition 1.23. Let C be a cone and C
0
= {x ÷ C : ÷x ÷ C].
(a) C
0
is the maximal vector subspace of C and the unique minimal face of C.
(b) Let U be a vector subspace complement of C
0
. Then C = C
0
˚(C ¨ U),
and C ¨ U is a pointed cone.
The proposition is just a special case of Proposition 1.11. By C = C
0
˚ C
t
we indicate that each element of C has a unique decomposition x = x
0
÷x
t
with
x
0
÷ C
0
and x
t
÷ C
t
.
Pointed cones can be characterized in terms of their polar cones:
Proposition 1.24. A cone C in the vector space V is pointed if and only if dimC
+
=
dimV .
Proof. A vector subspace U of V is contained in C if and only if C
+
is contained
in the polar vector space U
+
.
1.C. Finite generation of polyhedra 17
The proposition shows that the class of full-dimensional pointed cones is closed
under polarity. Very often we will use that a pointed cone has an essentially unique
minimal system of generators:
Proposition 1.25. Let C be a pointed cone, and choose an element x
i
= 0 from each
1-dimensional face R
1
. . . . . R
n
of C. Then x
1
. . . . . x
n
is, up to order and positive
scalar factors, the unique minimal system of generators of C.
Proof. By Proposition 1.24 one has dimC
+
= dimV = dimV
+
. By Theo-
rem 1.6 C
+
has a unique description as an irredundant intersection of halfspaces
H
÷
1
. . . . . H
÷
n
. So Theorem 1.20 implies that C has a unique minimal system of
generators, up to positive scalar factors, since the elements x
i
of V = V
++
with
H
÷
x
i
= H
÷
i
are uniquely determined up to positive scalar factors. Moreover, the
faces of C that are polar to the facets of C
+
are exactly the 1-dimensional faces of
C, and the face polar to C
+
¨ H
i
contains x
i
.
The 1-dimensional faces of a pointed cone C are called its extreme rays. As a
consequence of Proposition 1.25 we obtain that pointed cones have polytopes as
cross-sections:
Proposition 1.26. Let C be a pointed cone, z ÷ int(C
+
), and a ÷ R, a > 0. Then
the hyperplane H = {x : z(x) = a] intersects C in a polytope P, and one has
R
÷
P = C.
Proof. Since z ÷ int(P
+
), the hyperplane H
0
= {x : z(x) = 0] intersects C only
in 0, as follows from Theorem 1.22 and Proposition 1.24. Therefore az(x)
-1
x ÷ P
for all x ÷ C, x = 0, and C = R
÷
P.
It remains to be verified that P is bounded. Let R be an extreme ray of C. It
suffices to show that R intersects H in exactly one point. Then Proposition 1.25
implies that P is the convex hull of the finite set of the intersection points of H
with the extreme rays of C.
Let G be the facet of C
+
polar to R, and U the vector subspace generated by
G. Then the linear forms vanishing on R are exactly the elements of U (note that
dimU = dimC ÷ 1). Since z … U, the hyperplane H intersects R in exactly one
point.
In the situation of Proposition 1.26 we will say that H defines a cross-section
of C.
1.C. Finite generation of polyhedra
We have introduced polyhedra as the set of solutions of linear systems of inequal-
ities. It is an elementary fact from linear algebra that the set S of solutions of a
linear system of equations is an affine subspace of the form S = x ÷S
0
where S
0
is the vector space of solutions of the associated homogeneous systemof equations.
As we will see, polyhedra have a similar description that in fact characterizes them.
18 1. Polytopes, cones, and complexes
Recession cone and projectivization. In the same way as one associates a homoge-
neous system of equations with an inhomogeneous one, we associate a cone with a
polyhedron:
Definition 1.27. Let P = H
÷
1
¨ ¨H
÷
n
be a polyhedron. Then the recession cone
of P is
rec(P) =
¯
H
÷
1
¨ ¨
¯
H
÷
n
where
¯
H
÷
i
is the vector halfspace parallel to the affine halfspace H
÷
i
, i = 1. . . . . n
(so
¯
H
÷
i
= H
÷
i
÷ x for some x ÷ H
÷
i
). (Halfspaces are parallel if their bounding
hyperplanes are parallel.)
x
x ÷rec(P) 0
rec(P)
Figure 1.4. A polyhedron and its recession cone
It is easy to see and left to the reader that the recession cone can also be de-
scribed as the set of “infinite directions” in P:
Proposition 1.28. Let P be a polyhedron, x ÷ P, and : ÷ V . Then : ÷ rec(P) if
and only if x ÷a: ÷ P for all a ÷ R
÷
.
We have constructed the recession cone by passing from an inhomogeneous
system of linear equations to the associated homogeneous system. There is a sec-
ond construction leading from an inhomogeneous system to a homogeneous one,
namely projectivization (or homogenization): we introduce a new variable and
view the constant as its coefficient. In other words, with an affine form
˛ = z ÷ˇ. ˇ = ˛(0).
we associate the linear form
˛ : V ˚R ÷R. ˛(:. h) = z(:) ÷ˇh.
The affine hyperplane and the affine halfspaces associated with ˛ are then ex-
tended to the linear hyperplane and the linear halfspaces associated with ˛. We
indicate the extension by .
We have chosen the letter h for the auxiliary variable since we interpret h as the
“height” of the point (:. h) over :. In the following we have to embed subsets of V
into V = V ˚R at a specific height. Thus we set
(X. b) = X × {b] Ç V
for X Ç V .
1.C. Finite generation of polyhedra 19
Definition 1.29. Let P = H
÷
1
¨ ¨ H
÷
n
Ç V be a polyhedron. Then we define
the cone over P by
C(P) = H
÷
1
¨ ¨ H
÷
n
¨ H
÷
y
where y : V ÷R is given by y(:. h) = h.
h
Figure 1.5. Cross-sections of C(P)
Clearly (P. 1) = {(:. h) ÷ C(P) : h = 1]. Thus C(P) is just the cone with
cross-section P at height 1, a statement made precise by
Proposition 1.30. Let C be the conical hull of (P. 1) in V .
(a) Then C(P) ¨ (V. 0) = (rec(P). 0).
(b) C(P) ¨ (V. a) = C ¨ (V. a) for a = 0.
(c) C(P) is the closure of C in V .
Proof. (a) is trivial and, moreover, C(P) = C L (rec(P). 0), as is easily checked.
This proves (b). For (c) it is therefore enough to show that (rec(P). 0) is contained
in the closure of C.
Pick : ÷ rec(P) and choose y ÷ P. Then y ÷ t : ÷ P for all t ÷ R
÷
, and
W = {u(y ÷ t :. 1) : t. u ÷ R
÷
] Ç C. But W contains a point from every
neighborhood of (:. 0) in V . Therefore (:. 0) is contained in the closure of W,
and, a fortiori, in the closure of C.
Figure 1.5 illustrates the construction of C(P), showing the recession cone at
h = 0, (P. 1), and (2P. 2).
The theorems of Minkowski and Motzkin. According to Theorem 1.18 C(P) is
finitely generated, say C(P) = R
÷
y
1
÷ ÷R
÷
y
m
. If y
i
has height h > 0, then we
can replace it by h
-1
y
i
. After this operation {y
1
. . . . . y
m
] decomposes into a subset
Y
0
of elements of height 0 and a subset Y
1
of elements of height 1. Let x
i
= ¬(y
i
)
where ¬ is the projection V ˚R ÷V .
Choose x ÷ P. Then z = (x. 1) can be written as a linear combination
z =

y÷Y
0
a
y
y ÷

y÷Y
1
a
y
y. a
y
÷ R
÷
. y ÷ Y
0
L Y
1
.

y÷Y
1
a
y
= 1. (1.1)
A comparison of the last components shows that indeed

y÷Y
1
a
y
= 1. Further-
more ¬(y) ÷ P for y ÷ Y
1
and ¬(y) ÷ rec(P) for y ÷ Y
0
.
A first consequence of these observations is Minkowski’s theorem:
20 1. Polytopes, cones, and complexes
Theorem 1.31. Let P Ç V . Then the following are equivalent:
(a) P is a polytope;
(b) P is a polyhedron and P = conv(vert(P));
(c) P is the convex hull of a finite subset of V .
Proof. For the implication (a) == (b) one notes that rec(P) = 0 if P is a poly-
tope. So Proposition 1.30 implies that C(P) is a pointed cone with C(P) ¨(V. 0) =
0, and thus we can apply Proposition 1.25: the elements of height 1 in the extreme
rays of C(P) form a minimal system of generators of C(P), and they are evidently
the vertices of P. We have Y
0
= 0 in equation (1.1), and it follows immediately
that P = conv(vert(P)).
The implication (b) == (c) is trivial since every polyhedron has only finitely
many vertices.
As to (c) == (a), if P is the convex hull of a finite set, then it is certainly
bounded, and the conical hull C of (P. 1) is finitely generated. So Theorem 1.18
implies that C is a cone. But then C ¨(V. 1) is a polyhedron that can be identified
with P. To sum up: P is a polytope, and the proof of Minkowski’s theorem is
complete.
If we give up the restriction rec(P) = 0, then we obtain Motzkin’s theorem:
Theorem 1.32. Let P Ç V . Then the following are equivalent:
(a) P is a polyhedron;
(b) there exist a polytope Qand a cone C such that P = Q÷C.
Proof. For the implication (a) == (b) we choose Q = conv(¬(Y
1
)) and let C
be the cone generated by ¬(Y
0
). The notation is as for equation (1.1), and this
equation proves the claim.
For the converse we choose a finite system X of generators of C. Then the
conical hull of (X. 0) Lvert(Q. 1) is a cone. Its cross-section at height 1 is not only
equal to (Q÷C. 1), but also a polyhedron.
Since a polytope is the convex hull of its finite vertex set, and since a cone is
finitely generated, Theorem 1.32 allows us to say that a polyhedron is finitely gen-
erated.
If P has vertices, then we call the union of the bounded faces the bottom of P.
(In Figure 1.4 the bottom is indicated by a thick line.)With this notion we can give
a more precise statement about the generation of polyhedra.
Proposition 1.33. Let P be a polyhedron, A a maximal affine subspace contained
in P, A
t
a complementary affine subspace and P
t
= P ¨ A
t
. Furthermore let
¬ : V ÷A
t
denote the parallel projection along A, and B
t
the bottom of P
t
.
Then P = B
t
÷ rec(P). Moreover, if P = B ÷ C for a bounded set B and a
cone C, then B
t
Ç ¬(B) and C = rec(P).
Proof. Since P = P
t
÷rec(A) by Proposition 1.11, the equation P = B
t
÷rec(P)
follows from the analogous equation for P
t
. In proving the first statement we may
therefore assume that P = P
t
or, in other words, that P has vertices.
1.C. Finite generation of polyhedra 21
Let x ÷ P and choose y ÷ rec(P). Then x ÷ ty ÷ P for all t ÷ R
÷
, but the
line x ÷ Ry cannot be contained in P; see Proposition 1.9. So there exists t ¸ 0
for which x
t
= x ÷ ty belongs to a facet of P. But F has vertices, too, and we can
apply induction since the proposition is obviously true in dimension 0: there exists
x
tt
in the bottom of F and y
t
÷ rec(F) such that x
t
= x
tt
÷y
t
. But x
tt
belongs to
the bottom of P and y
t
÷ rec(P). So x = x
tt
÷(y
t
÷ty) with y
t
÷ty ÷ rec(P).
We leave the proof of the second assertion to the reader.
First of all, note that the proof of the theorem contains a new demonstration of
the implications (a) ==(b) of Minkowski’s and Motzkin’s theorems.
Second, the proposition has the following interpretation: B
t
is a minimal sys-
tem of generators of P as a rec(P)-“module”, and a subset B of P is a minimal
system of generators if and only if it differs from B
t
by “units” of rec(P), namely
elements of the maximal subgroup contained in rec(P). In particular, if P has ver-
tices, then the bottom of P is the unique minimal system of generators of P over
rec(B).
Remark 1.34. The separation theorem for cones in Remark 1.21 implies a separa-
tion theorem for polyhedra: Let P be a polyhedron and x … P; then there exists a
hyperplane separating P and x.
For completeness we include a comparison of the face lattices of P, rec(P), and
C(P).
Proposition 1.35. Let P be a polyhedron, and F a nonempty face of P.
(a) Then rec(F) is face of rec(P). Moreover, F is unbounded if and only if
rec(F) = 0.
(b) C(F) is a face of C(P) that is not contained in (V. 0).
(c) Every face of rec(P) is of the form rec(G) for a face G of P.
(d) The faces of C(P) ¨ (V. 0) are the faces of (rec(P). 0).
(e) The faces of C(P) not contained in (V. 0) are of the form C(G) for a face G
of P.
(f) The following are equivalent:
(i) P has vertices;
(ii) rec(P) is pointed;
(iii) C(P) is pointed.
Several of the statements are just relative versions of the theorems above, and
the easy proofs of the remaining claims are left to the reader.
Minkowski sums and joins. Minkowski’s and Motzkin’s theorems help us to un-
derstand the behavior of polyhedra under affine maps.
Proposition 1.36. Let V. V
t
be vector spaces over R, ç : V ÷ V
t
an affine map, P
a polyhedron in V , and P
t
a polyhedron in V
t
.
(a) Then ç(P) and ç
-1
(P
t
) are polyhedra.
(b) ç(P) is a polytope if P is a polytope.
(c) If ç is linear and P. P
t
are cones, then ç(P) and ç
-1
(P
t
) are cones.
22 1. Polytopes, cones, and complexes
Proof. Since ç = [ ÷:
t
for a linear map [ : V ÷V
t
and some :
t
÷ V
t
, we may
assume that ç is linear.
For the preimage we start from a description P = H
÷
1
¨ ¨H
÷
n
. The preim-
age of a (linear) halfspace under a linear map is a (linear) halfspace, and the preim-
age of the intersection is the intersection of the preimages.
The image ç(C) of a finitely generated conical set C Ç V is certainly conical
and finitely generated. Therefore, if P is a cone, then so is ç(P). For (b) we argue
similarly, using Minkowski’s theorem and convex sets.
In order to show that the image of a polyhedron is a polyhedron, we write it in
the formQ÷C where Qis a polytope and C is a cone. Then ç(P) = ç(Q)÷ç(C)
is a polyhedron, as follows from (b), (c) and Motzkin’s theorem.
As a consequence of Proposition 1.36 the Minkowski sum
P ÷P
t
= {x ÷x
t
: x ÷ P. x
t
÷ P
t
]
of polyhedra (polytopes) is a polyhedron (polytope):
Theorem 1.37. Let P. P
t
be polyhedra in V .
(a) Then P ÷P
t
is a polyhedron.
(b) P ÷P
t
is a polytope if (and only if) P and P
t
are polytopes.
(c) If P and P
t
are cones, then P ÷P
t
is a cone.
Proof. Consider P × P
t
in V ˚ V . Together with P and P
t
it is a polyhedron,
polytope or cone, respectively. Now we can apply Proposition 1.36 to the linear
map V ˚V
t
÷V , (:. :
t
) ÷: ÷:
t
.
Another natural construction associated with polytopes P and P
t
is the convex
hull Q = conv(P L P
t
) of their union. Evidently Q is the convex hull of the finite
set vert(P) L vert(P
t
), and thus it is a polytope by Minkowski’s theorem. It is
likewise easy to see that
D = conv(C L C
t
) = C ÷C
t
is a cone if C and C
t
are cones. However, conv(P L P
t
) need not be a polyhedron
if P and P
t
are just polyhedra. For example if P is a line in R
2
and P
t
consists of
a single point x not in P, then the smallest polyhedron containing P and P
t
is the
strip bounded by P and its parallel through x. Of all the points on the parallel,
only x is in conv(P L P
t
). But conv(P L P
t
) may fail to be a polyhedron only
because it need not be closed. For this reason we introduce the closed convex hull
conv(X) = conv(X)
as an additional construction.
Proposition 1.38. Let P and P
t
be polyhedra. Then conv(P L P
t
) is a polyhedron.
Proof. We write P = Q÷rec(P) and P
t
= Q
t
÷rec(Q) with polytopes Q and
Q
t
. Then
R = conv(QL Q
t
) ÷rec(P) ÷rec(Q)
is a polyhedron by Motzkin’s theorem and it contains conv(P L P
t
) as is easily
verified. On the other hand rec(T) Ç rec(P) ÷ rec(P
t
) for every polyhedron
1.D. Polyhedral Complexes 23
T ¸ P L P
t
, as follows immediately from Proposition 1.28. Therefore R is the
smallest polyhedron containing P L P
t
. Clearly R ¸ conv(P L P
t
).
It remains to show that R Ç conv(P L P
t
). Pick s ÷ R,
s = ax ÷(1 ÷ a)y ÷n ÷z. x ÷ Q. y ÷ Q
t
. a ÷ Œ0. 1|.
n ÷ rec(P). z ÷ rec(P
t
).
If a = 0. 1, then s ÷ conv(P L P
t
), and if a = 0, then each neighborhood of s
contains a point s
t
= a
t
x ÷(1 ÷a
t
)y ÷n÷z, a
t
> 0, with s
t
÷ conv(P LP
t
). For
a = 1 one argues similarly.
We leave it to the reader to find the exact conditions under which conv(P L
P
t
) = conv(P L P
t
).
As a last construction principle for polyhedra we introduce the join. Roughly
speaking, it is the “free convex hull” that we obtain by considering polyhedra in
positions independent of each other. Let P Ç V and Q Ç W be polyhedra. Then
we set V
t
= V ˚W ˚R,
P
t
= {(x. 0. 0) : x ÷ P]. Q
t
= {(0. y. 1) : y ÷ Q].
and
join(P. Q) = conv(P
t
L Q
t
).
If P and Q are polyhedra in R
d
such that aff(P) ¨ aff(Q) = 0, then conv(P ¨
Q) will also be called the join of P and Q. (This polyhedron is isomorphic to
conv(P
t
¨Q
t
).) By Proposition 1.38 the join of P and Qis a polyhedron. If P and
Qare polytopes, then join(P. Q) is a polytope.
1.D. Polyhedral Complexes
It is evident (in the true sense of the word) that a polygon, i. e. a 2-dimensional
polytope can be triangulated, as illustrated by Figure 1.6(a):
(a)
(b)
(c)
Figure 1.6. Polyhedral complexes
In higher dimensions triangles are generalized by simplices:
Definition 1.39. A d-polytope with exactly d ÷1 vertices is called a d-simplex. In
other words, a polytope is a simplex if its vertex set is affinely independent. The
“conical analogue” is a simplicial cone, generated by a linearly independent set of
vectors.
24 1. Polytopes, cones, and complexes
In the triangulation in Figure 1.6(a) the intersection of two triangles P
1
and P
2
and, more generally, of faces F
1
Ç P
1
, F
2
Ç P
2
, is a face of P
1
and of P
2
. This
compatibility condition defines a polyhedral complex. In the example above the
polygon P and all triangles are contained in a single affine space. This condition
is too strict in general. We weaken it as follows.
Definition 1.40. A polyhedral complex consists of (i) a finite family ˘ of sets, (ii)
a family P
p
, p ÷ ˘, of polyhedra, and (iii) a family ¬
p
: P
p
÷ p of bijections
satisfying the following conditions:
(a) for each face F of P
p
, p ÷ ˘, there exists ¡ ÷ ˘ with ¬
p
(F) = ¡ ;
(b) for all p. q ÷ ˘ there exist faces F of P
p
and G of P
q
such that p ¨ q =
¬
p
(F) = ¬
q
(G) and, furthermore, the restriction of ¬
-1
q
ı ¬
p
to F is an
isomorphism of the polyhedra F and G.
We speak of a polytopal complex if the polyhedra P
p
are polytopes, and of a conical
complex if they are cones. In the latter case we require that the restriction of ¬
-1
q
ı
¬
p
is an isomorphism of cones.
If all the polyhedra P are simplices, then the complex is simplicial. A simplicial
conical complex consists of simplicial cones.
We denote a polyhedral complex simply by ˘, assuming that the polyhedra P
p
and the maps ¬
p
, p ÷ ˘, are given implicitly. Moreover, [˘[ stands for the support
_
p÷˘
p of ˘.
A subcomplex of ˘ is a subset ˘
t
that is itself a polyhedral complex (with re-
spect to the families P
p
and ¬
p
, p ÷ ˘
t
).
An isomorphism of polyhedral complexes ˘. ˘
t
is a bijective map ç : [˘[ ÷

t
[ such that ˘
t
= {ç(p) : p ÷ ˘] and for each p ÷ ˘ the map (¬
t
p
0
)
-1
ı ç ı ¬
p
is an isomorphism of the polyhedra P
p
and P
t
p
0
, p
t
= ç(p). In the case of conical
complexes one requires that (¬
t
p
0
)
-1
ı ç ı ¬
p
is an isomorphism of cones.
Example 1.41. we choose the following 6 points in R
3
:
x
1
= (0. 0. 0). x
2
= (1. 0. 0). x
3
= (1. 1. 0).
x
4
= (0. 1. 0). x
5
= (0. 0. 1). x
6
= (0. 1. 1).
Let M be the subset of R
3
formed by the union of Q
1
= conv(x
1
. x
2
. x
3
. x
4
),
Q
2
= conv(x
1
. x
4
. x
5
. x
6
) and the set S that is obtained by mapping the unit
square Q
3
= conv((0. 0). (1. 0). (1. 1). (0. 1)) in R
2
so that the image is a scroll
turning by 180
ı
, identifying the left edge E
1
= Œ(0. 0). (0. 1)| with Œx
2
. x
3
| and the
right edge E
2
= Œ(1. 0). (1. 1)| with Œx
6
. x
5
|, reversing the orientation. (The image
is a topological realization of the Möbius strip, see Figure 1.6(b)). Then the family
consisting of
(i) the squares Q
1
. Q
2
, their 7 edges, their vertices x
1
. . . . . x
6
, 0, the set S,
and the images of Œ(0. 0). (1. 0)| and Œ(0. 1). (1. 1)|, and
(ii) the identity mappings for all the polytopes except Q
3
. E
1
. E
2
, for which
we choose the (restrictions of) the scroll map,
is a polyhedral complex.
1.D. Polyhedral Complexes 25
Example 1.42. Let ˘ be a polyhedral complex and c > 0. Then we obtain an
isomorphic complex c˘ as follows. The sets p ÷ ˘ remain unchanged, however,
each polytope P
p
is replaced by cP
p
, and ¬
p
is replaced by ¬
p
ıc
-1
, where c
-1
(as a
map) denotes multiplication by c
-1
÷ R. While this example is rather irrelevant at
this stage, it will be very important later, where an additional structure will break
the isomorphism.
An important class of polyhedral complexes is given by those with a “global”
embedding:
Proposition 1.43. Let ˘ be a finite family of polyhedra P Ç R
d
satisfying the fol-
lowing conditions:
(a) if F is a face of P ÷ ˘, then F ÷ ˘;
(b) P ¨ Qis a face of both P and Qfor all P. Q ÷ ˘.
Then ˘, together with the family ¬
P
= id
P
, P ÷ ˘, of bijections, is a polyhedral
complex.
In fact, the conditions for a polyhedral complex are obviously satisfied.
Definition 1.44. The polyhedral complexes given by Proposition 1.43 are called
embedded. An embedded conical complex is called a fan.
The face lattice of a polyhedron P is a polyhedral complex as we have seen in
Section 1.A. The set of all faces F = P of P is also a polyhedral complex, the
boundary complex of P. More generally, if X is the union of faces of P, then the
family of all faces F Ç X forms an embedded polyhedral complex. We call such
complexes boundary subcomplexes.
Let us now show that the family ˘, together with the operation (p. q) ÷p¨q,
has properties that we have found already in the face lattice of a polyhedron:
Proposition 1.45. Let ˘ be a polyhedral complex, and p. q. ¡ ÷ ˘ with ¡ Ç p
and p = q. Then
(a) there exists a (unique) face F of P
p
such that ¬
p
(F) = ¡ , and ¬
p
ı ¬
-1
(
maps F isomorphically onto P
(
;
(b) ¬
p
(int(P
p
)) ¨ ¬
q
(int(P
q
)) = 0;
(c) p ¨ q ÷ ˘.
Proof. (a) is just a specialization of 1.40(b) to the case in which q Ç p.
(b) we can assume that p ,Ç q. Then p¨q is a proper subset of p and the face F
of p with ¬
p
(F) = p¨q is a proper face of P. We have ¬
p
(int(P
p
))¨¬
q
(int(P
q
)) Ç
¬
p
(int(P
p
)) ¨ ¬
p
(F) = 0 – again we have used the injectivity of ¬
p
.
(c) There exists a face F of P
p
with ¬
p
(F) = p ¨ q, and ¬
p
(F) ÷ ˘ by
1.40(a).
Example 1.46. Let H
1
. . . . . H
n
be affine hyperplanes in R
d
. Such a set of hyper-
planes defines an (embedded) polyhedral complex ˘ consisting of 0 and all those
polyhedra H
÷
i
1
¨ ¨ H
÷
i
r
¨ H
-
j
1
¨ ¨ H
-
j
s
, 0 ¸ r ¸ n, 0 ¸ j ¸ n, such that
{i
1
. . . . . i
r
. j
1
. . . . . j
s
] = {1. . . . . n]. Such polyhedral complexes are called hyper-
plane dissections. The support of the complex is R
d
: with each x ÷ R
d
we associate
26 1. Polytopes, cones, and complexes
its “sign vector” ÷ {÷1. ÷1. 0]
n
in an obvious way, and the sign vector determines
the face of ˘ containing x.
Remark 1.47. In an embedded polyhedral complex the sets p ÷ ˘ coincide with
the polytopes P
p
. But also in the abstract setting this identification is allowed after
we have realized the polyhedra P
p
, p ÷ ˘ in disjoint affine spaces (vector spaces
in the case of a conical complex): we then identify each x ÷ P
p
with ¬
p
(x) ÷ p.
Under this identification, each face F of P
p
is identified with the set ¡ ÷ ˘ for
which ¬
p
(F) = ¡ . If p ¨ q = ¬
p
(F) = ¬
q
(G) for faces F of P
p
and G of P
q
,
then, under the identification of P with P
p
and Q with P
q
, P ¨ Q = p ¨ q.
Henceforth we identify p and P
p
.
Thus we find ourselves in a situation which is almost identical to that of an
embedded complex, except that taking convex (or conical) hulls, or linear combi-
nations or intersections with halfspaces does not make sense globally. However,
such combinations can be formed locally. We return to the formal framework of
Definition 1.40 for the argument to be given. Let x
1
. . . . . x
n
÷ [˘[ and suppose
that there exists p. q ÷ ˘ with x
1
. . . . . x
n
÷ p ¨ q. We have to show that
¬
p
(a
1
¬
-1
p
(x
1
) ÷ ÷a
n
¬
-1
p
(x
n
)) = ¬
q
(a
1
¬
-1
q
(x
1
) ÷ ÷a
n
¬
-1
q
(x
n
))
for all a
1
. . . . . a
n
÷ R
÷
with

i
a
i
= 1. But the equation follows immediately
upon the application of ¬
-1
q
to both sides, since ¬
-1
q
ı ¬
p
(where defined) respects
affine linear combination by definition (this is the first time we use this property).
In the case of conical complexes ¬
-1
q
ı ¬
p
respects linear combinations.
It is a useful observation that line segments Œx. y| do not “ramify”: suppose
that Œx. y| Ç P ÷ ˘ and Œx. y
t
| Ç Q ÷ ˘; if there exists z ÷ (x. y| ¨ (x. y
t
|,
then y. y
t
÷ P ¨ Q and Œx. y| Ç Œx. y
t
| or viceversa. In fact, we can replace
P and Q by their smallest faces containing Œx. y| and Œx. y
t
|, respectively. Then
(x. z) Ç int(P) ¨ int(Q). It follows that P = Q (by the disjoint decomposition
of [˘[ into relative interiors) so that we can argue in aff(P) where the claim is
obvious. (The reader should generalize this observation from line segments to
subpolytopes of higher dimension.)
The preceding discussion shows that it is justified to extend the notions intro-
duced for single polyhedra to polyhedral complexes. The polytopes P ÷ ˘ are
the faces of ˘. The 0-dimensional faces are called vertices, and the 1-dimensional
faces are the edges. The edges of a conical complex are also called its rays. The
maximal faces are the facets of ˘ – so the facets of a polyhedron Q are the facets
of its boundary complex. The dimension of ˘ is the maximum of the dimensions
of its faces.
Often it is useful to work with the skeletons of a polyhedral complex, obtained
as follows: the e-skeleton ˘
(e)
consists of all polyhedra in ˘ that have dimension
¸ e. The 1-skeleton is a graph (undirected, without multiple edges and loops).
Example 1.48. We continue the discussion of Example 1.41, claiming that the poly-
hedral complex ˘ constructed there is not isomorphic to any embedded polyhe-
dral complex. In fact, the 2-dimensional polyhedra underlying ˘ are unit squares;
1.D. Polyhedral Complexes 27
the polytopes (affinely) isomorphic to such squares are exactly the parallelograms,
and one sees immediately that there is no realization in 2-space.
But also in higher dimension a realization is impossible. In fact, we can assume
that (the points representing) x
1
. x
2
. x
4
. x
6
are affinely independent (otherwise the
realization would be “flat”). But then the affine space spanned by them contains all
the six vertices, so that we have a realization already in 3-space. Since we are free
to apply an affine transformation, we can now assume that the x
i
have exactly the
coordinates given. Then the third quadrangle realizing M is conv(x
2
. x
3
. x
5
. x
6
) –
a contradiction since x
2
and x
6
must be connected by an edge.
It is not much harder to see that even the combinatorial type of ˘ has no em-
bedded realization where the combinatorial type of a polyhedral complex is the
isomorphism class of the partially ordered set of its faces, ordered by inclusion.
Figure 1.6(c) shows a complex in R
2
that can no longer be embedded if we
“declare” all quadrangles to be squares (or just parallelograms).
Example 1.49. In a simplicial complex ˙ there exist no affine dependencies be-
tween the vertices of the simplices, and such a complex can always be embedded.
Let x
1
. . . . . x
n
be the vertices of ˙. Then we choose affinely independent points
:
1
. . . . . :
n
in R
n-1
, and let ˙
t
be the collection of simplices conv(:
i
1
. . . . . :
i
k
) for
which x
i
1
. . . . . x
i
k
span a simplex in ˙. Clearly ˙ Š ˙
t
.
This consideration shows that a simplicial complex is determined up to iso-
morphism by the sets {:
i
1
. . . . . :
i
k
] that appear as vertex sets of its simplices. An
abstract simplicial complex ˙ is a collection of subsets of a finite set V such that
F ÷ V implies F
t
÷ F for all subsets F
t
of V . Simplicial complexes can be treated
as combinatorial objects defined on their sets of vertices.
Similarly one sees that every simplicial conical complex can be realized as a fan.
Let now ˘ be a polyhedral complex. Then a function ˛ : [˘[ ÷ R
d
is affine,
(strictly) convex or (strictly) concave, respectively, if this attribute applies to ˛[P
for each P ÷ [˘[. Note that such a function need not be extendable to an affine,
convex or concave function on conv([˘[) if ˘ is an embedded polyhedral complex:
the property is only required to hold “piecewise”.
Example 1.50. Let us once more come back to example 1.41. Suppose ˛ is an affine
function on ˘, and set y
i
= ˛
i
. Then y
1
. . . . . y
6
must satisfy the following system
of equations resulting from the fact that the two diagonals in a quadrangle meet at
their common midpoint:
y
1
÷y
3
= y
2
÷y
4
. y
1
÷y
6
= y
4
÷y
5
. y
2
÷y
6
= y
3
÷y
5
.
It follows that y
1
= y
4
, y
2
= y
3
, y
5
= y
6
. On the other hand, every assignment of
values y
i
to x
i
, i = 1. . . . . 6, satisfying these identities can be extended to an affine
function on ˘: the vector space of affine functions on ˘ has dimension 3.
This shows again that ˘ is not embeddable: its image under an affine map
[˘[ ÷R
d
for some d spans an affine subspace of dimension ¸ 2.
Weighted conical complexes. There is a canonical way to associate a conical com-
plex to an arbitrary polytopal complex ˘: for every P ÷ ˘, P Ç R
d
we consider
28 1. Polytopes, cones, and complexes
the cone C(P) = R
÷
(P. 1) Ç R
d÷1
and then, clearly, these cones define a conical
complex C(˘). The cones in C(˘) are all pointed.
Conversely, it is sometimes possible to derive a polytopal complex from a con-
ical complex ˘ consisting of pointed cones. Namely, assume we are given a func-
tion n : [˘[ ÷ R
÷
, a weight, satisfying the following conditions: ¡ is linear on
each cone C ÷ ˘ and n
-1
(0) = 0 where 0 is the point of [˘[ represented by the
apex of any of the cones C ÷ ˘. Then the collection of the polytopes n
-1
(1) forms
a polytopal complex ˘
1
in obvious way. We call it the weight 1 cross section of ˘.
In this situation we say that ˘ is a weighted conical complex (with weight function
n).
For a single pointed cone C Ç R
d
there always exists a weight – a linear
function n : C ÷ R
÷
such that n
-1
(0) = 0. This can be seen as follows.
Assume C = H
÷
1
¨ ¨ H
÷
m
where ˛
i
: R
d
÷ R are linear mappings and
H
÷
i
= {x ÷ R
d
: ˛
i
(x) ¸ 0], i = 1. . . . . m. Then it is not difficult to see that
˛
1
÷ ÷˛
m
is a weight for C.
Moreover, a simplicial conical complex ˙ is always weighted: we choose a
nonzero vector : on each of its rays and assign a positive number n(:) to it. This
assignment extends uniquely to a weight of ˙. However, for a general complex ˘
of pointed cones a weight may not exist.
Example 1.51. Choose 6 rational non-coplanar points in R
3
as shown in Figure
1.7 where the top and bottom triangles are in parallel planes and at least one of the
Figure 1.7. Construction of a nonweighted complex
quadrangles is not “flat” (indicated by the fact that the 3 nonhorizontal lines do not
meet in a common point). Suppose that 0 ÷ R
3
lies in the interior of the convex
hull of these 6 points so that the cones with common apex 0 that are spanned by
the 2 triangles and the 3 quadrangles form a fan F of rational cones in R
3
with
[F[ = R
3
. We claim that the conical complex F cannot be weighted.
In fact, this could only be the case if all 3 quadrangles were flat. We leave the
proof of this general statement to the reader and content ourselves with a concrete
example. Choose the 6 points as follows:
:
1
= (1. 1. 1). :
3
= (÷1. 1. 1). :
5
= (0. ÷1. 1).
:
2
= (1. 1. ÷1). :
4
= (÷1. 1. ÷1). :
6
= (1. ÷3. ÷1).
Then we have the relations
:
1
÷:
4
= :
2
÷:
3
. 3:
2
÷4:
5
= 2:
1
÷:
6
. 2:
3
÷:
6
= :
4
÷2:
5
.
It is easy to show by hand that one cannot assign positive weights to the 6 vectors
such that these equations are satisfied.
1.E. Subdivisions and triangulations 29
1.E. Subdivisions and triangulations
We have motivated the notion of polyhedral complex by the example of a triangu-
lation in the plane (see Figure 1.6). Now we introduce the notion of triangulation,
and more generally that of subdivision, formally.
Definition1.52. Let ˘. ˘
t
be polyhedral complexes. We say that ˘
t
is a subdivision
of ˘ if [˘[ = [˘
t
[ and each face of ˘ is the union of faces P
t
÷ ˘
t
.
If Qis a polyhedron, then a subdivision of Qis a subdivision of its face lattice.
If ˘ and ˘
t
are polytopal (or conical) complexes and ˘
t
is simplicial, then it
is called a triangulation of ˘.
Let ˘ be a polyhedral complex and ˘
t
a subdivision. Since the decomposition
[˘[ =
_
P
0
÷˘
0
int(P
t
) is disjoint, it follows immediately that each nonempty P
t
÷
˘
t
is contained in (the uniquely determined) P ÷ ˘ for which int(P) ¨int(P
t
) =
0.
We want to show that every polytopal complex ˘ has a triangulation whose
vertices can be chosen freely, provided the vertices of ˘ are taken into account.
In order to construct it inductively we have to know the local structure of ˘ at a
subset X Ç [˘[. It is determined by two subcomplexes of ˘ (and their difference):
Definition 1.53. The closed star neighborhood of X is the subcomplex
star
˘
(X) = {F : F is a face of some G ÷ ˘ with X ¨ G = 0].
The link of X is the subcomplex
link
˘
(X) = {F ÷ star
˘
(x) : X ¨ F = 0].
and the open star neighborhood is
openstar
˘
(X) = star
˘
(X) \ link
˘
(X).
If there is no ambiguity about the complex ˘, then we suppress the reference to it.
x x x
Figure 1.8. A polytopal complex, star and link of a vertex
As we have seen in Section 1.Dthe notion conv(X) is well-defined if there exists
P ÷ ˘ with X Ç P. Similarly we are justified in writing x ÷ aff(F) or x … aff(F)
if x ÷ P and F Ç P for some P ÷ ˘.
Proposition 1.54. Let ˘ be a polytopal complex and x ÷ [˘[. Suppose that F. G ÷
link(x) and P ÷ ˘ such that x ÷ P. Then the following hold:
(a) x … aff(F);
(b) if F is the union of the sets M
i
Ç F, then conv(F. x) =
_
i
conv(M
i
. x);
(c) P =
_
{conv(F
t
. x) : F
t
÷ link(x). F
t
Ç P];
30 1. Polytopes, cones, and complexes
(d) [ star(x)[ =
_
{conv(F
t
. x) : F
t
÷ link(x)];
(e) conv(M. x) ¨ [ link(x)[ = M for every convex subset M of F;
(f) conv(M. x) ¨ conv(N. x) = conv(M ¨ N. x) for all convex sets M Ç
F and N Ç G (where conv(M ¨ N. x) is formed within a face R of ˘
containing x and F ¨ G).
Proof. (a) Let Q be a face of ˘ such that x ÷ Q and F is a face of Q. Then
x … aff(F) since F = Q¨ aff(F).
(b) Since F is convex, conv(F. x) is the union of the line segments Œx. y| with
y ÷ F. This observation immediately implies the claim.
(c) The containment ¸ follows from the convexity of P. For the converse let
y ÷ P, y = x. Passing to a face of P if necessary, we can assume that no proper
face of P contains x and y. The ray (in aff(P)) from x through y leaves P in a
point z ÷ P. The facet Q of P containing z cannot contain x – otherwise it would
contain y as well. So Q ÷ link(x), and y ÷ conv(Q. x).
(d) follows immediately from (c).
(e) We choose Q ÷ star(x) as in (a) and work in aff(Q). The line through x
and a point z ÷ M Ç F meets F only in z since x … aff(F).
(f) Note that we can find a suitable R ÷ ˘ since F ¨ G ÷ link(x). Let y ÷
conv(M. x)¨conv(N. x). If y = x, then certainly y ÷ conv(M¨N. x). Otherwise
there exist line segments (x. z|, z ÷ M, and (x. z
t
|, z
t
÷ N, both containing x.
According to Remark 1.47 one has z ÷ (x. z
t
|, or z
t
÷ (x. z|. We can assume that
z ÷ (x. z
t
|. It follows that z ÷ conv(N. x) ¨ [ link(x)[, and so z ÷ N by (e) (and
z = z
t
). But this implies z ÷ M ¨ N and y ÷ conv(M ¨ N. x).
Very often the notion of visibility is useful: we say that a point y ÷ R
d
is visible
from x ÷ R
d
with respect to the obstacle M Ç R
d
if the semi-open line segment
Œx. y) does not intersect M. For an interpretation of Proposition 1.54 in terms of
visibility imagine an embedded complex [˘[ as a subset of the surrounding space,
[˘[ filled with some opaque material. Suppose star
˘
)x) consists of d-dimensional
polytopes. Then [ link
˘
(x)[ is the set of all those points of ˘ that are invisible from
x, but become visible if we remove the interior of all faces containing x from[˘[.
Let P be a polytope and x … aff(P). Then one calls conv(P. x) the pyramid
over (the base) P with apex x. The faces of such a pyramid are easily described:
Proposition 1.55. Let Q be the pyramid over P with apex x. Then the faces of Q
are given by the faces of P and the pyramids over the faces of P with apex x.
We leave the very easy proof to the reader.
It is easily checked that a subdivision ˘
t
of ˘ can be restricted to a subcomplex
˘
tt
of ˘: the polytopes P ÷ ˘
t
with P Ç [˘
tt
[ form a subdivision of ˘
tt
. One
often constructs subdivisions by going the other way round. Then one has to patch
subdivisions of subcomplexes:
Proposition 1.56. Let ˘
1
and ˘
2
be subcomplexes of ˘, and let ˘
t
i
be a subdivision
of ˘
i
, i = 1. 2. If the subdivisions of ˘
1
¨˘
2
induced by ˘
t
1
and ˘
t
2
coincide, then
˘
t
1
L ˘
t
2
is a subdivision of ˘
1
L ˘
2
.
1.E. Subdivisions and triangulations 31
The straightforward proof is left to the reader.
We are mainly interested in polytopal and conical complexes, and will not try to
extend the results proved below to arbitrary polyhedral complexes. While we have
derived assertions about polytopes from their conical counterparts in the previous
section, we will take the opposite direction now.
An important technique for the construction of subdivisions is stellar subdivi-
sion of a polytopal complex ˘ with respect to some point x ÷ [˘[ as introduced
in
Lemma 1.57. Let ˘ be a polytopal complex and x ÷ [˘[. Suppose that ˘
t
is a
subdivision of {F ÷ ˘ : x … F]. Let ¹ be the set of polytopes F ÷ ˘
t
that
subdivide link
˘
(x). Then
˘
tt
= ˘
t
L {conv(F. x) : F ÷ ¹].
is a subdivision of ˘ with vert(˘
tt
) = vert(˘
t
) L {x].
If ˘
t
is a triangulation, then so is ˘
tt
.
x
˘ ˘
t
˘
tt
Figure 1.9. Stellar subdivision
Proof. Let us first deal with the very last assertion: if F is a simplex, then the
pyramid conv(F. x) is s simplex as well.
In view of the patching principle 1.56 it is enough for the main part of the
lemma to show that ˙ = {conv(F. x) : F ÷ ¹] L ¹ is a subdivision of star
˘
(x)
containing the vertex x, since (i) star
˘
(x) ¨ {F ÷ ˘ : x … F] = link
˘
(x), and
(ii) both ˙ and ˘
t
restrict to the same subdivision on link
˘
(x). Fact (ii) follows
from Proposition 1.54(e): conv(F. x) ¨ [ link
˘
(x)[ = F for all F ÷ ¹ (and F is a
face of conv(F. x)).
After this preparation one notes that {x] is a face of ˙, obtained by taking
F = 0. Therefore x ÷ vert(˙). Next it follows immediately from 1.54(a) and 1.55
that ˙ contains all the faces of its polytopes.
We have to check that the intersection of two polytopes P. Q ÷ ˙ is a face of
P as well as of Q. This is clear if x … P L Q, for both P and Q are members of ¹
in this case.
So assume that x ÷ P. Then P = conv(F. x) with F ÷ ¹. If x ÷ Q, too, then
Q = conv(G. x) with G ÷ ¹, and P ¨Q = conv(F ¨G. x) by Proposition 1.54(f).
Now F ¨G is a face of both F and G, and the pyramid conv(F ¨G. x) is a face of
both the pyramids P and Qwith apex x.
Let x … Q. Then Q Ç [ link
˘
(x)[. We claim that P ¨ Q = F ¨ Q. This
suffices since F ¨ Q is certainly a face of Q and of F, and therefore of P. That
P ¨ Q = F ¨ Qfollows from 1.54(e): P ¨ Q = P ¨ [ link
˘
(x)[ ¨ Q = F ¨ Q.
32 1. Polytopes, cones, and complexes
To sum up: ˙ is a polytopal complex. It remains to verify that every P ÷
star
˘
(x) is the union of those Q ÷ ˙ that are contained in P. This is clear if
x … P, and follows from 1.54(b) and (c) if x ÷ P.
Lemma 1.57 is a very powerful tool, as the following theorem shows:
Theorem 1.58. Let ˘ be a polytopal complex and V a finite set of points with
vert(˘) Ç V Ç [˘[. Then there exists a triangulation ˙ of ˘ with vert(˙) = V .
Proof. It is enough to prove the theorem for V = vert(˘) since all extra vertices
can be added by stellar subdivision as the lemma has shown.
However, the lemma solves the problem also for V = vert(˘). We use induc-
tion on V starting with the trivial case V = 0. For V = 0 we choose x ÷ V and
set V
t
= V \ {x]. The polytopal complex {F ÷ ˘ : x … F] in the lemma has V
t
as its set of vertices. By induction it has a triangulation ˙
t
, and the lemma tells us
how to extend it to a triangulation of ˘.
Remark 1.59. (a) Let us consider a single polytope P of dimension d. Then the tri-
angulation constructed in the proof of Theorem 1.58 does not include a d-simplex
before the very last vertex of P has been inserted. The emergence of this triangu-
lation is illustrated in Figure 1.10.
1 1
2
1
2
3
1
2
3
4
5
4
1
2
3
5
4
1
2
3
Figure 1.10. Two triangulations of a pentagon
Suppose that the elements of V ¸ vert(P) are given in some order, V =
{x
1
. . . . . x
m
]. Then it is perhaps more natural to successively build up a triangula-
tion of P by extending a triangulation of P
n-1
= conv(x
1
. . . . . x
n-1
) to a triangu-
lation of P
t
n
= conv(x
1
. . . . . x
n
), n = 1. . . . . m. This construction is illustrated by
the rightmost triangulation in Figure 1.10.
Simplifying notation we set P
t
= conv(x
1
. . . . . x
n-1
) and P = conv(x
1
. . . . .
x
n
). Let ˙ be a triangulation of P
t
. If x
n
in P
t
, we use stellar subdivision, refining
the triangulation of P
t
= P. (Or we omit x
n
if we are only interested in finding a
triangulation of P.) If x
n
… P
t
we must distinguish the cases (i) x
n
… aff(P
t
) and
(ii) x
n
÷ aff(P
t
).
In case (i) we add to ˙ all the simplices conv(F. x), F ÷ ˙; that we obtain a
triangulation in this way follows fromLemma 1.57. In case (ii) we augment ˙ by all
simplices P = conv(F. x) where F ÷ ˙ is contained in a support hyperplane H
of P
t
such that x
n
÷ H
-
. Now Lemma 1.57 does not cover our claim. It is however
not too difficult to extend it accordingly. One should prepare the extension by a
suitable variant of Proposition 1.54, in which link(x) is replaced by the set of all the
faces F of ˙ just mentioned.
(b) Obviously the triangulation constructed via Lemma 1.57 or the procedure
outlined in (a) (applied to V ¨P for each polytope in a complex) depends not only
1.E. Subdivisions and triangulations 33
on the set V , but also on the order in which the points of V are inserted as vertices.
On the other hand, if V is given, then it depends only on this order, and more
precisely, the triangulation of each P ÷ ˘ depends only on the order in which
the elements of V ¨ P are inserted. This shows that we could have carried out the
construction in each polytope separately (after fixing an order on the elements of
V ), patching together the individual triangulations via 1.56. In particular, part (a)
can be extended to polytopal complexes.
As an immediate consequence of Theorem 1.58 we obtain a classical result,
Carathéodory’s theorem:
Theorem 1.60. Let X Ç R
d
. Then for every x ÷ conv(X) there exists an affinely
independent subset X
t
of X such that x ÷ conv(X
t
).
Proof. There exists a finite subset X
tt
of X such that x ÷ conv(X
tt
). We may
assume X = X
tt
. Then conv(X) is a polytope. It has a triangulation ˙ with vertex
set X. We now choose X
t
= vert(F) where F ÷ ˙ contains x.
There also exists a very elementary proof of Carathéodory’s theorem, not using
the existence of triangulations; see [103, Prop. 1.15].
Now we transfer the theorem on triangulations from polytopal to conical com-
plexes. Note that there is a unique point in the support of a conical complex 1 that
is identified with 0 ÷ RC for each C ÷ 1 . We denote it simply by 0 as well.
Theorem 1.61. Let 1 be a conical complex, and V Ç [1 [ be a finite set of vectors
: = 0 such that V ¨ C generates C for each C ÷ 1 . Then there exists a triangula-
tion ˙ of 1 such that {R
÷
: : : ÷ V ] is the set of 1-dimensional faces of 1 .
Proof. For each cone C ÷ 1 we consider the convex hull P
C
= conv(V ¨ C. 0).
These polytopes (together with all their faces) form a polytopal complex ˘. The
essential point in proving this claim is that one has P
C
¨ F = P
F
if F is a face of
C.
Set ¹ = link
˘
(0). Let y ÷ [1 [, y = 0, and choose a face C of 1 containing y.
The ray R
÷
y leaves P
C
at a point y
t
which belongs to [¹[, as we have seen in the
proof of 1.54. We can replace each : ÷ V by :
t
, in other words, we can assume that
V Ç [¹[. By Theorem1.58 we find a triangulation ˙
t
of ¹such that vert(˙
t
) = V .
Lemma 1.57 tells us how to extend ˙
t
to a triangulation ˙
tt
of ˘. That ˙
tt
is a
triangulation is evidently equivalent to the fact that the cones R
÷
D
t
, D ÷ ˙
t
, form
a conical complex ˙, and since each D is a simplex, ˙ is a complex of simplicial
cones. Its 1-dimensional faces are exactly the rays R
÷
:, : ÷ V , and that it covers
[1 [ follows since [¹[ consists exactly of the points y
t
, y ÷ [1 [, y = 0. We finish
the construction by choosing ˙ as the collection of cones R
÷
P, P ÷ ˙
t
.
As in the case of a polytopal complex we have a Carathéodory theorem:
Theorem 1.62. Let X Ç R
d
. Then for every x ÷ R
÷
X there exists an linearly
independent subset X
t
of X such that x ÷ R
÷
X
t
.
34 1. Polytopes, cones, and complexes
1.F. Regular subdivisions
There is a distinguished class of subdivisions of polyhedral complexes which will
be important in our study of monoid rings and toric varieties – the class of regular
subdivisions.
Let P be a polyhedron and consider a function ¡ : P ÷ R. We say that a
connected subset W Ç P is a domain of linearity of ¡ if the restriction ¡ [W
can be extended to an affine function on aff(P) and there is no connected subset
V Ç P containing W strictly for which ¡ [V extends to an affine mapping. If ˘
is a polyhedral complex and ¡ : [˘[ ÷R an affine function, then a subset W is a
domain of linearity of ¡ if there exists a facet P ÷ ˘ such that W Ç P and W is
a domain of linearity of ¡ [P.
Definition 1.63. A subdivision ˘
t
of a polyhedral complex ˘ is called regular if
there is a convex function ¡ : [˘[ ÷ R whose domains of linearity are the facets
of ˘
t
. Such a function ¡ is called a support function for the subdivision ˘
t
. The
set of all support functions will be denoted by SF(˘. ˘
t
).
To give a trivial example: the constant 0 is a support function of the trivial
subdivision ˘ of ˘, and so ˘ is a regular subdivision of itself. Several construc-
tions of nontrivial regular subdivisions will be discussed below. An example of a
nonregular subdivisions will be given in Remark 1.70.
Remark 1.64. Regular subdivisions are called projective in [55] (perhaps the most
descriptive name in view of the discussion below). One also encounters the name
coherent (for instance, in [35]). Our choice ‘regular’ is common in the literature
dealing with applications to combinatorial commutative algebra [83] and (compu-
tational) geometry [103], [30].
Remark 1.65. Let P Ç R
d
and Q Ç R
d
× R
-
Ç R
d÷1
be polytopes such that
P is a proper face of Q and, moreover, the projection R
d÷1
÷ R
d
, (z
1
. . . . . z
d
.
z
d÷1
) ÷(z
1
. . . . . z
d
) maps Qonto P. (Here we identify R
d
and (R
d
. 0).) Then the
faces of Q that consist of lower endpoints of maximal line segments Œ(z. 0). (z. h)|,
h ¸ 0, in Q, form the bottom of Q below P. (In this section the term ‘bottom’ is
used in this sense. It is reconciled with the notion of bottom in Section 1.C if we
replace Qby the union of all vertical rays emerging from points in Q.) These faces
project up to P to certain polytopes inside P, defining a subdivision of P. This
P
Q
Figure 1.11. Projection of the bottom
subdivision is regular. It is supported by the function ¡ : P ÷R which assigns to
a point z ÷ P the height (i. e. the (d ÷1)th coordinate) of the point in the bottom
of Q right below z. In other words, the bottom is the graph of a support function.
Conversely, all regular subdivisions of P are obtained this way.
1.F. Regular subdivisions 35
Although we have defined regular subdivisions for arbitrary polyhedral com-
plexes, in practice we will only encounter two special cases – those of polytopal
complexes and conical complexes consisting of pointed cones. Essentially every-
thing that will be said on polytopal complexes admits a natural conical version,
usually not mentioned explicitly.
For a polytopal complex ˘ the set of functions ¡ : [˘[ ÷R that are affine on
the faces of ˘ is denoted by PA(˘) – the set of piecewise affine functions. Clearly,
PA(˘) is a finite-dimensional real vector space, and if ˘
t
is a subdivision of ˘,
then PA(˘) is a subspace of PA(˘
t
).
The convexity of a function ¡ ÷ PA(˘
t
) can be tested as follows. For each
facet P
t
÷ ˘
t
we let ¡
P
0 be the affine function (on aff(P) for the facet P of ˘ with
P
t
Ç P) such that ¡ [P
t
= ¡
P
0 [P
t
. Then ¡ is convex if and only if
¡ (x) = max{¡
P
0 (x) : P
t
Ç P]
for all x ÷ [˘[ and all facets P of ˘ containing x. Moreover, ¡ is a support
function if and only if for each facet P of ˘ all the functions ¡
P
0 , P
t
Ç P, are
pairwise different. These conditions can be tested by the application of suitable
linear forms on the space of all functions, as we will see now,
Let ˘
t
be an arbitrary subdivision of a polytopal complex ˘. Afinite nonempty
system D of real-valued linear functionals on the space of all functions R
[˘[
is a
regularity test system for the pair (˘. ˘
t
) if the following equivalence holds for all
functions ¡ ÷ PA(˘
t
):
ı(¡ ) ¸ 0 for all ı ÷ D ” ¡ is convex on ˘.
That regularity test systems exist and do in fact test regularity (and not only con-
vexity of functions) is shown in
Lemma 1.66. Let ˘
t
be a subdivision of a polytopal complex ˘.
(a) Then there exists a regularity test system for (˘. ˘
t
).
(b) The set of all functions in PA(˘
t
) that are convex with respect to ˘ form a
cone C.
(c) SF(˘. ˘
t
) is the absolute interior of C in PA(˘
t
), and so SF(˘. ˘
t
) = 0
if and only C is full-dimensional.
(d) Let D be a regularity test system for (˘. ˘
t
) and ¡ ÷ PA(˘
t
). Then ¡ ÷
SF(˘. ˘
t
) if and only ı(¡ ) > 0 for all ı ÷ D.
Proof. P ÷ ˘ is a facet and P
t
÷ ˘
t
a face of dimension dimP ÷ 1 such that
int(P
t
) Ç int(P). There are exactly two facets P
1
. P
2
÷ ˘
t
such that P
1
. P
2
Ç P
and P
1
¨ P
2
= P
t
. Fix points z
1
÷ int(P
1
) and z
2
÷ int(P
2
) in such a way that the
line segment Œz
1
. z
2
| intersects int(P
t
), say in a point z. We have the linear map
ı
P1,P
2
: R
[˘[
÷R. ı
P1,P
2
(¡ ) = g(z) ÷ ¡ (z).
where g is the affine map from the segment Œz
1
. z
2
| to R such that g(z
1
) = ¡ (z
1
)
and g(z
2
) = ¡ (z
2
). It is easily checked that g(z) depends linearly on ¡ .
By running P through the facets of ˘ and P
t
through the faces of ˘
t
, so that
int(P
t
) Ç int(P) and dimP
t
= dimP÷1, and fixing for each P
t
two points z
1
. z
2
÷
36 1. Polytopes, cones, and complexes
[˘[ on both sides of P
t
as above, we get a finite system D of linear functionals
R
[˘[
÷R.
z
1
z
2
z
¡
g
z
1
z
2 z
Figure 1.12. Construction of the regularity test system
In order to check that D is a regularity test systemassume first that ı
P
1
,P
2
(¡ ) <
0 for some ¡ ÷ R
[˘[
. Then ¡ is not convex along the line segment chosen for the
definition of ı
P
1
,P
2
. Conversely suppose that ı
P
1
,P
2
(¡ ) ¸ 0 for some ¡ ÷ PA(˘
t
)
and all functionals ı
P
1
,P
2
÷ D. Let L be a line segment in a facet P of ˘, and
let Q
1
. . . . . Q
t
be the facets of ˘
t
traversed by L. In checking the convexity of
¡ we are allowed to assume that Q
i
¨ Q
i÷1
is a face of dimension dimP ÷ 2
for i = 1. . . . . t ÷ 1; otherwise we vary the endpoints of L, and use a continuity
argument. Finally it is enough to consider the case t = 2. We can assume that ¡
is of constant value 0 on Q
1
and must show that it has nonnegative values on Q
2
.
This follows immediately from ı
Q
1
,Q
2
(¡ ) ¸ 0.
This shows (a), and (b) follows immediately from (a). Moreover, in view of (b),
every regularity test system D
t
defines the same cone, and its absolute interior is
given by {¡ : ı(¡ ) > 0 for all ı ÷ D
t
]. For (c) and (d) it is therefore enough to
consider the systemD constructed above. Now one observes that the graph of ¡ is
broken along the common boundary of P
1
and P
2
if and only if ı
P1,P
2
(¡ ) > 0, and
so ¡ is a support function for ˘
t
if the strict inequality holds for all pairs P
1
. P
2
that are contained in a facet P of ˘
t
and meet along a common face of dimension
dimP ÷ 1.
For a polytopal complex ˘ the evaluation of the functions at the vertices of ˘
gives rise to a linear embedding PA(˘) ÷ R
n
, n = vert(˘). In the special case
when ˘ is a simplicial complex this is an isomorphism of real vector spaces.
The existence of regularity test systems implies the following
Corollary 1.67. Let ˘ be a polytopal complex and ˘
t
a triangulation of ˘. If ˘
t
is
regular, then SF(˘. ˘
t
) is the interior of an n-dimensional cone in R
n
Š PA(˘
t
),
n = # vert(˘
t
).
In particular, the corollary says that regular triangulations ˘
t
are stable with
respect to small perturbations of the values ¡ (x), x ÷ vert(˘
t
).
Next we give a quick overview of several basic regular triangulations. More
specific triangulations will be considered in the next chapters.
Intersections and dissections. Let ˘
1
. . . . . ˘
k
be a system of subdivisions of a
polyhedral complex ˘. Then all possible intersections P
1
¨ ¨P
k
of faces P
i
Ç
˘
i
, i ÷ Œ1. k| form a subdivision of ˘ which we denote by ˘
1
¨ ¨ ˘
k
and call
1.F. Regular subdivisions 37
the intersection of the ˘
i
(by abuse of notation and terminology). The intersection
of regular subdivisions is regular because if ¡
i
÷ SF(˘. ˘
i
), i = 1. . . . . k, then
¡
1
÷ ÷¡
k
÷ SF(˘. ˘
1
¨ ¨ ˘
k
).
One special case of this construction deserves special consideration. Let P Ç
R
d
be a polytope and H Ç R
d
a hyperplane which cuts P in two parts of the
same dimension, say P = Q L R. Then the union of the face lattices of Q
and R forms a regular subdivision of P. In fact, if H is given by an equation
a
1
X
1
÷ ÷ a
d
X
d
= b, a
1
. . . . . a
d
. b ÷ R, then the function ¡ : P ÷ R,
¡ (z) = [a
1
z
1
÷ ÷ a
d
z
d
÷ b[, z = (z
1
. . . . . z
d
) is a support function of the
subdivision under consideration. Taking intersections of such subdivisions one
concludes that an arbitrary finite system of hyperplanes H
1
. . . . . H
k
Ç R
d
cuts P
into smaller polytopes that define a regular subdivision of P. Such subdivisions
are called dissections (also see Example 1.46).
Composition and patching. If ˘
t
is a subdivision of ˘ and ˘
tt
is one of ˘
t
, then
˘
tt
is a subdivision of ˘ in a natural way. In this situation ˘
tt
can be viewed
as a composite of two successive subdivisions. Compositions enjoy the following
transitivity of regularity.
Proposition 1.68. If ˘
t
is a regular subdivision of a polytopal complex ˘ and ˘
tt
is one of ˘
t
, then ˘
tt
is a regular subdivision of ˘.
Proof. We choose a regularity test systemD
tt
for the pair (˘. ˘
tt
) as in the proof
of Lemma 1.66. For each pair of facets of ˘
tt
that are contained in the same facet
P of ˘ and share a common face of dimension dimP ÷ 1 it contains a function
ı
P
1
,P
2
. We split D
tt
into two subsets, namely the set D
t
of those ı
P
1
,P
2
for which P
1
and P
2
are contained in the same facet of ˘
t
and its complement D = D
tt
\ D
t
.
Then D
t
is a regularity test system for (˘
t
. ˘
tt
) of the type constructed in the
proof of Lemma 1.66. Moreover, it is easily seen that D is a regularity test system
for (˘. ˘
t
). In fact the line segment used in the construction of ı
P
1
,P
2
÷ D con-
nects interior points in different facets Q
1
. Q
2
of ˘
t
and meets the interior of the
face Q
1
¨ Q
2
, dimQ
1
¨ Q
2
= dimQ
1
÷ 1 = dimQ
2
÷ 1.
Now choose ¡ ÷ SF(˘. ˘
t
) and g ÷ SF(˘
t
. ˘
tt
). Then ı(¡ ) > 0 for all ı ÷ D
and ı
t
(g) > 0 for all ı
t
÷ D
t
. Moreover, ı(¡ ) = 0 for all ı ÷ D
t
. Therefore, if the
real number c > 0 is small enough, then ı(¡ ÷cg) > 0 for all ı ÷ D
tt
.
The definition of regular triangulation contains the following patching princi-
ple.
Corollary 1.69. Let ˘ be a polytopal complex with regular subdivision ˘
t
. Further-
more let ˘
tt
P
be a regular subdivision of each facet P ÷ ˘
t
. Then the subdivisions
˘
tt
P
can be patched to a regular subdivision ˘
tt
of ˘ if and only if there are func-
tions ¡
P
÷ SF(P. ˘
tt
P
) (we simply write P for the face complex of P) such that

P
)[P ¨ Q = ¡
Q
[P ¨ Qfor all facets P. Q ÷ ˘
t
.
Proof. In view of Proposition 1.68 we can assume that ˘ = ˘
t
. If the ˘
tt
P
define
a global regular subdivision with support function ¡ , then we can take ¡
P
= ¡ [P
for each facet P ÷ ˘, and the condition (¡
P
)[P ¨ Q = ¡
Q
[P ¨ Q is obviously
38 1. Polytopes, cones, and complexes
satisfied for all facets P. Q ÷ ˘. Conversely, if the support functions ¡
P
satisfy
this condition, then (i) we can patch them up to a function on [˘[ and (ii) we can
patch up the subdivisions to a global subdivision ˘
tt
of ˘ which is automatically
regular. (Regularity is decided on each facet P ÷ ˘ separately.)
Remark 1.70. If the subdivisions ˘
tt
P
can merely be patched up to a global subdi-
vision ˘
t
of ˘, then ˘
t
need not be regular. The following example (taken from
[55]) in dimension 2 shows this. Let ˘
t
be the (visibly) regular subdivision of the
˘
t
C
t
A
t
B
t
C
A B
˘
tt
Figure 1.13. A regular subdivision and a nonregular refinement
triangle T with face lattice ˘ (see Figure 1.13) into 3 quadrangles and one triangle.
We refine it by cutting each of the quadrangles along a line so that each of them is
regularly subdivided. However, the resulting triangulation ˘
tt
of ˘ is not regular.
To this end we can assume that the support function ¡ vanishes on the entire
inner triangle. Let D be the intersection point of the line segments ŒA. B
t
| and
ŒA
t
. B|. Then D = zA ÷jB
t
= zB ÷jA
t
. Since ¡ (D) = z¡ (A) ÷j¡ (B
t
) <
z¡ (B) ÷ j(A
t
), one obtains ¡ (A) < ¡ (B). Running this argument around the
triangle, we arrive at the contradiction ¡ (A) < ¡ (A).
The nonregularity of ˘
tt
follows also from a criterion of Gelfand, Kapranov,
and Zelevinsky ([35, Ch. 7, Th. 1.7]): Let P be a polytope and A a finite subset
of ˘ containing vert(˘). Each triangulation ˘ of P with vertex set A defines a
function y
P
: A ÷R whose value at a ÷ A is the sum of the volumes of the facets
of ˘ adjacent to a. If ˘ and ˘
t
are different regular triangulations with vertex set
A, then y
˘
= y
˘
0 . The proof of the criterion requires the theory of the secondary
polytope developed in [35].
In our example ˘
tt
and its mirror image with respect to the vertical axis have
the same vertex set and the same characteristic function, so they cannot be regular,
since they are different.
Triangulations with a given set of vertices. Let ˘ be a polytopal complex ˘ and
V ¸ vert(˘) a finite subset of [˘[. By Theorem 1.58 there exists a triangulation ˙
of ˘ such that vert(˙) = V . Now we derive the regular version of this fact.
There are two ways of doing so: (i) by exhibiting a ‘fast’ process, based on the
existence of a global strictly convex function, and (ii) by constructing a triangu-
lation step by step, preserving the regularity. The first method is particularly well
suited for a single polytope but becomes a bit complicated for general complexes.
The second method is relatively ‘slow’ but makes no difference fromthe complexity
point of view between single polytopes and polytopal complexes. As we will see, it
uses the regularity of stellar triangulations.
1.F. Regular subdivisions 39
Let us first discuss the construction via a global strictly convex function.
Proposition 1.71. Let ˘ be a polytopal complex and ¡ : ˘ ÷Rbe a strictly convex
function with ¡ (x) ¸ 0 for all x ÷ [˘[. Let V be a subset of ˘ containing vert(˘).
For each face P ÷ ˘ let P
t
be the convex hull of the set (P. 0) L {(:. ¡ (:)) : : ÷
V ¨ P] Ç aff(P) × R. Then the projection of the bottom of P
t
onto P, P ÷ ˘,
defines a regular subdivision ˘
t
of ˘ such that V = vert(˘
t
).
Proof. In viewif Remark 1.65 and Corollary 1.69 the only critical point is the claim
that V = vert(˘
t
). In checking it, we can assume that ˘ is the face lattice of a
single polytope P. Then we have to check that (:. ¡ (:)) is a vertex of P
t
for each
: ÷ V . Since P
t
is a polytope, it is enough that (:. ¡ (:)) is an extreme point of
P
t
; see Proposition 1.15. This follows immediately from the definition of strict
convexity.
First we consider the case of a single polytope P and a finite subset V Ç P
containing vert(P). The existence of a strictly convex function ¡ on P is shown by
the following construction. Choose a large real number r ;0 such that the d-ball
{x ÷ R
d
[ [x[ ¸ r] contains P. Then the subset
˚
z = (z
1
. . . . . z
d÷1
) ÷ R
d÷1
: (z
1
. . . . . z
d
) ÷ P. z
d÷1
¸ 0. [z[ = r
_
of the standard d-sphere {z ÷ R
d÷1
: [z[ = r] of radius r is a graph of a strictly
convex function ¡ on P.
The subdivision defined by Proposition 1.71 need not be a triangulation. How-
ever, by a small perturbation of the values ¡ (:), : ÷ V , that replaces (:. ¡ (:))
by (:. ¡ (:) ÷ c
¡
) we neither lose any of the vertices nor create new ones in the
subdivision of P. If the perturbation is generic, we break the affine relations of the
vertices of the nonsimplicial facets in the bottom of Q(notation as in Remark 1.65)
and do not create new ones. Then the projection of the modified Qonto P defines
a regular triangulation ˙ of P such that vert(˙) = V .
The case of a general polytopal complex makes no difference once the existence
of a strictly convex function on such a complex has been established. To this end
we prove the following claim: for every polytope P and a function ç : dP ÷ R,
whose restriction to any face F Ç P is strictly convex, there exists a strictly convex
function ¡ : P ÷ R such that ¡ [dP = ç. Then a strictly convex function on ˘
can be constructed inductively via the skeleton filtration
˘
(0)
Ç ˘
(1)
Ç Ç ˘
(d)
. d = dim˘.
On the 0-skeleton every function is strictly convex.
In its turn the claim above follows from the existence of a function g : P ÷R
with g[dP = 0 that is convex on P and strictly convex on int(P). In fact, we can
extend ç to a convex function on P by considering the convex hull Rof the set
(P. 0) L {(x. ç(x)) : x ÷ dP] Ç R
d÷1
The bottom of Ris a graph of a convex function h : P ÷R, and (g ÷h) : P ÷R
is a strictly convex function on P extending ç.
40 1. Polytopes, cones, and complexes
We construct the function g : P ÷ R with the desired properties as follows.
Fix d linearly independent vectors :
1
. . . . . :
d
÷ R
d
, none of them parallel to a facet
of P. For each index i we consider all line segments L
x
= (x ÷R:
i
) ¨ P, x ÷ P.
They are parallel to :
i
(some of them may be degenerated into a point). For each
of these segments Œu. :| we consider the semicircle ; with diameter Œu. :| passing
through
_
(u ÷ :),2. ÷[u ÷ :[,2
_
÷ R
d÷1
. (A ‘semicircle’ is a figure congruent
to {(a. b) : a
2
÷ b
2
= r
2
. b ¸ 0] Ç R
2
for some r ¸ 0.) Again, some of these
semicircles may be degenerated into a point. The union of these semicircles (for
fixed i ) defines a graph of a function g
i
: P ÷ R. We claim that the function
g = g
1
÷ ÷g
d
has the desired property.
The vanishing of each g
i
on dP follows from its very construction. The con-
vexity of g
i
is not hard to see. We must only show its convexity on line segments
Œx. y| Ç P. If Œx. y| is parallel to :
i
, then the convexity follows from the convexity
of the semicircle. In the other case we consider the plane through Œx. y| parallel to
:
i
, and in it the quadrangle Q spanned by the endpoints of the line segments L
x
and L
y
. Let ¯ g be the function obtained by applying the same construction to Q.
Then ¯ g(x) = g
i
(x), ¯ g(y) = g
i
(y), and ¯ g is convex, as the reader may easily check.
Moreover, g
i
(z) ¸ ¯ g(z) since the semicircle under the segment (z ÷ R:
i
) ¨ Q,
z ÷ Qis contained in the semicircles under L
z
.
It follows that g vanishes on dP and is convex on P. The strict convexity on
int(P) can be checked by the criterion in Remark 1.16: g
i
is strictly convex along
line segments parallel to :
i
, i = 1. . . . . d, and this property is inherited by g.
Now we discuss the regularity of stellar triangulations. The proof of Theorem
1.58 gives an explicit construction of a triangulation ˙ of ˘ such that vert(˙) =
V . It is an inductive process consisting of successive stellar subdivisions, described
in Lemma 1.57. We have seen that the composite of regular triangulations is regu-
lar. Therefore, the regularity of the triangulation constructed in the proof of The-
orem 1.58 follows from
Lemma 1.72. Let ˘ be a polytopal complex and x ÷ [˘[. Then its subdivision
˘(x) = (˘ \ star
˘
(x)) L link
˘
(x) L {conv(F. x) : F ÷ link
˘
(x)].
is regular.
Proof. Let G ÷ star
˘
(x) \ link
˘
(x), say G Ç R
d
. Consider the pyramid P
G
=
conv(G. (x. ÷1)) Ç R
d÷1
with base G and apex x. Then the bottom of P
G
, viewed
as a graph, defines a function ¡
G
: G ÷ R. When G runs through star
˘
(x) \
link
˘
(x), these functions can be patched up to a function ¡ : [ star
˘
(x)[ ÷
R. Extending it by 0 to the whole set [˘(x)[ = [˘[ we get a function ¡ ÷
SF(˘. ˘(x)).
Projective fans. Recall that a fan F is an embedded conical complex. Consider
the situation when dimF = d and R
d
is the ambient space. If [F[ = R
d
, we
say F is a complete fan. A regular subdivision of R
d
, into pointed cones is called
a projective fan. A subcomplex of a projective fan is a quasiprojective fan. Thus a
1.F. Regular subdivisions 41
quasiprojective fan is projective if and only if it is complete. This terminology is
explained by the link to algebraic geometry via toric varieties.
For fans we will simply write SF(F) instead of SF(F. R
d
) .
Suppose Qis a d-polytope in the dual vector space (R
d
)
+
= Hom
R
(R
d
. R). For
each vertex : ÷ Q consider the cone C
¡
= R
÷
(÷: ÷ Q) Ç (R
d
)
+
, i. e. C
¡
is the
cone R
÷
(÷: ÷ Q) spanned by the shifted polytope ÷: ÷ Q. Then we form the
collection of cones N (Q) consisting of the (C
¡
)
+
, : ÷ vert(Q) and all their faces
(in the original vector space R
d
).
Proposition 1.73. Let F be a d-dimensional fan in R
d
. Then the following condi-
tions are equivalent:
(a) F is projective,
(b) there exists a d-polytope Q Ç R
d
such that 0 ÷ int(Q) and the cones in F
are exactly the cones over the faces of Q,
(c) F = N (P) for some d-polytope P Ç (R
d
)
+
.
In particular, N (P) is a fan, called the normal fan of P. Thus two d-polytopes
0
Figure 1.14. A polygon and its normal fan
P. Q Ç R
d
have the same normal fans if they have the same combinatorial type
and the corresponding pairs of facets span parallel affine hulls. Clearly, for every
projective fan there are infinitely many polytopes that support it.
Proof. (a) ==(b): For each maximal cone (i. e. facet) C ÷ max F there exists an
element ç ÷ SF(F) such that ç
-1
(0) = C. In fact, choose a function [ ÷ SF(F)
with [(0) = 0 and let [
C
denote the linear extension of [[C to the whole space
R
d
. Then the function [ ÷ [
C
is an element of SF(F) with the desired property.
For each C ÷ max F we choose an element ç
C
÷ SF(F) such that ç
-1
C
(0) = C
and set y =

C÷max F
ç
C
. Then the polytope
Q = y
-1
(Œ0. 1|) Ç R
d
has the property as in (b).
(b) == (a): Consider the polytope (Q. 1) = {(z. 1) : z ÷ Q] Ç R
d÷1
and
the cone C(Q) = R
÷
(Q. 1) over it. For each point x ÷ [F[ there is a unique point
z(x) ÷ dC(Q) which projects down to x. Let h(z(x)) denote the height of z(x)
above R
d
. Then the mapping x ÷h(z(x)), x ÷ R
d
, defines an element of SF(F).
(a) ” (c): The following correspondences between SF(F) and the set of
d-polytopes P Ç (R
d
)
+
for which N (P) = F are mutually inverse:
ç ÷P := {h ÷ (R
d
)
+
: h(x) ¸ ç(x) for all x ÷ R
d
].
P ֍ := max
g÷P
g(÷) : [F[ ÷R.
42 1. Polytopes, cones, and complexes
The support function constructed in the proof of (b) == (a) is nonnegative
on R
d
and has value 0 at the origin. It is also a weight on the projective fan, and
every projective fan, as well as every quasiprojective fan, is weighted. However not
every weighted complete fan of pointed cones is projective. As a counterexample
we may choose a tetrahedron in R
3
with the origin in its interior and subdivide
one of its faces by a nonregular triangulation (see Figure 1.13). It follows from
Proposition 1.73(b) that the cones over the resulting triangulation of the boundary
of the tetrahedron form a nonprojective simplicial fan.
1.G. Rationality and integrality
The polyhedra we are mainly interested in later on are defined over the field of
rational numbers, and especially the polytopes usually have their vertices in the
integral lattice. In the following we introduce these more special objects.
Rational polyhedra. Let V be an R-vector space. A rational structure on V is a
Q-subspace Qsuch that V = RQand dim
Q
Q = dim
R
V . We now define rational
structures on affine spaces:
Definition 1.74. Let A be an affine space over R with associated vector space V ,
and x ÷ A. Moreover, let Q be a rational structure on V . Then Q
t
= x ÷ Q is
called a rational structure on A.
Let A and B be affine spaces with rational structures Q and R. An affine map
˛ : A ÷B is rational if ˛(Q) Ç R. An affine subspace U of A is called rational if
A = aff(A¨ Q).
A halfspace is rational if its bounding hyperplane is rational, and a polyhedron
is rational if it is the intersection of rational closed halfspaces. (A polytope or a
cone is rational if it is rational as a polyhedron.) Later on, when we speak about
rational polyhedra, it is always understood that a rational structure has been in-
troduced previously.
A morphism ç : P ÷ Q of rational polyhedra P and Q is a map P ÷ Q that
can be extended to a rational affine map aff(P) ÷ aff(Q). (We will see below that
aff(P) is rational if P is rational.)
A polyhedral complex ˘ is rational if all its member polyhedra P
p
are rational
and the maps ¬
-1
q
ı ¬
p
(where defined) are morphisms of rational polyhedra.
A piecewise affine function is rational if its domains of linearity are rational
polyhedra and the restrictions of ¡ to them are rational (with respect to the fixed
rational structure Qon R).
Let A be an affine space with a rational structure Q. Then, upon the choice
of an origin in Q and a coordinate system with rational axes, the pair (A. Q) can
be identified with the pair (R
d
. Q
d
) where d = dimA. After this identification,
Gaussian elimination shows that an affine subspace U of A is rational if and only
if it is of the formU = ç
-1
(0) where ç : A ÷R
m
is a suitable rational affine map.
(We can choose m = dimA ÷ dimU). In particular, a rational halfspace is of the
from x ÷ A : ˛(x) ¸ 0] with a rational affine form ˛ : A ÷ R. It follows that
1.G. Rationality and integrality 43
a polyhedron P is rational if and only if it is given by affine rational inequalities
˛
i
(x) ¸ 0.
Let ˘ be a rational polyhedral complex. Then the subspace of all rational func-
tions in PA(˘) defines a rational structure on this vector space, because the values
¡ (:), : ÷ vert(˘), completely determine ¡ and are subject only to a rational sys-
tem of equations resulting from the affine dependence relations of the vertices of
the faces P ÷ ˘.
If we speak of a rational object in R
d
, then we always refer to the rational struc-
ture Q
d
Ç R
d
, unless stated otherwise.
In order not to overload the text of the next proposition we leave some details
to the reader, for example the definition of the rational structures considered in
part (d).
Proposition 1.75. Let V be a vector space and A be an affine space with rational
structures. Then the following hold:
(a) The affine hull of a rational polyhedron in Ais rational.
(b) The intersection of rational polyhedra is rational.
(c) Each face of a rational polyhedron P is rational.
(d) The recession cone rec(P) and the projectivization C(P) of a rational poly-
hedron P are rational cones.
(e) A cone C Ç V is rational if and only if the polar cone C
+
is rational (with
respect to the associated rational structure on V
+
).
(f) A cone C is rational if and only if it is generated by rational vectors.
(g) A polytope P Ç Ais rational if and only if its vertices are rational points.
(h) A polyhedron P is rational if and only if it is the Minkowski sum of a ratio-
nal polytope and a rational cone.
(i) Let ˘
t
be a regular rational subdivision of a rational polytopal (or conical)
complex. Then ˘
t
has a rational support function.
Proof. (a) follows immediately from Proposition 1.2: aff(P) is the intersection of
rational hyperplanes. (b) follows immediately from the definition. Thus the facets
of P are rational, and so is each face, being an intersection of facets. This proves
(c). Part (d) is again obvious.
For (e) and (f) we use Fourier-Motzkin elimination. By definition and by The-
orem 1.20, a cone is rational if the polar cone is generated by rational linear forms.
Then Fourier-Motzkin elimination yields a systemof rational generators for C
++
=
C.
(g) and (h) follow from (f) via projectivization, in the same way as Theorem
1.32 has been proved.
(i) follows from Proposition 1.66: SF(˘. ˘
t
) is the interior of a full-dimensio-
nal cone and therefore is the closure of its intersection with the set of rational piece-
wise affine functions.
For a subset X of Q
d
we can form the convex or conical hull within Q
d
, using
only rational coefficients in convex or conical linear combinations. However, these
notions yield nothing new:
44 1. Polytopes, cones, and complexes
Proposition 1.76. Let X Ç Q
d
. Then conv
Q
(X) = conv
R
(X) ¨ Q
d
and Q
÷
X =
R
÷
X ¨ Q
d
.
Proof. Let x ÷ conv
Q
(X). In view of Carathéodory’s theorem we can assume that
X is affinely independent. In that case the presentation of x ÷ conv
R
(X) ¨ Q
d
as a (convex) linear combination of X is uniquely determined, and therefore only
possible with rational coefficients. The same argument works for conical hulls.
The proposition can be transferred immediately to affine and vector spaces
with rational structures.
Lattices and lattice polytopes. In analogy to a rational structure on an R-vector
space V one defines a lattice structure: it is a finitely generated subgroup L with
V = RLand rank L = dimV .
If we speak of a lattice L in V , then we do not insist that rank L = dimV .
We only require that Lis a subgroup generated by R-linearly independent vectors.
Clearly, Lis a lattice structure on the subspace RLof V .
Definition 1.77. Let A be an affine space. Then a lattice structure on A is a subset
of type L
t
= x ÷ L where x ÷ A and L is a lattice structure on the associated
vector space V of A.
Let A and B be affine spaces with lattice structures L and M. Then an affine
map ç : A ÷B is called integral if ç(L) Ç M.
Similarly as in the case of vector spaces, an affine lattice in Ais a subset x ÷L
where Lis a lattice in V . Of course, x ÷Lis a lattice structure on aff(x ÷L).
A morphism of affine lattices L. M is a mapping ç : L ÷ M such that
ç(

i
a
i
x
i
) =

i
a
i
ç(x
i
) for all affine combinations (i. e.

i
a
i
= 1). In the
case M = Zwe say that ç is primitive if ç(L) = Z.
Remark 1.78. Let L be an affine lattice and H a hyperplane of A = aff(L) such
that H = aff(H ¨ L) (in other words, L ¨ H is a lattice structure on H). We
choose one of the two halfspaces of Awith respect to H as H
÷
and an affine form
˛ on Awith ˛(H) = 0.
We write L = x
0
÷ L
0
with a lattice L
0
and x
0
÷ L
0
¨ H. In H
>
we find a
point x ÷ Lsuch that the residue class of x÷x
0
generates L
0
,U where U = (H¨
L) ÷x
0
. In fact, L
0
,U Š Z. After division by ˛(x) we can assume ˛(x) = 1. This
affine form satisfies the following conditions: (i) ˛(L) = Z, and (ii) ˛(H) = {0].
We call it the L-height above H (with respect to the choice of H
÷
),
ht
L,H
(y) = ˛(y). y ÷ A.
If y ÷ L, then the absolute value of the height counts the number of proper paral-
lels to H through points of Lthat lie between y and H.
If A is an affine space with lattice structure L
t
= x ÷L, then we can identify
it with R
d
, d = dimA, and lattice structure Z
d
, by choosing x as the origin and
coordinate axes in direction e
i
, i = 1. . . . . d, where e
1
. . . . . e
d
is a basis of L.
Evidently every lattice structure x÷Lon Adefines a rational structure, namely
x ÷QL.
1.G. Rationality and integrality 45
Definition 1.79. Let Lbe an affine lattice in an affine space A. Then the convex hull
of P = conv(x
1
. . . . . x
m
) of finitely many points x
i
÷ L is called an L-polytope
or, less precisely, a lattice polytope. The set P ¨ L of L-points of P will also be
denoted by lat(P) if there is no ambiguity in regard to L.
Let P be an L-polytope and P
t
be an L
t
polytope. Then a morphism of lattice
polytopes ç : P ÷ P
t
is a morphism of polytopes with the additional condition
ç(aff(P) ¨ L) Ç L
t
.
Sometimes it will be necessary to restrict the lattice L to aff(P). Therefore we
introduce the notation
L
P
= L¨ aff(P).
The smallest lattice that admits P as a lattice polytope is
L(P) = :
0
÷

¡÷vert(P)
Z(: ÷ :
0
)
where :
0
is an arbitrary vertex of P.
Figure 1.15 shows a lattice polygon with lattice Z
2
. The lattice points of P are
marked by solid circles.
Figure 1.15. A lattice polygon and one of its lattice triangulations
Lattice polytopal complexes. The definition of a lattice polytopal complex needs
some care. In order to simplify it as much as possible, we identify the sets p ÷ ˘
with the associated polytopes P
p
and simply write P ÷ ˘. Correspondingly the
map ¬
P
, P ÷ ˘, is just the identity map on P, and we can consider aff(P ¨Q) as
an affine subspace of aff(P) and of aff(Q).
Definition 1.80. A lattice polytopal complex is a polytopal complex ˘ augmented
by a family ¹
P
, P ÷ ˘, of affine lattices such that the following conditions are
satisfied:
(a) every face P ÷ ˘ is a ¹
P
-polytope,
(b) ¹
P
¨ aff(P ¨ Q) = ¹
Q
¨ aff(P ¨ Q) for all faces P. Qof ˘.
In an embedded lattice polytopal complex consisting of polytopes P ÷ R
d
all the
affine lattices ¹
P
, P ÷ ˘, coincide with a fixed affine lattice L in R
d
(so that
condition (b) is automatically satisfied).
The system ¹ = {¹
P
: P ÷ ˘] is called a lattice structure on ˘. If we want
to emphasize the lattice structure being considered we will write (˘. ¹) instead of
just ˘. The collection of all the lattice points in the faces P ÷ ˘ is the set of lattice
points of ˘.
46 1. Polytopes, cones, and complexes
The simplest examples of lattice polytopal complexes are given by the face lat-
tices of lattice polytopes: the system
˚
(L
F
) : F a face of P
_
is a lattice structure on the face lattice of P.
Remark 1.81. Condition (b) of Definition 1.80 does in general not follow from (a).
The reason is that for an L-polytope one can not always reconstruct L ¨ aff(P)
from L ¨ P: the smallest affine lattice containing L ¨ P may be a proper subset
of L ¨ aff(P), as the following example shows. Let P = conv
_
(0. 0. 0). (0. 1. 0).
(1. 0. 0). (1. 1. 2)
_
Ç R
3
. Then the vertices are the only lattice points of the Z
3
-
polytope P, and the smallest lattice containing P ¨ Z
3
is Z
2
˚2Z.
Now we identify R
3
with (R
3
. 0) Ç R
4
and set Q = conv(P. (0. 0. 0. 1)). The
polytopal complex ˘ consisting of Qand all its faces is not a lattice polytopal com-
plex if we choose ¹
Q
= Z
4
and ¹
F
as the smallest affine lattice in R
4
containing
F for all proper faces F of Q, despite the fact that condition (a) is satisfied for this
collection of data.
Later on we will consider more general objects than those described in Defini-
tion 1.80.
Remark 1.82. Let L be an affine lattice and P. Q be L-polytopes. Then P ÷ Q
is again a lattice polytope, namely with respect to the affine lattice : ÷ L where
: ÷ vert(P). Similarly cP is a lattice polytope with respect to (c ÷ 1): ÷L.
If ˘ is a lattice polytopal complex, then we can endow c˘ with a lattice struc-
ture by applying the construction above to cP for every P ÷ ˘. Without the lattice
structure, c˘ is isomorphic to ˘ (see Example 1.42), but as lattice polytopal com-
plexes ˘ and c˘ are isomorphic only in trivial cases.
In a lattice subdivision we require that the associated lattice does not change if
we pass from a polytope to one of its subdividing parts:
Definition 1.83. A lattice subdivision of a lattice polytopal complex (˘. ¹) is a
lattice polytopal complex (˘
t
. ¹
t
) such that the following conditions are satisfied:
(a) ˘
t
is a subdivision of ˘,
(b) vert(˘
t
) consists of lattice points in ˘,
(c) for all faces P ÷ ˘ and P
t
÷ ˘
t
with P
t
Ç P we have ¹
P
0 = ¹
P
¨aff(P
t
).
An example of a lattice subdivision is given in Figure 1.15.
Exercises
1.1. Let P and Q be polyhedra, x ÷ P, y ÷ Q. Show that x ÷ y is a vertex of P ÷ Q if
and only x is a vertex of P, y is a vertex of Q, and there exists an affine form˛ that attains
its maximum on P exactly in x and its maximum on Qexactly in y.
Chapter 2
Affine monoids and their Hilbert bases
2.A. Affine monoids
In common usage, a monoid is a set M together with an operation M × M ÷M
that is associative and has a neutral element. We are mainly interested in a special
class of commutative monoids:
Definition 2.1. A monoid is affine if it is finitely generated and isomorphic to a
submonoid of a free abelian group Z
d
for some d ¸ 0.
Very often, especially in the commutative algebra literature, affine monoids are
called affine semigroups – a line of tradition followed in our joint papers. The usage
of ‘monoids’ in this book is more compatible with another tradition – that in the
K-theoretic literature.
In view of the definition above it is appropriate to use additive notation for the
operation in M. The condition on finite generation then just means that there exist
x
1
. . . . . x
n
÷ M for which
M = Z
÷
x
1
÷ ÷Z
÷
x
n
= {a
1
x
1
÷ ÷a
n
x
n
: a
i
÷ Z
÷
].
We are of course always free to consider an affine monoid as a submonoid of Z
d
for suitable d.
Within the class of commutative monoids, the affine monoids are characterized
by being (i) finitely generated, (ii) cancellative, and (iii) torsionfree.
Cancellativity means that an equation x ÷y = x ÷z for x. y. z ÷ M implies
y = z. Equivalently, M can be embedded into a group. More generally, for every
commutative monoid M there exists a monoid homomorphism | : M ÷ G to a
group G solving the following universal problem: every homomorphismç : M ÷
H to a group H factors in a unique way as ç = [ ı | with a group homomorphism
[ : G ÷ H. The group G is unique up to isomorphism. It is constructed as
follows. The group of differences gp(M) of M is the set of the equivalence classes
of pairs (x. y)
-
, x. y ÷ M with (x. y)
-
= (u. :)
-
if and only if x÷:÷z = u÷y÷z
for some z ÷ M. The addition is defined by (x. y)
-
÷(u. :)
-
= (x ÷u. y ÷:)
-
.
The map x ÷(x. 0) is a homomorphismof monoids | : M ÷gp(M) that satisfies
the universality condition. Obviously, when M is cancellative, then | is injective.
47
48 2. Affine monoids and their Hilbert bases
To be torsionfree for a monoid M means that ax = ay for a ÷ Nand x. y ÷ M
implies x = y. (It is not enough to require the torsion freeness condition for
gp(M); see Exercise 2.1). If M is cancellative, this condition is equivalent to the
torsion freeness of gp(M).
We agree on the following
Convention. In the remainder of this book the term monoid always means a com-
mutative, cancellative and torsionfree monoid, unless explicitly stated otherwise.
We see that if a monoid M is finitely generated, then it can be embedded into a
group which is finitely generated and torsion free, i. e. isomorphic to a free abelian
group Z
r
. If M is given as a submonoid of Z
d
, then gp(M) can be identified with
the subgroup ZM of Z
d
generated by M. The monoids that are isomorphic to
Z
r
÷
, r ÷ Z
÷
are called free monoids.
Definition 2.2. The rank of a monoid M is the rank of the abelian group gp(M).
In other words, it is the vector space dimension of Q˝gp(M) over Q.
Clearly, if M is affine and gp(M) Š Z
r
, then rank M = R. However, note that
the definition of rank is not restricted to finitely generated monoids.
Remark 2.3. Every submonoid of Z is finitely generated (Exercise 2.2). It is iso-
morphic to a submonoid of Z
÷
, unless it is a subgroup of Z. Submonoids of Z
÷
are often called numerical semigroups. A vast amount of research has been devoted
to them. See Barucci, Dobbs and Fontana [2].
In contrast, already Z
2
contains submonoids without a finite system of gener-
ators, for example {(0. 0)] L {(x. y) : x ¸ 1].
Subcones C of R
d
are examples of “continuous” monoids. Unless C = 0, such
a monoid is not finitely generated. Neither is C¨Q
d
finitely generated if it contains
a nonzero vector.
As in ring theory, it is useful to introduce the notion of module over a monoid
M: a set N with an (additively written) operation M × N ÷ N is called an M-
module if (a÷b) ÷x = a÷(b ÷x) and 0÷x = x for all a. b ÷ M and x ÷ M. A
typical example of a module over M Ç Z
d
is Z
d
itself. Then every subset U of Z
d
with M ÷ U Ç U is a submodule. The submodules of M itself are called ideals.
The empty set is considered an M-module.
Remark 2.4. (a) Let M Ç Z
d
be an affine monoid. Then we define the interior of
M by int(M) = M ¨ int(R
÷
M). Since x ÷ y ÷ int(R
÷
M) for x ÷ int(R
÷
M),
y ÷ R
÷
M, it follows that int(M) is an ideal.
One has 0 ÷ int(M) if and only if M is a group. In that case int(M) = M. In
all the other cases int(M) is not a monoid itself. However, M
+
= int(M) L {0] is
a submonoid of M that will play an important role later on. Note that M
+
= M
if and only if rank M ¸ 1 or M = int(M). Otherwise M
+
is not even finitely
generated, as the reader may show (Exercise 2.3).
(b) In (a) we have assumed that M Ç Z
d
. Nevertheless int(M) is defined
intrinsically in terms of M and does not depend on an embedding of M into Z
d
(or R
d
) for some d. The reason is that an isomorphism of monoids M Ç R
d
and
2.A. Affine monoids 49
N Ç R
e
induces isomorphisms (i) gp(M) Š gp(N), (ii) RM Š RN, and (iii)
R
÷
M Š R
÷
N.
Another solution for intrinsic definitions is to associate a vector space with a
commutative monoid M in an abstract way by taking V = gp(M) ˝
Z
R.
(c) Let M be an affine monoid and F a face of the cone R
÷
M. Then F ¨ M,
the intersection of M and the monoid F, is a submonoid of M. The submonoids
of M that arise in this way, are called extreme submonoids.
For an arbitrary monoid M and every commutative ring of coefficients R we
can form the monoid algebra RŒM| – the main vehicle in our study of the inter-
actions between discrete geometry, commutative ring theory, and algebraic K-
theory. As an R-module, RŒM| is free with a basis consisting of the symbols X
a
,
a ÷ M, and the multiplication on RŒM| is defined by the R-bilinear extension of
X
a
X
b
= X
a÷b
.
1
For an M-module N we can analogously define an RŒM|-module RN that
as an R-module is free on the basis X
n
, n ÷ N, and on which X
a
operates by
X
a
X
n
= X
a÷n
. We leave it to the reader to give the proofs of the statements
implicitly contained in the descriptions of RŒM| and RN.
Suppose ç : gp(M) ÷ Z
d
is an embedding of M. It induces an embedding
RŒM| ÷ RŒZ
d
|. The group algebra RŒZ
d
| is identified with the Laurent polyno-
mial ring RŒX
˙1
1
. . . . . X
˙1
d
| if we set
X
s
= X
s
1
1
X
s
d
d
. s = (s
1
. . . . . s
d
).
Under this identification RŒM| becomes a subalgebra of RŒX
˙1
1
. . . . . X
˙1
d
| that is
generated by finitely many monomials, and M itself can be considered as a finitely
generated monoid of monomials. Therefore we will call X
a
, a ÷ M, a monomial
in RŒM|.
Proposition 2.5. Let M be a monoid, N an M-module, and Ra ring. Then:
(a) M is finitely generated if and only if RŒM| is a finitely generated R-algebra;
(b) N is a finitely generated M-module if and only if RN is a finitely generated
RŒM|-module.
Proof. (a) The implication == is trivial. For the converse let ¡
1
. . . . . ¡
n
be a
system of generators of RŒM|. There exists a finite subset E of M such that every
¡
i
is an R-linear combination of the elements X
e
with e ÷ E. Let M
t
= Z
÷
E.
It follows immediately that any R-linear combination of the products ¡
a
1
1
¡
a
n
n
with a
1
. . . . . a
n
÷ Z
n
÷
is an R-linear combination of monomials X
b
, b ÷ M
t
.
Since ¡
1
. . . . . ¡
n
generate RŒM|, it follows that X
a
÷ RŒM
t
| for every a ÷ M. This
implies M = M
t
.
(b) is proved along the same lines.
In using Proposition 2.5 for proving assertions about M or N we are free to
choose R, for example we can take Rto be a field K.
1
Later on, when the use of monoid rings becomes essential, we will switch to a simpler notation
identifying the elements m ÷ M with X
m
and writing the monoid operation multiplicatively.
50 2. Affine monoids and their Hilbert bases
Proposition 2.6. Let M be a finitely generated monoid and N a finitely generated
M-module. Then every M-submodule of N is finitely generated.
Proof. We choose a field K. Then KŒM| is a finitely generated K-algebra and
therefore a Noetherian ring by Hilbert’s basis theorem (see [32, Section 1.4]). So
every submodule of the finitely generated KŒM|-module KN is finitely generated.
It follows that KU is finitely generated for every M-submodule U of X, whence
U is finitely generated over M by the previous proposition.
Gordan’s lemma. The first substantial result about affine monoids is Gordan’s
lemma:
Lemma 2.7. Let C be a rational cone in R
d
, and L Ç Q
d
a lattice. Then C ¨ L is
an affine monoid.
Proof. Set C
t
= C ¨ RL. Then C
t
is a rational cone as well. Moreover every
element x of Q
d
¨ RL is a rational linear combination of the elements of L and
so there exists a positive integer a with ax ÷ L. We choose a finite system of
generators x
1
. . . . . x
n
of C
t
. As just seen, we can assume that x
1
. . . . . x
n
÷ L. Let
M
t
be the affine monoid generated by x
1
. . . . . x
n
.
Every element x of C
t
¨ L has a representation x = a
1
x
1
÷ ÷ a
n
x
n
with
a
i
÷ R
÷
. (We can choose a
1
. . . . . a
n
÷ Q
÷
, but this is irrelevant here.) Therefore
x =
_
]a
1
¡x
1
÷ ÷]a
n
¡x
n
_
÷
_
q
1
x
1
÷ ÷q
n
x
n
_
.
0 ¸ q
i
= a
i
÷ ]a
i
¡ < 1. i = 1. . . . . n.
The first summand on the right hand side is in M
t
, the second is an element of
C
t
¨Lthat belongs to a bounded subset B of R
n
. It follows that C
t
¨Lis generated
as an M
t
-module by the finite set B ¨ C
t
¨ L. Being a finitely generated module
over an affine monoid, the monoid C ¨ Lis itself finitely generated.
The essential point in the hypothesis of Lemma 2.7 is the existence of a lattice
L
t
containing both Land the generators of C: after a change of coordinates we can
assume that L
t
Ç Z
d
; then C is rational, and we have arrived at the hypothesis of
Lemma 2.7.
The reader may show that the intersection of Z
2
with the cone generated by
(1. 0) and (1. a) is a finitely generated monoid if and only if a is rational (Exercise
2.5).
Corollary 2.8. Let M be a submonoid of R
d
, L a lattice in R
d
containing M, and
C = R
÷
M. Then the following are equivalent:
(a) M is an affine monoid;
(b)
¯
M
L
= C ¨ Lis an affine monoid;
(c) C is a cone.
Moreover, if M is affine, then
¯
M
L
is a finitely generated M-module.
Proof. It is obvious that both (a) and (b) imply (c), The implication (c) ==(b) is
essentially Gordan’s lemma. To complete the proof one notes that the affine monoid
2.A. Affine monoids 51
M
t
in the proof of the lemma can be chosen as a submonoid of M if the cone
under consideration is generated by M. Thus M is a submodule of the finitely
generated M
t
-module
¯
M
L
, and therefore itself finitely generated by 2.6. Since
¯
M
L
is a finitely generated module over M
t
, it is even more so over M.
The monoids of type
¯
M
L
will be further investigated in the next subsection.
Together with the polytopal monoids to be introduced below, they are the objects
of our primary interest.
Corollary 2.9. Let M and N be affine submonoids of R
d
, and let C be a cone gen-
erated by elements of gp(M). Then
(a) M ¨ N is an affine monoid;
(b) M ¨ C is an affine monoid;
(c) the extreme submonoids of M are affine.
Proof. (a) The group generated by M L N is a finitely generated torsionfree
abelian group and therefore isomorphic to Z
n
for some n (possibly, n > d). There-
fore we may right from the start assume that M and N are submonoids of Z
d
.
The conical set R
÷
M ¨ R
÷
N is the intersection of the rational cones R
÷
M and
R
÷
N. Such an intersection is itself a rational cone by Proposition 1.75. Since
R
÷
M ¨ R
÷
N = R
÷
(M ¨ N) (as is easily shown), the claim follows from the
previous corollary.
(b) By Corollary 2.7 gp(M) ¨ C is an affine monoid. Therefore M ¨ C =
M ¨ gp(M) ¨ C is affine by (a).
(c) is an immediate consequence of (b) since every face F of R
÷
M is generated
by M ¨ F.
These results provide us with a wealth of examples of affine monoids. Gordan’s
lemma has an inhomogeneous generalization. It can be considered as the discrete
analogue of Motzkin’s theorem 1.32 on the finite generation of polyhedra.
Theorem 2.10. Let P Ç R
d
be a rational polyhedron, C the recession cone of P,
and L Ç Q
d
a lattice. Then P ¨ L is a finitely generated module over the affine
monoid C ¨ L.
Proof. We form the cone C(P) Ç R
d÷1
over P and extend the lattice L to L
t
=
L˚Z. By Gordan’s lemma C(P)¨L
t
is an affine monoid. Its generators at ‘height’
0 generate C ¨ L, and its generators at ‘height’ 1 generate P ¨ L as a C ¨ L-
module.
The theorem contains (and is equivalent to) a statement about the set N of
solutions of a homogeneous system of linear diophantine inequalities and congru-
ences given as
a
i1
x
1
÷ ÷a
id
x
d
¸ u
i
. i = 1. . . . . m. a
ij
. u
i
÷ Z.
b
i1
x
1
÷ ÷b
id
x
d
÷ :
i
(n
i
). i = 1. . . . . n. b
ij
. :
i
. n
i
÷ Z.
(2.1)
(We can split equations into a pair of inequalities.) The inequalities define a ratio-
nal polyhedron P. The homogeneous congruences b
i1
x
1
÷ ÷b
id
x
d
÷ 0 (n
i
)
52 2. Affine monoids and their Hilbert bases
define a sublattice L of Z
d
, and we claim that N is a finite module over L. In fact,
the set of solutions of the inhomogeneous system of congruences is an affine lattice
of the formx
t
÷L(unless it is empty). Let P
t
= P ÷x
t
; then N = x
t
÷(P
t
¨L),
and to P
t
¨ L we can apply the theorem: there exists a finite set E generating
P
t
¨ L as an P ¨ L-module. To find the system of generators of N, we replace
each x ÷ E by x
t
÷x. To sum up:
Corollary 2.11. Let N Ç Z
d
be the set of solutions of the system (2.1), and let M Ç
Z
d
be the (affine normal) monoid of solutions of the corresponding homogeneous
system. Then N is a finitely generated M-module.
Irreducible elements, standard map and total degree. The nonzero elements of
an integral (say, Dedekind) domain form a commutative cancellative monoid with
respect to multiplication. In analogy to number-theoretic nomenclature, let us call
an element x of a monoid M a unit if x has an inverse in M, and irreducible if in
every decomposition x = y ÷z one of the summands y. z must be a unit.
It is not difficult to analyze affine monoids M in these terms. To this end, and
for many other purposes, we introduce the standard map on an affine monoid as
follows. The group gp(M) is isomorphic to Z
r
, r = rank M. We identify gp(M)
and Z
r
. Let C = R
÷
M Ç R
r
the cone generated by M. This cone has a represen-
tation
C = H
÷
o
1
¨ ¨ H
÷
o
s
as an irredundant intersection of halfspaces defined by linear forms on R
r
. Each of
the hyperplanes H
o
i
is generated as a vector space by integral vectors. Therefore
we can assume that o
i
is the Z
r
-height above H
o
i
(with nonnegative values on C;
see Remark 1.78). After this standardization we call the o
i
the support forms of M
and
o : M ÷Z
s
÷
. o(x) = (o
1
(x). . . . . o
s
(x))
the standard map on M. The hyperplanes H
o
i
are called the support hyperplanes
of M.
The standard map has a natural extension to R
r
with values in R
s
, also denoted
by o. It restricts to a Z-linear map gp(M) ÷ Z
s
. Furthermore o(C) Ç R
s
÷
.
Theorem 1.20 implies that o
1
. . . . . o
s
is a minimal set of generators of the polar
cone C
+
.
Note that the standard map depends only on the order of o
1
. . . . . o
s
. The o
i
(as
Z-linear forms on gp(M)) are defined intrinsically by M; see Remark 2.4(b).
We call t = o
1
÷ ÷ o
s
the total degree on M. The term “total degree” is
justified: for any ring R of coefficients, o induces an homomorphism RŒM| ÷
RŒY
1
. . . . . Y
s
| of R-algebras, namely the R-linear extension of the map X
m
÷
Y
o(m)
and t(m) is the total degree of the monomial Y
o(m)
.
Proposition 2.12. Let M be an affine monoid with standard map o. Then:
(a) the units of M are precisely the elements x with o(x) = 0, or, equivalently,
t(x) = 0;
2.A. Affine monoids 53
(b) every element x ÷ M has a presentation x = u ÷y
1
÷ ÷y
m
in which
u is a unit and y
1
. . . . . y
m
are irreducible;
(c) up to differences by units, there exist only finitely many irreducible elements
in M.
Proof. (a) Clearly if x. ÷x ÷ M, then o(x) = 0, since 0 is the only unit in Z
s
÷
.
Conversely, let o(x) = 0. Then o(÷x) = 0 as well, and so ÷x ÷ C, where
C = R
÷
M as above. Thus there exists a positive integer m with m(÷x) ÷ M.
Then x
t
= (m÷ 1)x ÷m(÷x) ÷ M, too, and x ÷x
t
= 0.
(b) Suppose that x is neither a unit nor irreducible. Then there exists a decom-
position x = y÷z in which neither y nor z is a unit, and since t(x) = t(y)÷t(z),
assertion (a) implies t(y). t(z) < t(x). So we are done by induction.
(c) By hypothesis, M is finitely generated, say by x
1
. . . . . x
p
. We apply (b) to
each of the x
i
and obtain a collection H of finitely many irreducible elements.
Evidently, every other irreducible element of M has the form h ÷ u with h ÷ H
and a unit u.
The total degree t is an example of a grading ; on M: we use this notion as a
synonym for “monoid homomorphism from M to Z”. The choice of this name is
motivated by the fact that a grading on M induces a grading on the algebra RŒM|
in which all elements of R have degree 0 and every monomial X
s
has degree ;(s)
(see Remark 4.7). The total degree is a nonnegative grading with values in Z
÷
.
Positive monoids. Suppose that 0 is the only unit in M. It follows immediately
from Proposition 2.12 that M has only finitely many irreducible elements in this
case. Not only do they constitute a systemof generators – they must be contained in
every system of generators. This observation justifies the definition of the Hilbert
basis:
Definition 2.13. A monoid is called positive if 0 is its only unit. The unique min-
imal system of generators of a positive affine monoid M given by its irreducible
elements is called the Hilbert basis of M and denoted by Hilb(M).
For positive affine monoids the standard map is injective, and therefore we call
it standard embedding in this case:
Proposition 2.14. Let M be an affine monoid with gp(M) = Z
r
and C = R
÷
M Ç
R
r
. Then the following are equivalent:
(a) M is positive;
(b) the standard map o is injective on M;
(c) o : R
r
÷R
s
is injective;
(d) C is pointed.
Proof. (a) == (d): Let U be the vector subspace of R
r
consisting of all the el-
ements x ÷ C with ÷x ÷ C. It is evidently the intersection of the support hy-
perplanes of C and therefore a rational vector subspace. It is enough to show that
x = 0 for x ÷ U ¨Q
r
. For such x there exists a ÷ N, with ax. a(÷x) ÷ M. Hence
x = 0.
54 2. Affine monoids and their Hilbert bases
(d) == (c): If C is pointed, the polar cone C
+
has dimension d (see Propo-
sition 1.24). Thus the support forms generate (R
r
)
+
, and a suitable collection
o
i
1
. . . . . o
i
r
of r support forms is linearly independent. If o
i
j
(x) = 0 for j =
1. . . . . r, then x = 0.
The remaining implications (c) ==(b) and (b) ==(a) are trivial.
The total degree t on a positive affine monoid is a grading under which only 0
has degree 0, as follows immediately from the injectivity of o: it is an example of a
positive grading ; on M, by definition a homomorphism ; : M ÷ Z
÷
such that
;(x) = 0 implies x = 0. A positive grading on M induces a positive grading on
the algebra RŒM| in which all elements of R have degree 0 and every monomial
X
s
, s = 0, has positive degree. Moreover, a positive grading on M induces a
weight (see p. 28) on the cone R
÷
M by linear extension to RM.
We can now characterize positive affine monoids as submonoids of Z
n
÷
, n ¸ 0:
Proposition 2.15. Let M be an affine monoid of rank r and with s support forms.
Then the following are equivalent:
(a) M is positive;
(b) M is isomorphic to a submonoid of Z
d
÷
for some d;
(c) M is isomorphic to a submonoid M
t
of Z
s
÷
such that the intersections
H
i
¨RM
t
of the coordinate hyperplanes H
1
. . . . . H
s
are exactly the support
hyperplanes of M
t
;
(d) M is isomorphic to a submonoid M
t
of Z
r
÷
such that such that the intersec-
tions H
i
¨ RM
t
of the coordinate hyperplanes H
1
. . . . . H
r
are among the
support hyperplanes of M
t
;
(e) M is isomorphic to a submonoid M
t
of Z
r
÷
with gp(M
t
) = Z
r
;
(f) M has a positive grading.
Proof. Each of (b)–(f) implies (a) for obvious reasons, and each of (c)–(e) implies
(b). That (a) ==(f) has already been observed.
For the implication (a) == (c) we use the injectivity of the standard embed-
ding o : M ÷ Z
s
. Set M
t
= o(M). Then M
t
= gp(M
t
) ¨ R
s
÷
, and since
M
t
Š M has s support hyperplanes, all the coordinate hyperplanes must be sup-
port hyperplanes of M
t
. For (a) == (d) we simply choose r linearly independent
ones among the support forms, like in the proof of Proposition 2.14(d) ==(c).
For the implication (a) == (e) we have to find r linear forms ,
1
. . . . . ,
r
with
integral coefficients in (R
÷
M)
+
that form a basis of (Z
r
)
+
. The existence of such
a basis will be shown in Theorem 2.67. (However, note that the additional con-
ditions in (d) and (e) cannot always be satisfied simultaneously, since ,
1
. . . . . ,
r
cannot always be chosen as support forms of M. In Exercise 2.6 is asked to find an
example.)
Polytopal monoids. We now introduce a special class of affine monoids, whose
investigation in combinatorial and algebraic terms will be very important for us.
2.A. Affine monoids 55
Definition 2.16. Let L be an affine lattice in R
d
and P an L-polytope. The poly-
topal affine monoid M(P) associated with P is the monoid
Z
÷
{(x. 1) : x ÷ lat(P)].
in R
d÷1
. The set {(x. 1) : x ÷ P ¨ L] generating M(P) is denoted by E(P).
Since the set lat(P) is finite, M(P) is an affine monoid. It is evidently positive
and its Hilbert basis is the set Z
÷
{(x. 1) : x ÷ lat(P)].
P
C(P)
Figure 2.1. Vertical cross-section of a polytopal monoid
The lattice L is not uniquely determined by P, and therefore the notations
E(P) and M(P) are somewhat ambiguous. It will however always be clear which
lattice Lis to be considered.
Remark 2.17. More generally, we can associate a submonoid of R
d
˚Z
÷
Ç R
d÷1
with every subset X of R
d
, replacing the generating set E(P) by {(x. 1) : x ÷ X].
The structure of this monoid depends only on the affine structure of X. In fact,
suppose that ˛ is an affine isomorphism of R
d
onto itself. Then ˛ induces a linear
isomorphism ˛
t
of the vector space R
d÷1
given as follows:
˛
t
(x. h) = (˛(x) ÷ ˛(0). h). x ÷ R
d
. h ÷ R.
The restriction of ˛
t
to the monoid M over X maps M isomorphically onto the
monoid over ˛(X).
In particular, if L
t
= x÷Lis an affine lattice associated with the lattice L. and
P
t
is an L
t
-polytope, then P = P
t
÷x is an L-polytope such that M(P
t
) Š M(P).
Whenever it should be convenient for the analysis of M(P), we can therefore as-
sume that P has its vertices in a lattice contained in a vector space (and not just in
an affine lattice).
Polytopal monoids are special instances of homogeneous affine monoids: such
monoids M are positive and admit a positive grading in which every irreducible
element has degree 1. The proof of the following proposition is left to the reader
(Exercise 2.7).
Proposition 2.18. Let M be an affine monoid. Then the following are equivalent:
(a) M is homogeneous;
(b) there exists a hyperplane H of R
d
, not passing through 0, such that M is
generated by elements of H;
(c) M is positive, and the number of summands in every representation of an
element x ÷ M as a sum of irreducibles is constant.
56 2. Affine monoids and their Hilbert bases
Prime and radical ideals. Let I be an ideal in a monoid M. The radical of I is
the ideal Rad(I) = {x ÷ M : ax ÷ I for some a ÷ N]. One calls I a radical ideal
if I = Rad(I), and a prime ideal if m ÷ n ÷ I for m. n ÷ M is only possible if
m ÷ I or n ÷ I. An affine monoid has only finitely many radical ideals and they
are completely determined by the geometry of R
÷
M:
Proposition 2.19. Let M be an affine monoid and I Ç M an ideal.
(a) I is a radical ideal if and only if I is the intersection of the sets M\F where
F is a face of R
÷
M with F ¨ I = 0.
(b) I is a prime ideal if and only if there exists a face F with I = M \ F.
Proof. Let F be a face of R
÷
M. Then M \ F is obviously a prime ideal, and since
the intersection of prime ideals is radical, the intersection of sets of type M \ F is
a radical ideal.
Conversely, suppose that I is a radical ideal, and let x ÷ M \ I. There is a
unique face F with x ÷ int(F), and it is enough to show that F ¨ I = 0. On the
contrary, assume there exists y ÷ F ¨ I. Let N = F ¨ M. Then kx ÷ y ÷N Ç
y÷M Ç I for some k > 0 by Corollary 2.33. Since I is a radical ideal, this implies
x ÷ I, a contradiction.
For (b) one must show that the set of faces F with F ¨ I = 0 has a unique
maximal element if I is a prime ideal. Assume that F
1
. . . . . F
n
are the maximal
faces in the complement of I. If n ¸ 2, then a sum x
1
÷ ÷ x
n
with x
i
÷
int(F
i
) ¨M for all i does neither belong to I if I is prime nor does it belong to the
union F
1
L L F
n
. This is a contradiction.
By Proposition 2.19(b) the extreme submonoids of M (compare Remark 2.4)
are exactly the complements of the prime ideals.
Special instances of prime ideals I are those generated by a single element x; in
this case I = M÷x. Then x is called a prime element. We say that prime elements
x. y are non-associated if M ÷x = M ÷y, or, in other words, x ÷ y … M.
Proposition 2.20. Let M be an affine monoid, and N a submonoid generated by
prime elements x
1
. . . . . x
n
that are pairwise non-associated. Then there exists a
single face F of R
÷
M such that x
i
… F for all i , and M = Z
÷
x
1
˚ ˚Z
÷
x
n
˚
(M ¨ F).
Proof. We use induction, starting with the case n = 1. So let x be a prime element
in M. Since x is not a unit, there exists a facet F not containing x. The ideal M\F
is a prime ideal that does not contain any other prime ideal of M, and since x … F,
we must have M ÷x = M \ F.
Let y be an element of M. Then there exists k ÷ Z
÷
such that y÷kx ÷ M, but
y ÷ (k ÷1)x … M (simply because o
F
(x) > 0). Then y ÷ kx ÷ F, for otherwise
y ÷kx ÷ M ÷x, and y ÷(k ÷1)x ÷ M. We conclude that M = Z
÷
÷(M ¨F).
Since Zx ¨ F = {0], we must even have M = Z
÷
x ˚(M ¨ F).
Now suppose that n > 1. Then we split M using x
1
: M = Z
÷
x
1
˚(M ¨F
1
).
Since x
2
. . . . . x
n
… M ÷ x, they all lie in F, and we can conclude by induction
since x
2
. . . . . x
n
are prime elements also in the monoid M ¨F
1
and pairwise non-
associated, as the reader may check.
2.B. Normal affine monoids 57
2.B. Normal affine monoids
Definition 2.21. Let M be a submonoid of a commutative monoid N. The satura-
tion or integral closure of M in N is the submonoid
¯
M
N
= {x ÷ N : mx ÷ M for some m ÷ N]
of N. One calls M saturated or integrally closed in N if M =
¯
M
N
.
The normalization
¨
M of a cancellative monoid M is the integral closure of M
in gp(M), and if
¨
M = M, then M is called normal.
The terms “integral closure” and “normalization” are borrowed from commu-
tative algebra, to which we will connect them in Section 4.D.
For submonoids of Q
d
the integral closure has a clear geometric interpretation.
It shows that the use of the notation
¯
M
L
in Corollary 2.8 was justified.
Proposition2.22. Let M Ç N be submonoids of Q
d
and C the conical set generated
by M. Then
¯
M
N
= C ¨ N.
If M and N are affine monoids, then so is
¯
M
N
.
Proof. The inclusion
¯
M
N
Ç C ¨ N is trivial.
For the converse inclusion we choose an element x ÷ C ¨ N. Then x has a
representation as a Q
÷
-linear combination of elements of M. Let m be the least
common multiple of the denominators of the coefficients. Then mx ÷ M and,
hence, x ÷
¯
M
N
.
For the last statement we set L = gp(N). Then
¯
M
L
is affine by Corollary 2.8,
and the intersection
¯
M
N
= N ¨
¯
M
L
is affine by Corollary 2.9.
Example 2.23. Consider the extension of affine monoids 2Z
÷
Ç Z
÷
. Then 2Z
÷
is
normal and, hence, smaller than its integral closure in Z
÷
, which is Z
÷
itself.
The most important consequence of Proposition 2.22 is the characterization of
affine normal monoids. It combines Proposition 2.22 with Gordan’s lemma.
Corollary 2.24. Let M Ç Z
r
be a monoid such that gp(M) = Z
r
. Then Z
r
¨R
÷
M
is the normalization of M.
Moreover, the following are equivalent:
(a) M is normal and affine;
(b) R
÷
M is finitely generated and M = Z
r
¨ R
÷
M;
(c) there exist finitely many rational halfspaces H
÷
i
Ç R
r
such that M =
_
i
H
÷
i
¨ Z
r
.
We will sometimes refer to the monoids H
÷
capZ
r
(with H
÷
a rational half-
space) as discrete halfspaces.
Before we record another useful corollary, we remind the reader that N
+
=
{0] L int(N) (see Remark 2.4).
Corollary 2.25. Let M be a (not necessarily affine) integrally closed submonoid of
an affine monoid N. If rank M = rank N, then gp(M) = gp(N). In particular,
gp(N
+
) = gp(N).
58 2. Affine monoids and their Hilbert bases
Proof. We can assume that gp(N) = Z
r
. Note that RM = RN = R
r
since
rank M = r. There is nothing to show if r = 0. Suppose now that r > 0, and
choose elements x
1
. . . . . x
r
÷ M generating R
r
as a vector space. Let M
t
=
Z
÷
x
1
÷ ÷ Z
÷
x
r
, and set M
tt
= R
÷
M
t
¨ N. Then M
tt
is the integral clo-
sure of M
t
in N and an affine monoid itself by Proposition 2.22. Since M
t
Ç M,
we obtain that M
tt
Ç M by hypothesis on M. We can now replace M by M
tt
, and
assume that M is itself affine.
Choose x ÷ int(M). Then all support forms of M have positive value on x.
For y ÷ N it therefore follows that y ÷ kx ÷ R
÷
M for k ; 0. Proposition
2.22 implies that y ÷kx is integral over M. So y ÷kx ÷ M by hypothesis, and
y ÷ gp(M).
In the previous subsection we have introduced the standard map o : M ÷Z
s
on an affine monoid M. Proposition 2.12(a) shows that o(x) = 0 if and only if x
is a unit in M. For normal monoids M we can say even more:
Proposition 2.26. Let M be a normal affine monoid, U(M) its subgroup of units,
and o : gp(M) ÷Z
s
the standard map on M. Then M is isomorphic to U(M) ˚
o(M).
Proof. Let L = gp(M). We claim that U(M) is the kernel of o. Clearly U(M) Ç
Ker o. Conversely let x ÷ Ker o. Then x ÷ C ¨ L, where C is the cone generated
by M. The normality of M then shows x ÷ M, and so x ÷ U(M).
Since U(M) is a direct summand of L, there exists a projection ¬ : L ÷
U(M), i. e. a surjective Z-linear map ¬ with ¬
2
= ¬. Let x ÷ M. Then
(¬(x). o(x)) ÷ U(M) ˚ o(M). Conversely, given (x
0
. y
t
) ÷ U(M) ˚ o(M) we
choose y ÷ M with y
t
= o(y). Then x
0
÷y ÷¬(y) ÷ M, ¬(x
0
÷y ÷¬(y)) = x
0
and o(x
0
÷y ÷ ¬(y)) = y
t
.
The direct sum of a monoid M with a group G can be considered a trivial
extension of M in almost every context treated in this book. Therefore one can
usually restrict the discussion of normal affine monoids to the positive ones.
In the proof of Proposition 2.26 the normality of M is only used to the extent
that U(M) is a direct summand of gp(M). However, if this condition is violated,
we cannot expect an isomorphism M Š U(M) ˚ o(M). For example, choose
M = {(x. y) ÷ Z
2
: y > 0. or y = 0 and x ÷ 0 (2)].
Polytopal monoids (see Definition 2.16) can be characterized in terms of their
normalizations:
Proposition 2.27. Let M be an affine monoid. Then the following are equivalent:
(a) M is polytopal;
(b) M is homogeneous and coincides with
¨
M in degree 1.
Proof. The implication (a) == (b) is (almost) the definition of polytopal mono-
id. In fact, let P be a lattice polytope. The “height” 1 lattice points of R
÷
M(P) are
exactly the generators of M(P), and so are contained in M
(
P).
For the converse let gp(M) = Z
d
. By hypothesis M has a grading ;. We
can first extend it to Z-linear form on Z
d
and then to a linear form on R
d
. Since
2.B. Normal affine monoids 59
M is generated by elements of degree 1, R
÷
M is generated by integral vectors of
degree 1. Their convex hull is the lattice polytope P = {x ÷ R
÷
M : ;(x) = 1].
Since all the lattice points of P correspond to elements of M by hypothesis, M Š
M(P).
Normal monoids as pure submonoids. We have seen that every positive affine
monoid can be realized as a submonoid of Z
n
÷
for suitable n. This statement can
be considerably strengthened if M is normal.
Theorem 2.28. The following are equivalent for a positive affine monoid M:
(a) M is normal;
(b) there exists m ÷ Z
÷
and a subgroup U of Z
m
of rank r such that M is
isomorphic to Z
m
÷
¨ U;
(c) there exist p. q ÷ Z
÷
and a Z-linear map z : Z
p
÷ Z
q
such that M is
isomorphic to Z
p
÷
¨ Ker z.
If M is normal of rank r and with s support forms, then one can choose m = r,
p = s, and q = r ÷s.
Proof. It is obvious that the monoids in (b) and (c) are normal – those in (c) are
even integrally closed in Z
p
.
Now suppose that M is normal. We identify gp(M) with Z
r
, apply the stan-
dard embedding o : M ÷ Z
s
÷
, and set M
t
= o(M), U = gp(o(M)). The
monoid isomorphism o : M ÷ M
t
extends to an isomorphism o : gp(M) ÷ U.
Proposition 2.22 shows that x ÷ M if and only if o(x) ÷ Z
s
÷
.
For the more difficult implication (b) == (c) we have to use the elementary
divisor theorem (for example, see [60]): let L Š Z
n
be a free abelian group, and L
t
a subgroup of rank m; then there exist a basis e
1
. . . . . e
n
of L and positive integers
d
1
. . . . . d
m
such that d
1
e
1
. . . . . d
m
e
m
is a basis of L
t
(and d
1
[ d
2
[ [ d
m
).
We can assume that M = Z
s
÷
¨ U, U = gp(M), r = rank M, and apply the
elementary divisor theorem with L = Z
s
and L
t
= U. Let ,
i
, i = 1. . . . . s, be
the linear form on Z
s
that assigns each vector its i th coordinate with respect to the
basis e
1
. . . . . e
s
. Then an element x = (x
1
. . . . . x
s
) ÷ Z
s
belongs to M if and only
if (i) x
1
. . . . . x
s
¸ 0, (ii) ,
i
(x) = 0 for i = r ÷ 1. . . . . s, (iii) ,
i
(x) ÷ 0 (d
i
) for
i = 1. . . . . r.
In order to achieve our goal we have to convert the congruences into linear
equations. We can change the coefficients of ,
1
. . . . . ,
r
(with respect to the given
coordinates of Z
s
) by adding multiples of d
i
, and so we may assume that ,
1
. . . . . ,
r
have nonnegative values on Z
s
÷
. For elements x ÷ Z
s
÷
the congruence ,
i
(x) ÷ 0
(d
i
) is then equivalent to the solvability of the equation ,
i
(x) = y
i
d
i
with y
i
÷ Z
÷
,
and the solution y
i
= y
i
(x) is uniquely determined by x.
Define the map ¬ : U ÷ Z
s÷r
by x ÷
_
x. y
1
(x). . . . . y
r
(x)
_
, and set V =
¬(U). Then ¬ is injective, and ¬(x) ÷ Z
s÷r
÷
if and only if x ÷ M. It remains to
show that V is the kernel of a suitable map Z
s÷r
÷Z
s
. This is equivalent (by the
elementary divisor theorem) to the torsion freeness of Z
s÷r
,V .
60 2. Affine monoids and their Hilbert bases
Let z ÷ Z
s÷r
, z = (z
t
. z
tt
), z
t
÷ Z
s
, z
tt
÷ Z
r
, and suppose that mz ÷ V for
some positive integer m. Then x = mz
t
÷ U. Moreover,
,
j
(z
t
) = m
-1
,
j
(x) = m
-1
y
j
(x)d
j
= m
-1
mz
tt
j
d
j
= z
tt
j
d
j
.
It follows that z
t
÷ U and (z
t
. z
tt
) ÷ V . This shows the torsion freeness of Z
s÷r
,V .

Suppose M is a submonoid of N. Then M is called pure in N if M = N ¨
gp M. Equivalently we can require that N \ M is an M-submodule of N. The
notion of purity develops its power in the context of monoid algebras, where the
purity of M in N means that RŒM| is a direct summand of RŒN| as an RŒM|-
module. Theorem 2.28 shows that a normal affine monoid cannot only be realized
as a pure submonoid of Z
n
÷
for suitable n (this is (b)), but also as a pure, integrally
closed submonoid of a free monoid (this is (c)).
Remark 2.29. Let M be an positive affine normal monoid. The standard embed-
ding o : M ÷ Z
s
realizes M as a pure submonoid as we have seen in Theorem
2.28(b). The quotient Z
s
,o(gp(M)) is an intrinsic invariant of M, namely the
divisor class group Cl(M) of M (or any monoid algebra KŒM| in which K is a
factorial ring). See Corollary 4.49. So o(gp(M)) is a direct summand of Z
s
if and
only if Cl(M) is torsionfree.
Let M be a pure submonoid of the monoid N. Then N decomposes naturally
into a disjoint union of M-submodules given by the intersections of N with the
cosets of gp(M) in gp(N):
N =
_
_
y ÷gp(M)
_
¨ N
where y runs through a system of representatives of the cosets z ÷gp(M), z ÷ N.
We call the modules
_
y÷gp(M)
_
¨N the coset modules of M in N. Under suitable
finiteness conditions they are finitely generated:
Proposition 2.30. Let M be a pure submonoid of an affine monoid N. Then M is
affine, too, and the coset modules of M in N are finitely generated over M. More-
over, each coset module is isomorphic to an M-submodule of gp(M).
Proof. Since M is the intersection of affine monoids, namely N and gp(M), it is
itself affine (Corollary 2.9).
Let us first assume that N is normal. Then M is likewise normal. After the
identification of gp(N) with Z
n
, we can understand N as the set of solutions of a
homogeneous linear diophantine system of inequalities, and M is cut out from N
by a system of such equations and congruences. The coset module is then given as
the set of solutions of an inhomogeneous variant of the homogeneous system, and
so we can apply Corollary 2.11.
Now let N be arbitrary, and observe that
¨
M Ç
¨
N is again pure. Since
¨
M is a
finitely generated M-module, the coset modules of
¨
M in
¨
N are finite also over M.
But each coset module of M in N is an M-submodule of a coset module of
¨
M in
¨
N, and the desired finiteness follows.
2.B. Normal affine monoids 61
The last statement is evident: if n ÷
_
y ÷ gp(M)
_
¨ N, then ÷n ÷
_
y ÷
gp(M)
_
¨ N Ç gp(M).
The coset modules of the standard embedding play a special role: they repre-
sent the divisorial ideals of M; see Theorem 4.55.
Adjoining inverse elements. In the case of a normal affine monoid M, certain ex-
tensions of M can be controlled by the support hyperplanes of M. For a monoid
M contained in a group G and N Ç M we let MŒ÷N| denote the smallest sub-
monoid of G containing M and all the elements ÷x, x ÷ N. (If N = {x], then we
write MŒ÷x| instead of MŒ÷N|.) One calls MŒ÷N| the localization with respect
to N.
Proposition 2.31. Let M be a normal affine monoid with gp(M) = Z
d
, x ÷ M,
and H
1
. . . . . H
s
its support hyperplanes. Furthermore let F be the face of R
÷
M
with x ÷ int(F). Then
MŒ÷x| =
_
{H
÷
i
: x ÷ H
i
] ¨ Z
d
.
Furthermore MŒ÷x| splits into a direct sumL˚M
t
where L Š Z
e
, e = dimF.
If M is positive, then M
t
is positive.
Proof. Let N be the normal affine monoid on the right hand side. Clearly
MŒ÷x| Ç N, since M Ç N and ÷x ÷ N. Conversely suppose that y ÷ N. For all
the support forms o
j
with x … H
j
we have o
j
(x) > 0. Therefore o
j
(y ÷kx) ¸ 0
for all such j if k ;0. It follows that y ÷kx ÷ M for k ;0, and so y ÷ MŒ÷x|.
Altogether we have MŒ÷x| = N.
Let U = RF ¨ Z
d
. Then U is a direct summand of Z
d
, and, moreover, U Ç
MŒ÷x|. In fact, U Ç H
i
for each support hyperplane H
i
of M with x ÷ H
i
. By
the same argument as in the proof of Proposition 2.26, U splits off M.
For the last assertion one has to show that U is the group of units of MŒ÷x|
if M is positive. Let y ÷ MŒ÷x| be an invertible element, say y = y
t
÷ px with
y
t
÷ M and p ÷ Z. Then ÷y = y
tt
÷ qx, y
tt
÷ M, q ÷ Z. The equation
y
t
÷y
tt
÷ (p ÷q)x = 0 shows that y
t
= 0 if p ÷q ¸ 0, since y
t
is a unit in M in
this case. If p ÷q > 0, then both y
t
and y
tt
must belong to the face F, and we are
again done.
The conductor. Let M be an affine monoid with normalization
¨
M. Then the set
of “gaps”
¨
M \ M is, in a sense, small.
Proposition 2.32. Let M be an affine monoid. Then the ideal
c(
¨
M,M) = {x ÷ M : x ÷
¨
M Ç M]
is nonempty.
Moreover, the set
¨
M \ M is contained in finitely many hyperplanes parallel to
the facets of M.
62 2. Affine monoids and their Hilbert bases
Proof. We have seen in Corollary 2.8 that
¨
M is a finitely generated M-module.
Since
¨
M Ç gp(M), there exist elements z
1
. . . . . z
n
÷ gp(M) with
¨
M = (z
1
÷
M) L L (z
n
÷M). Write z
i
= y
i
÷ x
i
with x
i
. y
i
÷ M, i = 1. . . . . n, and set
x = x
1
÷ ÷x
n
. Then x ÷ c(
¨
M,M), and c(
¨
M,M) = 0.
Pick x ÷ c(
¨
M,M). If z ÷
¨
M\M, then there exists at least one support formo
i
of M with o
i
(z) < o
i
(x). So z belongs to one of the hyperplanes {y : o
i
(y) = k],
k ÷ Z
÷
, k < o
i
(x).
The common name for c(
¨
M,M) is conductor
2
. The reader may check that
c(
¨
M,M) is the largest ideal of M that is also an ideal of
¨
M (Exercise 2.8).
Our first corollary shows that monoid elements in the interior of R
÷
M are
nilpotent modulo any nonempty ideal of M. More precisely:
Corollary 2.33. Let M be an affine monoid, L a lattice containing gp(M), x ÷ M,
and y ÷
¯
M
L
¨ int(R
÷
M). Then ky ÷ x ÷int(M) for some k ÷ N.
If, in addition, y ÷ gp(M), then ky ÷ x ÷int(M) for k ;0.
Proof. Suppose first that y ÷ gp(M). Then y ÷
¨
M = gp(M) ¨ R
÷
M. Choose
z ÷ c(
¨
M,M). Then x ÷z ÷
¨
M Ç x ÷M, and after replacing x by x ÷z we are
left with the case in which M is normal. Since y ÷ int(R
÷
M) we have o
i
(y) > 0
for all support forms o
i
of M. Thus o
i
(ky) > o
i
(x) for all i and k ; 0. It follows
m
x
2x
m÷R
÷
M
Figure 2.2. Nilpotence of the interior modulo a nonempty ideal
that ky ÷ x ÷ int(R
÷
M) ¨ gp(M) = M and therefore ky ÷ x ÷int(M).
In the general case there exists q ÷ N with qy ÷ M for some q ÷ N. Now we
apply to qy what has just been shown.
One can investigate the set
¨
M \ M in more detail. The next proposition can be
derived rather easily from standard facts of commutative algebra. We will obtain
it as a consequence a more general result on modules over affine monoids; see
Proposition 4.32.
Proposition 2.34. Let M be an affine monoid. Then:
(a) Rad
_
c(
¨
M,M)
_
is the set of all x ÷ M such that MŒ÷x| is normal.
2
In the classical German number theory literature, the conductor was called “Führer” and de-
noted by f; however, after a certain period in history, the use of the letter f has become less popular.
2.B. Normal affine monoids 63
(b)
¨
M \ M is the union of a finite family of sets x ÷(F ¨ M) where x ÷ M
and F is a face such that F ¨ c(
¨
M,M) = 0. Moreover, if F is maximal
among these faces, then at least one set of type x ÷(F ¨ M) must appear.
Unions of normal monoids. A monoid M that is a union of normal submonoids
M
i
with gp(M
i
) = gp(M) is evidently itself normal. If M is affine and only finitely
many submonoids are brought into play, then a much stronger statement is possi-
ble:
Theorem2.35. Let M be an affine monoid and M
1
. . . . . M
n
submonoids of M such
that M = M
1
L L M
n
and M
i
is normal whenever rank M
i
= rank M. Then
M is normal, and it is the union of those M
i
for which rank M
i
= rank M.
Proof. We can assume that gp(M) = Z
d
. It is our first goal to reduce the theorem
to its special case in which R
÷
M
i
= R
÷
M for all i with rank M
i
= d. To this end
we consider all hyperplanes H in R
d
that support one of the cones C
i
= R
÷
M
i
with rank M
i
= d. This set of hyperplanes induces a dissection 1 of the cone
C = R
÷
M into rational subcones (see Proposition 1.75). Each cone D ÷ 1 is
generated by D ¨ M, and M =
_
{D ¨ M : D ÷ 1. dimD = d].
Furthermore, if D ÷ 1 with dimD = d, then gp(D ¨ M) = Z
d
by Corollary
2.25. So the normality of M follows fromthe normality of the intersections D¨M.
Now we can consider all the monoids D ¨ M, D ÷ 1 , dimD = d, and their
submonoids D ¨ M
i
separately. By construction of 1 we may therefore assume
that R
÷
M
i
= R
÷
M for all submonoids M
i
with rank M
i
= d.
Renumbering the M
i
if necessary we can further assume that rank M
i
= d
for i = 1. . . . . p and rank M
i
< d for i > p. Set G
i
= gp(M
i
). If our claim
does not hold, then we can find x ÷
¨
M \
_
p
i=1
M
i
=
¨
M \
_
p
i=1
G
i
. Set E =
(x ÷
_
p
i=1
G
i
) ¨C. Then, on the one hand, E Ç
¨
M \
_
p
i=1
G
i
. On the other hand,
E is not contained in the union of finitely many hyperplanes. However,
¨
M\M (by
Proposition 2.32) and each submonoid M
i
with rank M
i
< d is contained in such
a union, and the conclusion is that E contains elements of M \
_
n
i=1
M
i
. This is a
contradiction.
Seminormal monoids. A property close to normality and of importance in K-
theory (see Chapter 8) is seminormality:
Definition 2.36. A monoid is seminormal if every element x ÷ gp(M) with
2x. 3x ÷ M (and therefore mx ÷ M for m ÷ Z
÷
, m ¸ 2) is itself in M. The
seminormalization sn(M) of M is the intersection of all seminormal submonoids
of gp(M) containing M.
It follows immediately from Corollary 2.8 that the seminormalization sn(M) of
an affine monoid M is also affine; in fact, it is contained in the normalization
¨
M,
and therefore R
÷
sn(M) = R
÷
M is a cone.
64 2. Affine monoids and their Hilbert bases
A normal monoid is obviously seminormal, but the converse does not hold.
There even exist seminormal, nonnormal polytopal monoids. We will give an ex-
ample in Remark 2.53(b). For affine monoids M, the relationship between nor-
mality and seminormality is made precise by the next proposition. Recall that
M
+
= int(M) L {0] (Remark 2.4).
Proposition 2.37. An affine monoid M is seminormal if and only if (M ¨ F)
+
is
a normal monoid for every face F of R
÷
M. In particular, M
+
=
¨
M
+
if M is
seminormal.
Proof. Suppose that M is seminormal. Then M ¨ F is obviously seminormal
for each face F of R
÷
M. Therefore it is enough to show that M
+
is normal. By
Corollary 2.25, gp(M
+
) = gp(M). Let x ÷ gp(M
+
), x = 0, with ax ÷ M
+
for
some a ÷ N. Then x ÷ int(R
÷
M) and, by Corollary 2.33, kx ÷ M for all k ; 0.
Let mbe the largest integer for which mx … M. If m > 0, we have 2mx. 3mx ÷ M,
and so mx ÷ M by seminormality, an obvious contradiction.
For the converse implication let x ÷ gp(M) be such that 2x. 3x ÷ M. Let F be
the unique face of R
÷
M with 2x ÷ int(F). Then certainly 3x ÷ int(F), too, and
so x ÷ gp
_
int(F) ¨ M
_
. By hypothesis x ÷ int(F) ¨ M.
If M is seminormal, then M
+
is normal, as just seen, and the normality of M
+
implies M
+
=
¨
M
+
.
While M
+
is almost never finitely generated, it is the filtered union of affine
submonoids, and if M is seminormal, these can be chosen to be normal. To be
more precise: there exists a family M
i
of affine submonoids, indexed by the ele-
ments of a set I such that (i) M =
_
i÷I
M
i
, and (ii) for all i. j ÷ I there exists
k ÷ I such that M
i
. M
j
Ç M
k
. In such a situation we will simply speak of a filtered
union. For simplicity we restrict ourselves to positive affine monoids.
Proposition 2.38. Let M be a positive affine monoid. Then M
+
is the filtered union
of affine submonoids. If M
+
is normal, then these submonoids can be chosen to be
normal.
Proof. We embed M into Z
r
= gp(M), choose a rational cross-section P of
R
÷
M, and a rational point z ÷ int(P). In the affine space aff(P) we consider the
homothety 0
z
with center z and factor z ÷ (0. 1) ¨ Q. Set M
z
= M ¨ R
÷
0
z
(P).
Then M
+
is the filtered union of the affine monoids M
z
, and M
z
is normal if M
+
is. We leave the detailed proof of this claim to the reader (Exercise 2.11).
We want to give another characterization of seminormal monoids that will turn
out useful in proving the seminormality of their associated algebras and is similar
in spirit to the characterization of normal affine monoids in Corollary 2.24.
Proposition 2.39. Let M Ç Z
r
be a monoid with gp(M) = Z
r
. Then the following
are equivalent:
(a) M is seminormal affine;
2.C. Generating normal affine monoids 65
(b) there exists finitely many rational halfspaces H
÷
i
and subgroups U
i
Ç H
i
¨
Z
r
such that rank U
i
= r ÷ 1 and
M =
_
i
U
i
L (H
>
i
¨ Z
r
).
We leave the detailed proof to the reader (Exercise 2.12), together with some
related results. However, we want to indicate the construction of the halfspaces and
the subgroups U
i
: for each face F we choose a rational hyperplane H
i
intersecting
R
÷
M exactly in F
i
, and set U
i
= gp(M ¨ F) ˚ V
i
where V
i
is a subgroup of
H
i
¨ Z
r
such that rank V
i
= r ÷ 1 ÷ rank U
i
and U
i
¨ V
i
= 0. Note that the
proposition includes the statement that the monoids U
i
L (H
>
i
¨ Z
r
) are affine.
2.C. Generating normal affine monoids
Let C Ç R
d
be a pointed rational cone, and L a sublattice of Q
d
. Then C ¨ L
is integrally closed submonoid of L. In this section we study the Hilbert basis of
C ¨ L. It is no restriction to assume that dimC = rank L. Otherwise we can
replace C by the cone C ¨ RL. However, we do not insist that d = dimC.
Simplicial cones and multiplicities. Let :
1
. . . . . :
r
be linearly independent vectors
in the lattice L. By
par(:
1
. . . . . :
r
) =
˚
q
1
:
1
÷ ÷q
r
:
r
: 0 ¸ q
i
< 1. i = 1. . . . . n
_
we denote the semi-open parallelotope spanned by :
1
. . . . . :
r
. Let U denote the
sublattice generated by :
1
. . . . . :
r
. It is clear that every residue class x ÷U for an
element x in the saturation
¨
U
L
= QU ¨ L = RU ¨ L of U in L has a unique
representant in par(:
1
. . . . . :
r
), namely
x
t
= (a
1
÷ ]a
1
¡):
1
÷ ÷(a
r
÷ ]a
r
¡):
r
.
where x =

r
i=1
a
i
:
i
, a
i
÷ Q. Furthermore, if x ÷ C ¨ L where C is the cone
generated by :
1
. . . . . :
r
, then a
1
. . . . . a
r
¸ 0, and x ÷ x
t
÷M where M = Z
÷
:
1
÷
÷Z
÷
:
r
. We have thus proved
Proposition 2.40. With the notation just introduced, the following hold:
(a) E = L¨par(:
1
. . . . . :
r
) is a system of generators of the M-module C ¨L;
(b) (x ÷M) ¨ (y ÷M) = 0 for x. y ÷ E, x = y;
(c) #E = ŒQU ¨ L : U|;
(d) Hilb(C ¨ L) Ç {:
1
. . . . . :
r
] L E.
In part (c) ŒQU ¨L : U| is the index of the subgroup U in the group QU ¨L.
The proposition can be interpreted as saying that C ¨ L is a free module over M
with basis E. The rank of this free module, namely ŒQU ¨ L : U|, measures how
far U is fromQU ¨ L. Before continuing the main theme, namely the generation
of normal affine monoids, we want to transfer this measure to lattice simplices,
which, unlike the simplex conv(0. :
1
. . . . . :
r
) in Proposition 2.40, may not have 0
among their vertices.
66 2. Affine monoids and their Hilbert bases
Let L = x ÷L
0
be an affine lattice with associated lattice L
0
and U an affine
sublattice, say U = y ÷ U
0
. Then we have y ÷ L and U
0
Ç L
0
. Then we set
L,U = L
0
,U
0
. Since L
0
and U
0
do not depend on the choice of x and y, this
definition is justified. However, it is reasonable beyond the formal justification: L
decomposes into a disjoint union of translates of V ÷y where V runs through the
cosets of U
0
in L
0
. Consequently we define the index of U in Lby
ŒL : U| = ŒL
0
: U
0
|.
Let L be an affine lattice and z an L-simplex (see Definition 1.79). Then the
roles of the lattices compared in Proposition 2.40(c) are played by the lattices in-
troduced after Definition 1.79, namely
L
z
= L¨ aff(z)
and its affine sublattice
L(z) = :
0
÷

¡÷vert(z)
Z(: ÷ :
0
)
where :
0
is an arbitrary vertex of z.
Definition 2.41. The multiplicity of z (with respect to L) is the index j
L
(z) =
ŒL
z
: L(z)|.
The multiplicity has a very useful geometric interpretation that will allow us to
define it for an arbitrary lattice polytope. In order to simplify the discussion we
choose the origin in a vertex :
0
of z so that we are back in the situation of Propo-
sition 2.40 and L
z
and L(z) are lattices of the same rank r. By the elementary di-
visor theorem we can find a basis e
1
. . . . . e
r
of L
z
and integers d
1
. . . . . d
r
¸ 1 such
that d
1
e
1
. . . . . d
r
e
r
is a basis of L(z). Then 2.40(c) shows that j
L
(z) = d
1
d
r
.
On V = RL
z
we introduce the r-dimensional volume function that gives volume
1 to the parallelotope spanned by e
1
. . . . . e
r
. This volume function depends only
on L
z
. We call it the L-volume vol
L
on V . In fact, every change of basis in L
z
is by an integral matrix of determinant ˙1. It follows that j
L
(z) is exactly the
L-volume of the parallelotope P spanned by :
1
. . . . . :
r
. As we will see in Section
3.B, the parallelotope P decomposes into r! simplices that have the same volume
as z:
Corollary 2.42.
j
L
(z) = r! vol
L
(z).
We can also pass to the monoid over the L-simplex z Ç R
d
and use it to
measure the multiplicity of z. If :
0
. . . . . :
r
are the vertices of z, then M(z) is
generated by the vectors :
t
i
= (:
i
. 1) ÷ R
d÷1
, i = 0. . . . . r. Together with 0 they
span again a simplex z
t
. Its multiplicity must be taken with respect to the lattice
over L, namely the sublattice L
t
of R
d÷1
generated by the vectors (x. 1), x ÷ L.
We will write L
t
= Z(L. 1). (If 0 ÷ L, then L
t
= L˚Z.)
Proposition 2.43. With the notation introduced, we have
j
L
(z) = j
L
0 (z
t
).
2.C. Generating normal affine monoids 67
Proof. We can assume that :
0
= 0 since all data are invariant with respect to the
choice of origin in the affine space containing z. Next we are allowed to assume
that aff(z) = R
r
, and finally that L = Z
r
. Then the L-volume is just the Euclidean
volume on R
r
, and L
t
= Z
r÷1
. Clearly
j(z) = r! vol(z) = [ det(:
1
. . . . . :
r
)[
= [ det(:
t
0
. . . . . :
t
r
)[ = (r ÷1)! vol(z
t
) = j(z
t
). (2.2)
Alternatively, one establishes a bijection between E = par(:
1
. . . . . :
r
) ¨ Z
r
and E
t
= par(:
t
0
. . . . . :
t
r
) ¨ Z
r÷1
using :
t
0
= (0. 1) to “lift” the elements of E to
their proper heights in E
t
.
Example 2.44. Let z Ç R
n
be the convex hull of the vectors x
0
= 0. x
1
=
e
1
. . . . . x
n-1
= e
n-1
(here e
i
is the i th unit vector) and x
n
= (1. . . . . 1. n). Then we
have j(z) = n (with respect to L = Z
n
), as formula (2.2) shows. We consider the
simplicial cone generated by x
t
i
= (x
i
. 1). Then par(x
t
1
. . . . . x
t
n
) contains exactly n
lattice vectors, namely 0 ÷ R
n÷1
and
(1. . . . . 1. a. n ÷ a). 1 ¸ a ¸ n,2.
(1. . . . . 1. a. n ÷ a ÷1). n,2 < a < n.
The reader should compute this example, including all details. In Example 2.55 it
will be used again.
In the situation of Proposition 2.40 the monoid C ¨Ldepends only on C (and
L), and if one wants to determine Hilb(C ¨ L) it is certainly advisable to take
:
1
. . . . . :
r
in such a way that #E becomes as small as possible. There is a unique
such choice: for each extreme ray R of (an arbitrary rational cone) C the monoid
R¨ L is normal and of rank 1. Therefore it is generated by a single element e; we
call e an extreme L-generator of C, or an extreme integral generator if L = Z
d
.
It is obvious that the extreme L-generators are irreducible elements of C ¨ L and
therefore elements of the Hilbert basis.
Definition 2.45. Let C be a simplicial rational cone. The simplex z
L
(C) with ver-
tices in 0 and the extreme L-generators of C is called the basic L-simplex of C.
The L-multiplicity of C is j
L
(C) = j
L
(z
L
(C)).
Simplices and simplicial cones with the smallest possible multiplicity have a
special name:
Definition 2.46. An L-simplex z is unimodular if j
L
(z) = 1. A simplicial ratio-
nal cone C is unimodular (with respect to L) if z
L
(C) is unimodular.
Proposition 2.47. Let z = conv(x
0
. . . . . x
r
) be an L-simplex where L Ç R
d
is an
affine lattice, and let L
t
= Z(L. 1). Then the following are equivalent:
(a) z is unimodular;
(b) the sublattice generated by x
1
÷ x
0
. . . . . x
r
÷ x
0
is a direct summand of
L÷ x
0
;
(c) the submonoid of L
t
generated by x
t
i
= (x
i
. 1), i = 0. . . . . r, is integrally
closed in L
t
;
68 2. Affine monoids and their Hilbert bases
(d) the cone C(z) is unimodular with respect to L
t
.
Cones over polytopes. We generalize Proposition 2.40 in two steps, first to cones
C(P) where P is an L-polytope, and later on to arbitrary rational cones.
There is a simple strategy how to make Proposition 2.40 useful for C(P): we
triangulate P, consider the simplicial cone over each simplex and unite the systems
of generators obtained in this way. Let us first introduce some special types of
triangulations:
Definition 2.48. An L-polytope P is empty (with respect to L) if it contains no
elements from L other than its vertices. A triangulation ˙ of an L-polytope P is
full (with respect to L) if vert(˙) = lat(P), or, equivalently, all its simplices are
empty.
A triangulation ˙ of an L-polytope is unimodular (with respect to L) if all
simplices in ˙ are unimodular.
It is immediately observed that every unimodular triangulation is full: a uni-
modular L-simplex contains no point of L different from its vertices, and every
point x ÷ P ¨ Lmust be contained in one of the simplices of ˙.
The natural degree on Z(L. 1) Ç R
d÷1
associated with a lattice L Ç R
d
is
given by the last coordinate. If P is an L-polytope, then the natural generators of
the polytopal monoid M(P) have degree 1. (Polytopal monoids have been intro-
duced in Definition 2.16.)
Proposition 2.49. Let P be an L-polytope and C = C(P). Then the integral closure
¯
M(P) = C ¨ Z(L. 1) of M(P) in Z(L. 1) is generated as an M(P)-module by
elements of degree ¸ dimP ÷ 1.
Proof. We choose a full triangulation ˙ of P. For each simplex z ÷ ˙ we ap-
ply Proposition 2.40 to the cone C(z). The union of the corresponding sets E
z
evidently generates
¯
M(P) =
_
z
¯
M(z) over M(P).
It remains to show that each element y ÷ E
z
has degree at most dimP ÷ 1.
Let x
1
. . . . . x
r
, r = dimP ÷1, be the vertices of z. Then y = q
1
x
t
1
÷ ÷q
r
x
t
r
,
0 ¸ q
i
< 1, (x
t
i
= (x
i
. 1), i = 1. . . . . r. Clearly
deg y = q
1
÷ ÷q
r
< r = dimP ÷1
and if we had chosen a “coarse” triangulation of zwe may not be able to reach the
bound dimP ÷ 1. However, the basic simplices of the cones C(z) are empty. If
deg y = dimP, then y
t
= (x
t
1
÷ ÷x
t
r
) ÷y has degree 1 and belongs to the basic
simplex of C(z) (more precisely, to its face opposite of 0). This is a contradiction,
since y
t
= x
t
i
for all i .
Remark 2.50. In the proof of Proposition 2.49 we have constructed a system of
generators of
¯
M(P) as the union of the sets E
z
. Note that

z
#E
z
=

z
vol
L
(z)(dimP)! = vol
L
(P)(dimP)!
2.C. Generating normal affine monoids 69
is independent of the triangulation. We call j
L
(P) = vol
L
(P)(dimP)! the multi-
plicity of P (with respect to L).
The sets E
z
are not disjoint (except if P = z) since 0 appears in each E
z
,
and within the category of monoids we cannot interpret j
L
(P) as the rank of a
module. However, j
L
(P) is the rank of the algebra KŒP| (Ka field) over a Noether
normalization. Therefore j
L
(P) is the multiplicity of KŒP| as a graded K-algebra.
In the simplicial case, the Noether normalization is a “discrete” object, namely
KŒx
t
1
. . . . . x
t
r
| Ç KŒP| where x
1
. . . . . x
r
are the vertices of P.
Proposition 2.49 has a strong consequence for dimP ¸ 2:
Corollary 2.51. Let P be an L-polytope of dimension ¸ 2. Then M(P) is integrally
closed in Z(L. 1).
An empty L-simplex of dimension ¸ 2 is unimodular, and every full triangula-
tion of an L-polytope of dimension ¸ 2 is unimodular.
That empty 2-simplices are unimodular is the essential argument in the proof
of Pick’s formula.
Remark 2.52. In dimension d ¸ 3 an empty lattice simplex need not be unimodu-
lar. For example, let d = 3 and consider the simplex
z
pq
= conv
_
(0. 0. 0). (0. 1. 0). (0. 0. 1). (p. q. 1)
_
with coprime integers 0 < q < p. Then z
pq
is empty (with respect to Z
3
), but
has multiplicity p. It has been proved by White [101] that every empty 3-simplex
is isomorphic to z
pq
for some p and q, and that z
pq
Š z

if and only if p = u
(an obviously necessary condition) and : = q or : = p ÷ q.
A similar classification of empty simplices in higher dimension is unknown.
The classification in dimension 3 follows readily if one has shown that there exists
a Z-linear form˛ and an integer msuch that zis sitting between the hyperplanes
given by ˛(x) = m and ˛(x) = m ÷ 1: z has lattice width 1. This is no longer
true in dimension > 3. See [48] and [73] for a discussion of empty simplices and
further references.
Example 2.53. (a) Let z
pq
be the simplex introduced in the preceding remark, p >
1. Then the basic simplex of C(z
pq
) is not unimodular, and the monoid M(z
pq
)
is generated by the vectors (x
i
. 1) where x
i
, i = 1. . . . . 4, runs through the vertices
of z
pq
. By Proposition 2.47 M(z
pq
) is not integrally closed in Z
4
. Thus M(P)
need not be integrally closed if dimP ¸ 3. However, note that z
pq
, being empty,
is unimodular with respect to L(z
pq
). Therefore M(z
pq
) is normal.
(b) Let :
1
. . . . . :
4
be the vertices of z = z
21
and n = (:
1
÷ ÷ :
4
),2.
Furthermore we set u
ij
= (:
i
÷ :
j
), 1 ¸ i ¸ j ¸ 4. Then the u
ij
and n are
the lattice points of 2z, an integrally closed polytope according to Corollary 2.54
below. Now we form a 4-dimensional polytope P as the convex hull of F
0
= (z. 0)
and F
1
= (2z. 1). So z and 2z are two parallel facets of P.
Set M = M(P). A run of normaliz shows that Hilb(
¯
M) consists of the 15
elements corresponding to the lattice points of P, namely (:
i
. 0. 1), (u
ij
. 1. 1),
(n. 1. 1), and one more additional element (n. 0. 2) … (P. 1). (This element must
70 2. Affine monoids and their Hilbert bases
appear since M(z) is not integrally closed.) We claim that M is seminormal (see
Definition 2.36) but not normal. In fact, suppose that 2x. 3x ÷ M. Then x ÷
¯
M,
and so x can be represented by the 16 elements of Hilb(
¯
M). If (n. 0. 2) appears
with the coefficient 0, then x ÷ M. If 2x. 3x in
¯
M(F
0
), then x ÷ gp(M(F
0
)),
and so x ÷ M(F
0
) Ç M(P) since M(F
0
) is normal (albeit not integrally closed).
The only possibility remaining is that (n. 0. 2) and at least one of the generators
(u
ij
. 1. 1) or (n. 1. 1) of M(F
1
) appear in the presentation of x with positive coef-
ficients. But since (n. 0. 2) ÷(u
ij
. 1. 1) ÷ M(P) as well as (n. 0. 2) ÷(n. 1. 1) ÷
M(P) (as the reader may check), we finally conclude that x ÷ M(P). Therefore
the polytopal monoid M is seminormal. It is not normal since gp(M) = Z
4
, and
so M =
¯
M = M.
(c) There exist 3-dimensional non-(semi)normal polytopes P. Let
P = {x ÷ R
3
÷
: x
1
,2 ÷x
2
,3 ÷x
3
,5 ¸ 1].
Then P is a simplex with vertices (0. 0. 0), (2. 0. 0), (0. 3. 0), and (0. 0. 5). Obvi-
ously Z
3
is the smallest lattice containing the vertices of P. The monoid M =
M(P) is not normal: (1. 2. 4. 2) belongs to C(P), but not to M. Since normality is
violated in M
+
, M is not even seminormal (see Proposition 2.37).
We refer the reader to [18] for a detailed discussion of the normality of sim-
plices like P.
Nevertheless, Corollary 2.51 can be generalized if we consider multiples of lat-
tice polytopes. (The lattice associated with a multiple cP, c ÷ N, is defined in
Remark 1.82.)
Corollary 2.54. Let P be an L-polytope and : ÷ vert(P). Then M(cP) is integrally
closed in Z((c ÷ 1): ÷L. 1) for c ÷ N, c ¸ dimP ÷ 1.
Proof. After a parallel translation of P we can assume that : = 0. Then the
surrounding lattice is simply L˚Z.
The integral closure of M(cP) in L˚Zis isomorphic to the intersection of the
integral closure of M(P) with L˚cZ. Under the isomorphismL˚Z ÷L˚cZ,
(x. h) ÷ (x. ch) the set of generators E(cP) of M(cP) is identified with the set
of degree c elements in
¯
M(P). Therefore we must show that every element x in
¯
M(P) of degree nc, n ÷ Z
÷
, is the sumof elements of degree c. There is nothing to
show for n ¸ 1. Suppose that n ¸ 2. Then, by Proposition 2.49, x = x
t
÷x
tt
where
x
t
÷ M(P) and deg x
tt
¸ c. Since x
t
decomposes into a sum of degree 1 elements,
we can modify the representation of x to x = x
1
÷x
2
where x
2
has degree c. The
summand x
1
can be “subdivided” into a sum of degree c elements.
The proof shows a slightly stronger statement than stated in Corollary 2.54:
M(cP) is generated by elements in E(cP) as a module over its submonoid M(P)
c
,
the “Veronese” submonoid of M(P) generated by all elements (x
1
÷ ÷x
c
. 1),
x
i
÷ P, i = 1. . . . . c.
Example 2.55. Suppose that P ÷ R
n
is a lattice polytope with z = 0 ÷ vert(P).
Then M(P)Œ÷(z. 1)| = M
t
˚ Z where M
t
is the submonoid of Z
n
generated by
2.C. Generating normal affine monoids 71
the lattice points in P (considered as vectors). If M(P) is integrally closed in Z
n÷1
,
then the monoid M
t
is integrally closed in Z
n
. In other words, the monoid C ¨Z
n
,
C = R
÷
P, is generated by the vectors x ÷ C ¨ P.
We apply this argument to P = cz where z is the simplex constructed in
Example 2.44, n ¸ 3, c ¸ n ÷ 2. The vector n = (1. . . . . 1) belongs to C, but it is
not contained in M
t
. First, it is not contained in P, since one has

a
i
= n ÷ 1
in a representation n = a
1
x
1
÷ ÷ a
n
x
n
(take a = 1 in 2.44), and, moreover,
every lattice vector in C with last coordinate 1 has all its coordinates positive: n is
irreducible in C ¨ Z
n
.
This example shows that Proposition 2.49 or Corollary 2.54 cannot be im-
proved. It is not even possible to improve the bound dimP ÷ 1 if one replaces
“integrally closed in L ˚Z” by “normal”. In fact, gp(M(cz)) = Z
n÷1
for c ¸ 2,
as the reader may show (take a = n ÷ 1 in 2.44).
We transfer the attributes “integrally closed” and “normal” from monoids to
polytopes:
Definition 2.56. Let P be an L-polytope. We say that P is L-integrally closed if
M(P) is integrally closed in Z(L. 1) and that it is normal if M(P) is normal.
The integral closedness of 2-dimensional lattice polytope P has been derived
from the fact that the simplices in a full triangulation of P are unimodular. In the
next section we will discuss to what extent this argument can be reversed: does
the integral closedness of P imply the existence of a triangulation into unimodular
simplices? What can be said about cP in this respect?
Pointed rational cones. It is not difficult to generalize the arguments above to ar-
bitrary pointed rational cones. First we have to find a replacement for the polytope
P, or rather (P. 1), spanning the cone. Now the notion of bottom introduced in
Section 1.C is very useful. We set
C
t
L
= conv(x ÷ L¨ C. x = 0].
Then C
t
L
= conv(Hilb(C ¨ L)) ÷C, and so C
t
L
is a polyhedron by Theorem 1.32.
Its bottom B
L
(C) is the union of its bounded faces (since C does not contain a
line). Moreover, the line segment connecting 0 with a point in C
t
L
intersects B
L
(C)
(all the nonbounded faces are faces of C). Let F be a facet of B
L
(C) and H
0
the
hyperplane through 0 and parallel to F. Then the L-height above H
0
is a linear
form; we denote it by ˇ
F
and call it a basic L-grading of C; the basic F-degree of
C is the constant value b
F
= ˇ
F
(x), x ÷ aff(F).
Example 2.57. Let C Ç R
3
be generated by x
1
= (0. 1. 0), x
2
= (0. 0. 1), and
x
3
= (2. 1. 1). Together with (0. 0. 0) these vectors span the empty simplex z
21
of multiplicity 2 (see Remark 2.71). Thus the simplicial cone C has a single basic
grading, corresponding to the single facet F of the bottom, the triangle spanned
by x
1
. x
2
. x
3
. As the reader may check, one has b
F
= 2.
One needs one more element for Hilb(C), namely y = (1. 1. 1): the cones
generated by y and two of x
1
. x
2
. x
3
are unimodular so that C is the union of 3
unimodular cones.
72 2. Affine monoids and their Hilbert bases
C
t
L
Figure 2.3. The bottom
Proposition 2.58. Let C Ç R
d
be a pointed rational cone of dimension d, La lattice
in Q
d
. Furthermore, let B = B
L
(C) ¨ Land M = Z
÷
B. Then:
(a) if x ÷ C ¨ L, x = 0, and there exists a facet F of B
L
(C) with ˇ
F
(x) <
2b
F
, then x ÷ Hilb(C); in particular B Ç Hilb(C);
(b) for each element y in the minimal system of generators G of C ¨ L as an
M-module there exists a facet F of B
L
(C) such that ˇ
F
(y) < (d ÷ 1)b
F
.
Proof. For (a) it is enough to note that an element x ÷ C ¨ L, x = 0, with
ˇ
F
(x) < 2b
F
must be irreducible since b
F
is the minimal value of ˇ
F
on C
t
L
.
For (b) we choose a triangulation ˙ of C whose rays constitute the set {R
÷
x :
x ÷ B] (Theorem 1.61). Let D ÷ ˙ be a simplicial cone. Then the basic L-simplex
z
L
(D) is empty. As in the proof of Proposition 2.49 (where b
F
= 1) we obtain that
D ¨ L is generated by elements y with ˇ
F
(y) < (d ÷ 1)b
F
as a module over the
monoid spanned by D¨B. It only remains to unite all these vectors y to a system
of generators of C ¨ Lover B.
In dimension 2, the lattice points in the bottom form the Hilbert basis be-
cause all the cones over the segments of the unique triangulation ˙ of B
L
(C) with
vert(˙) = B
L
(C) ¨ Lare unimodular by Corollary 2.51:
Corollary 2.59. If dimC = 2, then Hilb(C ¨ L) = B
L
(C) ¨ L.
In dimension 3, B
L
(C) ¨Lneed not be the Hilbert basis of C ¨L, as Example
2.57 shows. Nevertheless one can easily describe the Hilbert basis: Hilb(C ¨ L)
consists exactly of B = B
L
(C) ¨ L and the nonzero elements in the minimal
system of generators of C ¨Las a module over Z
÷
B, as follows immediately from
Proposition 2.58. For later application we record
Corollary 2.60. Let C be a cone of dimension 3, and D Ç C a rational subcone.
(a) If B
L
(D) is contained in B
L
(C), then Hilb(D ¨ L) = D ¨ Hilb(C ¨ L).
(b) If D is generated by elements :
1
. :
2
. :
3
÷ B
L
(C) ¨ L that span an empty
simplex in a facet F of B
L
(C), then
Hilb(D ¨ L) = {:
1
. :
2
. :
3
] L
_
par(:
1
. :
2
. :
3
) ¨ L\ {0]
_
.
2.C. Generating normal affine monoids 73
Proof. (a) An element of D¨Lthat is irreducible in C ¨L, is certainly irreducible
in D ¨ L. This shows the containment ¸, which holds whenever D Ç C.
If dimD ¸ 2, then the converse containment follows from the previous corol-
lary. So let dimD = 3 and consider x ÷ Hilb(D ¨ L). Then there exists a basic
grading ˇ
F
of D with ˇ
F
(x) < 2b
F
. But ˇ
F
is a basic grading of C as well, and so
x ÷ Hilb(C).
(b) This follows from the argument in the proof of the proposition: ˇ
F
(x) <
2b
F
for each x ÷ par(:
1
. :
2
. :
3
) ¨ L.
We can give up some precision in measuring the size of the elements of Hilb(C),
using only the extreme L-generators of C. Since the bottom B
L
(C) is contained
in the convex hull of 0 and the extreme generators, we obtain
Corollary 2.61. Let X be the set of extreme L-generators of the cone C. Then
Hilb(C ¨ L) Ç (d ÷ 1) conv(0. X) if d = dimC ¸ 2.
Carathéodory rank. So far we have tried to bound the systems of generators or the
Hilbert basis in terms of degree. Another type of measure is given by the minimal
number of elements in the Hilbert basis of a monoid M that, given x ÷ M, are
needed to represent x:
Definition 2.62. Let M be a positive affine monoid. The representation length ,(x)
of x ÷ M is the smallest number k of elements x
1
. . . . . x
k
÷ Hilb(M) such that x
is a Z
÷
-linear combination of x
1
. . . . . x
k
.
The Carathéodory rank of M, CR(M), is the maximum of the representation
lengths ,(x), x ÷ M.
Since M is finitely generated, CR(M) is finite, namely CR(M) ¸ # Hilb(M).
Without further hypothesis a better bound for CR(M) is impossible.
Example 2.63. Let p
1
. . . . . p
n
pairwise different prime numbers, and set q
i
=

jyi
p
j
. Then q
1
. . . . . q
n
are coprime, and every m ÷ Z, m ; 0 is in the ad-
ditive monoid generated by q
1
. . . . . q
n
. However, unless mis divisible by one of the
prime numbers p
i
, all the generators q
1
. . . . . q
n
appear in every presentation of m.
If M is normal, then one can bound Carathéodory rank by a linear function of
rank:
Theorem 2.64 (Seb˝ o). Let M be a positive normal affine monoid of rank r. Then
CR(M) = r if r ¸ 3, and r ¸ CR(M) ¸ 2r ÷ 2 if r ¸ 4.
Proof. We identify gp(M) with Z
r
and set C = R
÷
M.
It is easy to see that CR(M) ¸ r. In fact, the elements of M that are Z-linear
combinations of at most r ÷ 1 elements of Hilb(M) are contained in the union of
finitely many hyperplanes in R
r
, and such a union cannot contain M.
If r = 1, then M Š Z
÷
, and there is nothing to show. If r = 2, then the
bottomB of C
t
(constructed with respect to L = Z
2
) decomposes into a sequence
of line segments Œx. y| such that (x. y) ¨ Z
2
= 0. As we have seen in the proof of
74 2. Affine monoids and their Hilbert bases
Proposition 2.58, every lattice point in the cone generated by Œx. y| belongs to the
monoid generated by x and y. The case r = 3 will be discussed in Theorem 2.70.
So let r ¸ 4 and x ÷ M. As in the proof of 2.58 (and the case r = 2 above) we
choose a triangulation of C whose rays pass through all lattice points in the bottom
B
L
(C). The line segment Œ0. x| intersects a facet F of the bottom, and we have seen
that we can write x = x
t
÷x
tt
where (i) x
t
is a linear combination of those (at most
r) elements of B ¨ Z
r
that generate the smallest cone in ˙ containing x, and (ii)
ˇ
F
(x
t
) ¸ rb
F
÷ 2. Therefore x
tt
can be written as a Z
÷
-linear combination of at
most r ÷ 2 elements in Hilb(M).
We will continue the discussion of Carathéodory rank in the next section. At
this point it is not easy to find monoids M as in the theorem with CR(M) >
rank M.
Remark 2.65. In [17] two variants of CR have been discussed, asymptotic and vir-
tual Carathéodory rank. The asymptotic Carathéodory rank CR
a
(M) is the small-
est number msuch that the proportion
#{x ÷ M. [x[ ¸ t : ,(x) > m]
#{x ÷ M. [x[ ¸ t ]
goes to 0 with t ÷ o. The virtual Carathéodory rank CR
v
(M) is the smallest
number msuch that ,(x) ¸ mwith only finitely many exceptions.
One can show that CR
a
(M) ¸ 2 rank(M) ÷3 if M is a normal monoid of rank
¸ 3. For a proof and further results we refer the reader to [17].
2.D. Normality and unimodular covering
In the previous sections we have used triangulations in order to construct systems
of generators for normal monoids. In this section we will discuss the existence
of triangulations into, and more generally, of covers by unimodular simplices and
unimodular cones. While the notions to be introduced below make sense with
respect to arbitrary lattice structures, we will work only with the lattice Z
d
in this
section. This does not restrict the generality in an essential way, but simplifies the
formulations somewhat. Consequently, we set Hilb(C) = Hilb(C ¨ Z
d
) if C is a
pointed rational cone.
Definition 2.66. Let C be a pointed rational cone. A triangulation ˙ of C is called
unimodular (with respect to Z
d
) if all simplicial cones D of ˙ are unimodular.
Theorem 2.67. Let M Ç Z
d
be a positive affine monoid. Then the following are
equivalent:
(a) M is integrally closed in Z
d
;
(b) there exists a triangulation ˙ of R
÷
M such that each D ÷ ˙ is a unimod-
ular simplicial cone generated by elements of M.
Proof. We start with an arbitrary triangulation ˙ of C = R
÷
M whose simpli-
cial cones are generated by elements of M. Such a triangulation exists according
2.D. Normality and unimodular covering 75
to Theorem 1.61. It remains to refine the triangulation if there exists a nonuni-
modular cone D ÷ ˙. We use induction on (i) the maximal multiplicity of a
d-dimensional cone D in ˙, and (ii) on the number of such D with maximal mul-
tiplicity.
Let m = max{j(D) : D ÷ ˙]. If m > 1, we choose D with j(D) = m. Let
y
1
. . . . . y
d
be the extreme integral generators of D. Since M = C ¨Z
d
the vectors
y
1
. . . . . y
d
belong to M. By Proposition 2.40 there exists x ÷ par(x
1
. . . . . x
d
)¨Z
d
.
Now we apply stellar subdivision by x to ˙. Figure 2.4 shows a typical situation
after 2 generations of subdivisions in the cross-section of a 3-cone.
Figure 2.4. Successive stellar triangulations
In this process every d-dimensional cone D ÷ ˙ with x ÷ D is replaced by
the union of subcones D
tt
= R
÷
x ÷D
t
, D
t
a facet of D with x … D
t
. Then there
exists j such that the y
i
with i = j generate D
t
. One has
j(D
tt
) ¸ [ det(y
1
. . . . . y
j-1
. x. y
j÷1
. . . . . y
d
)[ < [ det(y
1
. . . . . y
d
)[ = j(D)
since the coefficient of y
j
in the representation of x is in the interval Œ0. 1). Since at
least one D ÷ ˙ with j(D) = mis properly subdivided, we are done by induction.
The implication (b) ==(a) is a trivial special case of Theorem 2.35.
Despite the simplicity of its proof, Theorem 2.67 has very important applica-
tions as we will see later on.
Hilbert triangulations and covers. In the previous section we have seen much
stronger results for special types of cones. For example, if P is a 2-dimensional
lattice polytope, then C(P) has a unimodular triangulation with rays through the
Hilbert basis E(P) of M(P). The following definition covers this additional con-
dition.
Definition 2.68. We say that ˙ is a Hilbert triangulation of a pointed rational cone
C if each ray of ˙ is of type R
÷
x with x ÷ Hilb(C). It is a full Hilbert triangulation
if all rays R
÷
x with x ÷ Hilb(C) are rays of ˙.
A very strong unimodularity condition for Hilb(C) is: Every full Hilbert trian-
gulation of C is unimodular. All cones of dimension 2 and polytopal cones C(P)
with dimP = 2 have this property. Another class is given by the cones over direct
products of two unimodular simplices of arbitrary dimension (see [?]).
Remark 2.69. Relaxing the Hilbert condition, one may ask whether every positive
rational cone C contains a finite subset X of Z
d
such that each triangulation using
all the rays R
÷
x, x ÷ X, is unimodular. We leave it to the reader that X has this
property if and only if it is supernormal in the sense of [62]: D ¨ X contains a
76 2. Affine monoids and their Hilbert bases
Hilbert basis of D for every cone D generated by a subset X
t
of X. In general, a
cone C does not contain a supernormal subset; see [62] for a counterexample.
A substantially weaker and much more important property is
(UHT) C has a unimodular Hilbert triangulation.
Theorem 2.70 (Seb˝ o). Let C Ç R
3
be a positive rational cone. Then C has a uni-
modular Hilbert triangulation.
The crucial step in the proof is the following lemma.
Lemma 2.71. Let :
1
. :
2
. :
3
be vectors in Z
3
, for which conv(0. :
1
. :
2
. :
3
) is an empty
3-simplex. Furthermore let ˇ be the primitive linear form on Z
3
that has constant
value b > 0 on :
1
. :
2
. :
3
. Then:
(a) ˇ is the (unique) basic grading of C;
(b) Ker ˇ = Z(:
2
÷ :
1
) ÷Z(:
3
÷ :
1
);
(c) b = j(C);
(d) on par(:
1
. :
2
. :
3
) ¨ Z
3
the linear form ˇ takes exactly the values 0. b ÷
1. . . . . 2b ÷ 1.
Proof. (a) is clear from the choice of ˇ. For (b) we observe that Z(:
2
÷ :
1
) ÷
Z(:
3
÷ :
1
) Ç Ker ˇ. On the other hand 0, :
2
÷ :
1
, and :
3
÷ :
1
are the vertices of
an empty lattice triangle. Such a triangle is unimodular, or, in other words, :
2
÷:
1
and :
3
÷ :
1
generate a direct summand of Z
3
.
Now (c) is clear: since Ker ˇ Ç U = Z:
1
÷ Z:
2
÷ Z:
3
, we have Z,ˇ(U) Š
Z
3
,U, and j(C) = #(Z
3
,U). For part (d) one notes that the elements in
par(:
1
. :
2
. :
3
) ¨ Z
3
represent the residue class modulo U. On the one hand,
they have separate values under ˇ and, simultaneously, only the values given are
possible since z is empty.
Proof of Theorem 2.70. We start with a triangulation of the bottom B of C into
empty simplices z. By Corollary 2.60(a) we have Hilb(R
÷
z) = Hilb(C) ¨ R
÷
z,
and this allows us to replace C by R
÷
z. In other words, we can assume that C is
generated by :
1
. :
2
. :
3
÷ Z
3
for which conv(0. :
1
. :
2
. :
3
) is an empty simplex.
Assume C is not unimodular. With the notation of the lemma, we choose n ÷
par(:
1
. :
2
. :
3
) such that ˇ(n) = b÷1. Write n = q
1
:
1
÷q
2
:
2
÷q
3
:
3
with rational
numbers q
i
, 0 ¸ q
i
< 1. Then ˇ(n) = (q
1
÷q
2
÷q
3
)b. Furthermore let C
i
, i =
1. 2. 3 be the cone spanned by n and the :
j
with j = i . Since conv(0. :
1
. :
2
. :
3
) is
empty and ˇ(n) is a generator of Z
3
,U, all three cones are full dimensional. We
have j(C
i
) = q
i
b, and

3
i=1
j(C
i
) = b ÷1.
Set sdiv(C) = par(:
1
. :
2
. :
3
) ¨ (Z
3
\ {0]). With analogous notation for C
i
we
claim
sdiv(C) = {n] L sdiv(C
1
) L sdiv(C
2
) L sdiv(C
3
).
If this equation holds, we are done: the basic simplices of the cones C
i
are empty
(there is no integer between b and b ÷ 1), and an application of the same argu-
ment to C
i
, i = 1. 2. 3, shows that all further subdividing vectors can be chosen in
sdiv(C) Ç Hilb(C) (compare 2.60(b)).
2.D. Normality and unimodular covering 77
Note that the sets on the right hand side are disjoint. Since # sdiv(C) = b ÷ 1
and # sdiv(C
i
) = q
i
b ÷1, the choice of n guarantees that both sides have the same
cardinality. Thus it is enough to show that sdiv(C) is contained in the right hand
side R. This holds since sdiv(C) Ç Hilb(C) by Corollary 2.60(b), and R, together
with :
1
. :
2
. :
3
, generates the monoid C ¨ Z
3
.
Remark 2.72. Theorem 2.70 cannot be extended to higher dimension. A polytopal
counterexample in dimension 4 can be constructed as follows. We choose an empty
3-simplex z
pq
(see Remark 2.52) with 1 < q < p ÷ 1 (and p. q coprime) and set
P = 2z. By Corollary 2.54 the polytopal monoid M(P) is integrally closed in
Z
4
so that E(P) is a Hilbert basis of C(P). If C(P) admits a unimodular Hilbert
triangulation, then P has a triangulation into unimodular simplices. However,
such does not exists according to [53]. (The first, nonpolytopal counterexample
was given in [12].)
While the example just discussed shows the failure of (UHT) in dimension ¸
4, it does not exclude that a weaker unimodularity condition is always satisfied,
namely
(UHC) Let C be a positive rational cone. Then C is the union of the unimodular
simplicial cones generated by elements of Hilb(C).
As we will see below, also (UHC) fails: there exists an integrally closed polytope
P of dimension 5 such that C(P), a cone of dimension 6, violates (UHC); equiva-
lently, P is not covered by its unimodular subsimplices. However, we do not know
of a counterexample C to (UHC) with dimC = 4 or dimC = 5.
While the unimodular subcones generated by elements of Hilb(C) may fail to
cover C, there always exist such unimodular subcones.
Proposition 2.73. Let C be a pointed rational cone and 0 Ç F
1
Ç Ç F
d
= C a
complete flag of faces of C. Then there exist x
1
. . . . . x
d
÷ Hilb(C) with the following
properties:
(a) R
÷
x
1
÷ ÷R
÷
x
d
is unimodular;
(b) x
1
. . . . . x
i
÷ F
i
for all i = 1. . . . . n.
Proof. By induction we can assume that x
1
. . . . . x
d-1
÷ F
d-1
have been found
such that they satisfy the conditions (a) and (b) for F
d-1
. Let o be the support
form of C associated with the facet F
d-1
. Then there exists y ÷ C ¨ Z
d
with
o(y) = 1. Therefore o(x) = 1 for at least one element x ÷ Hilb(C). Since
Hilb(C) ¨ F
i
= Hilb(F
i
), the choice x
d
= x satisfies our needs.
Isomorphism classes of lattice polytopes with given multiplicity. One interesting
consequence of Proposition 2.73 (together with Proposition 2.54) is the following
finiteness result for polytopes of bounded multiplicity.
Corollary 2.74. Let d and j be natural numbers. Then there are only finitely many
isomorphism classes of lattice polytopes P Ç R
d
such that j(P) ¸ j.
78 2. Affine monoids and their Hilbert bases
Proof. In view of the definition of morphism of lattice polytopes (see Definition
1.79), we can assume that aff(P) = R
d
, replacing R
d
by aff(P) and Z
d
by aff(P) ¨
Z
d
if necessary.
It is sufficient to consider only normal d-polytopes P with gp(M(P)) = Z
d÷1
.
In fact, the (d ÷ 1)th multiple of a lattice d-polytope in R
d
always satisfies these
conditions (Corollary 2.54) and, therefore, the multiplication by d ÷ 1 injects the
class of lattice polytopes P in R
d
of multiplicity ¸ j into that of normal lattice
polytopes Qin R
d
of multiplicity ¸ j(d ÷1)
d
with gp(M(Q)) = Z
d÷1
. We leave
it to the reader to show that P and P
t
are isomorphic if and only if (d ÷ 1)P and
(d ÷ 1)P
t
are isomorphic.
According to Proposition 2.73 every polytope P from our class contains a uni-
modular d-simplex. This simplex is isomorphic to z = conv(0. e
1
. . . . . e
d
), and
therefore we can assume that z Ç P. But then the condition j(P) ¸ j implies
that P is contained in the cube
˚
(x
1
. . . . . x
d
) : [x
i
[ ¸ j. i = 1. . . . . d
_
Ç R
d

Even the derivation of a good upper bound for the number of isomorphism
classes of lattice d-simplices of multiplicity jseems to be a highly nontrivial prob-
lem.
The integral Carathéodory property. If a pointed rational cone C Ç R
d
has
(UHC), then it satisfies the integral Carathéodory property, a name explained by
the obvious analogy to Carathéodory’s theorem 1.62:
(ICP) CR(C ¨ Z
d
) = dimC.
The counterexample presented below not only refutes (UHC), but (ICP) too.
The validity of (UHC) or (ICP) seems to be open in dimensions 4 and 5. Moreover,
no cone C satisfying (ICP), but violating (UHC) seems to be known.
We can formulate all the conditions above for positive affine monoids M and
their Hilbert bases: M has (UHT) if R
÷
M has a triangulation ˙ whose simplices
are generated by unimodular subsets of Hilb(M) (with respect to gp(M)). In a
similar manner, (UHC) is transferred to monoids, and (ICP) simply means that
CR(M) = rank M. While the flexibility of the monoid language will be useful in
the next subsection, it does not lead to a proper generalization. In fact, the weakest
of these properties implies the normality of M:
Proposition 2.75. Let M be an affine monoid and X a finite subset of M such that
every element of M can be represented as a Z
÷
-linear combination of r = rank M
elements of X. Then M is normal, and every element of M is a Z
÷
-linear combi-
nation of linearly independent elements of X.
In particular, M is normal if CR(M) = r.
Proof. We consider the submonoids N of M generated by at most r elements of
X. If N has rank r, then its Hilbert basis is linear independent, and so N is normal.
By hypothesis M is the union of all these submonoids N. Therefore it is normal
by Theorem 2.35 and the union of those N that have rank r.
2.D. Normality and unimodular covering 79
Tight cones. We introduce the class of tight cones and monoids and show that
they play a crucial role for (UHC) and (ICP).
Definition 2.76. Let M be a positive normal affine monoid, x ÷ Hilb(M), and M
t
the monoid generated by Hilb(M) \ {x]. We say that x is nondestructive if M
t
is
normal and gp(M
t
) is a direct summand of gp(M) (and therefore equal to gp(M)
if rank gp(M) = rank gp(M
t
)). Otherwise x is destructive. We say that M is tight
if every element of Hilb(M) is destructive. A pointed rational cone C Ç R
d
is tight
if C ¨ Z
d
is tight.
x
C
t
C
Figure 2.5. Cross-section of tightening of a cone
It is clear that only extreme elements of Hilb(M) can be nondestructive. Sup-
pose that x is an extreme element of Hilb(M). Then the localization MŒ÷x| splits
into a product Zx ˚M
x
where M
x
is again a positive normal affine monoid (see
Proposition 2.31).
Lemma 2.77. Let M be a normal positive affine monoid and x ÷ Hilb(M) a
nondestructive element. Let M
t
be the monoid generated by Hilb(M) \ {x].
(a) If M
t
and M
x
both satisfy (UHC), then so does M.
(b) One has CR(M) = max(CR(M
t
). CR(M
x
) ÷1).
Proof. Suppose M
t
and M
x
both satisfy (UHC). Since gp(M
t
) is a direct sum-
mand of gp(M) and Hilb(M
t
) = Hilb(M) \ {x] by the hypothesis on x, it is clear
that all elements of M
t
are contained in submonoids of M generated by unimodu-
lar subsets X
i
of Hilb(M) (with respect to gp(M)). We have to show that this holds
for elements of M \ M
t
, too.
Let z ÷ M \ M
t
. By hypothesis on M
x
, the residue class of z in M
x
has a
representation ¨ z = a
1
¨ y
1
÷ ÷a
m
¨ y
m
with a
i
÷ Z
÷
and ¨ y
i
÷ Hilb(M
x
) for i =
1. . . . . m such that ¨ y
1
. . . . . ¨ y
m
span a direct summand of gp(M
x
). Next observe
that Hilb(M) is mapped onto a system of generators of M
x
by the residue class
map. Therefore we may assume that the preimages y
1
. . . . . y
m
belong to Hilb(M)\
{x]. Furthermore, z = a
1
y
1
÷ ÷a
m
y
m
÷bx with b ÷ Z.
It only remains to show that b ÷ Z
÷
. There is a representation of z as a
Z
÷
-linear combination of the elements of Hilb(M) in which the coefficient of x
is positive. Thus, if b < 0, z has a Q
÷
-linear representation by the elements of
Hilb(M) \ {x]. This implies z ÷ R
÷
M
t
, and hence z ÷ M
t
, a contradiction.
This proves (a), and (b) follows similarly.
80 2. Affine monoids and their Hilbert bases
We say that a monoid M as in the lemma shrinks to the monoid T if there is a
chain M = M
0
¸ M
1
¸ ¸ M
t
= T of monoids such that at each step M
i÷1
is
generated by Hilb(M
i
)\{x] where x is nondestructive. An analogous terminology
applies to cones.
Corollary 2.78. A counterexample to (UHC) that is minimal, first with respect to
dimension and then with respect to # Hilb(C), is tight. A similar statement holds
for (ICP).
In fact, suppose that the cone C is a minimal counterexample to (UHC) with
respect to dimension, and that C shrinks to D. Then D is also a counterexample
to (UHC) according to Lemma 2.77. For (ICP) the argument is the same. It is
therefore clear that one should search for counterexamples only among the tight
cones. In fact, experiments have shown that such counterexamples are extremely
rare, and without the restriction to tight cones such may never have been found.
Remark 2.79. It is not hard to see that there are no tight cones of dimension ¸ 2.
However, we cannot prove that all 3-dimensional cones C are nontight; in general,
an extreme element of Hilb(C) can very well be destructive, even if dimC = 3. In
dimension 4 there exist tight cones but none of the examples we have found is of
the form C
P
with a 3-dimensional lattice polytope P. In dimension ¸ 5 one can
easily describe a class of tight cones: let W be a cube whose lattice points are its
vertices and its barycenter; then the cone C
W
is tight if dimW ¸ 4.
The shrinking of cones and the test for (UHC) have been implemented as com-
puter programs. See [17] or [20] for a description of the search for a counterexam-
ple.
A counterexample to (UHC) and (ICP). Let C
6
Ç R
6
, be the cone generated by the
vectors z
1
. . . . . z
10
:
z
1
= (0. 1. 0. 0. 0. 0). z
6
= (1. 0. 2. 1. 1. 2).
z
2
= (0. 0. 1. 0. 0. 0). z
7
= (1. 2. 0. 2. 1. 1).
z
3
= (0. 0. 0. 1. 0. 0). z
8
= (1. 1. 2. 0. 2. 1).
z
4
= (0. 0. 0. 0. 1. 0). z
9
= (1. 1. 1. 2. 0. 2).
z
5
= (0. 0. 0. 0. 0. 1). z
10
= (1. 2. 1. 1. 2. 0).
The cone C
6
and the monoid M
6
= C
6
¨ Z
6
have several remarkable proper-
ties:
(a) Hilb(C
6
) = {z
1
. . . . . z
10
].
(b) C
6
has 27 facets, of which 5 are not simplicial.
(c) The automorphism group Aut(M
6
) of M
6
has order 20, and it operates
transitively on Hilb(M
6
). In particular this implies that z
1
. . . . . z
10
are all
extreme generators of C
6
.
(d) The embedding C
6
Ç R
6
above has been chosen in order to make vis-
ible the subgroup U of those automorphisms that map each of the sets
Exercises 81
{z
1
. . . . . z
5
] and {z
6
. . . . . z
10
] to itself; U is isomorphic to the dihedral
group of order 10. However, C
6
can even be realized as the cone over a
0-1-polytope in R
5
.
(e) The vector of lowest degree disproving (UHC) is t
1
= z
1
÷ ÷ z
10
.
Evidently t
1
is invariant for Aut(M
6
), and it can be shown that its multiples
are the only such elements.
(f) The Hilbert basis is contained in the hyperplane H given by the equation
÷5ç
1
÷ ç
2
÷ ÷ ç
6
= 1. Thus z
1
. . . . . z
10
are the vertices of a normal
5-dimensional lattice polytope P
5
that is not covered by its unimodular
lattice subsimplices (and contains no other lattice points).
(g) If one removes all the unimodular subcones generated by elements of
Hilb(C
6
) from C
6
, then there remains the interior of a convex cone N.
While P
5
has volume 25,120, the intersection of N and P
5
has only vol-
ume 1,(1080 × 120).
(h) The vector
z
1
÷3z
2
÷5z
4
÷2z
5
÷z
8
÷5z
9
÷3z
10
can not be represented by 6 elements of Hilb(M) (and it is “smallest” with
respect to this property.) Moreover, one has CR(C
6
) = 7 (as can be seen
from a triangulation containing only two nonunimodular simplices).
In particular, C
6
is even a counterexample to (ICP).
3
See [23] for more detailed
information.
So far the authors have only found one essentially different example of a cone C
violating (UHC) and (ICP). It is has dimension 6, too, but its Hilbert basis consists
of 12 elements.
Exercises
2.1. Give an example of a monoid with torsion for which gp(M) is torsionfree.
2.2. Show that every submonoid of Zis finitely generated.
2.3. Prove the following claims for an affine monoid M:
(a) M = M
+
if and only if rank M ¸ 1 or M = int(M).
(b) If M = M
+
, then M
+
is not affine.
2.4. Let M be an affine monoid M. Show that the following are equivalent:
(i) M is a group;
(ii) the normalization
¨
M is a group;
(iii) R
÷
M is a vector space.
2.5. Showthat the intersection of Z
2
with the cone generated by (1. 0) and (1. a) is a finitely
generated monoid if and only if a is rational.
2.6. Show by means of an example that the additional conditions in Proposition 2.15(d)
and (e) cannot always be satisfied simultaneously. (Rank 2 is enough.)
3
The authors, having been fooled by a faulty random number generator, missed to notice that
C
6
does not have (ICP). This fact was then discovered by Henk, Martin,and Weismantel.
82 2. Affine monoids and their Hilbert bases
2.7. Prove Proposition 2.18.
2.8. Show that c(
¨
M,M) is the largest ideal of M that is also an ideal of
¨
M.
2.9. Let M be an affine monoid and x ÷ int(M). Show:
(a) mx ÷ c(
¨
R,R) for m ;0;
(b) MŒ÷x| = gp(M).
Hint: (a) reduces (b) to the case in which M =
¨
M.
2.10. Let M be a monoid. Show sn(M) consists of all x ÷ gp(M) such that nx. (n ÷1)x
for some n ÷ Z
÷
.
2.11. Provide the details in the proof of Proposition 2.38.
2.12. Give a detailed proof of Proposition 2.39.
2.13. Let M be an affine monoid and let N be an overmonoid of M contained in sn(M).
Show that there exists a chain
M = M
0
Ç M
1
Ç Ç M
n
= N
of monoids such that M
i
= M
i-1
L (M
i-1
÷ x) for some x with 2x. 3x ÷ M
i-1
, i =
1. . . . . n.
Hint: first construct a strictly ascending chain of extensions as required, and note that
M-submodules of
¨
M satisfy the ascending chain condition.
(In ring-theoretic terminology the extension M
i-1
Ç M
i
is called elementary subintegral
and the extension M Ç N is subintegral; see p. 141.)
Chapter 3
Multiples of lattice polytopes
3.A. Knudsen-Mumford triangulations
In this section we describe the unimodular triangulation of R
d
constructed by
Knudsen and Mumford [55, Chapter III]. It is basic for the results in Sections 3.B
and 3.C on the triangulation and covering of multiples of lattice polytopes by uni-
modular lattice simplices.
Weyl chambers. Let e
1
. . . . . e
d
÷ R
d
be the standard basis of R
d
. With each
permutation o of the permutation group S
d
of {1. . . . . d] we associate the simplex
z
o
= conv
_
0. e
o(1)
. e
o(1)
÷e
o(2)
. . . . . e
o(1)
÷e
o(2)
÷ ÷e
o(d)
_
.
Obviously z
o
is unimodular (with respect to Z
d
) and contained in the unit cube
spanned by e
1
. . . . . e
d
. One has
z
o
= {(x
1
. . . . . x
d
) ÷ R
d
: 1 ¸ x
o(1)
¸ x
o(2)
¸ ¸ x
o(d)
¸ 0].
The order
0 < e
o(1)
< e
o(1)
÷e
o(2)
< e
o(1)
÷ ÷e
o(d)
is called the canonical order of vert (z
o
). For the vertices of z
1
, the simplex cor-
responding to the identity permutation, we use the notation z
0
= 0 and z
i
=
e
1
÷ ÷e
i
, i = 1. . . . . d.
We want to show that the simplices z
o
tile the entire space R
d
in a natural way.
Consider the system of hyperplanes
H
i,j
=
_
{(x
1
. . . . . x
d
) ÷ R
d
: x
j÷1
= x
i÷1
] if 0 ¸ j < i < d.
{(x
1
. . . . . x
d
) ÷ R
d
: x
j÷1
= 0] if 0 ¸ j < i = d.
Note that H
i,j
is the hyperplane through the origin 0 which is parallel to the com-
plementary faces
conv(z
0
. z
1
. . . . . z
j
. z
i÷1
. z
i÷2
. . . . . z
d
). conv(z
j÷1
. z
j÷2
. . . . . z
i
) Ç z
1
.
For instance, H
i,i-1
= aff(z
0
. z
1
. . . . . ¨ z
i
. . . . . z
d
), i = 1. . . . . d.
83
84 3. Multiples of lattice polytopes
More generally, for every number k ÷ Zwe put
H
i,j
k
=
_
{(x
1
. . . . . x
d
) ÷ R
d
: x
j÷1
÷ x
i÷1
= k] if 0 ¸ j < i < d.
{(x
1
. . . . . x
d
) ÷ R
d
: x
j÷1
= k] if 0 ¸ j < i = d.
In particular, H
i,j
0
= H
i,j
.
Proposition 3.1. Every point x = (x
1
. . . . . x
d
) that is not contained in one of the
hyperplanes H
i,j
k
lies in the interior of exactly one of the integral parallel translates
of the simplices z
o
, o ÷ S
d
. In particular, the hyperplanes H
i,j
k
, 0 ¸ j < i ¸ d,
k ÷ Z, dissect R
d
into simplices which are the integral parallel translates of the
simplices z
o
, o ÷ S
d
.
Proof. Put a
i
= ]x
i
¡ and t
i
= {x
i
] = x
i
÷ a
i
. Since x …
_
H
i,j
k
we have t
i
= t
j
whenever i = j . Consequently, there is a unique element o ÷ S
d
such that
1 > t
o(1)
> t
o(2)
> > t
o(d)
> 0.
Moreover, the connected component of x in R
d
\
_
H
i,j
k
is the set
{(x
t
1
. . . . . x
t
d
) ÷ R
d
: 1 > x
t
o(1)
÷ a
o(1)
> > x
t
o(d)
÷ a
o(d)
> 0].
and this is exactly the interior of the simplex (a
1
. . . . . a
d
) ÷z
o
.
Figure 3.1. Weyl chambers of the plane
The representatives of the infinite collection of unimodular d-simplices thus
obtained will be called Weyl chambers. Applying integral parallel translations we
extend the notion of canonical order to the vertex set of any of the Weyl chambers.
Figure 3.1 shows the tiling of the plane. We have marked the chambers z
1
and
(÷2. ÷1) ÷z
(1 2)
.
Remark 3.2. Our Weyl chambers are actually the affine Weyl chambers of type A
d
.
In other words, each of the simplices z
o
, o ÷ S
d
, is bounded by hyperplanes
orthogonal to the elements of a root system of type A
d
. Since we only need this
type of Weyl chambers the attribute “of type A
d
” will be omitted. (See [11] or [51]
for a general discussion of Weyl groups and Weyl chambers.)
The Weyl chambers do not depend on the order of the basis vectors since the
system of hyperplanes H
i,j
k
is invariant under permutations of the coordinates.
Lemma 3.3. For all Weyl chambers z
1
and z
2
the canonical orders of vert(z
1
) and
vert(z
2
) agree on vert(z
1
¨ z
2
).
3.A. Knudsen-Mumford triangulations 85
Proof. We realize the canonical orders as restrictions of a linear order on Z
d
.
Then the claim follows immediately.
Consider positive real numbers ˛
1
. . . . . ˛
d
that are linearly independent over
Q. Then the linear form z =

˛
i
X
i
separates the points of Z
d
and, simulta-
neously, has positive values at the standard basis vectors e
i
. Then the canonical
order on the vertices z of every Weyl chamber is induced by the order of the values
z(z) ÷ R.
To a Weyl chamber a ÷z
o
we associate the mapping
¬ : {1. . . . . d] ÷Z. i ÷a
o(i)
.
Conversely, suppose we are given a mapping ¬ : {1. 2. . . . . d] ÷ Z. Then there is
a unique permutation o ÷ S
d
satisfying the following conditions:
¬(o(i )) ¸ ¬(o(j )) for i ¸ j (3.1)
and
¬(o(i )) = ¬(o(j )) ==
_
i < j ”o(i ) < o(j )
_
. (3.2)
Actually, one has the formula
o(k) = #
˚
i : ¬(i ) > ¬(k) or ¬(i ) = ¬(k) and i ¸ k
_
.
To ¬ we associate the Weyl chamber a÷z
o
1 where a =
_
¬(o(1)). . . . . ¬(o(d))
_
.
Although different mappings ¬ define different Weyl chambers, the two associa-
tions are not mutually inverse. For instance, all chambers z
o
, o ÷ S
d
, yield the
same constant map {1. 2. . . . . d] ÷ {0], and all Weyl chambers associated with
maps ¬ lie in a certain region in R
d
. However, on this region the correspondence
is bijective:
Lemma 3.4. The correspondence above establishes a bijection between the maps ¬ :
{1. 2. . . . . d] ÷Zand the Weyl chambers in the region
Z =
˚
(x
1
. . . . . x
d
) ÷ R
d
: x
1
¸ x
2
¸ ¸ x
d
_
Ç R
d
Proof. Given a map ¬, let o be the permutation derived from it. We want to
show that the linear forms X
i÷1
÷ X
i
, i = 1. . . . . d ÷ 1 (which correspond to the
hyperplanes H
i,i-1
), cannot have positive values on the chamber (¬(o(1)). . . . .
¬(o(d))) ÷z
o
1. But this follows from the conditions (3.1) and (3.2):
_
¬(o(i ÷1)) ÷ ¬(o(i ))
_
÷(x
i÷1
÷ x
i
) =
_
¬(o(i ÷1)) ÷ ¬(o(i ))
_
÷(x
o
1
(o(i÷1))
÷ x
o
1
(o(i))
) ¸ 0
if (x
1
. . . . . x
d
) satisfies the inequalities 1 ¸ x
o
1
(1)
¸ ¸ x
o
1
(d)
¸ 0.
Conversely, given a Weyl chamber z Ç Z, z = (a
1
. . . . . a
d
) ÷z
o
1 for some
a
i
÷ Z and o ÷ S
d
, we have 1 > x
o
1
(1)
> > x
o
1
(d)
> 0 for an interior
point (x
1
. . . . . x
d
) ÷ int(z
o
1). The containment z Ç Z is equivalent to the
condition that for each i ÷ {1. . . . . d ÷ 1] either a
i
> a
i÷1
or a
i
= a
i÷1
and
x
i
> x
i÷1
. But the latter inequality is the same as x
o
1
(o(i))
> x
o
1
(o(i÷1))
, that is
o(i ) < o(i ÷1). Therefore, if ¬ is the map associated to z, we indeed recover z
from ¬ if z Ç Z.
86 3. Multiples of lattice polytopes
The set Z is bounded by the hyperplanes H
i,i-1
, i = 1. . . . . d÷1. In particular,
it is tiled by the Weyl chambers contained in it.
Corollary 3.5. For every natural number j ¸ 1 there is a natural bijective corre-
spondence between the mappings ¬ : {1. . . . . d] ÷ {0. 1. . . . . j ÷ 1] and the Weyl
chambers in the multiple jz
1
. Moreover, these chambers triangulate the simplex
jz
1
.
The triangulations in this corollary will be called the canonical triangulations
(with respect to the enumeration e
1
. . . . . e
d
of the standard basis vectors e
i
).
We introduce a notation for the correspondence between Weyl chambers and
define some additional data. For a map ¬ : {1. 2. . . . . d] ÷Zwe let
v z
t
denote the corresponding Weyl chamber,
v z
t
0
. z
t
1
. . . . . z
t
d
denote the vertices of z
t
in the canonical order,
v o
t
÷ S
d
denote the permutation derived from ¬,
v 0
t
= z
t
0
and e
t
i
= (e
i
)
t
= z
t
i
÷ z
t
i-1
, i = 1. . . . . d,
v (x
t
1
. . . . . x
t
d
) be the coordinate vector of (x
1
. . . . . x
d
) in the new affine
coordinate system with origin 0
t
and basis e
t
1
. . . . . e
t
d
.
The equation z
t
=
_
¬(o
t
(1)). . . . . ¬(o
t
(d))
_
÷z
(o

)
1 implies the following
rule for the coordinate change:
Lemma 3.6. x
i
= ¬(o
t
(i )) ÷x
t
o

(i)
for every i = 1. . . . . d.
We have used the inverse permutations in defining the correspondence between
mappings and Weyl chambers in order to arrive at this nice formula.
Canonical and mixed triangulations. Let z Ç R
d
be a d-simplex together with
an order p
0
< p
1
< < p
d
of its vertices. Then the bijection z
1
÷ z, z
i
÷
p
i
, i = 1. . . . . d, extends to a unique affine transformation R
d
÷ R
d
. Using
this transformation we can transfer to z all data we have associated to z
1
. The
resulting tiling of R
d
is called canonical with respect to z and the given order on
vert(z).
Let ı Ç z be a face. The order on vert(ı), induced by that of vert(z), is called
z-canonical. Using this order we can develop similar notions within the affine
subspace aff(ı) Ç R
d
. Furthermore, we can do this in each of the parallel translates
z ÷aff(ı), z ÷ L(z), where, as usual, L(z) Ç R
d
is the affine lattice
p
0
÷
d

i=0
Z(p
i
÷ p
0
) Ç R
d
.
The simplices obtained are the ı-canonical ones. It follows from Lemma 3.3 that
all canonical orders (with respect to z) agree in the sense that if ı
t
and ı
tt
are two
canonical simplices (not necessarily of the same dimension) then ı
t
¨ ı
tt
is again
a canonical simplex and the canonical order on vert(ı
t
¨ ı
tt
) agrees with those on
vert(ı
t
) and vert(ı
tt
).
Let j and v be natural numbers. Then we can construct a triangulation of
the multiple jvz as follows. For each member ı of the z-canonical triangula-
tion of jz we consider the ı-canonical triangulation of vı (with respect to the
3.A. Knudsen-Mumford triangulations 87
z-canonical order). Since the canonical orders are compatible, these triangula-
tions can be patched up to a global triangulation T of jvz. However, we could
have obtained T in a simpler way:
Lemma 3.7. T is the z-canonical triangulation of jvz.
Proof. Let ı be a z-canonical d-simplex in jz and ; a ı-canonical d-simplex in
vı. Let ¬ : {1. . . . . d] ÷ {1. . . . . j ÷ 1] and , : {1. . . . . d] ÷ {1. . . . . v ÷ 1] be
the corresponding mappings, the latter taken with respect to the coordinates x
t
i
(Corollary 3.5). By Lemma 3.6 we then have x
i
= ¬(o
t
(i )) ÷ x
t
o

(i)
and x
t
i
=
,(o
p
(i )) ÷x
p
o

(i)
for i = 1. . . . .d. Thus
x
i
= ¬(o
t
(i )) ÷,
_
o
p
(o
t
(i ))
_
÷x
p
o

(o

(i))
. i = 1. . . . .d.
and this means that ; is a z-canonical simplex in jvz with corresponding per-
mutation (o
p
ı o
t
)
-1
.
Let z, j and v be as above. Suppose V is an ambient Euclidean space, strictly
containing aff(z), and q a point in V \ aff(z). Put
¯
z = conv(z. q). We con-
struct a triangulation of the simplex jv
¯
z as follows. By Corollary 3.5 an element
from the z-canonical triangulation of jz is of the type z
t
for some mapping ¬ :
{1. . . . . d] ÷{0. . . . . j÷1]. For each ¬ we consider the canonical triangulation of
the multiple v conv(z
t
. jq) with respect to the order p
t
0
< p
t
1
< < p
t
d
< jq
on vert(conv(z
t
. jq)). It follows from Lemma 3.3 that these triangulations can
be patched up to a global triangulation, which we denote by (vj
¯
z)
µ,¡
. Note that
¯
z
j = 2
j = 2
v = 3
j = 3
j = 3
v = 2
Figure 3.2. Mixed (j. v)-triangulations
(vj
¯
z)
µ,¡
= (vj
¯
z)
¡,µ
if j = v. Nevertheless, by Lemma 3.7, the induced triangu-
lations on vjz coincide (with the z-canonical one).
Next we introduce the concept of mixed triangulation for simplicial complexes,
generalizing the triangulations (vj
¯
z)
µ,¡
constructed above.
Suppose we are given the following data:
(1) a simplicial complex ˘,
(2) a decomposition vert(˘) = {p
1
. . . . . p
r
. q
1
. . . . . q
s
] with openstar
˘
(q
i
) ¨
openstar
˘
(q
j
) = 0 for 1 ¸ i < j ¸ s,
(3) an order p
1
< < p
r
< q
1
< < q
s
on vert(˘),
(4) two systems of natural numbers ¨ j = (j
1
. . . . . j
s
) and ¨ v = (v
1
. . . . . v
s
)
such that the products j
i
v
i
, i = 1. . . . . d, are all equal to, say, j.
88 3. Multiples of lattice polytopes
For every simplex z ÷ ˘ we consider the following triangulation of jz:
T
z
=
_
(jz)
µ
i

i
if q
i
÷ z for some i.
the z-canonical triangulation else.
All the triangulations are formed with respect to the order of vertices we have fixed.
By Lemma 3.7 these triangulations can be patched up to a global triangulation of
j˘.
Definition 3.8. The global triangulation of j˘ is called the mixed ( ¨ j. ¨ v)-triangu-
lation with respect to the fixed order of vert(˘) where ¨ j = (j
1
. . . . . j
s
) and ¨ v =
(v
1
. . . . . v
s
). This triangulation is denoted by (j˘)
¯ µ, ¯ ¡
.
q
1
q
2
Figure 3.3. Triangulation (j˘)
¯ µ, ¯ ¡
with ¨ j = (2. 2), ¨ v = (1. 1)
3.B. Unimodular triangulations of multiples of polytopes
The main result of this section, due to Knudsen and Mumford [55, Chapter III], is
that every lattice polytope P has a multiple cP, c ÷ N, that admits a unimodular
triangulation. The only known proof of this result is through its generalization to
lattice polytopal complexes (Theorem 3.17).
Let ˘ be a lattice polytopal complex with lattice structure ¹. Each polytope
Q ÷ ˘ has a multiplicity j(Q) = j
/
Q
(Q) with respect to its associated lattice
¹
Q
. The multiplicity of ˘ is
max{j(Q) : Q ÷ ˘].
For the proof of the main result we have to find c ÷ N and a lattice triangulation
˙ of cP with j(˙) = 1.
Comparing multiplicities. Let L Ç R
d
be an affine lattice, P Ç R
d
an L-polytope,
and F a facet of P.
We choose H = aff(F) and set
ht
L,F
(x) = ht
L,H
(x). x ÷ aff(P).
where we have chosen the sign in such a way that P Ç H
÷
(see Remark 1.78).
Proposition 3.9. Let z be an L-simplex, and : ÷ vert(z). Let F be the facet of z
opposite to :. Then
j
L
(z) = ht
L,F
(:)j
L
(F).
If G is an arbitrary face of z, then j
L
(G) [ j
L
(z).
3.B. Unimodular triangulations of multiples of polytopes 89
Proof. The simplest proof of the formula is by introducing coordinates in a suit-
able way and computing determinants as in the proof of Proposition 2.43 (which is
essentially the case ht
L,F
(:) = 1). We leave the details to the reader.
The assertion on faces follows by induction and transitivity.
The simple formula in Proposition 3.9 hides a relationship between the lattices
used in the definition of the multiplicities, namely
L
z
= L¨ aff(z). L(z) = :
0
÷

¡÷vert(z)
Z(: ÷ :
0
)
(:
0
an arbitrary vertex of z) and the corresponding lattices defined by F. Since
algebraic arguments play a key role in the following, it may be useful to make the
relationship explicit.
We choose a vertex :
0
of F as the origin. Then L
F
. L
z
. L(F). L(z) are lat-
tices (and not only affine lattices). Moreover, z = ht
L,F
is now a linear form. One
has a commutative diagram
0
L(F) L(z)
z
Z
h
0
0 L
F
L
z
z
Z 0
The map h : Z ÷ Z is the multiplication by h = ht
L,F
(:). The diagram has
exact rows and the vertical maps are injective. (Note that L(F) = L(z) ¨L
F
=
L(z) ¨ Ker z since z is a simplex.) So we obtain an exact sequence of cokernels
0 ÷L
F
,L(F) ÷L
z
,L(z) ÷Z,hZ ÷0
whose numerical essence is the formula j
L
(z) = hj
L
(F).
Lemma 3.10. Let L
t
Ç L be lattices in R
d
for which the index ŒL : L
t
| is finite.
Furthermore let z. ı be L
t
-simplices, and suppose that ı is L
t
-unimodular and that
aff(ı) Ç aff(z). Then the following hold:
(a) j
L
(ı) [ j
L
(z);
(b) j
L
(ı) [ ŒL,L
t
|.
Proof. We can augment ı to an L
t
-unimodular simplex ı
t
Ç aff(z) such that
dimı
t
= dimz, or equivalently aff(ı
t
) = aff(z). Then L
ı
0 = L
t
ı
0
= L
t
z
, and in
view of Proposition 3.9 we may assume that ı = ı
t
.
One has j
L
(z) = ŒL
z
: L(z)| = ŒL
z
: L
t
z
|ŒL
t
z
: L(z)|. The first factor is
exactly j
L

t
). This shows (a).
For (b) we note that L
ı
,L
t
ı
is a subgroup of L,L
t
in a natural way, since L
t
ı
=
L
t
¨ L
ı
. Furthermore j
L
(ı) = ŒL
ı
: L
t
ı
| since L
t
ı
= L
ı
.
Multiplicity and the distribution of lattice points. Let zbe an L-simplex with ver-
tices p
0
. . . . . p
d
. In the following we will have to use lattice points in the semiopen
parallelotope
par(z. p
0
) = p
0
÷par(p
1
÷ p
0
. . . . . p
d
÷ p
0
).
90 3. Multiples of lattice polytopes
for the triangulation of suitable multiples cz. To this end we need some informa-
tion about the distribution of these lattice points.
Each point in aff(z) has a unique representation
x = :
0
÷
d

i=1
˛
i
(p
i
÷ p
0
). ˛
i
÷ R.
We define the affine form , by ,(x) =

d
i
1
˛
i
. In other words,
,(x) = ÷ht
L(z),F
0
(x) ÷1. (3.3)
where F
0
is the facet of z opposite p
0
and the height is measured with respect to
the lattice L(z). Note that the heights, measured with respect to L and L(z)
differ only by a positive factor, more precisely
ht
L,F
0
(x) = ht
L,F
0
(p
0
) ht
L(z),F
0
(x). (3.4)
This follows from the fact that two affine forms that have value 0 on aff(F
0
) differ
by a constant factor.
A very useful tool is the Waterman map defined by
w(x) = p
0
÷
x ÷ p
0
{,(x)¦
. x = p
0
.
Note that w(x) ÷ z for all x ÷ par(z. p
0
). The points w(x), x ÷ lat(par(z. p
0
)),
x = p
0
, are called p
0
-Waterman points of z.
Let x ÷ lat(par(p
0
. z)) and suppose that w(x) does not belong to F
0
. In the
following we will use that this condition is equivalent to ,(x) … Z.
We construct a triangulation T of hz, h = {,(x)¦, as follows. First we set
¯ x = hp
0
÷(x ÷ p
0
) = (h ÷ 1)p
0
÷x.
Then we choose the z-canonical triangulation on link
hz
( ¯ x), and finally we extend
it by stellar subdivision with respect to ¯ x to a triangulation of hz. Figure 3.4 illus-
trates T in dimension 2 (with h = 2, p
0
= 0, ¯ x = x).
w(x)
x
Figure 3.4. The triangulation T
Lemma 3.11. j(T ) < j(z).
Proof. Since the construction is invariant under parallel translations we can as-
sume that p
0
= 0. Then ¯ x = x. We let F
i
be the facet of z opposite p
i
.
Consider a d-simplex ı ÷ T . By construction x is one of its vertices, and its
facet G opposite to x lies in one of the facets hF
i
of hz. By construction j(G) =
3.B. Unimodular triangulations of multiples of polytopes 91
j(F
i
), and for the claim it is enough that the L-height of x above G is smaller than
the L-height of p
i
above F
i
; compare Lemma 3.9.
Suppose first that i = 0. Then, since h ÷ ,(x) < 1, the height of x above hF
0
is smaller than the height of p
0
above F
0
.
Now let i > 1. Since x ÷ par(z. 0), the height of x above hF
i
= F
i
is smaller
than the maximal height of a point of P above F
i
, and that is attained in p
i
. So the
height of x above F
i
is smaller than that of p
i
above F
i
.
Remark 3.12. For each face F of z such that x … aff(F) the triangulation T
t
on
conv(hF. ¯ x) induced by T is the first step in the construction of the mixed (h. g)-
triangulation of conv(F. w(x)), with g ÷ Narbitrary.
The mixed triangulation (hg conv(F. w(x))
h,g
then has multiplicity < j(z),
since it is constructed by the canonical extension of T
t
.
Now we compare z to its facets
Lemma 3.13. Let z = conv(p
0
. . . . . p
d
) be an L-simplex.
(a) Let ı be a face of z. Then j(ı) = j(z) if and only if every p
0
-Waterman
point of z lies in ı.
(b) Suppose G = L,L(z) is a cyclic group and j(ı) < j(z) for any proper
face ı Ç z. If z is a point in lat(par(z. p
0
)) such that the residue class of
z ÷ p
0
generates G, then w(z) ÷ int(z).
Proof. We can assume that p
0
= 0 and write par(z) for par(z. 0). Moreover, L
and L = L(z) are groups.
(a) First consider the case 0 ÷ ı. The multiplicity of z is the number of lattice
points in par(z), and an analogous statement holds for par(ı). Thus both condi-
tions in the lemma are equivalent to lat(par(z)) = lat(par(ı)).
Now suppose that 0 … ı, and let ı
t
= conv(0. ı). In view of the case already
settled and since j(ı) ¸ j(ı
t
) ¸ j(z), we can replace zby ı
t
. In other words, we
can assume that ı is the face F
0
of z opposite to 0.
One has w(x) ÷ F
0
if and only if ,(x) ÷ Z. Each point of L differs from a
suitable L-point in par(z) by an element of L. Moreover, ,(y) ÷ Zfor all y ÷ L.
So we conclude that w(x) ÷ F
0
for all L-points x = 0 in par(z) if and only if
,(y) ÷ Z for all y ÷ L. By equation (3.3) we obtain the equivalent condition
ht
F
0
,L
(y) ÷ Zfor all y ÷ L.
In view of equation (3.4) and since there exists y ÷ L with ht
F
0
,L
(y) = 1, we
finally see that our hypothesis is equivalent to ht
L,F
0
(p
0
) = 1, and this in turn is
equivalent to j(z) = j(F
0
) by Lemma 3.9.
(b) If w(z) … int(z), then there exists a facet ı of z with w(z) ÷ ı.
If 0 ÷ ı, then z and all multiples of z lie in R
÷
ı ¨ L, and each lattice point in
par(z) is already in par(ı). In fact, every x ÷ L is congruent to a multiple of z
modulo L, since z generates G. But then j(ı) = j(z).
If z ÷ F
0
, then, by the same argument, all L-points of z have integral value
under ,. Again we have reached a contradiction with (a).
Corollary 3.14. Let ı
t
and ı
tt
be faces of z. Then j(ı
t
) = j(z) = j(ı
tt
) ”
j(ı
t
¨ ı
tt
) = j(z).
92 3. Multiples of lattice polytopes
For the proof of the main theorem we have to analyze the structure of the union
of the simplices z in a lattice simplicial complex ˘ with j(z) = j(˘). Let ˘ be
a lattice simplicial complex and put
U =
_
z÷˘
µ(z)=µ(˘)
int(z).
From Corollary 3.14 we derive the following decomposition result:
Lemma 3.15. The connected components of U are the open star neighborhoods of
the relative interiors of the minimal (with respect to inclusion) simplices ı ÷ ˘ with
multiplicity j(ı) = j(˘):
U =
_
min¦ı÷˘ : µ(ı)=µ(˘)¦
openstar
˘
(int(ı)).
the union being disjoint.
Finally, Corollary 3.15 and Lemma 3.13 imply
Corollary 3.16. Let ˘ be a lattice simplicial complex such that j(˘) is a prime
number. Suppose {z
1
. . . . . z
s
] = min{ı ÷ ˘ : j(ı) = j(˘)] (defined as in
Lemma 3.15). Then for every index j = 1. . . . . s and every vertex p ÷ vert(z
j
)
there exists a p-Waterman point x ÷ int(z
j
).
Unimodular triangulations of multiples of complexes. The main result of this sec-
tion is
Theorem 3.17 (Knudsen-Mumford). Let ˘ be a lattice polytopal complex. Then
there exists a natural number c such that c˘ has a unimodular triangulation.
Corollary 3.18. For every lattice polytope P there exists a natural number c such
that cP has a unimodular triangulation.
Remark 3.19. Suppose that c˘ admits a unimodular triangulation T , and fix an
order on vert(T ). By Lemma 3.7 the iterated use of canonical triangulations with
respect to this order shows that all multiples c
t
c˘, c
t
÷ N, have unimodular trian-
gulations.
Proof of Theorem 3.17. By Theorem 1.58, which guarantees the existence of a
lattice triangulation of ˘, there is no loss of generality in assuming that ˘ is a
simplicial complex.
The proof is by induction on j(˘). If j(˘) = 1, then ˘ is already unimodu-
lar. So we can assume that we have proved Theorem 3.17 for simplicial complexes
of multiplicity < k = j(˘).
Since we will consider different lattice structures on the geometric realization
[˘[ we will always specify the lattice structure at issue, ¹ = {¹
z
: z ÷ ˘] being
the original lattice structure on ˘.
We fix a total order of the vertices of ˘.
3.B. Unimodular triangulations of multiples of polytopes 93
Case A: k is a composite number. Let z
1
. . . . . z
s
÷ ˘ be the simplices of mul-
tiplicity k and P
t
j
= par(z
j
. p
j
), j = 1. . . . . s, be the semi-open parallelotope
with respect to the minimal vertex of z
j
(“minimal” refers to the fixed order of the
vertices of ˘). For each index j we choose a point z
j
÷ lat(par(z
j
. p
j
)) such that
the residue class of z
j
÷ p
j
has order r
j
< k in the quotient group ¹
z
j
,L(z
j
).
(We can take r
j
to be a fixed prime divisor of k independently of j , but that is
irrelevant.)
We consider the sequence of lattice structures
¹
0
. ¹
1
. . ¹
s
. ¹
on the geometric realization [˘[, defined recursively as follows:
v ¹
0
is generated by vert(˘), more precisely: (¹
0
)
z
= L(z) for each
simplex z ÷ ˘. In particular, (˘. ¹
0
) is a unimodular lattice simplicial
complex. Therefore (˘. ¹
0
) is an embedded lattice polytopal complex.
This follows by the same argument that shows that a simplicial complex
(without lattice structure) is embedded; see Example 1.49. Let L
0
be the
ambient lattice of the embedding.
v For each j = 1. . . . . s we set L
j
= L
j-1
÷Z(z
j
÷ p
j
), and consider the
lattice structure ¹
j
on ˘ induced by it: ¹
j
is composed of the restric-
tions of (L
j
)
z
to the affine spaces aff(z), z ÷ ˘. So all lattice simplicial
complexes (˘. ¹
0
), (˘. ¹
1
). . . . . (˘. ¹
s
) are embedded.
Let z be a simplex in ˘. By Lemma 3.10, the multiplicity j
L
1
(z) divides the
index r
1
of L
0
in L
1
. Since r
1
< k, we can apply the induction hypothesis, and find
a factor c
1
, for which c
1
˘ has a ¹
1
-unimodular triangulation. The same argument
can be applied to the extension ¹
2
of ¹
1
etc. We end up with a multiple (c
1
c
s

that has a ¹
s
-unimodular triangulation.
We claim that every L
s
-unimodular simplex ı in (c
1
c
s
)˘ has multiplicity
< k in ¹. If this holds, then we are done because we can apply the induction
hypothesis again to obtain a multiple (c
t
c
1
c
s
)˘ with a ¹-unimodular triangu-
lation.
Suppose first that ı is an L
s
-unimodular simplex contained in aff(z) where z
is a simplex in ˘ with j
/
(z) < k. Applying Lemma 3.10 to the lattices L(z) and
¹
z
we see that j
/
(ı) is < k.
Now suppose that j(z) = k, say z = z
j
. Then j
/
(ı) divides the index
Œ¹
z
: (L
j
)
z
|. Next
k = j
/
(z) = Œ¹
z
: (L
0
)
z
| = Œ¹
z
: (L
j
)
z
|Œ(L
j
)
z
: (L
0
)
z
|.
and Œ(L
j
)
z
: (L
0
)
z
| > 1 by the construction of L
j
. In fact, z
j
÷ p
j
… (L
0
)
z
.
Case B: k = p is a prime number. Let U be the union of the interiors of the
simplices in (˘. ¹) of multiplicity p. By Lemma 3.15 U decomposes into a disjoint
union
U =
_
j
openstar(int(z
j
))
and, by Corollary 3.16, each of these z
j
has an interior Waterman point q
j
= w(z
j
)
(with respect to a vertex of ı
j
). Let (˘
t
. ¹) denote the lattice simplicial complex
94 3. Multiples of lattice polytopes
obtained from ˘ via stellar triangulation with respect to the new vertices q
j
. We
order the vertices of ˘
t
in such a way that the order of the vertices of ˘ remains
the same and the points q
j
are greater than the vertices of ˘.
Let r be the least common multiple of the numbers r
j
= {,(z
j
)¦ and choose
natural numbers s
j
such that
r
1
s
1
= = r
j
s
j
= = r.
and let ˘
tt
be the mixed (r
j
. s
j
)-triangulation of the lattice simplicial complex
(r˘
t
. ¹) with respect to the fixed order of vert(˘
t
). Then (˘
tt
. ¹) is a lattice tri-
angulation of (r˘. ¹) in which all simplices of multiplicity p have been refined;
see Remark 3.12. Since j(˘
tt
. ¹) < p. Moreover, no new simplices of multiplicity
¸ p have been created, and so the induction hypothesis applies. (We have chosen
r as the least common multiple in order to keep the factor r as small as possible—
any other common multiple would suffice.)
Remark 3.20. (a) One could try to give an effective upper bound for the number c
in Theorem 3.17 by tracing its proof. Evidently the bound obtained does not only
depend on j(˘) and d = dim˘, but also on the number of simplices in ˘.
(b) An easy observation is that c can be chosen in such a way that its prime di-
visors are bounded by {d,2¦. In fact, one only needs to show this in the case where
k is prime. Every point z ÷ lat(par(z. p
0
)) has ,(z) < d. If ,(z) > {d,2¦, then we
use the symmetry trick that we have first encountered in the proof of Proposition
2.49. Especially, if d = 3 or d = 4, then one can take c to be a power of 2.
A lower bound for c is d ÷ 1, as is shown by Example 2.55.
Remark 3.21. More precise results are known in dimension 3. Kantor and Sarakaria
[53] have shown that 4˘ has a unimodular triangulation if ˘ is a lattice polytopal
complex of dimension 3. On the other hand, 2z
pq
does not admit such a triangula-
tion if q = 1 and q = p ÷1 (see Remark 2.52 for the definition of z
pq
). Moreover,
Lagarias and Ziegler [58] have proved that cz
pq
has a unimodular triangulation
for all c ¸ 4.
But even here some open questions remain: has c˘ a unimodular triangulation
for all c ¸ 4? What can be said about 3z
pq
?
Regularity. It is proved in [55] that the unimodular triangulation of c˘ con-
structed above is regular. The proof of the regularity requires a rather technical
and involved consideration of different geometrical situations.
We only explain where the difficulty lies. In the proof of Theorem 3.17 we
have only used stellar and mixed triangulations, and compositions of triangula-
tions already obtained. Therefore, in view of Proposition 1.68 and Lemma 1.72, it
is enough to prove the regularity of mixed triangulations.
We know that dissections are regular (Section 1.F). Therefore, by Proposition
3.1, the canonical triangulations of multiples of simplices are regular. Next we show
the regularity of the mixed triangulation (vj
¯
z)
µ,¡
of a simplex z(with an ordered
set of vertices), j. v ÷ N. To this end we need appropriate support functions for
the canonical triangulations of multiples of simplices.
3.C. Unimodular covers of multiples of polytopes 95
Let z = conv(p
0
. . . . . p
n
) Ç R
d
be an n-simplex, p
0
< < p
n
. Let
x
0
. . . . . x
n
be the corresponding barycentric coordinates in aff(z). Then the
canonical triangulation of a multiple ×z, × ÷ N is obtained by dissecting ×z
by hyperplanes
H
i,j
p
=
_
(x
0
. . . . . x
n
) :

j¸s¸i
x
i
= ,
_
. 0 ¸ j < i ¸ n. 1 ¸ , ¸ × ÷ 1.
It follows that the function
¡ =

0¸j~i¸n
1¸p¸×-1
ˇ
ˇ
ˇ

j¸s¸i
x
s
÷ ,
ˇ
ˇ
ˇ : aff(z) ÷R
supports the canonical triangulation of ×z.
Suppose that j. v ÷ Nand q ÷ R
d
\ aff(z). Put
¯
z = conv(p
0
. . . . . p
n
. q) and
order the vertices by p
0
< < p
n
< q, and consider the mixed triangulation
(vj
¯
z)
µ,¡
. Let z
t
be the facet of the canonical triangulation of jz, corresponding
to a mapping ¬ : {1. . . . . d] ÷ {0. 1. . . . . j ÷ 1] (Corollary 3.5). Then vj
¯
z is
triangulated into simplices v conv(z
t
. jq). This is a regular triangulation because
it is the vth multiple of the stellar subdivision
conv(jz. jq) =
_
t:¦1,...,d¦-¦0,1,...,µ-1¦
conv(z
t
. jq)
The functions ¡ that support the canonical triangulations of v conv(z
t
. jq) agree
on common faces. Therefore Corollary 1.69 implies the regularity of (vj
¯
z)
µ,¡
.
However, no longer do these support functions agree when one considers
mixed triangulations of general lattice polytopal complexes, and here is where the
difficulty lies.
It seems there is no known application of the regularity of the triangulations
constructed in the proof of Theorem 3.17, neither in the context of semi-stable re-
ductions [55, Page ???] nor in the context of Koszul properties of polytopal monoid
rings. For this reason we have not included the details.
3.C. Unimodular covers of multiples of polytopes
Let P Ç R
d
be a lattice polytope. The union of all unimodular d-simplices inside
a d-polytope P is denoted by UC(P). In this section we investigate for which mul-
tiples cP of a lattice d-polytope one can guarantee that cP = UC(cP). To this end
we let c
pol
d
denote the infimum of the natural numbers c such that c
t
P = UC(c
t
P)
for all lattice d-polytopes P and all natural numbers c
t
¸ c. The following seem
to be the only known values of these numbers: c
pol
1
= 1 for trivial reasons, c
pol
2
= 1
by Corollary 2.51, and the equality c
pol
3
= 2 is proved in Kantor and Sarkaria [53].
A priori, it would not be excluded that c
pol
d
= o.
The main result of this section is the following upper bound:
96 3. Multiples of lattice polytopes
Theorem 3.22. As d ÷oone has
c
pol
d
= O
_
d
2.1d÷5
_
.
Theorem 3.22 is proved by passage to cones, for which we establish a similar
result on covers by unimodular subcones.
We define c
cone
d
to be the infimum of all natural numbers c such that every ratio-
nal d-dimensional pointed cone C Ç R
d
admits a unimodular cover C =
_
k
j=1
C
j
for which
Hilb(C
j
) Ç cz
C
j = 1. . . . . k.
Here z
C
denotes the polytope spanned by 0 and the extreme integral generators
of C.
As we already know, the unique full Hilbert triangulation in dimension two is
unimodular. In particular, c
cone
2
= 1. Since Hilb(C) Ç 2z
C
in dimension 3 and C
has a unimodular Hilbert triangulation, we have c
cone
3
= 2. (See Corollary 2.61 and
Theorem 2.70.)
We can now formulate the main result for unimodular covers of rational cones:
Theorem 3.23. For all d ¸ 2 one has
c
cone
d
¸
(d ÷1)d
2

_
d ÷ 1
_
(d ÷ 1)
_
(2ln 2)d÷1
.
Slope independence. As in Section 3.A we consider the system of simplices
z
o
Ç Œ0. 1|
d
. o ÷ S
d
.
where S
d
is the permutation group of {1. . . . . d]. We have already seen that the
integral parallel translates of the simplices z
o
cover the cone
R
÷
e
1
÷R
÷
(e
1
÷e
2
) ÷ ÷R
÷
(e
1
÷ ÷e
d
) ~ R
d
÷
.
Suppose we are given a nonzero real linear form
˛(X
1
. . . . . X
d
) = a
1
X
1
÷ ÷a
d
X
d
.
The width of a polytope P Ç R
d
in direction (a
1
. . . . . a
d
), denoted by width
˛
(P),
is defined to be the Euclidean distance between the two extreme hyperplanes that
are parallel to the hyperplane a
1
X
1
÷ ÷a
d
X
d
= 0 and intersect P. Since Œ0. 1|
d
is inscribed in a sphere of radius
_
d,2, we have width
˛
(z
o
) ¸
_
d for all linear
forms ˛ and all permutations o. We arrive at
Proposition 3.24. All integral parallel translates of z
o
, o ÷ S
d
, that intersect a
hyperplane H Ç R
d
are contained in the
_
d-neighborhood of H.
Let L be an affine lattice. The union of all L-unimodular simplices inside
a polytope P Ç R
d
is denoted by UC
L
(P) (we do not require that P is an L-
polytope). If z = conv(n
0
. . . . . n
e
) Ç R
d
is an e-simplex, then UC
z
(P) denotes
UC
L(z)
(P) where, as usual, L(z) = n
0
÷

e
i=0
Z(n
i
÷n
0
) is the smallest affine
lattice containing vert(ı).
3.C. Unimodular covers of multiples of polytopes 97
Let z Ç z
t
be two d-simplices in R
d
for which the origin 0 is a common vertex
and the two simplicial cones R
÷
z and R
÷
z
t
. The following lemma says that the
L(z)-unimodularly covered area in a multiple cz
t
, c ÷ N, approximates cz
t
with
a precision independent of zand z
t
. The precision is therefore independent of the
“slope” of the facets of z and z
t
opposite to 0. The lemma will be critical both in
the passage to cones (Proposition 3.28) and in the proof of Theorem 3.23.
Lemma 3.25. For a pair of d-simplices z Ç z
t
, having 0 as a common vertex at
which they span the same cone, and a real number c ÷ (0. 1) one has
(1 ÷ c)cz
t
Ç UC
z
(cz
t
).
whenever c ¸
_
d,c.
Proof. Let :
1
. . . . . :
d
be the vertices of zdifferent from0, and let n
i
, i = 1. . . . . d
be the vertex of z
t
on the ray R
÷
:
i
. By a rearrangement of the indices we can
achieve that
[n
1
[
[:
1
[
¸
[n
2
[
[:
2
[
¸ ¸
[n
d
[
[:
d
[
¸ 1.
where [ [ denotes Euclidean norm. Moreover, the assertion of the lemma is invari-
ant under linear transformations of R
d
. Therefore we can assume that
z = conv(0. e
1
. e
1
÷e
2
. . . . . e
1
÷ ÷e
d
).
Then L(z) = Z
d
. The ratios above are also invariant under linear transforma-
tions. Thus
[n
1
[
[e
1
[
¸
[n
2
[
[e
1
÷e
2
[
¸ ¸
[n
d
[
[e
1
÷ ÷e
d
[
¸ 1.
Now Lemma 3.27 below shows that the distance h from 0 to the affine hyperplane
H through n
1
. . . . . n
d
is at least 1.
By Proposition 3.24, the subset
(cz
t
) \ U
_
d
(cH) Ç cz
t
is covered by integral parallel translates of the simplices z
o
, o ÷ S
d
that are con-
tained in cz. (U
ı
(M) is the ı-neighborhood of M.) In particular,
(cz
t
) \ U
_
d
(cH) Ç UC
z
(cz
t
). (3.5)
Therefore we have
_
1÷c
_
cz
t
Ç
_

_
d
c
_
cz
t
Ç
_

_
d
ch
_
cz
t
=
ch ÷
_
d
ch
cz
t
= (cz
t
)\U
_
d
(cH).
and the lemma follows from (3.5).
Remark 3.26. One can derive an analogous result using the trivial tiling of R
d
÷
by
the integral parallel translates of the unit cube Œ0. 1|
d
and the fact that Œ0. 1|
d
itself
is unimodularly covered. The argument would then get simplified, but the estimate
obtained is c ¸ d,c, and thus worse than c ¸
_
d,c.
98 3. Multiples of lattice polytopes
We have formulated the Lemma 3.25 only for full dimensional simplices, but it
is clear that the claim holds for simplices of smaller dimension as well: one simply
chooses all data relative to the affine subspace generated by z
t
.
Above we have used the following
Lemma 3.27. Let e
1
. . . . . e
d
be the canonical basis of R
d
and set n
i
= z
i
(e
1
÷ ÷
e
i
) where z
1
¸ ¸ z
d
> 0. Then the affine hyperplane H through n
1
. . . . . n
d
intersects the set Q = z
d
(e
1
÷ ÷e
d
)÷R
d
÷
only in the boundary dQ. In particular
the Euclidean distance from0 to H is ¸ z
d
.
Proof. The hyperplane H is given by the equation
1
z
1
X
1
÷
_
1
z
2
÷
1
z
1
_
X
2
÷ ÷
_
1
z
d
÷
1
z
d-1
_
X
d
= 1.
The linear form ˛ on the left hand side has nonnegative coefficients and n
d
÷ H.
Thus a point whose coordinates are strictly smaller than z
d
cannot be contained
in H.
Passage to cones. In this section we want to relate the bounds for c
pol
d
and c
cone
d
.
This allows us to derive Theorem 3.22 from Theorem 3.23. We will use the corner
cones of a polytope P: the corner cone C of P at a vertex : is given by
C = R
÷
(P ÷ :) =

u÷vert(P)
R
÷
(n ÷ :).
Proposition 3.28. Let d be a natural number. Then c
pol
d
is finite if and only if c
cone
d
is
finite, and, moreover,
c
cone
d
¸ c
pol
d
¸
_
d(d ÷1)c
cone
d
. (3.6)
Proof. Suppose that c
pol
d
is finite. Then the left inequality is easily obtained by
considering the multiples of the polytope z
C
for a cone C: the cones spanned by
those unimodular simplices in a multiple of z
C
that contain 0 as a vertex constitute
a unimodular cover of C.
Now suppose that c
cone
d
is finite. For the right inequality we first triangulate a
polytope P into lattice simplices. Then it is enough to consider a lattice d-simplex
z Ç R
d
with vertices :
0
. . . . . :
d
.
Set c
t
= c
cone
d
. For each i there exists a unimodular cover (D
ij
) of the corner
cone C
i
of z with respect to the vertex :
i
such that c
t
z ÷ c
t
:
i
contains z
D
ij
for
all j . Thus the simplices z
D
ij
÷ c
t
:
i
cover the corner of c
t
z at c
t
:
i
, that is, their
union contains a neighborhood of c
t
:
i
in c
t
z.
We replace z by c
t
z and can assume that each corner of z has a cover by
unimodular simplices. It remains to show that the multiples c
tt
zare unimodularly
covered for every number c
tt
¸
_
d(d ÷1) for which c
tt
P is an integral polytope.
Let
o =
1
d ÷1
(:
0
÷ ÷:
d
)
3.C. Unimodular covers of multiples of polytopes 99
be the barycenter of z. We define the subsimplex z
i
Ç z as follows: z
i
is the
homothetic image of zwith respect to the center :
i
so that o lies on the facet of z
i
opposite to :
i
. The factor of the homothety that transforms zinto z
i
is d,(d ÷1).
In particular, the simplices z
i
are pairwise congruent. It is also clear that
d
_
i=0
z
i
= z. (3.7)
The construction of o and the subsimplices z
i
commutes with taking multiples
of z. It is therefore enough to show that c
tt
z
i
Ç UC(c
tt
z) for all i . In order to
simplify the use of dilatations we move :
i
to 0 by a parallel translation.
In the case in which :
i
= 0 the simplices c
tt
z and c
tt
z
i
are the unions of their
intersections with the cones D
ij
. This observation reduces the critical inclusion
c
tt
z
i
Ç c
tt
z to
c
tt
(z
i
¨ D
ij
) Ç c
tt
(z¨ D
ij
)
for all j . But now we are in the situation of Lemma 3.25, with the unimodular
simplex z
D
ij
in the role of the zof 3.25 and z¨D
ij
in that of z
t
. For c = 1,(d÷1)
we have c
tt
¸
_
d,c and so
c
tt
(z
i
¨ D
ij
) = c
tt
d
d ÷1
(z¨ D
ij
) = c
tt
(1 ÷ c)(z¨ D
ij
) Ç UC(z¨ D
ij
).
as desired.
The steps in the proof Proposition 3.28 are illustrated by Figure 3.5 where we
have marked the barycenter of c
tt
c
t
z.
:
0
:
2
:
1
z
c
t
z
:
0
:
2
:
1
c
t
z
c
tt
c
t
z
o
Figure 3.5. Extension of the corner cover into a multiple of z
At this point we can deduce Theorem3.22 fromTheorem3.23. Using the bound
for c
cone
d
, given in Theorem 3.23, and that {
_
d ÷ 1¦(d ÷1) = O
_
d
1.5
_
as d ÷o,
100 3. Multiples of lattice polytopes
we get
c
pol
d
¸
_
d(d ÷1)c
cone
d
¸
_
d(d ÷1)
(d ÷1)d
2

_
d ÷ 1
_
(d ÷ 1)
_
(2ln 2)d÷1
= O
_
d
2.1d÷5
_
.
as desired. (The left inequality in (3.6) has only been stated for completeness; it
will not be used later on.)
Bounding triangulations of simplicial cones. Let C be a simplicial rational d-
cone. Recall that the simplex z
C
is the convex hull of 0 and the extreme integral
generators of C. Let t(C) be the minimum of all real numbers c such that C has
a unimodular triangulation by simplicial cones D
i
, i ÷ I, such that Hilb(D
i
) Ç
cz
c
. t(C) is well-defined since C can be triangulated into unimodular cones by
Theorem 2.67. Moreover, there exist only finitely many isomorphism classes of
simplices z
C
of given multiplicity and dimension by Corollary 2.74 so that we can
set
t
d
(j) = max
˚
t(C) : dimC = d. j(C) ¸ j
_
.
Exponential factors enter the bounds for c
cone
d
and c
pol
d
only through t
d
, and
any improvement of Proposition 3.30 below would critically affect their orders of
magnitude.
We want to apply stellar subdivision as in the proof of Theorem 2.67, but with
numerical control on the lengths of the vectors used for it. The following lemma is
the first step towards this goal. We use the notation
(:
1
. . . . . :
e
) = {z
1
:
1
÷ ÷z
e
:
e
: 0 ¸ :
1
. . . . . :
e
¸ 1]
for linearly independent elements :
1
. . . . . :
e
÷ R
d
.
Lemma 3.29. Let :
1
. . . . . :
d
÷ Z
d
be linearly independent elements which do not
form a basis of Z
d
. Then there is a lattice point n in the subset
d
d ÷1
(:
1
. . . . . :
d
) \ {0] Ç par(:
1
. . . . . :
d
).
Proof. We use induction on d, the case d = 1 being obvious. Suppose we have
shown the claim in dimension < d. If there is an index i with 1 ¸ i ¸ d such
that {:
1
. . . . . :
i-1
. :
i÷1
. . . . . :
d
] is not a basis of the group Z
d
¨
_
jyi
R:
j
_
then,
by the induction assumption, (d ÷ 1),d(:
1
. . . . . :
i-1
. :
i÷1
. . . . . :
d
) contains a
nonzero lattice point and we are done because
d ÷ 1
d
(:
1
. . . . . :
i-1
. :
i÷1
. . . . . :
d
) Ç
d
d ÷1
(:
1
. . . . . :
d
).
Therefore we can assume that {:
1
. . . . . :
i-1
. :
i÷1
. . . . . :
d
] is a basis of the group
Z
d
¨
_
jyi
R:
j
_
, i = 1. . . . . d. It follows from Proposition 2.40 that there is
3.C. Unimodular covers of multiples of polytopes 101
a lattice point : ÷ int((:
1
. . . . . :
d
)). The following implication holds for each
i = 1. . . . . d:
a ÷ Z. a: ÷ Z:
i
÷

jyi
R:
j
== a: ÷
d

j=1
Z:
j
.
In fact, there exists b ÷ Zwith a: ÷ b:
i
÷ Z
d
¨
_
jyi
R:
j
_
=

jyi
Z:
j
.
By a rational change of coordinates we can achieve that :
i
= e
i
, i = 1. . . . . d.
Then : becomes a rational point in the interior of the standard unit cube Œ0. 1|
d
,
say : =
_
p
1
q
1
. . . . .
p
d
q
d
_
where 0 < p
i
< q
i
and gcd(p
i
. q
i
) = 1, i = 1. . . . . d.
The implication above means that if some coordinate of an integral multiple of :
is integral then all coordinates of : are integral. This is the same as saying that
q
1
= = q
d
> 1. Let q be the common value of the q
i
.
If q ¸ d ÷ 1,then 1 ÷ p
i
q
-1
¸ (d ÷ 1)
-1
for all i = 1. . . . . d and therefore
: ÷ d(d ÷1)
-1
Œ0. 1|
d
.
We are left with the case q ¸ d ÷2. For every index i the numbers p
i
. 2p
i
. 3p
i
.
. . . . (q ÷1)p
i
represent different residues modulo q. In particular, for every index
i the two systems of numbers
_
p
i
q
_
.
_
2p
i
q
_
. . . . .
_
(q ÷ 1)p
i
q
_
and q
-1
. 2q
-1
. . . . . (q ÷ 1)q
-1
coincide up to permutation. Here for a real umber a we put {a] = a÷]a¡. Since at
most (q ÷1)(d ÷1)
-1
among the numbers q
-1
. . . . . (q ÷1)q
-1
can be larger than
1 ÷ (d ÷1)
-1
there must exist a natural number c ¸ q such that
_
cp
i
q
_
÷
_
0.
d
d ÷1
_
. i = 1. . . . . d.
But then
{0] =
_
c: ÷Z
d
_
¨
_
d(d ÷1)
-1
Œ0. 1|
d
_
= 0.
For simplicity we exclude the trivial case d = 1 in the bound for t
d
:
Proposition 3.30. Let d > 1. Then
t
d
(j) ¸
d
2
j
(2ln 2)d
Proof. We have proved Theorem 2.67 by iterated stellar subdivision. It only re-
mains to be shown that we arrive at a unimodular triangulation ˙ whose cones D
satisfy the condition Hilb(D) Ç (d,2)j(C)
(2ln 2)d
, provided we choose the vectors
in the stellar subdivisions according to Lemma 3.29.
A stellar subdivision by a vector n as in Lemma 3.29 splits a cone D into sub-
cones D
t
such that j(D
t
) ¸ d(d ÷1)
-1
j(C) (compare the argument in the proof
of Theorem 2.67). Therefore we arrive at a unimodular cone after at most g = ]x¡
steps where x solves the equation
_
d
d ÷1
_
x
j(C) = 1.
102 3. Multiples of lattice polytopes
The Taylor series expansion ln(1÷t ) = t ÷t
2
,2÷t
3
,3÷ shows that ln(1÷t ) >
1
2
t for 0 < t < 1 so that g ¸ ]2d ln j(C)¡.
For the control of the lengths of the subdividing vectors n we use the sequence
h
k
, k ¸ ÷(d ÷ 1), of integers defined recursively as follows:
h
k
= 1. k ¸ 0. h
k
= h
k-1
÷ ÷h
k-d
. k ¸ 1.
One sees easily that this sequence is increasing, and that h
k
¸ d2
k-1
for all k ¸ 0
(note that d ¸ 2 by hypothesis).
We say that a triangulation ˙ of C is h-restricted if j(D) ¸ j(C) for each d-
cone D ÷ ˙ and, moreover, D satisfies the following condition: let j be the largest
integer such that
j(D) ¸
_
d
d ÷1
_
j
j(C):
then, in a suitable enumeration, the extreme integral generators u
1
. . . . . u
d
of D
satisfy the bound u
i
÷ h
j-i÷1
z
C
. (Note that j ¸ 0.)
The trivial triangulation of C is h-restricted, since j = 0 for C itself and
h
-(d-1)
. . . . . h
0
= 0. If we apply a stellar subdivision to an h-restricted trian-
gulation ˙ using Lemma 3.29, and the cone D with extreme integral generators
u
1
. . . . . u
d
is subdivided into subcones D
t
, then
v j(D
t
) ¸ d(d ÷1)
-1
j(D),
v n = q
1
:
1
÷ ÷q
d
:
d
÷ h
j
z
C
÷ ÷h
j-d÷1
z
C
= h
j÷1
z
C
.
In D
t
one of the extreme integral generators of D is replaced by n, and we order
the generators of D
t
by choosing n as the first and then the remaining ones in the
same order as in D. It follows immediately that the stellar subdivision of ˙ is again
h-restricted.
Iterated stellar subdivision therefore yields an h-restricted unimodular trian-
gulation. The result follows since
h
g
¸ d2
g-1
¸
d
2
2
2d ln µ(C)
=
d
2
j(C)
(2ln 2)d
.
Remark 3.31. (a) One can improve the bound for t
d
slightly if one observes that
(i) it would be sufficient to choose h
k
= d(d ÷1)
-1
(h
k-1
÷ ÷h
k-d
) and (ii)
ln(1 ÷t ) > l
t
t with liml
t
= 1. However, both measures do not decrease the order
of magnitude significantly, since they only replace the constant 2 ln 2 by a factor e
d
with lim
d-o
e
d
= ln 2.
(b) Proposition 3.30 improves [21, Th. 4.1] (for j sufficiently large). The for-
mulation of the proof given in [21] is not completely correct; it must be replaced
by an argument following the lines of the proof of Proposition 3.30.
Corner covers. Let C be a rational cone and : one of its extreme generators. We
say that a system {C
j
]
k
j=1
of subcones C
j
Ç C covers the corner of C at : if : ÷
Hilb(C
j
) for all j and the union
_
k
j=1
C
j
contains a neighborhood of : in C.
Lemma 3.32. Suppose that c
cone
d-1
< o, and let C be a simplicial rational d-cone
with extreme generators :
1
. . . . . :
d
.
3.C. Unimodular covers of multiples of polytopes 103
(a) Then there is a system of unimodular subcones C
1
. . . . . C
k
Ç C covering
the corner of C at :
1
such that Hilb(C
1
). . . . . Hilb(C
k
) Ç (c
cone
d-1
÷1)z
C
.
(b) Moreover, each element n = :
1
of a Hilbert basis of C
j
, j = 1. . . . . k, has
a representation n = ¸
1
:
1
÷ ÷¸
d
:
d
with ¸
1
< 1.
Proof. For simplicity of notation we set c = c
cone
d-1
. Let C
t
be the cone generated by
n
i
= :
i
÷ :
1
, i = 2. . . . . d, and let V be the vector subspace of R
d
generated by
the n
i
. We consider the linear map ¬ : R
d
÷ V given by ¬(:
1
) = 0, ¬(:
i
) = n
i
for i > 0, and endow V with a lattice structure by setting L = ¬(Z
d
). (One has
L = Z
d
¨ V if and only if Z
d
= Z:
1
÷(Z
d
¨ V ).) Note that :
1
. z
2
. . . . . z
d
with
z
j
÷ Z
d
form a Z-basis of Z
d
if and only if ¬(z
2
). . . . . ¬(z
d
) are a Z-basis of L.
This holds since Z:
1
= Z
d
¨R:
1
, and explains the unimodularity of the cones C
j
constructed below.
Note that n
i
÷ L for all i . Therefore z
C
0 Ç conv(0. n
2
. . . . . n
d
). The cone
C
t
has a unimodular cover (with respect to L) by cones C
t
j
, j = 1. . . . . k, with
Hilb(C
t
j
) Ç cz
C
0 . We “lift” the vectors x ÷ Hilb(C
t
j
) to elements ¯ x ÷ C as follows.
Let x = ˛
2
n
2
÷ ÷ ˛
d
n
d
(with ˛
i
÷ Q
÷
). Then there exists a unique integer
n ¸ 0 such that
¯ x := n:
1
÷x = n:
1
÷˛
2
(:
2
÷ :
1
) ÷ ÷˛
d
(:
d
÷ :
1
)
= ˛
1
:
1
÷˛
2
:
2
÷ ÷˛
d
:
d
with 0 ¸ ˛
1
< 1. If x ÷ cz
C
0 Ç c conv(0. n
2
. . . . . n
d
), then ¯ x ÷ (c ÷1)z
C
.
We now define C
j
as the cone generated by :
1
and the vectors ¯ x where x ÷
Hilb(C
t
j
). It only remains to show that the C
j
cover a neighborhood of :
1
in C.
To this end we intersect C with the affine hyperplane H through :
1
. . . . . :
d
. It is
enough that a neighborhood of :
1
in C ¨ H is contained in C
1
L L C
k
.
For each j = 1. . . . . k the coordinate transformation fromthe basis n
2
. . . . . n
d
of V to the basis x
2
. . . . . x
d
with {x
2
. . . . . x
d
] = Hilb(C
t
j
) defines a linear operator
on R
d-1
. Let M
j
be its [ [
o
norm.
Moreover, let N
j
be the maximum of the numbers n
i
, i ÷= 2. . . . . d defined by
the equation ¯ x
i
= n
i
:
1
÷x
i
as above. Choose c with
0 < c ¸
1
(d ÷ 1)M
j
N
j
. j = 1. . . . . k.
and consider
y = :
1
÷ˇ
2
n
2
÷ ÷ˇ
d
n
d
. 0 ¸ ˇ
i
< c.
Since the C
t
j
cover C
t
, one has ˇ
2
n
2
÷ ÷ ˇ
d
n
d
÷ C
t
j
for some j , and
therefore
y = :
1
÷;
2
x
2
÷ ÷;
d
x
d
.
where {x
2
. . . . . x
d
] = Hilb(C
t
j
) and 0 ¸ ;
i
¸ M
j
c for i = 2. . . . . d. Then
y =
_
1 ÷
d

i=2
n
i
;
i
_
:
1
÷;
2
¯ x
2
÷ ÷;
d
¯ x
d
104 3. Multiples of lattice polytopes
and
d

i=2
n
i
;
i
¸ (d ÷ 1)N
j
M
j
c ¸ 1.
whence
_
1 ÷

d
i=2
n
i
;
i
_
¸ 0 and y ÷ C
j
, as desired.
The bound for cones. The next lemma contains the argument that we need to
finish the proof of Theorem 3.23. Let us call a simplicial cone C empty if z
C
is an
empty simplex.
Lemma 3.33. Let C be an empty simplicial rational d-cone. Then C has a unimod-
ular cover by unimodular rational subcones D such that
Hilb(D) Ç (d ÷1) ;
d
t
d
(;
d
). ;
d
=

_
d ÷ 1
_
(d ÷ 1)
_
. (3.8)
In Theorem 3.23 we claim the same inequality as in Lemma 3.33, however for
arbitrary rational pointed d-cones. Therefore it is enough that C can be triangu-
lated into empty simplicial cones C
t
such that z
C
0 Ç z
C
. In fact, one first triangu-
lates C into simplicial cones generated by extreme generators of C. After this step
one can assume that C is simplicial with extreme generators :
1
. . . . . :
d
. If z
C
is
not empty, then we use stellar subdivision along a ray through some : ÷ z
C
¨Z
d
,
: = 0. :
1
. . . . . :
d
, and for each of the resulting cones C
t
the simplex z
C
0 has a
smaller number of integral vectors than z
C
. This completes the proof of 3.23.
Before we embark on the proof of Lemma 3.33, we single out a technical step.
Let {:
1
. . . . . :
d
] Ç R
d
be a linearly independent subset. Consider the hyperplane
H =
d

i=2
R:
1
÷(d ÷ 1):
i
Ç R
d
It cuts a simplex ı off the simplex conv(:
1
. . . . . :
d
) so that :
1
÷ ı. Let ˚ denote the
closure of
R
÷
ı \
__
(1 ÷R
÷
):
1
÷R
÷
e
2
÷ ÷R
÷
:
d
_
L z
_
Ç R
d
.
where z = conv(0. :
1
. . . . . :
d
). The polytope
˚
t
= ÷
1
d ÷ 1
:
1
÷
d
d ÷ 1
˚
is the homothetic image of the polytope ˚ under the dilatation with factor d,(d ÷
3:
2
3:
1
:
2
:
1
H
˚
ı
3z
Figure 3.6
3.C. Unimodular covers of multiples of polytopes 105
1) and center :
1
. (See Figure 3.6 for the case d = 2.) We will need that
˚
t
Ç (d ÷1)z. (3.9)
The easy proof is left to the reader.
Proof of Lemma 3.33. The lemma holds for d = 2 since empty simplicial 2-cones
are unimodular and the right hand side of (3.8) is 3,2 for d = 2. By induction we
can now assume that the lemma has been shown for all dimensions < d (the right
hand side of (3.8) is increasing with d). We set
; = ;
d
=
˙
_
d ÷ 1
_
(d ÷ 1) and × = (d ÷1) ;
d
t
d
(;
d
).
Let us first outline the course of the somewhat involved arguments following
now. They are subdivided into four major steps. The first three of them are very
similar to their analogues in the proof of Proposition 3.28. In Step 1 we cover the
d-cone C by d ÷ 1 smaller cones each of which is bounded by the hyperplane
that passes through the barycenter of conv(:
1
. . . . . :
d
) and is parallel to the facet
of conv(:
1
. . . . . :
d
) opposite of :
i
, i = 1. . . . . d. We summarize this step in Claim
A below.
In Step 2 Lemma 3.32 is applied for the construction of unimodular corner
covers. Claim B states that it is enough to cover the subcones of C “in direction” of
the cones forming the corner cover.
In Step 3 we extend the corner cover far enough into C. Lemma 3.25 allows us
to do this within a suitable multiple of z
C
. The most difficult part of the proof is
to control the size of all vectors involved.
However, Lemma 3.25 is applied to simplices 1 = conv(n
1
. . . . . n
e
) where
n
1
. . . . . n
e
span a unimodular cone of dimension e ¸ d. The cones over the uni-
modular simplices covering c1 have multiplicity dividing c, and possibly equal to
c. Nevertheless we obtain a cover of C by cones with bounded multiplicities. So we
can apply Proposition 3.30 in Step 4 to obtain a unimodular cover.
A convention: in the course of the proof we will sometimes have to deal with
sets of the form : ÷C where : ÷ R
d
which will be called a cone with apex :.
Step 1. The facet conv(:
1
. . . . . :
d
) of z
C
is denoted by 1
0
. (We use the letter 1 for
(d ÷ 1)-dimensional simplices, and z for d-dimensional ones.) For i = 1. . . . . d
we put
H
i
= aff(0. :
i
÷(d÷1):
1
. . . . . :
i
÷(d÷1):
i-1
. :
i
÷(d÷1):
i÷1
. . . . . :
i
÷(d÷1):
d
)
and
1
i
= conv
_
:
i
. 1
0
¨ H
i
_
.
Observe that :
1
÷ ÷:
d
÷ H
i
. In particular, the hyperplanes H
i
, i = 1. . . . . d
contain the barycenter of 1
0
, i. e. (1,d)(:
1
÷ ÷ :
d
). In fact, H
i
is the vector
subspace of dimension d÷1 through the barycenter of 1
0
that is parallel to the facet
of 1
0
opposite to :
i
. Clearly, we have the representation
_
d
i=1
1
i
= 1
0
, similar to
(3.7) above. In particular, each of the 1
i
is homothetic to 1
0
with factor (d ÷1),d.
To prove (3.8) it is enough to show the following
106 3. Multiples of lattice polytopes
Claim A. For each index i = 1. . . . . d there exists a system of unimodular cones
C
i1
. . . . . C
ik
i
Ç C
such that Hilb(C
ij
) Ç ×z
C
, j ÷= 1. . . . . k
i
, and 1
i
Ç
_
k
i
j=1
C
ij
.
The step from the original claim to the reduction expressed by Claim A seems
rather small – we have only covered the cross-section 1
0
by the 1
i
, and state that
it is enough to cover each 1
i
by unimodular subcones. The essential point is that
these subcones need not be contained in the cone spanned by 1
i
, but just in C.
This gives us the freedom to start with a corner cover at :
i
and to extend it far
enough into C, namely beyond H
i
. This is made more precise in the next step.
Step 2. To prove Claim A it is enough to treat the case i = 1. The induction
hypothesis implies c
cone
d-1
¸ × because the right hand side of the inequality (3.8) is
an increasing function of d. Thus Lemma 3.32 provides a system of unimodular
cones C
1
. . . . . C
k
Ç C covering the corner of C at :
1
such that
Hilb(C
j
) \ {:
1
. . . . . :
d
] Ç
_
×z
C
_
\ z
C
. j = 1. . . . . k. (3.10)
Here we use the emptiness of z
C
– it guarantees that Hilb(C
j
) ¨ (z
C
\ 1
0
) = 0
which is crucial for the inclusion (3.12) in Step 3.
With a suitable enumeration {:
j1
. . . . . :
jd
] = Hilb(C
j
), j = 1. . . . . k we have
:
11
= :
21
= = :
k1
= :
1
and
0 ¸ (:
jl
)
¡
1
< 1. j = 1. . . . . k. l = 2. . . . . d. (3.11)
where (÷)
¡
1
is the first coordinate of an element of R
d
with respect to the basis
:
1
. . . . . :
d
of R
d
(see Lemma 3.32(b)).
Now we formulate precisely what it means to extend the corner cover beyond
the hyperplane H
1
. Fix an index j = 1. . . . . k and let D Ç R
d
denote the simpli-
cial d-cone determined by the following conditions:
(i) C
j
Ç D,
(ii) the facets of D contain those facets of C
j
that pass through 0 and :
1
,
(iii) the remaining facet of D is in H
1
.
Figure 3.7 describes the situation in the cross-section 1
0
of C.
D
H
1
:
1
1
0
C
2
C
1
Figure 3.7
By considering all possible values j = 1. . . . . k, it becomes clear that to prove
Claim A it is enough to prove
3.C. Unimodular covers of multiples of polytopes 107
Claim B. There exists a system of unimodular cones D
1
. . . . . D
T
Ç C such that
Hilb(D
t
) Ç ×z
C
. t = 1. . . . . T and D Ç
T
_
t =1
D
t
.
Step 3. For simplicity of notation we put z = z
C
j
, H = H
1
. (Recall that z
is of dimension d, spanned by 0 and the extreme integral generators of C
j
.) The
vertices of z, different from0 and :
1
are denoted by n
2
. . . . . n
d
in such a way that
there exists i
0
, 1 ¸ i
0
¸ d, for which
(i) n
2
. . . . . n
i
0
÷ D \ H (‘bad’ vertices, on the same side of H as :
1
),
(ii) n
i
0
÷1
. . . . n
d
÷ C
j
\ D (‘good’ vertices, beyond or on H),
neither i
0
= 1 nor i
0
= d being excluded. In the situation of Figure 3.7 the cone
C
2
has two bad vertices, whereas C
1
has one good and one bad vertex. (Of course,
we see only the intersection points of the cross-section 1
0
with the rays from 0
through the vertices.)
If all vertices are good, there is nothing to prove since D Ç C
j
in this case. So
assume that there are bad vertices, i. e. i
0
¸ 2. We now show that the bad vertices
are caught in a compact set whose size with respect to z
C
depends only on d, and
this fact makes the whole proof work.
Consider the (d ÷ 1)-dimensional cone
E = :
1
÷R
÷
(n
2
÷ :
1
) ÷ ÷R
÷
(n
d
÷ :
1
).
In other words, E is the (d ÷ 1)-dimensional cone with apex :
1
spanned by
the facet conv(:
1
. n
2
. . . . . n
d
) of z opposite to 0. It is crucial in the follow-
ing that the simplex conv(:
1
. n
2
. . . . . n
d
) is unimodular (with respect to Z
d
¨
aff(:
1
. n
2
. . . . . n
d
)), as follows from the unimodularity of C
j
.
Due to the inequality (3.11) the hyperplane H cuts a (d ÷1)-dimensional (usu-
ally nonlattice) simplex off the cone E. We denote this simplex by 1 . Figure 3.8
illustrates the situation by a vertical cross-section of the cone C.
By (3.10) and (3.11) we have
1 Ç ˚ = R
÷
1
1
\
_
(:
1
÷C) L z
C
_
.
Let 0 be the dilatation with center :
1
and factor d,(d ÷ 1). Then by (3.9) we have
the inclusion
0(1 ) Ç (d ÷1)z
C
. (3.12)
One should note that this inclusion has two aspects: first it shows that 1 is not
too big with respect to z
C
. Second, it guarantees that there is some ç > 0 only
depending on d, namely ç = 1,(d ÷ 1), such that the dilatation with factor 1 ÷ç
and center :
1
keeps 1 inside C. If ç depended on C, there would be no control on
the factor c introduced below.
Let ˙
1
= conv(:
1
. n
2
. . . . . n
i
0
) and ˙
2
be the smallest face of 1 that contains
˙
1
. These are d
t
-dimensional simplices, d
t
= i
0
÷ 1. Note that ˙
2
Ç 0(˙
2
).
We want to apply Lemma 3.25 to the pair
;:
1
÷(˙
1
÷ :
1
) Ç ;:
1
÷(˙
2
÷ :
1
).
108 3. Multiples of lattice polytopes
C
j
H
D \ C
j
E
C \ D
z
C
n
i
1
:
1
R
÷
:
1
R
÷
:
2
÷ ÷R
÷
:
d
Figure 3.8
of simplices with the common vertex ;:
1
. The lattice of reference for the unimod-
ular covering is
L = L(;:
1
÷(˙
1
÷ :
1
)) = ;:
1
÷
i
0

j=2
Z(n
j
÷ :
1
).
Set
c =
1
d
and c =
d
d ÷ 1
; =
˙
_
d ÷ 1
_
d.
Since d
t
¸ d ÷1, Lemma 3.25 (after the parallel translation of the common vertex
to 0 and then back to ;:
1
) and (3.12) imply

2
Ç UC
L
_
;0(˙
2
)
_
Ç ;(d ÷1)z
C
. (3.13)
Step 4. Consider the i
0
-dimensional simplices spanned by 0 and the unimodular
(i
0
÷ 1)-simplices appearing in (3.13). Their multiplicities with respect to the i
0
-
rank lattice ZL
˙
1
are all equal to ;, since ˙
1
, a face of conv(:
1
. n
2
. . . . . n
d
) is
unimodular and, thus, we have unimodular simplices o on height ;. The cones
R
÷
o have multiplicity dividing ;. Therefore, by Proposition 3.30 we conclude that
the i
0
-cone R
÷
˙
2
is in the union ı
1
L L ı
T
of unimodular (with respect to the
lattice ZL
˙
1
) cones such that
Hilb(ı
1
). . . . . Hilb(ı
T
) Ç t
d
(;)z
R
C
˙
2
Ç (d ÷1) ; t
d
(;) z
C
= ×z
C
.
In view of the unimodularity of conv(:
1
. n
2
. . . . . n
d
), the subgroup ZL
˙
1
is a di-
rect summand of Z
d
. It follows that
D
t
= ı
t
÷R
÷
n
i
0
÷1
÷ ÷R
÷
n
d
. t = 1. . . . . T.
is the desired system of unimodular cones.
Part II
Affine monoid algebras
Chapter 4
Monoid algebras
4.A. Monoid algebras and graded rings
In this section we relax our standard convention on monoids: they are only assumed
to be commutative, but not necessarily cancellative or torsionfree.
Let M be a monoid. Furthermore let R be a commutative ring. In Section 2.A
we have already introduced the monoid algebra RŒM|. As an R-module it is free
with basis M. Therefore every element of ¡ = RŒM| has a unique representation
¡ =

x÷M
¡
x
x. ¡
x
÷ R.
in which ¡
x
= 0 for all but finitely many x ÷ M. In order to introduce the multi-
plication in RŒM| we write the monoid operation multiplicatively. With this con-
vention,
¡g =

x÷M

yz=x
¡
y
g
z
yz.
Very often the ring R of coefficients will be a field k. The elements x ÷ M are
called the monomials of RŒM|, whereas the elements of the form rx, r ÷ R, x ÷
M, are called terms.
Let M. N be monoids, and ç : M ÷ N a monoid homomorphism. Then ç
induces an R-algebra homomorphism ç
R
: RŒM| ÷RŒN| via
ç
R
_

x÷M
¡
x
x
_
=

x÷M
r
x
ç(x).
In other words: the R-linear extension of ç is an R-algebra homomorphism.
There is a natural augmentation of the R-algebra RŒM|, i. e. an R-algebra ho-
momorphism c
1
: RŒM| ÷ Rfor which the composition with the natural embed-
ding R ÷RŒM|, r ÷r1, is the identity on R. It is given by
c
1
_

x÷M
¡
x
x
_
=

x÷M
¡
x
.
111
112 4. Monoid algebras
Note that c
1
is induced by the monoid homomorphism M ÷ {1], x ÷ 1 for all
x ÷ M. If M has no nontrivial invertible element, for example if M is a positive
affine monoid, then
c
0
_

x÷M
¡
x
x
_
= ¡
1
.
is also an R-algebra augmentation of RŒM|.
We have already shown in Proposition 2.5 that RŒM| is finitely generated as an
R-algebra if and only if M is a finitely generated monoid. Therefore, if M is affine
and R is a noetherian ring, then RŒM| is a noetherian ring by the Hilbert basis
theorem.
Remark 4.1. The converse is also true: RŒM| is noetherian if and only if R is noe-
therian and M is finitely generated. See [36, Theorem 7.7]. That R is noetherian
along with RŒM|, follows readily from the fact that Ris a retract of RŒM|. One has
to work harder to show that M is finitely generated if RŒM| is noetherian.
If N is an M-module (see page 48), then RN, the free R-module with basis
N, is an RŒM|-module in a natural way.
The very easy proof of the next proposition is left to the reader. More or less it
uses only that tensor product commutes with direct sums.
Proposition 4.2. Let Rbe a ring, and M. N be monoids. Then
RŒM ˚N| Š RŒM| ˝
R
RŒN| Š
_
RŒM|
_
ŒN| Š
_
RŒN|
_
ŒM|
as R-algebras.
Graded rings and modules. We first introduce graded abelian groups, despite the
fact that the grading monoid only plays the role of an index set and its monoid
operation is not yet used:
Definition 4.3. Let G be an commutative monoid (written additively). A G-graded
abelian group Ais an abelian group together with a decomposition A =

g÷G
A
g
as a direct sum of abelian subgroups A
g
.
The abelian group A
g
are called the degree g graded or homogeneous compo-
nent of A, and the summand a
g
in the decomposition a =

g÷G
a
g
, a
g
÷ A
g
for
all g ÷ G, is the degree g homogeneous component of a ÷ A.
A graded subgroup of Ais a subgroup U such that U =

g÷G
U ¨ A
g
.
Let A. B be G-graded groups. A homomorphism ç : A ÷ B is homogeneous
(or graded) if ç(A
g
) Ç B
g
for all g ÷ G.
Note that we do not require A
g
= 0 for all g ÷ G. Therefore, a G-graded group
is a G
t
-graded group for every monoid containing G.
By Exercise 4.2 a subgroup is graded if it contains all the homogeneous compo-
nents of each of its elements.
Proposition 4.4. Let A. B be G-graded groups, and ç : A ÷ B a homogeneous
homomorphism. Then the kernel ç is a graded subgroup of A, and its image is
a graded subgroup of B. Moreover, Coker ç is naturally graded with components
B
g
,(Imç)
g
, g ÷ G.
4.A. Monoid algebras and graded rings 113
Definition 4.5. A G-graded ring is a ring Awhose underlying additive group is G-
graded in such a way that A
g
A
h
Ç A
g÷h
for all g. h ÷ G. A graded A-module is an
A-module E together with a decomposition M =

g÷G
E
g
such that A
g
M
h
Ç
M
g÷h
for all g. h ÷ G.
A ring or module homomorphism is homogeneous if it is homogeneous as a
homomorphism of abelian groups. Sub-objects such as ideals, subrings and sub-
modules, are graded if they are graded as subgroups.
Usually our rings are algebras over a ring Rof coefficients. A graded R-algebra
is an R-algebra A such that the structure morphism R ÷ A maps R to A
0
. Note
that the homogeneous components of Aare R-modules in this situation.
The monoid algebra RŒM| is the simplest example of an M-graded R-algebra.
Its homogeneous elements are exactly the terms rx with r ÷ R, x ÷ M.
Remark 4.6. Suppose that k is a field. Then every M-graded component of kŒM|
has k-dimension 1, and one may ask oneself whether there exist M-graded k-
algebras A different from kŒM| such that dimA
x
= 1 for all x ÷ M, and if so,
how to describe this family of algebras. See [83, Ch. 10] for more information on
this subject. However, it is easily seen that A Š kŒM| if all non-zero homogeneous
elements of Aare non-zero-divisors of A(Exercise 4.4).
Suppose that A is an abelian group with a fixed decomposition A =

i÷I
A
i
,
like a graded ring or module. Then we set
supp(a) = {i ÷ I : a
i
= 0]. a ÷ A. a =

i÷I
a
i
.
and call this set the support of a. The support of a subset B Ç A is supp B =
_
b÷B
supp(b).
Convention. The most common gradings are Z-gradings, and therefore we call
them simply gradings. A graded ring or module is Z-graded if not specified other-
wise.
While the rings under consideration are often Z
÷
-graded, it is useful to con-
sider them as Z-graded, since one cannot avoid negative degrees as soon as homo-
geneous elements are inverted.
Remark 4.7. Note that a grading ; on a monoid M (by definition a monoid ho-
momorphism fromM to Z) induces a grading on the monoid algebra RŒM|: take

;(x)=i
Rx as the graded component of degree i .
If ; is positive, then RŒM| is a positively graded R-algebra: all homogeneous
elements x have nonnegative degree and RŒM|
0
= R.
In particular, if M is a positive affine monoid, then M has a positive grading
(Proposition 2.15), and RŒM| can be considered as a positively graded algebra.
Monoid domains. That we call the elements x ÷ M monomials in RŒM| is justi-
fied by the following proposition.
Proposition 4.8. Let M be an affine monoid, gp(M) = Z
r
.
114 4. Monoid algebras
(a) Then there is an injective R-algebra homomorphismRŒM| ÷RŒX
˙1
1
. . . . .
X
˙1
r
| mapping the elements of M to (Laurent) monomials in the variables
X
1
. . . . . X
r
.
(b) If M is positive, then there exists an injective R-algebra homomorphism
RŒM| ÷ RŒX
1
. . . . . X
r
| mapping the elements of M to monomials in the
variables X
1
. . . . . X
r
.
Proof. Let e
1
. . . . . e
r
be the standard unit vectors of Z
r
. Then the assignment
e
i
÷ X
i
induces an isomorphism of Z
r
and the group of monomials X
a
1
1
X
a
r
r
,
(a
1
. . . . . a
r
) ÷ Z
r
. This group isomorphism induces an R-algebra isomorphism
RŒgp(M)| Š RŒX
˙1
1
. . . . . X
˙1
r
|, and therefore an embedding of RŒM| into the
Laurent polynomial ring.
If M is positive, then there exists an embedding M ÷ Z
r
÷
by Proposition
2.15. It induces an embedding RŒM| ÷ RŒX
1
. . . . . X
r
| that maps monomials to
monomials.
Corollary 4.9. Let R be an integral domain and M an affine monoid. Then RŒM|
is an integral domain.
In fact, the cancellative and torsionfree monoids have a simple characterization
in terms of their monoid algebras:
Theorem 4.10. Let M be a monoid which is not necessarily torsionfree or cancella-
tive, and let Rbe a ring. Then the following are equivalent:
(a) Ris a domain and M is torsionfree and cancellative.
(b) RŒM| is a domain;
Proof. (a) == (b) Let ¡. g ÷ RŒM|, ¡. g = 0. Since supp(¡ ) and supp(g) are
finite, there exists a finitely generated submonoid N of M such that ¡. g ÷ RŒN| Ç
RŒM|. Since RŒN| is an integral domain by Corollary 4.9, we have ¡g = 0.
(b) ==(a) Since R Ç RŒM| in a natural way, Ris a domain along with RŒM|.
If M is not cancellative, say xy = xz with x. y. z ÷ M, y = z, then x(y ÷ z) = 0
in RŒM|, but x. y ÷ z = 0. If M is cancellative, but not torsionfree, say x
n
= y
n
for x. y ÷ M, x
k
= y
k
for k < n, then 0 = x
n
÷ y
n
= (x ÷ y)(x
n-1
÷x
n-2
y ÷
÷y
n-1
), and the second factor is nonzero since all the monomials appearing in
it are pairwise different (M is cancellative).
In Chapter 8 we will need the following result on group rings. It is inserted
here, despite the fact that it concerns only the property of being reduced.
Theorem 4.11. Let R be an integral domain of characteristic 0 and G an abelian
group. Then RŒG| is reduced.
Proof. Since the nonzero elements of R are non-zerodivisors of RŒG|, we can in-
vert them and restrict ourselves to group rings kŒG| where k is a field of charac-
teristic 0. Moreover, since G is the union of its finitely generated subgroups, it is
enough to consider finitely generated groups.
Then G is the direct sum of cyclic groups, and kŒG| is the tensor product of
their group rings. It is enough to show that the factors are reduced since then the
4.A. Monoid algebras and graded rings 115
tensor product is reduced as well. In fact, each of the factors embeds into a direct
product of finitely many fields. Therefore it is enough that field extensions preserve
reducedness in characteristic 0; see Zariski and Samuel [102, Vol. II, p. 226].
Now let G be cyclic. Then kŒG| = kŒX
˙1
| if G Š Z, and kŒG| Š kŒX|,(X
m
÷1)
if G Š Z,mZ. But X
m
÷ 1 splits into a product of pairwise prime irreducible
polynomials if char k. Thus kŒG| is a tensor product of integral domains. Hence it
is reduced by what has been said above.
Groups of units. In general it is very difficult to determine the units in monoid
algebras and even in group rings (see Passman [68] for this classical problem), but
in the case in which we are mainly interested the task is easy.
Proposition 4.12. Let Rbe reduced and M a cancellative, torsionfree monoid. Then
the units in RŒM| are exactly the terms ux where u is a unit in R and x is an in-
vertible element of M, in other words, one has a natural isomorphism U(RŒM|) Š
U(R) × U(M).
Proof. A reduced ring R can be embedded into a product of fields. Therefore we
may assume that R is an integral domain. Let ¡ be a unit in RŒM|. Then there
exists a finitely generated submonoid M
t
of M such that ¡ and ¡
-1
are both con-
tained in the subalgebra RŒM
t
| of RŒM|. This argument reduces the nontrivial
implication to the case of an affine monoid M. As the last reduction step, we re-
place M by gp(M), which is isomorphic to Z
r
, r = rank M.
We are now in the situation where RŒM| is the Laurent polynomial ring
RŒX
˙1
1
. . . . . X
˙1
r
|, and an inductive argument shows that we may assume r = 1.
Then it is enough to observe that a product of two Laurent polynomials has at least
two terms if one of the factors has more than one term.
As a corollary we obtain that the group of units of a monoid is determined by
the monoid algebra RŒM|:
Corollary 4.13. Let Rbe reduced and M. N be cancellative, torsionfree monoids. If
RŒM| and RŒN| are isomorphic as R-algebras, then U(M) Š U(N).
In fact, we have an isomorphism U(RŒM|) Š U(RŒN|) that maps U(R) Ç
U(RŒM|) to U(R) Ç U(RŒN|). Therefore it induces an isomorphism
U(M) Š U(RŒM|), U(R) Š U(RŒN|), U(R) Š U(N).
Corollary 4.13 is the first sign of the isomorphism theorem to be proved in Chap-
ter 5.
Monomial orders and Newton polytopes. The group Z
r
can be ordered in such a
way that it becomes an ordered group. A suitable such order is the lexicographic
one: (a
1
. . . . . a
r
) <
lex
(b
1
. . . . . b
r
) if and only the first nonzero component of a ÷b
is negative. Since the elements of Z
r
often appear as monomials we will call such
an order monomial.
If A is an abelian group graded by a group G Š Z
r
, then we may speak of the
leading term or component of an element of A, once a monomial order <on G has
116 4. Monoid algebras
been introduced: the leading term of a =

g÷G
a
g
is a
h
with h = max(supp a),
provided a = 0.
The following lemma is trivial, but very useful:
Lemma 4.14. Let R be a G-graded ring and < a term order on G. Let R be a G-
graded ring and M a graded R-module, r an element of R with leading term r
g
and x an element of M with leading termx
h
.
(a) If r
g
x
h
= 0, then r
g
x
h
is the leading term of rx.
(b) In particular, if Ris a domain and M is a torsionfree R-module, then r
g
x
h
is the leading term of rx.
Remark 4.15. By a theoremof Levi, every torsionfree abelian group can be ordered;
see [36, Section 3]. We will not need this “transfinite” fact, since we can usually
restrict ourselves to considering finitely generated subgroups of the grading group.
A more refined approach to the combinatorial structure is provided by Newton
polytopes. If a ring Ror a module M is Z
r
-graded as above, we define the Newton
polytope of an element a ÷ Ror M by
N(a) = conv(supp a).
The figure shows the Newton polytope of a Laurent polynomial in two variables.
N(¡ )
¡ = X
-1
÷Y
-1
÷XY
-
1
÷X
2
÷XY
2
÷X
-1
Y
Figure 4.1. A Newton polytope
The next lemma connects the Newton polytopes of r and x with the Newton
polytope of their product. We use the information on the vertices of Minkowski
sums of polytopes given in Exercise 1.1.
Lemma 4.16. Let G be an abelian group, G Š Z
r
. Let R be a G-graded ring and
M a graded R-module, a an element of Rand x an element of M.
(a) Suppose g is a vertex of N(a) and h is a vertex of N(x) such that g ÷h is a
vertex of N(a) ÷N(x). If a
g
x
h
= 0, then g ÷h is a vertex of N(ax).
(b) In particular, if R is a domain and M is a torsionfree R-module, then
N(ax) = N(a) ÷N(x).
Proof. One has supp(ax) Ç supp(a) ÷supp(x), and conv(supp(a) ÷supp(a)) =
N(a) ÷N(x). If g ÷h = g
t
÷h
t
for g
t
÷ supp(a) and h
t
÷ supp(x), then g
t
= g,
h
t
= h. Therefore g ÷h ÷ supp(ax), and g ÷h is a vertex of N(ax), since it is a
vertex of the polytope N(a) ÷N(x) ¸ N(ax).
Every vertex of the Minkowski sum of polytopes is the sum of vertices of the
summands. Under the hypothesis all vertices of N(a) ÷N(x) belong to N(ax), so
that N(ax) = N(a) ÷N(x).
4.B. Representations of monoid algebras 117
Krull dimension. We recall the most important notions of dimension theory. The
Krull dimension of a ring R is the supremum of the numbers n for which there
exists a strictly ascending chain p
0
Ç p
1
Ç Ç p
n
. The height ht p of a prime
ideal p is the Krull dimension of the local ring R
p
. The height of an arbitrary ideal
I is the minimum of the heights of the prime ideals p ¸ I. If I is an ideal in a
noetherian ring R generated by m elements, then ht p ¸ m for all minimal prime
overideals p ¸ I by Krull’s principal ideal theorem. (See Bourbaki [10, Ch. VIII]
for the basic notions of dimension theory.)
If the ring R of coefficients is a field k, then it is very easy to determine the
Krull dimension of an affine monoid domain kŒM|:
Proposition 4.17. Let k be a field and M an affine monoid. Then dimkŒM| =
rank M.
Proof. We have already shown that kŒM| is an affine k-domain. The Krull dimen-
sion of such a domain is the transcendence degree of its field of fractions over k
(see Kunz [57, Ch. 2, §3]). Let r = rank M. Then QF(kŒM|) = QF(kŒgp(M)|) Š
QF(kŒZ
r
|) Š QF(kŒX
1
. . . . . X
n
|), and so dimkŒM| = r.
Theorem 4.18. Let R be a noetherian ring, and M an affine monoid. Then
dimRŒM| = dimR÷rank M.
Proof. Set S = RŒM|. Consider a strictly ascending chain of prime ideals p
0
Ç
p
1
Ç Ç p
n
in R. Then the ideals p
i
S form a strictly ascending chain in RŒM|,
too, and they are prime ideals since S,p
i
S Š (R,p
i
)ŒM|.
Suppose that m = p
n
is a maximal ideal, and set k = R,m. Then kŒM| Š
S,mS has Krull dimension r = rank M, and we can extend the chain of prime
ideals in S by r more members, which we obtain as preimages in S of the mem-
bers of a chain 0 = q
1
Ç Ç q
r
of prime ideals in kŒM|. Therefore dimS ¸
dimR÷r.
It remains to show that dimS ¸ dimR÷r if dimR < o. Choose a maximal
ideal M of S, and set p = M¨ R. Then all elements in R\ p do not belong to M,
and therefore S
M
is a localization of R
p
ŒM|. We thus have dimS
M
¸ dimR
p
÷
dimS
M
,pS
M
. See [63, Theorem15.1] for this inequality; it follows very easily from
Krull’s principal ideal theorem.
Therefore it remains to show that dimS
M
,pS
M
¸ r. This follows again from
Proposition 4.17, since we can view S
M
,pS
M
as a localization of (R
p
,pR
p
)ŒM|.
But R
p
,pR
p
is a field.
Remark 4.19. Theorem4.18 cannot be generalized to arbitrary rings of coefficients.
In fact, it can happen that dimRŒX| > dimR÷1 (see Seidenberg [74]). For more
information on the Krull dimension of monoid rings we refer the reader to Gilmer
[36].
4.B. Representations of monoid algebras
In this section we relax our standard convention on monoids: they are only assumed
to be commutative, but not necessarily cancellative or torsionfree.
118 4. Monoid algebras
So far we have described affine monoids M as submonoids of lattices, and usu-
ally this description is very efficient, especially for normal monoids. However, it is
often necessary to use a representation of M as a quotient of a free commutative
monoid whose basis corresponds to a system N of generators of M. Such a rep-
resentation of M gives rise to a representation of the monoid algebra RŒM| as a
residue class ring of a polynomial ring RŒX
x
: x ÷ N| over R. The R-algebra ho-
momorphism¬ : RŒX
x
: x ÷ N| ÷RŒM| is induced by the substitution X
x
÷x.
If M is given, say, as a submonoid of a lattice Z
n
, then the main task is to calcu-
late the kernel of ¬. In this section we write the monoid operation multiplicatively,
unless indicated otherwise.
Congruences. Quotients of monoids are determined by congruence relations. A
congruence relation on a monoid M is a set C of pairs (x. y) ÷ M
2
such that (i) C
is an equivalence relation, and (ii) (zx. zy) ÷ C for all (x. y) ÷ C, z ÷ M. Instead
of (x. y) ÷ C we will also write x - y.
The set of congruence classes of C form a monoid if we set ¨ x ¨ y = xy. This
monoid is denoted by M,C. (It need not be cancellative or torsionfree, even if M
is so.)
Let ç : M ÷ N be a monoid homomorphism. Then ç induces a congruence
by x - y if and only if ç(x) = ç(y). As in the case of abelian groups we have
Imç Š M,C in a natural way.
Asubset E of C generates C if C is the smallest congruence relation containing
E. (This definition makes sense since M
2
is a congruence relation, and since the
intersection of congruence relations is again a congruence relation.)
The congruence relation generated by E can also be described constructively.
In order to capture symmetry let us set
¯
E = EL{(y. x) : (x. y) ÷ E]. The reader
may check (Exercise 4.6(a)) that E generates C if and only if for each (x. y) ÷ C,
x = y, there exist n ÷ N, (u
1
. :
1
). . . . . (u
n
. :
n
) ÷
¯
E, and n
1
. . . . . n
n
÷ M such
that x = n
1
u
1
, n
1
:
1
= n
2
:
2
. . . . . n
n-1
:
n-1
= n
n
u
n
, n
n
:
n
= y.
When M is an abelian group, we are on familiar territory:
Proposition 4.20. Let ç : M ÷ N be a homomorphism of abelian groups and E a
subset of G. Then the pairs (e. 1), e ÷ E, generate the congruence relation defined
by ç if and only if E generates Ker ç.
The proof is left to the reader (Exercise 4.6(b)).
Congruences and binomial ideals. We now relate the congruence of a monoid
homomorphism to the kernel of the associated algebra homomorphism ç
R
:
Proposition 4.21. Let ç : M ÷ N be a monoid homomorphism and R a ring.
Then the kernel of ç
R
: RŒM| ÷RŒN| is generated by the binomials z
1
÷z
2
where
z
1
. z
2
÷ M and ç(z
1
) = ç(z
2
).
Proof. Let ¡ =

x÷M
¡
x
x ÷ RŒM|. Then we have
ç
R
(¡ ) =

x÷M
¡
x
ç(x) =

y÷N
_

ç(x)=y
¡
x
_
y.
4.B. Representations of monoid algebras 119
Therefore ç
R
(¡ ) = 0 if and only if

ç(x)=y
¡
x
= 0 for all y ÷ N.
It is impossible that ¡
x
= 0 for a single x. If ¡
x
= 0 for at least two monomials
x
1
. x
2
, then we use induction, passing from ¡ to ¡ ÷ ¡
x
1
(x
1
÷ x
2
).
The description of Ker ç
R
in the proposition depends solely on M, and not on
the ring Rof coefficients. This observation can be extended further:
Proposition 4.22. With the notation of Proposition 4.21 let S be a set of binomials
x
1
÷ x
2
÷ Ker ç
R
. Set
¯
S = S L ÷S. Then the following are equivalent:
(a) S generates the ideal Ker ç
R
.
(b) For every binomial y
1
÷ y
2
÷ Ker ç
R
, y
1
÷ y
2
= 0, there exist n ÷ N,
binomials x
11
÷ x
12
. . . . . x
n1
÷ x
n2
÷
¯
S and monomials z
1
. . . . . z
n
such
that y
1
= z
1
x
11
, z
1
x
12
= z
2
x
21
. . . . . z
n-1
x
n-1,2
= z
n
x
n1
, z
n
x
n2
= y
2
.
(c) The pairs (x
1
. x
2
) with x
1
÷ x
2
÷ S generate the congruence relation on
M defined by ç.
Proof. The implication (b) ==(a) is very easy. First, ÷S is contained in the ideal
generated by S, and, second, we have the telescope sum
y
1
÷ y
2
=
n

i=1
z
i
(x
i1
÷ x
i2
).
This shows that the ideal generated by S contains all the binomials in Ker ç
R
, and
by Proposition 4.21 this is sufficient.
For the implication (a) ==(b) let us write
y
1
÷ y
2
=
n

i=1
¡
i
(x
i1
÷ x
i2
) (4.1)
with x
i1
÷x
i2
÷
¯
S and ¡
i
÷ RŒM|. Using each binomial several times if necessary,
we can assume that every coefficient ¡
i
is in fact a term in RŒM|, say ¡
i
= r
i
z
i
with r
i
= 0. Now we define a graph G whose vertices are the monomials z
1
x
i1
and
z
i
x
i2
, i = 1. . . . . n, and whose edges correspond to the binomials z
i
(x
i1
÷ x
i2
),
connecting z
i
x
i1
and z
i
x
i2
.
We can split the sum on the right hand side in (4.1) into sums each of which
only contains all the binomials corresponding to edges in a connected component
of G. If a connected component contains neither y
1
nor y
2
, then the correspond-
ing sum is 0 and can be omitted completely. After this simplification only one
connected component remains. In fact, if not, then the sum over one component
would yield y
1
and that over the other ÷y
2
. However, neither y
1
nor y
2
belongs to
Ker ç
R
.
We choose a path from y
1
to y
2
in the graph, and this is exactly what (b) re-
quires us to do. (Since b. ÷b ÷
¯
S for b ÷ S, we can use an edge of G in both
directions.)
The equivalence of (b) and (c) is evident. In fact, (b) is only a reformulation of
(c) in terms of binomials.
120 4. Monoid algebras
Corollary 4.23. With the notation of Proposition 4.21, if M is finitely generated,
then the congruence defined by ç, equivalently, the ideal Ker ç
R
is finitely generated.
Proof. In fact, the proposition shows that we are free in choosing the ring of co-
efficients in proving the corollary, and so we choose R = k to be a field. Then
RŒM| is a noetherian ring, Ker ç is a finitely generated ideal, and every system of
generators of Ker ç contains a finite system of generators.
Toric ideals. Suppose that M is affine and x
1
. . . . . x
n
is a finite system of gen-
erators of M. Then Ker ç
R
is called a toric ideal (see [64]). As just noted, it is
generated by finitely many binomials. However, in general it is a nontrivial prob-
lem to compute such a system of binomials, or even a minimal one.
One can split this task into two steps. The first is to extend ç
R
to a homomor-
phism of Laurent polynomial rings ç
-
R
: RŒgp(M)| ÷ RŒgp(M)|. This extension
exists for two (equivalent) reasons: (i) from the ring-theoretic point of view, the
monoid elements form a multiplicatively closed set, and so we may pass to the
ring of fractions by adjoining their inverses; (ii) from the monoid-theoretic point
of view, the monoid homomorphism ç : M ÷ N can be extended to a group
homomorphism ç
-
: gp(M) ÷ gp(N), which then gives rise to the R-algebra
homomorphismç
-
R
. (Note that the homomorphismM ÷gp(M) injective only if
M is cancellative.)
Proposition 4.24. Let ç : M ÷ N be a homomorphism of monoids, and R a ring.
Suppose that M is generated by x
1
. . . . . x
n
and that Ker ç
-
be generated by the
elements n
1
. . . . . n
m
÷ gp(M), n
i
= y
i
z
-1
i
, with y
i
. z
i
÷ M. Let I be the ideal in
RŒM| generated by the binomials b
i
= y
i
÷ z
i
. Then:
(a) IRŒgp(M)| is the kernel of the induced homomorphismç
-
R
: RŒgp(M)| ÷
RŒgp(N)|.
(b)
Ker ç
R
= {¡ ÷ RŒM| : (x
1
x
n
)
k
¡ ÷ I for some k ÷ N].
Proof. Let us first conclude (b) from(a). By (a) we have Ker ç
-
R
= IRŒgp(M)|. On
the other hand, by general localization arguments, Ker ç
-
R
= (Ker ç
R
)RŒgp(M)|,
and Ker ç
R
= Ker ç
-
R
¨ RŒM|. Thus Ker ç
-
R
= IRŒgp(M)| ¨ RŒX|, and now
(b) follows by the general formula for the contraction of an ideal extended to a
localization, since RŒgp(M)| = RŒM|Œ(x
1
x
n
)
-1
|.
For (b) we note first that IRŒgp(M)| is generated by the binomials b
t
i
= n
i
÷1,
since we can multiply by z
-1
i
in RŒgp(M)|.
On the other hand, by Proposition 4.20, the pairs (n
i
. 1) generate the congru-
ence relation defined by ç
-
. Then Proposition 4.22 implies that b
t
1
. . . . . b
t
m
generate
Ker ç
-
R
.
Let us consider the case in which M and N are affine. We switch to additive
notation in the monoids and exponential notation in the monoid algebras. Iden-
tifying gp(M) with Z
n
and gp(N) with Z
r
, we can easily compute a basis of the
kernel U of the group homomorphism Z
n
÷ Z
r
, for example using the Gaußian
4.B. Representations of monoid algebras 121
algorithm over Q and the elementary divisor algorithm over Z. In this way one
finds the binomials b
1
. . . . . b
m
, m = n ÷r, very quickly. Then it remains the much
more difficult task to compute the ideal I : (X
X
1
X
x
n
)
o
where we have used
a standard notation from commutative algebra to denote the ideal in (b). Several
special algorithms have been devised for this task; see Bigatti, La Scala and Rob-
biano [7].
Example 4.25. Let N be the submonoid of Z
2
generated by (3. 0). (2. 1). (1. 2).
(0. 3). Then rank N = 2, and the kernel U of the homomorphism Z
4
÷ Z
2
that
maps the unit vectors in Z
4
to the generators of M in the given order has rank 2.
Clearly e
1
÷2e
2
÷e
3
. e
2
÷2e
3
÷e
4
is a basis of U. With the notation of Proposition
4.24 we can choose I = (X
2
2
÷ X
1
X
3
. X
2
3
÷ X
2
X
4
). However, I is by no means
the full kernel of the homomorphism kŒX
1
. X
2
. X
3
. X
4
| ÷ kŒM|: we have to add
X
1
X
4
÷ X
2
X
3
to I.
We conclude this section by a characterization of those ideals I in polynomial
rings P = kŒX
1
. . . . . X
n
| for which P,I is an affine monoid algebra kŒM| in such
a way that M is the set of the residue classes of the monomials in P.
Theorem 4.26. Let k be a field and I an ideal in the polynomial ring P =
kŒX
1
. . . . . X
n
|. Let Lbe the Laurent polynomial ring kŒX
˙1
1
. . . . . X
˙1
n
|.
Let M be the set of monomials in P and ¬ : P ÷ A = P,I the natural
homomorphism. Then the following are equivalent:
(a) Ais an affine monoid algebra with underlying monoid ¬(M);
(b) I is a prime ideal generated by binomials X
u
÷ X
¡
;
(c) I = IL¨P, and the subgroup U of Z
n
generated by the differences u ÷:,
u. : ÷ Z
n
÷
, with X
u
÷ X
¡
÷ I is a direct summand of Z
n
.
Proof. (a) == (b) Since A is a domain, I must be a prime ideal. That I is gener-
ated by binomials has been shown above.
(b) == (c) Consider the congruence relation C on cM defined by the pairs
(u. :) with X
u
÷ X
¡
÷ I. It defines the factor monoid M,C (that, a priori, need
not be cancellative or torsionfree). Proposition 4.22 shows that I is in fact the
kernel of the natural map P ÷ kŒM,C|. we can identify M,C with ¬(M).
Therefore P,I is the free R-module with basis ¬(M).
The last statement implies that all monomials X
a
are nonzero modulo I. Since
I is a prime ideal, they are non-zerodivisors modulo I. Therefore I = J ¨ P
where J = IL.
Since I is an ideal, U is a subgroup, as the reader may check. The extension of
a prime ideal to a ring of fractions is again prime. Therefore J is a prime ideal in
L.
Let n ÷ Z
n
and m ÷ N such that mn ÷ U. We must show n ÷ U. Clearly
X
mu
÷ 1 ÷ J. Let p = char k. Suppose first that p > 0, p [ m. Then X
mu
÷ 1 =
(X
mu{p
÷ 1)
p
÷ J, and so X
mu{p
÷ 1 ÷ J since J is a prime ideal. Replacing m
by m,p, we argue by induction.
Suppose now that p = 0 or p > 0, p m. We have
X
mu
÷ 1 = (X
u
÷ 1)(X
(m-1)u
÷X
(m-2)u
÷ ÷1).
122 4. Monoid algebras
and the second term is not in the maximal ideal (X
1
÷ 1. . . . . X
n
÷ 1) ¸ J. A
fortiori, it does not belong to the prime ideal J, and we conclude that X
u
÷1 ÷ J.
But then n ÷ U.
(c) == (a) Consider the composition P ÷ L ÷ kŒZ
n
,U|. Its kernel is I.
Furthermore the image of M in Z
n
,U is an affine monoid.
The reader may check that the theorem remains true if P is an arbitrary affine
monoid domain and Lis replaced by kŒgp(M)| (Exercise 4.7).
4.C. Monomial prime and radical ideals
We start with a general result on prime ideals in graded rings. For the case G = Z
it can be found in [24, 1.5.6]. The generalization to arbitrary grading groups is not
difficult since all groups Z
n
can be ordered.
As usual, we set
Supp M = {p ÷ Spec R : M
p
= 0].
Ass M = {p ÷ Spec R : p = Ann x for some x ÷ M].
Here M is a module over a ring R, and Spec Rdenotes the set of prime ideals of R.
If M = 0 is a finitely generated module over a noetherian ring, then Supp M = 0,
and the minimal elements of Supp M belong all to Ass M. Moreover, Supp(M)
consists exactly of the prime ideals that contain the annihilator Ann M.
Lemma 4.27. Let G be a torsionfree abelian group and Rbe a G-graded ring. Let p
be a prime ideal in R, and q the ideal generated by all homogeneous elements in p.
(a) q is a prime ideal.
(b) Let M be a graded R-module.
(i) If p ÷ Supp M, then q ÷ Supp M.
(ii) If p ÷ Ass M, then p = q. Furthermore p is the annihilator of a
homogeneous element of M.
The lemma is the basis for the characterization of graded rings in terms of
their graded prime ideals; see [24] for results of this type. Its part (a) can easily be
extended to radical ideals and primary ideals (Exercise 4.8).
The monomial prime or radical ideals in monoid algebras can be described in
purely combinatorial terms.
Proposition 4.28. Let M be a monoid and R an integral domain (a reduced ring).
Then a monomial ideal p Ç RŒM| is a prime (radical) ideal if and only if I = M¨p
is a prime (radical) ideal in the monoid M.
Proof. We give the proof for prime ideals; that for radical ideals is even simpler.
Suppose p is a prime ideal. Since it is a monomial ideal, p =

x÷I
Rx. Let
y. z ÷ M with yz ÷ I. Then yz ÷ p and so y ÷ p or z ÷ p, which in turn implies
y ÷ I or z ÷ I, and so I is a prime ideal.
Conversely, let I Ç M be a prime ideal. Evidently p =

x÷I
Rx is an ideal.
Let ¡. g ÷ R such that ¡g ÷ p. As in the proof of the lemma we pass to a finitely
4.C. Monomial prime and radical ideals 123
generated submonoid of M, and can therefore assume that M is ordered (by re-
stricting the lexicographic order on gp(M) Š Z
r
to M). Stripping off the terms
of ¡ and g that belong to monomials in I, we can assume that the smallest mono-
mials of ¡ and g lie not in I. But then the smallest monomial of ¡g is not in I,
and it follows that ¡g … p.
The prime ideals and radical ideals in affine monoids M have already been
determined: they are just the complements of faces and unions of faces of R
÷
M,
respectively (see Proposition 2.19). For a face F of R
÷
M we set
p
F
= R{M \ F].
Corollary 4.29. Let R be an integral domain and M an affine monoid. Then the
ideals p
F
are exactly the monomial prime ideals in RŒM|, and their intersections
are exactly the radical monomial ideals.
Moreover, for each face F the embedding RŒM ¨ F| induces the decomposition
RŒM| = RŒM ¨ F| ˚p
F
as an RŒM ¨ F|-module, and RŒM ¨ F| is a retract of
RŒM|.
Proof. The first part has been proved already. For the second we note that
RŒM| = RŒM ¨ F| ˚ p
F
as an R-module, and even as a RŒM ¨ F|-module.
Since p
F
is an ideal, the projection RŒM| ÷ RŒM ¨ F| with kernel p
F
is a
R-algebra morphism that restricts to the identity on RŒM ¨ F|.
The reader should note that despite of its naturalness the face projection ¬
F
:
RŒM| ÷ RŒM ¨ F| is not induced by a monoid morphism. However, it maps
terms of RŒM| to terms of RŒM ¨ F|, and is therefore homogeneous with respect
to the M-grading.
Corollary 4.30. With the notation of Corollary 4.29 let I be an ideal in RŒM| gen-
erated by monomials. Then the minimal prime ideals of I are the ideals p
F
where
F runs through the set of faces that are maximal with respect to F ¨ I = 0.
Proof. We only need to combine Lemma 4.27 with Corollary 4.29.
Finally we determine the heights of the ideals p
F
.
Proposition 4.31. Let R be an integral domain, and M an affine domain. Then
ht p
F
= rank M÷dimF. If Ris noetherian, then dimRŒM|,p
F
= dimR÷dimF.
Proof. One has ht p
F
= dimRŒM|
p
F
. Since R ¨ p
F
= 0, we can first invert all
nonzero elements of R, and replace R by its field k of fractions. This reduces the
first equation to a field k of coefficients. In an affine algebra A over a field one
has dimA = ht p ÷ dimA,p for all prime ideals ([57, Ch. II, 3.6]), and the first
equation follows from the second since dimRŒM| = dimR ÷ rank M. But the
second equation is just this formula, if one uses that dimF = rank M ¨ F and
RŒM|,p
F
Š RŒM ¨ F|.
Filtrations and decompositions. Let N = 0 be a finitely generated module over a
noetherian domain. Then N has a filtration
0 = N
0
Ç N
1
Ç Ç N
n
= N
124 4. Monoid algebras
in which the successive quotients N
i
,N
i-1
are of type R,p
i
for prime ideals p
i
,
i = 1. . . . . n (see [10, Ch. IV, §1, Th. 1]). Moreover, each associated prime of N
appears among the p
i
(and this fact proves the finiteness of Ass(N)). In the graded
case Lemma 4.27 allows one to find such a filtration in which each N
i
is a graded
submodule generated by a single homogeneous element modulo N
i-1
, and the p
i
are (therefore) graded prime ideals. In fact, one starts the filtration by choosing N
1
being generated by a homogeneous element x whose annihilator is an associated
prime of N, and applies the same principle to N,N
1
etc. The ascending chain
condition on submodules ensures that the process stops after finitely many steps.
Now let M be an affine monoid, N a finitely generated M-module (see 48),
and U an M-submodule of N. Then we can choose a field k of coefficients and
apply the filtration argument to the kŒM|-module (kN),(kU).
Proposition 4.32. Let G be a torsionfree group, M Ç G an affine monoid, N Ç G
a finitely generated M-module, and U a submodule of N (possibly U = 0). Then
N \U decomposes into a disjoint union of subsets of type x÷(M¨F) where x ÷ N
and F is a face of R
÷
M.
Moreover, if F is maximal with respect to the condition F ¨{y ÷ M : y ÷N Ç
U] = 0, then a subset of type x ÷F appears in the decomposition.
We leave the detailed verification of the proof to the reader. The essential points
to observe are
(a) (kN),(kU) has a filtration by submodules of type (kU
i
),kU where U
i
¸
U is an N-submodule of N;
(b) the filtration induces a decomposition N \ U =
_
U
i
\ U
i-1
;
(c) U
i
\ U
i-1
is of type F ÷x if (kU
i
),kU
i-1
is of type kŒM|,p
F
;
(d) the faces F that are maximal with respect to the condition F ¨ {y ÷
M : y ÷ N Ç U] = 0 correspond to the minimal prime ideals in
Supp
_
(kN),(kU)
_
, and these prime ideals belong to Ass
_
(kN),(kU)
_
.
When applied to an affine monoid M and its normalization
¨
M, the proposi-
tion yields Corollary 2.34(b). Part (a) of this corollary follows easily from the fact
that NŒx
-1
| = 0 if and only if x ÷ Rad(Ann M), provided N is a finitely gen-
erated module over a noetherian ring R. This follows easily from the fact that
Supp N is the set of prime ideals containing Ann N, and that Supp NŒx
-1
| = {p ÷
Supp N. x … p].
Stanley-Reisner rings and polyhedral algebras. Let k be a field, and I a mono-
mial radical ideal in kŒX
1
. . . . . X
n
|. As we have just seen, the set of monomials in I
forms a radical ideal in the monoid Z
n
÷
, and in its turn a radical ideal in Z
n
÷
cor-
responds to a union of faces of the cone R
n
÷
. The set of all faces F with F ¨ I = 0
evidently forms a subfan F of the face lattice of R
n
÷
. Each of the cones in F is sim-
plicial, namely generated by the standard unit vectors e
i
it contains. In each face
of F these unit vectors span a simplex, and these simplices together form a sim-
plicial complex z, embedded in R
n
. The simplicial complex zis the most efficient
combinatorial representation of I.
4.C. Monomial prime and radical ideals 125
Conversely, let z be an abstract simplicial complex (see Example 1.49) on a
subset of V = {1. . . . . n]. Then z determines a monomial radical ideal I(z) in
kŒX
1
. . . . . X
n
|, namely that generated by all squarefree monomials X
i
1
X
i
r
, 1 ¸
r ¸ n, i
1
< < i
r
that do not belong to z. This shows
Proposition 4.33. The assignment z ÷ I(z) yields a bijective correspondence be-
tween the abstract simplicial complexes whose vertex set is contained in {1. . . . . n]
and the monomial radical ideals in kŒX
1
. . . . . X
n
|.
It is trivial that I(z) is generated by those monomials X
i
1
X
i
k
that are min-
imal with respect to divisibility among the monomial in I(z). They correspond
to the minimal “nonfaces” of z, i. e. the minimal subsets of {1. . . . . n] that do not
belong to z.
One calls kŒz| = kŒX
1
. . . . . X
n
|,I(z) the face ring or Stanley-Reisner ring of
z. Stanley-Reisner rings form one of the backbones of combinatorial commutative
algebra. We refer to Bruns and Herzog[24], Miller and Sturmfels [64] and Stanley
[80] for their theory.
More generally, let us consider rational cones C in R
n
. Then M = C ¨ Z
n
is an integrally closed submonoid in Z
n
, and the monomial radical ideals in kŒM|
correspond to the subfans of the fan of all faces of C. Let us call such fans rational
boundary fans. Conversely each monomial radical ideal I in kŒM| for an integrally
closed submonoid M of Z
n
determines a rational boundary fan, and again there
is a bijective correspondence between the residue class rings kŒM|,I and ratio-
nal boundary fans. (Residue class rings with the same monomial basis must be
identified.)
These constructions can be varied and generalized in several ways. If ˘ is a
lattice polytopal complex, then one can “glue” the polytopal algebras kŒP|, P ÷ ˘,
in a natural way to a k-algebra kŒ˘|. The monomial basis of kŒ˘| is formed by the
union of the monomials in the monoids M(P), identified along common faces,
and the product of monomials x and y that belong both to one of the monoids
M(P), P ÷ ˘, is simply xy ÷ M(P). If no such polytope exists, then xy = 0.
The algebras obtained in this way are the polyhedral algebras, introduced in [19].
If F is a fan of rational cones C in R
n
, then one analogously “glue” the algebras
kŒC ¨ Z
n
| along common faces. This class has been discussed by Stanley in [79].
From an abstract point of view all these constructions can be seen as special
cases of inverse limits of diagrams of rings over finite posets (˘. ¸). Such a di-
agram is a family of rings R
p
, p ÷ ˘, together with a family of ring homomor-
phisms ¡
pq
: R
q
÷ R
p
for all p. q ÷ ˘, p ¸ q such that ¡
pp
= 1
R
p
and
¡
pq
¡
qr
= ¡
pr
whenever p ¸ q ¸ r. The ring constructed is R = lim
÷÷
R
p
. See
Brun, Bruns and Römer [14] and Brun and Römer [15] for a discussion of such
rings.
In the “combinatorial” cases mentioned above, each ring R
p
is an affine
monoid algebra kŒM
p
|, and the morphisms ¡
pq
are face projections.
126 4. Monoid algebras
4.D. Normality
We have introduced the integral closure and normalization of monoids in Section
2.B. For domains we are using them in the standard way of commutative algebra
that we will now outline. A very good basic reference is Bourbaki [10, Ch. V].
Let R Ç S be an extension of commutative rings. Then an element x ÷ S is
integral over S if it satisfies an equation
x
n
÷a
n-1
x
n-1
÷ ÷a
1
x ÷a
0
. a
i
÷ R.
Clearly, x is integral over Rif and only if the subalgebra RŒx| of S is a finitely gen-
erated R-module. It is however sufficient for the integrality of x that it is contained
in an R-subalgebra of S which is finitely generated as an R-module. This criterion
is fundamental for the theory of integral dependence.
The set of elements x ÷ S that are integral over R form a subring
¨
R
S
of S. If
¨
R
S
= S, then S is integral over R. If R =
¨
R
S
, then R is integrally closed in S.
The integral closure
¨
Rof an integral domain Rin its field of fractions is called the
normalization of R, and R is normal if
¨
R = R. The classical examples of normal
domains are factorial domains, i. e. domains in which every nonzero and nonunit
elements can be written as a product of prime elements. (We call an element prime
if it generates a prime ideal; a decomposition into prime elements is unique up to
order and unit factors.)
We summarize the basic facts about integral closures:
Theorem 4.34.
(a) A reduced ring Ris integrally closed in RŒX| and RŒX
˙1
|.
(b) Let R Ç S Ç T be a tower of ring extensions. Then T is integral over R if
and only if S is integral over Rand T is integral over S.
(c) The integral closure of RŒX| in SŒX| is
¨
R
S
ŒX|.
(d) If T Ç R is a multiplicatively closed set, then T
-1
¨
R is the integral closure
of T
-1
Rin T
-1
S.
(e) The intersection
_
R
i
of integrally closed subrings R
i
of S is integrally
closed in S.
(f) If Ris a normal domain, then RŒX| is normal.
Fact (a) follows immediately from the definition of integral dependence: a Lau-
rent polynomial over Rcannot be a zero of a nonzero polynomial over R(Exercise
4.9). The nontrivial implication of fact (b) is readily derived from the transitivity
of module finiteness, whereas (e) is trivial. Although RŒX| is a monoid algebra
over R with monoid Z
÷
, we refer for the proof of (c) to [10, Ch. V, §1, Prop. 12].
Together with the (simpler) fact (d), for which we refer to [10, Ch. V, §1, Prop. 16],
it is the basis of our discussion below. The last assertion (f) follows from (c). In-
deed, let Q = QF(R). Then QŒX| is integrally closed in its field of fractions Q(X),
since QŒX| is factorial, and since R is normal, we conclude from (c) that RŒX| is
integrally closed in QŒX|.
We start with the most basic and most important result.
4.D. Normality 127
Theorem 4.35. Let R be a domain, and let M be a monoid. Then RŒM| is normal
if and only if Rand M are normal.
Proof. The implication == is very easy. Within the field of fractions of RŒM|
(which is a domain by Corollary 4.9) we have R = RŒM| ¨ QF(R). Therefore, if
an element of QF(R) is integral over R, then it is integral over RŒM| and belongs
to RŒM| if and only if it belongs to R. Moreover, every element x ÷ gp(M) with
mx ÷ M for some m > 0 represents a monomial X
x
with X
mx
= 1 ÷ RŒM|. If
RŒM| is normal, then X
x
÷ RŒM|, equivalently, x ÷ M.
For the converse implication we consider first the case in which M is affine. Let
gp(M) = Z
r
. We write R
÷
M Ç R
r
as an intersection of rational halfspaces H
÷
i
(Theorem 1.18 and Proposition 1.75). Then M =
_
i
Z
r
¨ H
÷
i
, and
RŒM| =
_
i
RŒZ
r
¨ H
÷
i
|.
Since the intersection of normal domains is normal, it is enough to treat the case
M = Z
r
¨H
÷
, where H
÷
is a rational halfspace. But then M Š Z
r-1
˚Z
÷
, and
RŒM| Š RŒZ
r-1
˚Z
÷
|. After choosing a basis of Z
r-1
, we can identify RŒM| with
the partial Laurent polynomial ring RŒX
˙1
1
. . . . . X
˙1
r-1
. X
r
|, which, as a localization
of a polynomial extension of R, is normal.
Now let M be an arbitrary monoid, and let ¡ ÷ QF(RŒM|) be integral over
RŒM|. Then there exists an affine submonoid N such that ¡ ÷ QF(RŒN|) and ¡
is integral over RŒN|. Since the normalization of N is contained in M, we are done
by the affine case.
We want to give a relative version that takes into account both ring and monoid
extensions. We first prove an auxiliary result.
Lemma 4.36. Let Rbe a domain, M an affine submonoid of Z
n
and
¯
M
Z
n the inte-
gral closure of M in Z
n
. Then RŒ
¯
M
Z
n| is the integral closure of RŒM| in RŒZ
n
|.
Proof. We set
¯
M =
¯
M
Z
n. Since every element of
¯
M is integral over RŒM| as a
monomial, the extension RŒM| Ç RŒ
¯
M| is integral. Let U be the smallest direct
summand of Z
n
containing M. Then U is the integral closure of gp(M) in Z
n
,
and RŒZ
n
| is a Laurent polynomial extension of RŒU|. Therefore RŒU| is integrally
closed in RŒZ
n
|. So it remains to show that RŒ
¯
M| is integrally closed in RŒU|.
Let k = QF(R). If an element of RŒU| is integral over RŒM|, the it is integral
over kŒM|, too. Since RŒ
¯
M| = kŒ
¯
M| ¨ RŒU|, it is enough to show that kŒ
¯
M| is
integrally closed in kŒU|. By Corollary 2.25 we have gp(
¯
M) = U, and by Theorem
4.35 kŒ
¯
M| is integrally closed in its field of fractions. Afortiori it is integrally closed
in kŒU|.
Theorem 4.37. Let R Ç S be an extension of domains and let M Ç N be an
extension of monoids.Then
¨
R
S
Œ
¯
M
N
| is the integral closure of RŒM| in SŒN|.
Proof. We deal with one extension at a time. Suppose first that S = R. Since each
element of
¯
M =
¯
M
N
is integral over RŒM|, it follows immediately that RŒ
¯
M| is
integral over RŒM|. It remains to show that RŒ
¯
M| is integrally closed in RŒN|.
128 4. Monoid algebras
Consider an element ¡ of RŒN| that is integral over RŒM|. In a monic polyno-
mial over RŒM| that has ¡ as a zero only finitely many elements of M can occur.
They generate an affine submonoid of M, and together with the monomials ap-
pearing in ¡ they generate an affine submonoid of N. Therefore we may at this
point assume that M and N are themselves affine. Let M
t
denote the integral clo-
sure of M in gp(N). Then RŒM
t
| is the integral closure of RŒM| in RŒgp(N)| by
Lemma 4.36.
Altogether we conclude that the integral closure of RŒM| in RŒN| is contained
in RŒN| ¨ RŒM
t
| = RŒ
¯
M|, since
¯
M = M
t
¨ N, as is easily checked.
Set
¨
R =
¨
R
S
. In the second step we only need to show that
¨
RŒM| is integrally
closed in SŒM|. Since
¨
RŒM| =
¨
RŒgp(M)| ¨ SŒN|, it is enough that
¨
RŒgp(M)|
is integrally closed in SŒgp(M)|. As above, we can reduce this claim to the case
of a finitely generated torsionfree group. Then it amounts to the claim that the
integral closure of RŒX
˙1
1
. . . . . X
˙1
n
| in SŒX
˙1
1
. . . . . X
˙1
n
| is
¨
RŒX
˙1
1
. . . . . X
˙1
n
|. But
taking integral closures commutes with polynomial extensions and localizations
(Theorem 4.34(c) and (d) above).
Some of the results above hold for more general rings of coefficients. The
reader may explore to what extent Theorem 4.37 can be generalized.
Normality and purity. In Theorem 2.28 we have discussed the embedding of pos-
itive affine monoids as pure submonoids of “positive orthants” Z
n
÷
. That theorem
can be extended to an analogous result on monoid algebras.
Recall that a ring extension R ÷ S is pure if for every R-module M, the map
M ÷ M ˝S, x ÷ x ˝1, is injective. Applied to an ideal I Ç R, or rather the
module R,I, purity yields I = IS ¨R. A sufficient condition for purity is that R
is a direct summand of S as an R-module.
Theorem 4.38. Let R be a normal domain, and M a positive affine monoid. Then
the following are equivalent:
(a) M is normal;
(b) there exists n such that RŒM| is isomorphic to a pure subalgebra S of
RŒX
1
. . . . . X
n
|;
(c) there exists n such that RŒM| is isomorphic to a subalgebra S of RŒX
1
. . . . .
X
n
|, for which S is a direct summand of RŒX
1
. . . . . X
n
| as an S-module;
(d) in addition to the properties in (c), S can be chosen to be integrally closed
in kŒX
1
. . . . . X
n
|.
Proof. Let M be normal. Then there is an embedding ç : M ÷ Z
n
such that
ç(M) is pure in Z
n
and integrally closed. The embedding of monoids extends to
an embedding of monoid algebras, and we conclude from Theorem 4.35 that RŒM|
is integrally closed in RŒZ
n
÷
|.
For the implication (a) == (d) we can now identify M with its image under
ç and have only to show that RŒM| is a direct summand of RŒZ
n
÷
|. Let U be the
R-submodule generated by N = Z
n
÷
\ M. Since M = gp(M) ¨Z
n
÷
, we conclude
that M ÷ N Ç N (writing the monoid operation additively). This amounts to
4.E. Divisorial ideals and the class group 129
the fact that RN is a module over RŒM|, and since RŒZ
n
÷
| = RŒM| ˚RN as an
R-module, we conclude that this splitting is in fact one of RŒM|-modules.
The implications (d) == (c) == (b) are trivial, and for (b) == (a) we only
need that normality is passed on to pure subalgebras (and from RŒM| to M by
Theorem 4.35). See Exercise 4.10.
Let M be a pure submonoid of N. In Section 2.B we have studied the decom-
position of N into a disjoint union of coset modules. After the introduction of a
ring R of coefficients, one obtains a splitting of RŒN| into a direct sum of RŒM|-
submodules, which we call the coset modules of RŒM| in RŒN|. Proposition 2.30
then translates as follows:
Proposition 4.39. Let M be a pure submonoid of an affine monoid N. Then the
coset modules of RŒM| in RŒN| are finitely generated (by monomials) over RŒM|
and of rank 1.
Factorial and regular monoid domains. There exist only “trivial” examples of fac-
torial or regular affine monoid algebras. We combine both these properties in a
single proposition.
Proposition 4.40. Let R be a domain and M an affine monoid. Then the following
are equivalent:
(a) RŒM| is factorial (regular).
(b) Ris factorial (regular) and M Š Z
m
˚Z
n
÷
for suitable m. n ÷ Z
÷
.
The proof is left to the reader (Exercises 4.12, 4.21). The implication (b) ==
(a) amounts to the standard fact that (Laurent) polynomial extensions preserve
both factoriality and regularity.
4.E. Divisorial ideals and the class group
Our basic references for this section are [10, Ch. VII] and Fossum [33].
The class group. Let R be an integral domain. A fractional ideal in R is an R-
submodule I of QF(R) for which there exists a ÷ R such that aI Ç R. If R is
noetherian, then it is sufficient that I is finitely generated, since we can choose a
as a common denominator of the generators. The inverse ideal of I is
I
-1
= {a ÷ QF(R) : aI Ç R].
Clearly I
-1
is also a fractional ideal, and I Ç (I
-1
)
-1
. If I = (I
-1
)
-1
, then I is
called divisorial. More generally we set
I : J = {a ÷ QF(R) : aJ Ç I].
It is easy to check that I : J and IJ are fractional. In particular, the fractional
ideals form a monoid under multiplication.
Principal fractional ideals I = aR, a ÷ QF(R), are evidently divisorial, since
I
-1
= a
-1
R. One even has II
-1
= R. This equality distinguishes the invertible
130 4. Monoid algebras
ideals to be considered in connection with the Picard group in Section 4.F. An
invertible ideal is certainly divisorial.
The formation of I : J can be viewed as a standard operation of linear algebra:
Lemma 4.41. Let Rbe an integral domain and let I. J be fractional ideals.
(a) The homomorphism I : J ÷ Hom
R
(I. J), which maps a ÷ I : J to the
multiplication by a, is an isomorphism.
(b) I is reflexive as an R-module if and only if I is a divisorial ideal.
(c) (I
-1
)
-1
is the smallest divisorial ideal containing R.
(d) if I and J are divisorial, then I : J is also divisorial.
Now suppose that Ris a normal noetherian domain. Though the theory can be
developed more generally for Krull domains (and Krull monoids [36]), we restrict
ourselves to the noetherian case. The following theorem characterizes the normal
noetherian domains and lays the foundation for the definition of divisors:
Theorem 4.42. Let Rbe a noetherian domain. Then the following are equivalent:
(a) Ris normal.
(b) Rsatisfies the following conditions:
(i) (Serre’s (R
1
)) For each height 1 prime ideal p the localization R
p
is a
discrete valuation domain;
(ii) (Serre’s (S
2
))
R =
_
ht p=1
R
p
.
Not only normality, but both conditions (R
1
) and S
2
) individually are combi-
natorial properties of M (over a suitable ring of coefficients). We will come back
to this point in ??.
Let us denote the discrete valuation on QF(R) associated with R
p
by :
p
. The
maximal ideal of R
p
is generated by a single element t , and if I is a fractional ideal
of R, I
p
= t
k
R
p
for a unique k ÷ Z. We set :
p
(I) = k. One has :
p
(I) = 0 for
all but finitely many p. (This follows easily from the fact that every element of Ris
contained in only finitely many height 1 prime ideals.) Therefore we can define the
divisor of I by
div(I) =

p
:
p
(I) div(p) ÷ Div(R) =

p
Zdiv(p)
where Div(R) is the free abelian group on the set of divisorial prime ideals and
div(p) is the basis element representing p. It follows immediately from the defini-
tion that
div(IJ) = div(I) ÷div(J) and div(I
-1
) = ÷div(I).
Theorem 4.43. Let Rbe a normal noetherian domain. Then:
(a) Every height 1 prime ideal of Ris divisorial.
(b) The divisorial fractional ideals I form a group D with product (I. J) ÷
((IJ)
-1
)
-1
.
(c) The map I ÷div(I) is an isomorphism of D and Div(R).
4.E. Divisorial ideals and the class group 131
(d) Divisorial ideals I and J are isomorphic R-modules if and only if I : J is
a principal fractional ideal.
For the proof of the theorem one observes that div
_
(I
-1
)
-1
_
= div(I), simply
because the formation of the inverse ideal commutes with localization, and every
ideal of R
p
is divisorial if ht p = 1. Moreover,
I =
_
ht p=1
I
p
” I is divisorial.
The principal divisors, i. e. divisors form a subgroup Princ(R). Therefore we
can define the (divisor) class group of Rby
Cl(R) = Div(R), Princ(R).
By Theorem 4.43(d) div(I) and div(J), I. J divisorial, differ by a principal divisor
if and only if I and J are isomorphic as R-modules. Therefore Cl(R) is the set of
isomorphism classes of divisorial fractional ideals of R. We will denote the class of
div(I) by ŒI|. Since Div(R) is generated by the prime divisors, Cl(R) is generated
by their classes.
The divisor class group measures the deviation of a normal domain from fac-
toriality.
Theorem4.44. Let Rbe a noetherian normal domain. Then the following are equiv-
alent:
(a) Ris factorial;
(b) every height 1 prime ideal of Ris principal;
(c) Cl(R) = 0.
The behavior of the divisor class group under polynomial extensions and lo-
calization is rather easy to control:
Theorem 4.45. Let Rbe a normal noetherian domain. Then:
(a) (Gauß) The extension I ÷ IRŒX| of divisorial ideals induces an isomor-
phism Cl(R) Š Cl(RŒX|).
(b) (Nagata) Let T be a multiplicatively closed subset of R. Then the extension
I ÷T
-1
I induces an exact sequence
0 ÷U ÷Cl(R) ÷Cl(T
-1
R) ÷0
in which U is the subgroup of Cl(R) generated by the classes of the diviso-
rial prime ideals p such that S ¨ p = 0.
Class groups of monoid algebras. We first show that the monoid of fractional
monomial ideals over an affine monoid domain is closed under the operation I : J
and identify the condition under which fractional monomial ideals are isomorphic.
Lemma 4.46. Let Rbe a domain and M an affine monoid. Let I and J be monomial
fractional ideals of RŒM|, and set I
t
= I ¨ M, J
t
= J ¨ M. Then
(a) I : J is a monomial ideal.
(b) The following are equivalent:
132 4. Monoid algebras
(i) I and J are isomorphic as RŒM|-modules;
(ii) there exists a monomial x ÷ gp(M) such that I = xJ;
(iii) there exists a monomial x ÷ gp(M) such that I
t
= xJ
t
.
Proof. We first show that I : J Ç RŒgp(M)|. In fact, if ¡J Ç I, then
¡JRŒgp(M)| Ç IRŒgp(M)|. But JRŒgp(M)| = RŒgp(M)|, and similarly for
I, so that ¡ ÷ RŒgp(M)|.
Now let ¡ ÷ RŒM| such that ¡J Ç I, and consider a monomial x representing
a vertex of the Newton polytope of ¡ . Let z ÷ J
t
. Then xz represents a vertex of the
Newton polytope of ¡ z. Since ¡ z ÷ I and I is monomial, it follows that xz ÷ I as
well. Passing from ¡ to ¡ ÷ ¡
x
x and using induction on # supp(¡ ), we conclude
that all monomials x ÷ supp ¡ belong to I : J. This proves (a).
For (b) we first observe the trivial implications (iii) == (ii) == (i), and for
(i) == (iii) it is enough to find a monomial x with I = xJ, since (ii) and (iii) are
clearly equivalent.
If I and J are isomorphic as RŒM|-modules, then, by (a), there exists ¡ ÷
RŒgp(M)| such that I = ¡J. Then we also have J = ¡
-1
I. Using (a) again, we
conclude that ¡
-1
÷ RŒgp(M)|, too. The only units in Rgp(M)| are the terms rx
where r is a unit in R(Proposition 4.12). If rxJ = I, then xJ = I as well, and we
are done.
For the rest of this subsection R is a normal noetherian domain and M is a
normal affine monoid. We denote the set of facets of R
÷
M by F. Note that the
ideals p
F
, F ÷ F, introduced in Section 4.C, are exactly the monomial prime
ideals of height 1 (see 4.31).
Let us first show that the valuation :
p
F
is the unique extension of the support
form o
F
of R
÷
M associated with the facet F. More generally let us consider an
arbitrary Z-linear t form on gp(M) (M an arbitrary affine monoid). We extend t
to RŒgp(M)| by
t
_

x÷gp(M)
¡
x
x
_
= max{t(x) : ¡
x
= 0]
and t(0) = o(0 ÷ R!). For all ¡. g one then has
t(¡g) = t(¡ ) ÷t(g).
t(¡ ÷g) ¸ max
_
t(¡ ). t(g)
_
.
t(¡ ÷g) = max
_
t(¡ ). t(g)
_
if t(¡ ) = t(g).
This shows that t is a valuation on RŒM|. It extends to a valuation on QF(RŒM|)
by t(¡,g) = t(¡ ) ÷ t(g).
Let us return to o
F
. We can now consider it as a valuation on QF(RŒM|). In
order to identify it with :
p
F
we have only to check that
p
F
= {¡ ÷ RŒgp(M)| : o
F
(¡ ) ¸ 1]
and that there exists ¡ ÷ RŒM| with o
F
(F) = 1. Both properties are clearly
satisfied. In fact, they imply that the valuation ring of o
F
is really RŒM|
p
F
, so that
o
F
= k:
p
F
for some k ÷ N, and obviously we must have k = 1.
4.E. Divisorial ideals and the class group 133
Let us now compute the divisor of a monomial x ÷ gp(M). In multiplicative
notation, x = y,z with y. z ÷ M. Therefore div(x) = div(y) ÷ div(z). The
minimal prime ideals of y nd z are all monomial by Corollary 4.30. Therefore
div(x) =

F÷F
:
p
F
(y)p
F
÷

F÷F
:
p
F
(z)p
F
=

F÷F
o
F
(x)p
F
. (4.2)
Similarly we compute the divisor of an arbitrary monomial ideal I:
div(I) =

F÷F
o
F
(I) div(p
F
)
where
o
F
(I) = min{o
F
(x) : x ÷ I ¨ gp(M)]. (4.3)
Let p be a divisorial prime ideal in a normal domain A, and :
p
the associated
valuation. Then
p
(k)
= {x ÷ A : :
p
(x) ¸ k]. k ÷ Z.
is called the kth symbolic power of p. Clearly div(p
(k)
) = k div(p). Then the divi-
sorial ideal I with div(I) = k
1
div(p
1
) ÷ ÷k
m
div(¡p
m
) is given by p
(k
1
)
1
¨ ¨
p
(k
m
)
m
. Equation (4.3) therefore implies
Theorem 4.47. Let R be a normal noetherian domain and M an affine normal
monoid. Then the divisorial monomial ideals of R are exactly the R-submodules of
RŒgp(M)| whose monomial basis is determined by a system
o
F
(x) ¸ s
F
. x ÷ gp(M).
of inequalities with s
F
÷ Z, F ÷ F.
In Figure 4.2 we illustrate the formation of monomial divisorial ideals. With
H
1
H
2
0
p
(2)
1
¨ p
(-2)
2
Figure 4.2. A divisorial monomial ideal
Theorem 4.47 we have identified the sets of solutions of the inhomogeneous sys-
tems of linear diophantine inequalities whose associated homogeneous system de-
fines a normal monoid irredundantly.
Remark 4.48. the condition “irredundantly” has been emphasized for good reason.
If we add to the system in Theorem 4.47 further inequalities t(x) ¸ t , where t is a
Z-linear form on gp(M) such that t(x) ¸ 0 for all x ÷ M, then the R-submodule
spanned by monomial solutions is still a fractional ideal of RŒM|, but it is no longer
134 4. Monoid algebras
divisorial in general. Since it is nevertheless the intersection of fractional valuation
ideals, it is integrally closed. See [22] for a study of this class of ideals.
The theorem shows that the monomial divisorial ideals depend essentially only
on M – the ring R only provides the coefficients. Therefore we call the subsets of
gp(M) that form the R-bases of the monomial divisorial ideals of RŒM| divisorial
ideals of M. Moreover, we are justified to denote the subgroup of Div(RŒM|) gen-
erated by the monomial divisorial prime ideals simply by Div(M). Similarly we
write Princ(M) for its subgroup of the principal monomial ideals.
Corollary 4.49 (Chouinard). The elements of the group
Cl(M) = Div(M), Princ(M)
represent the isomorphism classes of divisorial monomial ideals of RŒM|.
If F
1
. . . . . F
s
are the facets of R
÷
M and o
1
. . . . . o
s
are the corresponding sup-
port forms, then
Cl(M) Š Z
s
,o(gp(M)).
the isomorphism being induced by the map I ÷(o
i
(I). . . . o
s
(I)).
Proof. Lemma 4.46 shows that two monomial divisorial ideals I and J are iso-
morphic if div(I) ÷ div(J) is a monomial principal divisor.
For the second statement we have just numbered the basis elements of Div(M).
The identification of Princ(M) with o(gp(M)) is justified by equation (4.2).
Remark 4.50. We call Cl(M) the class group of M. This notion is also justified
from a purely monoid-theoretic point of view, since the notions of multiplicative
ideal theory introduced so far, can be developed “without coefficients”. See [36]
and [49].
Example 4.51. Let M be the polytopal monoid M(P) where P is the triangle
spanned by (÷1. ÷1). (1. 0). (0. 1) ÷ R
2
. Then M(P) has 4 generators, the fourth
Figure 4.3. The polytope P
one coming from (0. 0) ÷ P. Evaluating the 3 support forms on the generators
of M yields the vectors u
i
= 3e
i
, i = 1. 2. 3, and u
4
= e
1
÷ e
2
÷ e
3
. Clearly
Cl(M) = Z
3
,

Zu
i
Š
_
Z,(3)
_
2
.
That Cl(M) is finite in this example, is not surprising. In fact, the finiteness of
Cl(R) characterizes the normal monoids that are essentially simplicial:
Corollary 4.52. Let M be a normal monoid, and consider the canonical splitting
M = U(M) ˚M
t
. Then:
4.E. Divisorial ideals and the class group 135
(a) Cl(M) Š Cl(M
t
), and
rank Cl(M) = s ÷ rank M
t
= s ÷ (rank M ÷ rank U(M)).
where s is the number of facets of R
÷
M.
(b) In particular, Cl(M) is finite if and only if M
t
is simplicial.
(c) On has Cl(M) = 0 if and only if M Š Z
m
˚Z
s
÷
, m = rank U(M).
Proof. One has Cl(N
1
˚ N
2
) Š Cl(N
1
) ˚ Cl(N
2
) for all normal affine monoids
(Exercise 4.14). Since Cl(G) = 0 if G is a free group, it follows that Cl(M) Š
Cl(M
t
). The rest of (a) and (b) is clear. For (c) we have only to observe that M
t
Š
o(gp(M) ¨ Z
s
÷
. If Cl(M) = 0, then o(gp(M)) = Z
s
, and M
t
Š Z
s
÷
.
Part (c) of the Corollary is only a reformulation of Proposition 4.40, at least in
combination with the next theorem.
We have now collected all the tools that are necessary for the computation of
the class groups of affine monoid algebras.
Theorem 4.53 (Chouinard). Let Rbe a normal noetherian domain, and let M be a
normal affine monoid. Then
Cl(RŒM|) = Cl(R) ˚Cl(M)
where the embedding Cl(R) ÷ Cl(RŒM|) is induced by the extension of divisorial
ideals, and the embedding Cl(M) ÷ Cl(RŒM|) is induced by the map I ÷ ŒI| on
the set of divisorial monomial ideals.
Proof. It is very easy to check that the extension JRŒM| Ç QF(R)ŒM| of a di-
visorial ideal J of R is again divisorial. Therefore we obtain an induced map
Cl(R) ÷ Cl(RŒM|). Now let x ÷ int(M). Then RŒM. x
-1
| = RŒgp(M)|; see
Exercise 2.9 or the ring-theoretic argument in the proof of Lemma 5.31. By Na-
gata’s theorem we have an exact sequence
0 ÷U ÷Cl(RŒM|) ÷Cl
_
RŒgp(M)|
_
÷0.
Since RŒgp(M)| is the localization of a polynomial ring over R with respect
to a multiplicative set generated by prime elements, Gauß’ and Nagata’s theo-
rem show that the extension map J ÷ JRŒgp(M)| induces an isomorphism
Cl(R) ÷ Cl
_
RŒgp(M)|
_
which factors through Cl(RŒM|). It only remains to
identify the group U. By Nagata’s theorem it is generated by the classes of the
minimal prime ideals of x. These are exactly the monomial prime ideals p
F
, and
so U is the subgroup generated by them. Since two monomial divisorial ideals are
isomorphic if and only they represent the same class in Cl(M), we are done.
Corollary 4.54. Every divisorial ideal H in RŒM| is isomorphic to one of the ideals
IJ where I is a divisorial ideal in Rand J is a divisorial monomial ideal.
Proof. According to the theorem we can find a divisorial ideal I in R and a di-
visorial monomial ideal J such that H is isomorphic to the smallest divisorial
ideal containing IJ. But IJ is itself divisorial, as the reader may check (Exercise
4.13).
136 4. Monoid algebras
In particular, if R is factorial, for example a field, then every divisorial ideal in
RŒM| is isomorphic to a monomial such ideal.
The class ring. We have seen in Corollary 4.49 that the standard map determines
the class group of a normal monoid M. This is true far beyond the group structure
of Cl(R). For the sake of a clarity we consider only the case in which the ring of
coefficients is a field. (The generalization is immediate.)
Theorem 4.55. Let k be a field, M a normal affine monoid, and o : gp(M) ÷ Z
s
the standard map of M. For a class c ÷ Cl(kŒM|) Š Z
s
,o(gp(M)) let
D
c
= k{y ÷ Z
s
÷
: y ÷ c].
Then D
c
is a kŒM|-submodule (via o) of kŒZ
s
| Š kŒY
1
. . . . . Y
s
| isomorphic to a
divisorial ideal of class ÷c, and
kŒY
1
. . . . . Y
s
| =

c÷Cl(R)
D
c
.
Proof. In view of Corollary 4.52 we may assume Ker o = U(M) = 0. Then we
can consider M as a pure submonoid of Z
s
÷
via o, and gp(M) as a subgroup of Z
s
.
As stated in Proposition 4.39, kŒY
1
. . . . . Y
s
| splits into the coset modules of kŒM| in
kŒY
1
. . . . . Y
s
|, and by definition these are the modules D
c
.
For the rest of the proof it is advisable to use exponential notation for the
monoid algebras. The monomials in gp(M) will be denoted by X
x
and those in
Z
s
by Y
y
. The set of all monomials with exponent vector in c is denoted by Y
c
.
It only remains to show that D
c
is isomorphic as a kŒM|-module to a divisorial
ideal of class ÷c. To this end we choose z ÷ ÷c and multiply D
c
by Y
-z
. The image
I is a kŒM|-submodule of kŒgp(M)| spanned by all the monomials X
x
such that
o(x) ÷z ÷ Z
s
÷
. In other words: by all monomials X
x
such that o(x) ¸ z. Hence
I is divisorial with divisor z
1
div(p
1
) ÷ ÷z
s
div(p
s
), and we are done.
Thus the polynomial ring kŒY
1
. . . . . Y
s
| bundles all the divisor classes of kŒM|
into a single, well-understood k-algebra. For analogy with the number-theoretic
situation we are justified in calling kŒY
1
. . . . . Y
s
| the class ring of kŒM|. See [22] for
a study of certain properties of the divisorial ideals of kŒM| using the class ring.
Remark 4.56. The class group of M limits the pure embeddings of a positive nor-
mal affine monoid in the following way: if M is a pure submonoid of Z
n
÷
, then
Cl(R) is a subquotient of Z
n
, gp(M). See [22, Corollary 2.3]. This reflects the fact
that, roughly speaking, multiples of the support forms must appear in every system
of inequalities cutting out M from a group.
Positively graded algebras. Let R be a normal noetherian graded ring, and sup-
pose that the homogeneous nonunits in R generate a proper ideal p. For example,
R is a positively graded algebra over a field k: R =

o
i=0
R
i
and R
0
= k. In
this case p =

o
i=1
R
i
is even a maximal ideal. Since monoid algebras kŒM| with
a positive affine monoid have a positive grading, the theorem below applies espe-
cially to them. We will need it in Chapter 5.
4.F. The Picard group and seminormality 137
In the general case p is a prime ideal. This follows from Lemma 4.27: by con-
struction all homogeneous elements = 0 of R,p are units, and on the other hand
the minimal prime ideals of R,p are graded. So we may consider the localization
R
p
.
Theorem4.57. Under the above hypothesis on R, the extension R ÷R
p
induces an
isomorphism Cl(R) ÷Cl(R
p
).
Proof. Let us first show that every divisorial ideal in R is isomorphic to a graded
divisorial ideal. In fact, let T be the multiplicatively closed system of all homo-
geneous nonzero elements in R. Then it is not hard to see that all homogeneous
nonzero elements in T
-1
R (with its induced grading) are units. It follows readily
that T
-1
Ris either a field k or of the formkŒX. X
-1
|, depending on whether there
exist homogeneous elements of degree = 0. In any case, T
-1
Ris factorial.
Nagata’s theorem implies that Cl(R) is generated by the classes of the graded
prime ideals, and so every divisorial ideal is isomorphic to a graded one.
Applying Nagata’s theorem to the extension R ÷ R
p
yields the surjectivity of
the homomorphism Cl(R) ÷ Cl(R
p
). Thus it remains to be shown that a non-
principal graded ideal of R extends to a nonprincipal ideal in R
p
. This is part of
Exercise 4.20.
The advantage of Theorem 4.57 is that it makes the linear algebra over the local
ring R
p
available for the divisorial investigation of R.
4.F. The Picard group and seminormality
Let R be a domain. Then most elementary description of seminormality for R is
through Hamann’s criterion: R is seminormal if every element x of QF(R) such
that x
2
. x
3
÷ Rbelongs to R. It immediately explains the notion of seminormality
for monoids. However, it does not reveal the importance of seminormality which
rests on its role in K-theory, and for K-theory it is useful to generalize the notion
of seminormality as follows (Swan [89]):
Definition 4.58. Let Rbe a reduced ring. Then Ris called seminormal if the equa-
tion x
3
= y
2
for elements x. y ÷ R is only solvable with x = z
2
, y = z
3
for
z ÷ R.
The reader may check that z is uniquely determined, and the elementary proof
of the next proposition is also left to the reader:
Proposition 4.59. Let R be seminormal and S Ç R multiplicatively closed. Then
S
-1
Ris also seminormal
Let
PQF(R) =

p
QF(R,p)
where p runs through the minimal prime ideals of R. Since the zero ideal of R is
the intersection of the minimal prime ideals, the natural map R ÷ PQF(R) that
138 4. Monoid algebras
assigns each a ÷ R the family of its residue classes in PQF(R) is injective. We
consider Ras a subring of PQF(R).
Proposition 4.60. Let Rbe a reduced ring. Then the following are equivalent:
(a) Ris seminormal;
(b) if u
2
. u
3
÷ Rfor u ÷ PQF(R), then u ÷ R;
(c) if R Ç S is an extension of reduced rings and u
2
. u
3
÷ R for u ÷ S, then
u ÷ R.
Proof. (a) == (c) Set x = u
2
and y = u
3
. By seminormality we find z ÷ R with
x = z
2
and y = z
3
. The uniqueness of z (in S) implies u = z ÷ R.
(c) == (b) is trivial, and for (b) == (a) we let ¬
p
: R ÷ R,p be the natural
epimorphism. Suppose that x
3
= y
2
for x. y ÷ R. If p is a minimal prime of
R, then either both x. y ÷ p or x. y … p. Setting u
p
= 0 if x. y ÷ p, and u
p
=
¬
p
(y),¬
p
(x) ÷ QF(R,p) otherwise, we obtain ¬
p
(x) = u
2
p
, ¬
p
(y) = u
3
p
for all p.
Set u = (U
p
). Then u
2
. u
3
÷ R, and so u ÷ Rby (b).
Thus PQF(R) is a replacement for the quotient field of an integral domain, as
far as seminormality is concerned. If Rhas only finitely many minimal primes, for
example if R is noetherian, then PQF(R) is the total ring of fractions of R (this
follows easily by prime avoidance; see [89, 3.6]).
The Picard group. Let R be a ring, and P an R-module. We say that P is a
projective module if there exists an R-module Q such that P ˚ Q is isomorphic
to a free R-module R
n
of finite rank. Of course, Q is projective as well, and it
follows immediately that P is finitely presented. We include finite generation in the
definition of projective module since only this class is relevant for us. For the basic
theory of projective modules covering the unproved statements below we refer the
reader to Bass [4, Ch. III, §7] or Bourbaki [10, Ch. 2, §5].
Projective modules over local rings are free, as follows from Nakayama’s
lemma. Therefore, given a prime ideal p Ç R, the localization P
p
is a free R
p
-
module. In this section we are only interested in projective modules of rank 1,
i. e. projective modules for which rank P
p
= 1 for all prime ideals p. Projective
modules of arbitrary rank will be discussed in Chapter 8.
Proposition 4.61. Let Rbe a ring and P. Qprojective R-modules. Then:
(a) P ˝Qis also projective;
(b) rank P = 1 if and only if the evaluation map P ˝ Hom
R
(P. R) ÷ R,
x ˝ç ÷ç(x), is an isomorphism.
The proposition allows us to introduce the Picard group of R. Its underlying
set is the set of isomorphism classes of projective rank 1 modules, and addition is
induced by the assignment ŒP| ÷ŒQ| = ŒP ˝Q|. Clearly Pic(R) = 0 if and only if
every projective rank 1 module is isomorphic to R.
The following theorem will allow us to construct nontrivial elements in Pic(R).
Theorem 4.62. Let R Ç S be a ring extension and let I Ç S be an R-submodule.
Then the following are equivalent:
4.F. The Picard group and seminormality 139
(a) there exists an R-submodule J of S such that IJ = R;
(b) I is a projective R-module of rank 1 and IS = S.
An R-submodule I of S satisfying the conditions in the theorem is called in-
vertible with (uniquely determined) inverse I
-1
= J Š Hom
R
(I. R). This notion
generalizes that of invertible ideals in the field of fractions of a domain. In the case
of a domain all rank 1 projective modules can be realized as invertible ideals:
Proposition 4.63. Let Rbe a domain and P a rank 1 projective module. Then:
(a) there exists an embedding P ÷ QF(R), and every such embedding maps
P onto an invertible ideal I of R.
(b) Conversely, every invertible ideal of Ris a projective module of rank 1.
(c) If I and J are invertible ideals, then IJ is invertible, too, and I ˝J Š IJ
via the multiplication map.
We want to pursue the question under which conditions the extension R ÷
RŒM| induces an isomorphism of Picard groups. In fact, every extension R ÷ S
induces a natural map Pic(R) ÷ Pic(S) given by the extension P ÷ P ˝S, and
if an S-module Q is of type P ˝
R
S for some R-module P, then we say that Q
is extended from R. Since we have a retraction RŒM| ÷ R, the map Pic(R) ÷
Pic(RŒM|) is an injection if M is an arbitrary monoid, but the determination of
the cokernel is a difficult task.
If R and M are normal, then the situation is very simple, as we will see now.
It is immediate from the definition of the Picard group and Theorem 4.43 that one
has a natural embedding Pic(A) ÷ Cl(A) if R is normal noetherian domain. For
A = RŒM| this embedding allows us to compute Pic(A).
Theorem 4.64. Let R be a noetherian normal domain, and M a normal affine
monoid. Then the natural map Pic(R) ÷Pic(RŒM|) is an isomorphism.
Proof. Injectivity has already been stated. So let H be an invertible ideal of RŒM|.
According to Corollary 4.54 H Š IJ where I is a divisorial ideal in R and J is a
divisorial monomial ideal. We can assume that H = IJ. After multiplication by
(the extension of ) I
-1
we have to show that J is principal.
One checks immediately that J
-1
is the R-module generated by all monomials
y with xy ÷ M for all x ÷ M. On the one hand, JJ
-1
= RŒM|, and on the other
hand, it is an R-module generated by the monomials xy, x ÷ J, y ÷ J
-1
. Thus
xy = 1 for some x ÷ J, y ÷ J
-1
, and J is generated by x.
While we have just seen that in the presence of normality the extension
Pic(R) ÷ Pic(RŒM|) is surjective, let us now show that surjectivity fails if RŒM|
is not seminormal. The following lemma gives examples of rank 1 nonfree and
nonextended projective modules of rank 1.
Lemma 4.65 (Schanuel). Let R Ç T be an extension of reduced rings and suppose
that x ÷ T, x … R, but x
2
. x
3
÷ R.
(a) If U(RŒx|) = U(R), then the R-submodule I = (1 ÷x)R÷x
2
R Ç T is
a nonfree projective module of rank 1.
140 4. Monoid algebras
(b) Let M be a monoid, and t ÷ M, t = 1. Then the RŒM|-submodule I =
(1 ÷xt )RŒM| ÷(x
2
t
2
)RŒM|) of T ŒM| is a projective rank 1 module that
is not extended fromR.
Proof. (a) Set S = RŒx|. We have IJ = Rfor the R-submodule J = R(1 ÷x) ÷
Rx
2
Ç T . Thus I is an invertible module, and therefore projective of rank 1.
If I = cR for some element c ÷ T, then c ÷ S since I Ç S. We have SI = S
and, hence, c ÷ U(S) = U(R). It follows I = Rand x ÷ R– a contradiction.
(b) As in (a) we see that I is invertible with inverse J = (1 ÷ xt )RŒM| ÷
(x
2
t
2
)RŒM|). Suppose that I is extended from R. In order to derive a contradic-
tion we first choose a localization R
p
with x … R
p
. Replacing Rby the localization
we may assume that Ris local. Then, if I is extended as a projective module, it is a
free module, and therefore of the form cRŒM| with c ÷ I. Since ITŒM| = TŒM|
and U(T ŒM|) = U(T) × U(M) (Proposition 4.12), one concludes that c = du
with d ÷ U(T ) and u ÷ U(M). Now du(1 ÷ xt ) ÷ RŒM|, and so d ÷ R and
dx ÷ R(we have used that u = ut .)
If I = cRŒM|, then J = c
-1
RŒM|, and d
-1
÷ Rfollows by similar arguments.
But then x ÷ R– again a contradiction.
By Exercise 4.5 AŒM| is a reduced ring for a reduced ring A. Therefore it makes
sense to discuss the seminormality of AŒM|.
Corollary 4.66. Let A be a reduced ring and M a monoid. If AŒM| is not seminor-
mal, then the natural map Pic(A) ÷Pic(AŒM|) is not surjective.
Proof. We will see in Theorem4.69 that at least one of Aand M is not seminormal
if AŒM| is not seminormal. If A is not seminormal, then Proposition 4.60 and
Lemma 4.65(b) yield a nonextended projective module.
Suppose that M is not seminormal, and let x ÷ gp(M), x … M, but x
2
. x
3
÷
M. Then we define I as in part (a) of the lemma. Assume I is extended from
A, and choose a maximal ideal m of A. Then IA
m
ŒM| is extended from A
m
, and
therefore free. However the formation of I commutes with the passage from A to
A
m
. Therefore it is enough to show that I is always nonfree.
Let N = x
Z
C
M Ç gp(M). Then U(M) = U(N) because N = M L (xM)
and no element in xM can be a unit of N: otherwise we would have xyz = 1 or
xyxz = 1 with y. z ÷ M. In either case x
2
would be a unit of M so x = x
3
(x
2
)
-1
would lie in M. By Proposition 4.12 we have U(AŒM|) = U(AŒN|), and Lemma
4.65(a) shows that I is nonfree.
In Remark 8.45 we will show the converse of Corollary 4.66 for positive mono-
ids, and give references for the case in which M is not positive.
Let us now state Swan’s version of Traverso’s theorem. It is the converse to
Corollary 4.66 for the case M = Z
n
÷
:
Theorem 4.67. Let Rbe a reduced ring. Then the following are equivalent:
(a) Ris seminormal;
(b) Pic(R) ÷Pic(RŒX|) is an isomorphism;
(c) Pic(R) ÷Pic(RŒX
1
. . . . . X
n
|) is an isomorphism for all n.
4.F. The Picard group and seminormality 141
We refer to Traverso [93] for the case of a noetherian reduced ring with finite
integral closure (in its total ring of fractions) and to Swan [89] for the general case.
Dedekind domains. The existence of a nonfree projective rank 1 modules in
Lemma 4.65 is explained by the lack of (semi)normality, but there also exist classi-
cal types of normal domains with nonfree rank 1 projective modules, namely the
Dedekind domains that are not principal ideal domains. A regular domain R of
Krull dimension ¸ 1 is called a Dedekind domain. Thus Ris a Dedekind domain if
and only if is a field or is a noetherian domain such that R
m
is a discrete valuation
domain for every maximal ideal m of R. Classical examples of Dedekind domains
are the rings of integers in number fields.
Since a finitely generated module over a discrete valuation domain is free, every
such R-module M is projective if Ris Dedekind. Using a theorem of Steinitz, one
shows M is of the form free ˚rank one [65, §1]. More precisely, M is isomorphic
to a direct sum I
1
˚ ˚ I
n
of invertible fractional ideals I
1
. . . . . I
n
Ç R and,
furthermore, I
1
˚ ˚ I
n
Š R
n-1
˚ I
1
I
n
. In particular, one has Pic(R) =
Cl(R), and R has nonfree projective modules if and only if it is not a principal
ideal domain. A well-known example is R = QŒ
_
÷5|.
Seminormality and subintegral extensions. Seminormality has not only been ex-
plored from the viewpoint of K-theory, but also as an interesting notion in ring
theory, and to some extent it has been developed in analogy to normality. There-
fore we want to derive results for seminormality analogous to those for normality.
Let R Ç S be an extension of rings. Then one calls S subintegral over R, if (i)
S is integral over R, (ii) for each prime p of Rthere exists exactly one prime ideal q
of S such that p = q ¨R, and, (iii) the extension R ÷S induces an isomorphism
R
p
,pR
p
Š S
q
,qS
q
if p = q ¨ R.
If R Ç S is a ring extension, then there exists a unique maximal R-subalgebra
sn
S
(R) which is subintegral over R. It is called the seminormalization or subinte-
gral closure of R in S. One says that R is subintegrally closed or seminormal in S
if R = sn
S
(R). The “absolute” seminormalization sn(R) of a reduced ring is its
subintegral closure in PQF(R). By [89, 4.2] this is in accordance with the definition
in [89], and according to [89, 4.1] the attribute “absolute” is justified: if ç : R ÷S
is a ring homomorphism to a seminormal ring S, then ç can be extended to a
homomorphism [ : sn(R) ÷S; if ç is injective, then so is [.
Hamann’s criterion [50, 89] gives an “elementary” description of subintegrality
for arbitrary rings: if R Ç S is a proper subintegral extension, then there exists
x ÷ S \ R such that x
2
. x
3
÷ R. This criterion leads to a constructive description
of the subintegral closure sn
S
(R) as follows. Call R Ç Aan elementary subintegral
extension if A = RŒx| such that x
2
. x
2
÷ R. Then sn
S
(R) is the filtered union of
all subrings of S that arise from R by a finite number of elementary subintegral
extensions.
While there seems to be no element-wise description of subintegral depen-
dence in general, such is given for Q-algebras by a theorem of Roberts and Singh
[72]: an extension R Ç S of Q-algebras is subintegral if and only if for every b ÷ S
142 4. Monoid algebras
there exist c
1
. . . . . c
p
÷ S such that
b
n
÷
p

i=1
_
n
i
_
c
i
b
n-i
÷ R for n ;0.
(Compare this to the description of the seminormalization of monoids in Exercise
2.10.)
The basic facts about subintegral closures are summarized as follows:
Theorem 4.68.
(a) A reduced ring Ris subintegrally closed in RŒX| and RŒX
˙1
|.
(b) Let R Ç S Ç T be successive ring extensions. Then T is subintegral over
Rif and only if S is subintegral over Rand T is subintegral over S.
(c) The subintegral closure of RŒX| in SŒX| is sn
S
(R)ŒX|.
(d) If T Ç R is a multiplicatively closed set, then T
-1
(sn
S
(R)) is the subinte-
gral closure of T
-1
Rin T
-1
S.
(e) The intersection
_
R
i
of subintegrally closed subrings R
i
of S is subinte-
grally closed in S.
(f) If a reduced ring Ris seminormal, then RŒX| is seminormal.
Fact (a) is obvious since R is even integrally closed in RŒX
˙1
|. Fact (b) is [89,
2.3], whereas (e) is an easy consequence of Hamann’s criterion. For fact (c) we refer
to [13, Prop. 1], and (d) is [89, 2.3]. The last assertion (f) follows immediately from
the Traverso-Swan theorem.
Again we start with the most basic and most important theorem. Its proof
essentially uses only the Traverso-Swan theorem and Proposition 4.59.
Theorem4.69. Let Rbe a reduced ring and M a monoid. Then RŒM| is seminormal
if and only if Rand M are seminormal.
Proof. The implication ==follows immediately from Proposition 4.60 or simply
from the definition of seminormality.
For the converse, we consider the affine case first. Let gp(M) = Z
r
. As we
have seen in Proposition 2.39, a seminormal affine monoid can be represented in
the form
M =
s
_
i=1
M
i
.
where M
i
is of the form
M
i
= U
i
L (H
>
i
¨ Z
r
)
with a subgroup U
i
Ç H
i
¨ Z
r
(of rank r ÷ 1). It is clearly enough that RŒM
i
| is
seminormal for each i . Therefore we may assume that s = 1, M = M
1
H = H
1
,
U = U
1
.
We first note that RŒgp(M)| is seminormal, since it is a localization of a poly-
nomial extension of R. Suppose that x
3
= y
2
for x. y ÷ RŒM|. Then we find
¡ ÷ RŒgp(M)| such that x = ¡
2
, y = ¡
3
. Write
¡ = ¡
÷
÷¡
0
÷¡
-
Exercises 143
where ¡
÷
collects all terms of ¡ whose monomials lie in H
>
, ¡
0
collects all terms
with monomials in H and ¡
-
the remaining ones.
Clearly, if ¡
-
= 0, then ¡ … RŒgp(M) ¨ H
÷
|. This is impossible, since
RŒgp(M) ¨ H
÷
| is even integrally closed in RŒgp(M)|.
It follows that (¡
0
)
2
= x
0
and (¡
0
)
3
= y
0
. By the seminormality of RŒU| we
conclude that ¡
0
÷ RŒU| Ç RŒM|. Since ¡
÷
÷ RŒM| anyway, we are done for M
affine.
Now let M be an arbitrary monoid. Then an equation x
3
= y
2
for x. y ÷
RŒM| “lives” in an affine submonoid N of M. So it is enough to note that sn(N) Ç
M and to use that sn(N) is affine.
We draw a consequence for the seminormalization of a monoid ring:
Corollary 4.70. Let R be a reduced ring and M a monoid. Then sn(R)Œsn(M)| is
the seminormalization of RŒM|.
Proof. By the theorem sn(R)Œsn(M)| is seminormal. On the other hand, it is a
subintegral extension of RŒM|. The claim now follows from [89, 4.2].
We give the analogues of Lemma 4.36 and Theorem 4.37 for relative subintegral
closures. The proofs are analogous as well, and the details are left to the reader.
Lemma 4.71. Let R be a reduced ring and M an affine submonoid of Z
n
. Then
RŒsn(M)| is the subintegral closure of RŒM| in RŒZ
n
|.
Theorem 4.72. Let R Ç S be an extension of reduced rings and let M Ç N be an
extension of monoids. Set sn
N
(M) = N ¨ sn(M). Then sn
S
(R)
_
sn
N
(M)
_
is the
subintegral closure of RŒM| in SŒN|.
Exercises
4.1. Prove Proposition 4.2.
4.2. Let A be a G-graded abelian group. Show that a subgroup is graded if and only if it
contains the homogeneous components of each of its elements.
4.3. Let R be a Z
r
-graded ring for some r ¸ 0. Show: if all nonzero homogeneous ele-
ments of R are units, then R
0
is a field, and R is a Laurent polynomial ring over R
0
in at
most r indeterminates.
4.4. Let k be a field and R a Z
r
-graded k-algebra that, as a vector space, is isomorphic
to kŒM| for some submonoid of Z
r
. Prove: R and kŒM| are isomorphic as k-algebras if
(and only if) every nonzero homogeneous element in Ris not a zero-divisor. Hint: pass to
RŒS
-1
| where S is the set of all nonzero homogeneous elements, and use Exercise 4.3.
4.5. Let R be a reduced ring and M a (cancellative torsionfree) monoid. Show RŒM| is
reduced.
4.6. (a) Let M be a monoid (not necessarily cancellative or torsionfree), let E Ç M
2
and
¯
E = E L {(y. x) : (x. y) ÷ E]. Show that E generates the congruence relation C ¸ E if
and only if for each (x. y) ÷ C, x = y, there exist n ÷ N, (u
1
. :
1
). . . . . (u
n
. :
n
) ÷
¯
E, and
n
1
. . . . . n
n
÷ M such that x = n
1
u
1
, n
1
:
1
= n
2
:
2
. . . . . n
n-1
:
n-1
= n
n
u
n
, n
n
:
n
= y.
144 4. Monoid algebras
(b) Prove Proposition 4.20.
4.7. Extend Theorem 4.26 to ideals in affine monoid rings kŒM|.
4.8. Extend Lemma 4.27(a) to radical and primary ideals. Deduce that the radical of a
graded ideal is graded, and that a graded ideal on a graded noetherian ring is the intersec-
tion of graded primary ideals.
4.9. Prove that a reduced ring Ris integrally closed in RŒX
˙1
|.
4.10. Let R be a pure subring of the normal domain S. Show that R is also normal. Hint:
R = S ¨ QF(R).
4.11. Let R be a ring and M a monoid. Show the following formula for conductor ideals:
c(
¨

¨
M|) = c(
¨
R,R)c(
¨
M,M).
4.12. Let M be an affine monoid. Prove that RŒM| is factorial if and only if R is factorial
and M Š Z
m
˚Z
n
÷
for suitable mand n.
4.13. Let R be an integral domain, I a divisorial ideal of R and J a divisorial monomial
ideal of RŒM|. Show IJ is divisorial.
4.14. Let M and N ba normal affine monoids. Show Cl(M ˚N) = Cl(M) ˚Cl(N).
4.15. Let M be a normal affine monoid and let F be a face of R
÷
M. Show Cl(M ¨ F)
is a subquotient of Cl(M), i. e. Cl(M ¨ F) Š U,V where U. V are subgroups of Cl(M),
V Ç U.
Find an example for which Cl(M ¨ F) is not isomorphic to a subgroup of Cl(M).
4.16. Let MZ
n
be a simplicial normal monoid with support forms o
1
. . . . . o
n
. ShowCl(M)
is finite of order det(o
i
(e
j
)
i,j=1,...,n
).
4.17. Let z
1
. . . . . z
n
be positive integers and z = z(z
1
. . . . . z
n
) be the rectangular sim-
plex introduced in Exercise ??. Compute the divisor class group of the normalization of
kŒz|. (Both the statement and the proof of the result in [18] are overly complicated.)
4.18. Let z
1
. . . . . z
n
be positive integers. we now consider the simplex P with vertices
:
i
= z
i
e
i
where e
1
. . . . . e
n
are the unit vectors in R
n
.
(a) Compute the divisor class group of the normalization of M(P).
(b) Let Q be lattice polytope. Show that Q is isomorphic to one of the polytopes P just
introduced if and only if the normalization N of M(Q) has cyclic divisor class group.
Hint: Consider the standard embedding o : N ÷Z
s
and showthat there exists a primitive
Z-linear form ˛ on R
s
such that N = {x ÷ Z
s
÷
: ˛(x) ÷ 0 (d)] where d = # Cl(N).
4.19. Let R be a Z
d
-graded ring and suppose that the homogeneous nonunits in R gen-
erate a proper ideal p. Above Theorem 4.57 we have seen that p is a prime ideal. Set
k(p) = R
p
,pR
p
. Let M be finitely generated graded module, and let x
1
. . . . . x
n
be a
homogeneous system of generators of M.
Consider the epimorphism ¬ : R
m
÷ M that sends the i th basis vector e
i
to x
i
. This
epimorphism becomes homogeneous if one makes R
n
a graded R-module by choosing
the degree of e
i
as the degree of x
i
: R
n
=

i
R(÷deg x
i
) in the usual “shift” notation.
Show that Ker ¬ is generated by homogeneous elements and that the following are equiva-
lent:
(i) x
1
. . . . . x
m
is a basis of the k(p)-vector space M ˝k(p).
Exercises 145
(ii) x
1
. . . . . x
m
is a minimal system of generators of M;
(iii) Ker ¬ Ç pM.
4.20. With notation as in Exercise 4.19 show that the following are equivalent:
(i) M is a free R-module;
(ii) M is a projective R-module;
(iii) M
p
is a free R
p
-module.
4.21. Let R be a noetherian domain and M an affine monoid. Show that RŒM| is regular
if and only if Ris regular and M Š Z
m
˚Z
n
÷
for suitable mand n.
Hint: the main point is to show that the regularity of RŒM| implies that M Š Z
m
˚Z
n
÷
.
But regularity implies normality and so M = U(M) ˚ M
t
. Replace R by RŒU(M)| and
show that M
t
is generated by rank M
t
elements, using Exercise 4.20.
Alternative: Show that the regularity of kŒM| for a field k implies factoriality (Proposition
4.57) and use Exercise 4.12.
4.22. Translate the notions related to subintegral ring extensions into the language of
monoids and prove directly that every subintegral extension N of M is the filtered union
of submonoids N
i
that are obtained from M by a finite chain of elementary subintegral
extensions.
4.23. Prove Theorem 4.72.
Chapter 5
Isomorphisms and Automorphisms
5.A. Linear algebraic groups
Our standard reference for the theory of linear algebraic groups is Borel [9], to
which we refer the reader for proofs of the general facts surveyed below. Since
arithmetic issues never come up in our considerations, we remark right away that
in this section the ground field k is supposed to be algebraically closed.
An algebraic group over k is a group endowed with the structure of k-variety
such that the maps G × G ÷ G, (g
1
. g
2
) ÷ g
1
g
2
, and G ÷ G, g ÷ g
-1
, are
algebraic morphisms. A(homo)morphismof algebraic groups is a group homomor-
phism which is also a morphism of varieties. Every (Zariski) closed subgroup of
an algebraic group is an algebraic group in a natural way. If G is an affine variety,
we denote its coordinate ring by O(G). (In [9] the coordinate ring is denoted by
kŒG|, which we have reserved for the group ring.)
An algebraic group G is the disjoint union of its irreducible components of
which exactly one is a subgroup. It is called the unity component of G and denoted
G
0
. Furthermore, G
0
is a normal subgroup of G. Thus the quotient G,G
0
is a
finite group. If G = G
0
then G is called connected. In fact, G is connected if and
only if it is irreducible as a variety, but the term “irreducible” is avoided, since it
appears with a different meaning in the representation theory of G.
In the following we need the following facts about closedness:
Proposition 5.1. Let G. G
t
be algebraic groups.
(a) If ¡ : G ÷ G
t
is a morphism of algebraic groups, then Ker ¡ and Im¡
are closed subgroups of G and G
t
respectively. Moreover, ¡ (G
0
) is a closed
connected subgroup of (G
t
)
0
.
(b) If U
1
. . . . . U
n
are closed connected subgroups of G, then the subgroup gen-
erated (group theoretically) by their union is a closed connected subgroup.
Part (b) follows immediately from [9, 2.2]. Usually it is applied in order to show
that a group is connected, but we will need it to prove that a certain group is closed.
A classical example of an algebraic group is the general linear group of order n
for some n ÷ N, denoted GL
n
(k). It is naturally viewed as the affine k-subvariety of
147
148 5. Isomorphisms and Automorphisms
the affine space A
n×n÷1
with coordinates X
ij
, i. j = 1. . . . . n and Y , cut out by the
equation Y det
_
X
ij
_
= 1. It follows that the group GL(V ) of k-automorphisms
of a finite-dimensional vector space V carries the structure of a linear algebraic
group.
We are mainly interested in the special case of linear (algebraic) groups. They
are defined as closed subgroups of GL
n
(k) (or GL(V ), dimV < o).
The subgroup of diagonal matrices in GL
n
(k) is an example of a linear group:
its defining equations in GL
n
(k) are X
ij
= 0 for i = j . This group is denoted
by T
n
(k), or simply by T
n
if the ground field is clear from the context. It is nat-
urally identified with (k
+
)
n
. An algebraic torus over k is an algebraic group that
is isomorphic to T
n
(k) for some n so that we can write its elements as n-tuples

1
. . . . . ¸
n
) with ¸
i
÷ k
+
. It is a connected group whose coordinate ring is isomor-
phic to kŒX
˙1
1
. . . . . X
˙1
n
|.
The tautological action of T
n
on O(T
n
) = kŒX
˙1
1
. . . . . X
˙1
n
|, i. e. the one in-
duced by the group structure on T
n
, is given by (¸
1
. . . . . ¸
n
) : X
i
÷ ¸
i
X
i
, ¸
i
÷ k
+
,
i = 1. . . . . n.
We will use the following classical result of Borel [9, 11.3] in order to compare
monoid structures on rings.
Theorem5.2. All maximal (with respect to inclusion) algebraic tori in a linear group
G are conjugate.
Characters and diagonalizable groups. Acharacter of an algebraic group is a mor-
phism y : G ÷ k
+
. The characters form a group X(G) under multiplication, the
character group of G. Note that the characters form a linearly independent set of
functions fromG to k.
It is not hard to show that the monomials in O(T
n
) = kŒX
˙1
1
. . . . . X
˙1
n
| are
exactly the characters of T
n
(Exercise 5.1). Therefore the character group of T
n
is
isomorphic to Z
n
.
An algebraic group that is isomorphic to a closed subgroup of T
n
is called di-
agonalizable. The following theorem justifies this terminology.
Theorem 5.3. Let G be an algebraic group. Then the following are equivalent:
(a) G is diagonalizable;
(b) for every morphism G ÷ GL
n
(k), n ÷ N, the image of G is conjugate to a
subgroup of T
n
Ç GL
n
(k);
(c) G is isomorphic to a direct product T
m
× A, where A is a finite abelian
group without p-torsion if p = char k > 0;
(d) the characters form a k-basis of O(G).
The character group of a finite group A as in (c) is isomorphic to A, as fol-
lows quickly by decomposition into a product of cyclic groups. Together with the
fact that X(T
n
) Š Z
n
this shows that every finitely generated abelian group B
without p-torsion appears as the character group of a diagonalizable group (up
to isomorphism). The assignments G ÷ X(G) and B ÷ Hom(B. k
+
) induce
an equivalence between the category of diagonalizable groups and the category of
5.A. Linear algebraic groups 149
finitely generated abelian groups without p-torsion. (In Exercise 5.4 the reader is
asked to work out the details.)
Corollary 5.4. A sequence 0 ÷ G
t
÷ G ÷ G
tt
÷ 0 of diagonalizable groups is
exact if and only if the induced sequence 1 ÷ X(G
tt
) ÷ X(G) ÷ X(G
t
) ÷ 1 is
exact.
The crucial point in the proof of the corollary is the surjectivity of X(G) ÷
X(G
t
). It follows from the surjectivity of O(G) ÷ O(G
t
) and property 5.3(d) of
diagonalizable groups.
Rational representations. A rational representation of an algebraic group G is a
morphism of algebraic groups , : G ÷ GL(V ) where V is a finite-dimensional
k-vector space. One says that G acts rationally on V and sets
g: = (,(g))(:). g ÷ G. : ÷ V.
Since the vector space underlying a k-algebra is usually of infinite dimension, we
extend the notion of rational representation as follows: A group homomorphism
, : G ÷GL(V ) is called a rational representation if there exists a partially ordered
directed set I and a system V
i
, i ÷ I, of finite-dimensional subspaces such that
V =
_
i÷I
V
i
and g: ÷ V
i
for all i and : ÷ V
i
.
Suppose that G acts rationally on V , and let y be a character of G. Then the
weight space (or space of semi-invariants) of the character y is
V
y
= {: ÷ V : g: = y(g): for all g ÷ G].
In other words, V
y
consists of all those vectors in V that are eigenvectors for each
g ÷ G with corresponding eigenvalue y(g). The characters y such that V
y
= 0 are
called the weights of G on V .
Lemma 5.5. The weight spaces of a rational representation are linearly independent.
In other words, the linear map

y÷X(G)
V
y
÷V. (:
y
) ÷

:
y
.
is injective.
The rational representations of diagonalizable groups split into their weight
spaces.
Proposition 5.6. Let D be a diagonalizable group, acting rationally on V . Then
V =

y÷X(D)
V
y
.
In view of Lemma 5.5 it is enough to prove V =

V
y
, and this follows from
the the finite-dimensional case by passage to the limit. By Theorem 5.3 we can
then choose an isomorphism V Š k
n
such that D is mapped onto a subgroup
of the torus T
n
. Each weight space is now generated by the basis vectors e
i
that
belong to it, and each basis vector belongs to a weight space.
150 5. Isomorphisms and Automorphisms
Torus actions on monoid rings. In the following we let
G
k
(A)
denote the group of k-algebra automorphisms of a k-algebra A.
Let M be an affine monoid and A = kŒM|. An element t ÷ G
k
(A) is called a
toric automorphism (with respect to M) if t(x) = c
x
x, c
x
÷ k
+
. for all x ÷ M.
Evidently, the toric automorphisms form a subgroup T
k
(M) of G
k
(A). It is of
course enough for t to be toric that t(x
i
) = c
i
x
i
for some system of generators
x
1
. . . . . x
n
of M.
We can describe the toric automorphisms in two ways, first via an isomorphism
kŒgp(M)| Š kŒY
˙1
1
. . . . . Y
˙1
r
|, r = rank M.
Lemma 5.7. Let M be an affine monoid of rank r and e
1
. . . . . e
r
a basis of gp(M).
Consider the k-algebra isomorphism kŒY
˙1
1
. . . . . Y
˙1
r
| Š kŒgp(M)| induced by the
assignment Y
i
÷ e
i
. It induces a homomorphism T
r
÷ G
k
(kŒM|), which maps
T
r
isomorphically onto T
k
(M).
Proof. The tautological action of T
r
on kŒY
˙1
1
. . . . . Y
˙1
r
| is transferred via the
isomorphism to kŒgp(M)|. Evidently t
_
kŒM|
_
= kŒM| for all t ÷ T
r
, and each
monomial in M is an eigenvector for t since it corresponds to a monomial in the
Laurent polynomial ring. This shows that t acts as a toric automorphism on kŒM|.
Conversely, suppose that o is a toric automorphism of kŒM|. Then o extends to
an automorphism of kŒgp(M)|, since kŒgp(M)| arises by the inversion of all terms
uj, u ÷ k
+
, j ÷ M. Moreover, the set of terms is mapped onto itself by o since o
is toric.
But o extends to a toric automorphismof kŒgp(M)|, and so it is enough to show
that every toric automorphism of the Laurent polynomial ring belongs to T
r
. This
is clear since it is completely determined by its values on the indeterminates. (See
Exercise 5.14 for a description of the automorphismgroup of a Laurent polynomial
ring.)
The lemma allows us to consider T
k
(M) as a (linear) algebraic group. In fact
we can use the isomorphism T
r
Š T
k
(M) to transfer the geometry of T
r
to
T
k
(M). While the isomorphism depends on the choice of basis for gp(M), the
geometry obtained does not. Every choice of basis is via an element of GL
r
(Z),
which can be understood as an algebraic automorphism of T
r
.
Second, we characterize the toric automorphisms in terms of the defining ideal
of kŒM| as a residue class ring of a polynomial ring.
Lemma 5.8. Let M be an affine monoid, A = kŒM|, let x
1
. . . . . x
n
be a system of
generators of M and I the kernel of the epimorphismkŒX
1
. . . . . X
n
| ÷A, X
i
÷x
i
.
Then an assignment x
i
÷ c
i
x
i
, c
i
÷ k
+
, i = 1. . . . . n, can be extended to
a (toric) automorphism t of A if and only if ¡ (c
1
. . . . . c
n
) = 0 for all binomials
¡ ÷ I.
Proof. (a) Suppose first that t ÷ T
k
(M). Then t(x
i
) = c
i
x
i
, c
i
÷ k
+
, for all i by
the definition of T
k
(M). We have to show that ¡ (c
1
. . . . . c
n
) = 0 for all ¡ ÷ I.
Since I is generated by the binomials ¡ = X
a
1
1
X
a
n
n
÷ X
b
1
1
X
b
n
n
÷ I (see
5.A. Linear algebraic groups 151
Proposition 4.21), it is enough to consider such binomials. Since ¡ (x
1
. . . . . x
n
) =
0 and t is an endomorphism of A, we also have ¡ (c
1
x
1
. . . . . c
n
x
n
) = 0, and so
c
a
1
1
c
a
n
n
x
a
1
1
x
a
n
n
= c
b
1
1
c
b
n
n
x
b
1
1
x
b
n
n
.
This is only possible if c
a
1
1
c
a
n
n
= c
b
1
1
c
b
n
n
, whence ¡ (c
1
. . . . . c
n
) = 0.
Let now c
1
. . . . . c
n
÷ k
+
be given such that ¡ (c
1
. . . . . c
n
) = 0 for all binomials
¡ ÷ I. Consider the toric automorphism o ÷ T
n
of kŒX
1
. . . . . X
n
| given by the
assignment X
i
÷c
i
X
i
.
Since ¡ (c
1
x
1
. . . . . c
n
x
n
) = 0 for all binomials in I and since I is generated by
binomials, it follows that o maps I into itself. Hence it induces an endomorphism
t of A = kŒX
1
. . . . . X
n
|,I. Evidently t is a toric automorphism of A.
Corollary 5.9. With the notation of Lemma 5.8, the map T
k
(M) ÷ T
n
, t ÷
(c
1
. . . . . c
n
), is an algebraic isomorphism of T
k
(M) with the closed subgroup of T
n
defined by the ideal I kŒX
˙1
1
. . . . . X
˙1
n
|.
It will be very important in the following that we can identify the monoid M
by the action of T
k
(M) on A = kŒM|. Apart from part (e), the following lemma
is hardly more than a reformulation of what has already been proved, but it adds a
representation-theoretic flavor to it.
Lemma 5.10. With notation as in Lemma 5.8 we have the following:
(a) For each x ÷ M the function y
x
: T
k
(M) ÷ k
+
, y
x
(t) = t(x),x, is a
character.
(b) If x = y, then y
x
= y
y
.
(c) Let z ÷ A, z = 0, and x ÷ M. If t(z) = y
x
(t)z for all t ÷ T
k
(A), then
z = cx for some c ÷ k
+
.
(d) The characters y
x
form a monoid isomorphic to M.
(e) T
k
(A) is a maximal torus in G
k
(A).
Proof. (a) We can use the identification of T
k
(M) with T
r
given by Lemma 5.7.
Suppose that x = e
a
1
1
e
a
r
r
with a
1
. . . . . a
r
÷ Z. Then y
x
= Y
a
1
1
Y
a
r
r
as a
function on T
r
, and so it is a character.
(b) This is obvious, since x and y correspond to different monomials in O(T
r
).
(c) If t(z) = y
x
(t)z for all t ÷ T
k
(A), then z belongs to the weight space of
the character y
x
on A. It is the one-dimensional subspace spanned by x.
(d) The assignment x ÷y
x
is clearly a monoid homomorphism. It is injective
by (b).
(e) Suppose g is a k-automorphism that commutes with all elements t ÷ T .
Let x ÷ M. Then
t(g(x)) = g(t(x)) = g(y
x
(t)x) = y
x
(t)g(x)
for all t ÷ T. It follows from(c) that g(x) = c
x
x for all x ÷ M, c
x
÷ k
+
. Therefore
g is a toric automorphism.
An important consequence of Lemma 5.10 is that we can identify the monoid
M already from the characters y
x
where x is running through a system of gener-
ators of M. This shows that M is completely determined by the action of T
k
(M)
152 5. Isomorphisms and Automorphisms
on the vector space spanned by a system of generators of M. The next lemma
indicates how Borel’s theorem on conjugate tori may be applied to isomorphism
problems for monoid algebras.
Lemma 5.11. Let V be a k-vector space, and let T
1
. T
2
Ç GL(V ) be tori acting
rationally on V . If there exists g ÷ GL(V ) such that g
-1
T
1
g = T
2
, then the
weights of T
1
and T
2
occurring in V generate isomorphic submonoids of X(T
1
) and
X(T
2
), respectively.
Proof. The group isomorphism T
1
÷ T
2
, t ÷ g
-1
tg, induces a group isomor-
phism v : Hom(T
2
. k
+
) ÷ Hom(T
1
. k
+
). Since we can similarly consider v
-1
, it
is enough to show that v(y) is a character of T
1
occurring in V if y is a character
of T
2
occurring in V . Let y be a character of T
2
occurring in V .
Let n ÷ V
y
, n = 0, t ÷ T
1
, and o = g
-1
tg. Then
t(g(n)) = g(o(n)) = g(y(o)n) = y(o)g(n) =
_
v(y)(t)
_
(g(n)).
Thus the one-dimensional space kg(n) is invariant under the action of T
1
. Since
T
1
acts rationally on V , it acts rationally on every subspace of V , and so kg(n)
must be contained in a weight space of T
1
. Its associated character is v(y).
Groups of graded automorphisms. While G
k
(A) is not a linear algebraic group in
general, this holds for the group
1
k
(A)
of graded automorphisms under suitable hypotheses on A:
Proposition 5.12. Let A = k ˚ A
1
˚ A
2
˚ If A a finitely generated, positively
graded k-algebra. Then 1
k
(A) a linear group.
Proof. First of all, the finite generation of A over k implies that dim
k
A
i
< ofor
all i ÷ N. we choose d ÷ Nsuch that A
1
˚ ˚A
d
generates Aas a k-algebra. Put
n = dimA
1
÷ ÷dimA
d
and fix a k-basis B = {z
1
. . . . . z
n
] Ç A
1
˚ ˚A
d
such
that each z
i
belongs to some subspace A
j
. We have a group embedding 1
k
(A) ÷
GL
n
(k), defined as follows:
; ÷(a
ij
). ;(z
j
) =
n

i=1
a
ij
z
i
. j = 1. . . . . n.
Let F
s
(X
1
. . . . . X
n
) ÷ kŒX
1
. . . . . X
n
|, s = 1. . . . . t , be a finite systemof polynomials
generating the relations between the elements z
1
. . . . . z
n
÷ A. Then the entries of
the matrix (a
ij
), corresponding to ;, satisfy the equations
F
s
(a
11
z
1
÷ ÷a
n1
z
n
. . . . . a
1n
z
1
÷ ÷a
nn
z
n
) = 0. s = 1. . . . . t.
a
ij
= 0. deg z
i
= deg z
j
.
Conversely, if the entries of an invertible n ×n matrix satisfy these equations, then
the matrix corresponds to an element of ; ÷ 1
k
(A). In other words, the image of
1
k
(A) in GL
n
(k) is the set of solutions of a system of polynomial equations, and
therefore a closed subgroup.
5.B. Invariants of diagonalizable groups 153
Let M be an affine positive monoid. Then a positive grading on M induces a
positive grading on kŒM|, and it follows fromLemma 5.10 that T
k
(M) is a maximal
torus in the linear group 1
k
(A).
5.B. Invariants of diagonalizable groups
Let k be an algebraically closed field as in the previous section. We consider the
group GL
n
(k) of invertible n × n matrices over k as the group of homogeneous
(with respect to total degree) k-algebra automorphism of the polynomial ring R =
kŒX
1
. . . . . X
n
|, where each matrix (a
ij
) ÷ GL
n
(k) acts by a linear substitution:
g : R ÷R. X
j
÷
n

i=1
a
ij
X
i
. (5.1)
The reader should note that this equation does not describe the action of GL
n
(k)
on R = O(k
n
) induced by the standard isomorphism GL
n
(k) Š GL(k
n
). In fact,
the induced action yields an antihomomorphism due to the contravariance of the
functor O. The action considered in (5.1) is the composition of the induced action
with the transposition of matrices.
Let D be a subgroup of GL
n
(k). Then ¡ ÷ Ris an invariant of D, if g(¡ ) = ¡
for all g ÷ D. The ring of invariants of D is
R
D
= {¡ ÷ R : g(¡ ) = ¡ for all g ÷ D].
For fixed ¡ , the equation g(¡ ) = ¡ imposes a system of polynomial equations on
the entries of a matrix of g. Therefore we can always assume that D is closed. Then
the ring of invariants is nothing but the weight space of the trivial character.
Note that such an element ¸ = (¸
1
. . . . . ¸
n
) ÷ T
n
transforms the monomial
X
u
1
1
X
u
n
n
into the term
¸
u
1
1
¸
u
n
n
X
u
1
1
X
u
n
n
.
Therefore ¸(¡ ) = ¡ for ¡ ÷ Rif and only if each monomial occurring in ¡ is left
invariant by ¸.
Let D be a subgroup of T
n
. Then it follows that R
D
is a monomial subalgebra
of R. Much more is true, however:
Theorem 5.13. Let D be a subgroup of T
n
. Then kŒX
1
. . . . . X
n
|
D
is a normal affine
monoid algebra kŒM| where M is a pure submonoid of Z
n
÷
.
Proof. We have already seen that the ring of invariants is of the form kŒM|. We
must show that M = gp(M) ¨ Z
n
÷
. Then M is obviously normal and finitely
generated by Gordan’s lemma. Let x. y ÷ M such that xy
-1
= z ÷ Z
s
÷
. We have
x = yz where all three monomials lie in the polynomial ring. But if x and y are
invariant under D, then z is, too. Thus z ÷ M, and we are done.
Remark 5.14. Instead of a subgroup of T
n
we can more generally consider a diag-
onalizable group D that acts rationally on the vector space kX
1
÷ ÷kX
n
since
the image of D in GL
n
(k) is conjugate to a subgroup of T
n
; see Theorem 5.3.
154 5. Isomorphisms and Automorphisms
Our next goal is the converse of this theorem, namely to show that every pure
subalgebra of kŒX
1
. . . . . X
n
| that is generated by finitely many monomials can be
realized as a ring of invariants. However, one sees very quickly that this goal can
be reached only if k has the right characteristic.
Example 5.15. Consider A = kŒX
2
| Ç R = kŒX|. If char k = 2, it is impossible
to realize A as a ring of invariants of R. The only element of T
1
= GL
1
(k) leaving
X
2
invariant, is 1 since ¸
2
X
2
= X
2
implies ¸ = 1. But the ring of invariants of {1]
is Ritself. On the other hand, if char k = 2, then A = R
¦1,-1¦
.
Let G be a subgroup of Z
n
. Then we can find a system of diophantine ho-
mogeneous linear equations and congruences such that G is its set of solutions:
(x
1
. . . . . x
n
) ÷ G if and only if
a
i1
x
1
÷ ÷a
in
x
n
= 0. i = 1. . . . . p. a
ij
÷ Z.
b
i1
x
1
÷ ÷b
in
x
n
÷ 0 (n
i
). i = 1. . . . . q. b
ij
. n
i
÷ Z.
(5.2)
Moreover, we can assume that p = n÷rank G and that the congruences are chosen
in such a way that n
1
n
q
is the order of the torsion group of Z
n
,G. Then the
monoid M = G ¨ Z
n
÷
consists of the nonnegative solutions of this system.
Let k an algebraically closed field whose characteristic does not divide the order
of an element of G – in other words, char k does not divide n
1
n
q
. We use the
coefficients of the system and the moduli of the congruences in order to define a
subgroup D of T
n
by specifying a set of generators. Let ç
i
is a primitive n
i
th root
of unity. Then the generators of D are given by

a
i1
. . . . . ¸
a
i n
). i = 1. . . . . p. ¸ ÷ k
+

b
j1
j
. . . . . ç
b
jn
j
). j = 1. . . . . q.
(5.3)
Theorem 5.16. Let M be a pure submonoid of Z
n
÷
, and let G = Z
n
, gp(M). Sup-
pose that G contains no element of order char k. Then there exists a closed subgroup
D Ç T
n
such that kŒM| = kŒX
1
. . . . . X
n
|
D
.
More precisely, if M is the set of nonnegative solutions of the system (5.2), then
D can be chosen as the subgroup generated by the elements listed in (5.3).
Proof. It is enough to check that exactly the monomials in M are invariant under
each of the generators of D. The torus element (¸
a
i1
. . . . . ¸
a
i n
) sends X
u
1
1
X
u
n
n
to
¸
a
i1
u
1
÷···÷a
i n
u
n
X
u
1
1
X
u
n
n
and the coefficient is equal to 1 for all ¸ ÷ k
+
if and only a
i1
u
1
÷ ÷a
in
u
n
= 0,
simply because k has infinitely many elements.
To complete the argument one has only to check that ç
b
i1
u
1
÷···÷b
i n
u
n
j
= 1 if and
only if b
i1
u
1
÷ ÷b
in
u
n
÷ 0 (n
i
), and this is clear by the choice of ç
j
.
Though we have used a specific description of M, the group D only depends
on M:
Corollary 5.17. With the notation of Theorem 5.16, the group D is the largest sub-
group of T
n
such that kŒM| = R
D
. Moreover, X(D) Š Z
n
, gp(M).
5.B. Invariants of diagonalizable groups 155
Proof. For a monomial x = X
a
1
1
X
a
n
n
let y
x
be the characters of T
n
given by
x ÷ O(T
n
). That the monomials x ÷ M are invariants of D can equivalently
expressed by the relation
D Ç U =
_
x÷M
Ker y
x
.
We have
kŒM| Ç R
U
Ç R
D
= kŒM|.
It follows that kŒM| = R
U
, and U is clearly the largest subgroup of T
n
having
kŒM| as its ring of invariants. It remains to show that D = U.
The chain O(T
n
) ÷O(U) ÷O(D) induces a chain of epimorphisms
X(T
n
) Š Z
n
÷X(U) ÷X(D).
By the construction of D, exactly the characters y
x
, x ÷ gp(M), vanish on D, so
that X(D) Š Z
n
, gp(M) as claimed. on the other hand, all these characters vanish
also on U. It follows that the epimorphism X(U) ÷ X(D) is an isomorphism.
Hence O(U) ÷O(D) is an isomorphism, and so is the embedding D Ç U.
It follows fromthe corollary in conjunction with Theorem5.10 that the hypoth-
esis on char k in Theorem 5.16 is not only sufficient, but also necessary. This is also
a consequence of Exercise 5.6.
We know from Theorem 2.28 that every positive affine monoid can be embed-
ded into a monoid Z
n
÷
for suitable n as a pure submonoid. The theorem shows
that every such embedding yields a realization of kŒM| as an invariant subring (if
the hypothesis on char k is satisfied). If we embed M in such a way that it is even
integrally closed in Z
n
÷
, and this is likewise possible by Theorem 2.28, then only a
torus action is needed for a representation of kŒM| as a ring of invariants, and the
hypothesis on the characteristic of k can be dropped.
Corollary 5.18. Let M be an integrally closed and pure submonoid of Z
n
÷
. Then
there exists a subtorus T of T
n
such that kŒM| = kŒX
1
. . . . . X
n
|
T
.
We conclude the section by a statement about weight spaces (or semi-invari-
ants).
Theorem 5.19. Let D Ç T
n
be a closed subgroup, and R = kŒX
1
. . . . . X
n
|. Then
R is the direct sum of the weight spaces R
y
, y ÷ X(D). For every y ÷ X(G) the
weight space R
y
is a finite direct sum of finitely generated R
D
-modules of rank 1.
If X(D) = Z
n
, gp(M) where M is the monoid of D-invariant monomials in
R, then R
y
is an R
D
-module of rank ¸ 1 for every y ÷ X(D).
If, in addition M contains a monomial X
a
1
1
X
a
n
n
with a
i
> 0 for all i =
1. . . . . n, then rank R
y
= 1 for all y ÷ X(D).
Proof. The character group of D is a quotient Z
n
,U where U is a subgroup of Z
n
(see Corollary 5.4). Clearly R
y
is spanned as a k-vector space by all monomials in
R that belong to a single residue class modulo U. Each such residue class is given
by the set of solutions of an inhomogeneous version of the system (5.2). Therefore
we can apply Corollary 2.11. It implies that R
y
is a finitely generated R
D
-module.
156 5. Isomorphisms and Automorphisms
Each residue class modulo U splits into residue classes modulo gp(M), and the
subspace spanned by the monomials in a residue class modulo gp(M) is a rank 1
module over R
D
. In fact it appears as a coset module of the pure submonoid M of
Z
n
÷
(see Proposition 4.39). Clearly, R
y
splits into its coset submodules.
If U = gp(M), then R
y
has rank 1, provided it is nonzero. If M contains
a monomial with strictly positive coefficients, then every residue class modulo
gp(M) contains a monomial in R, and the corresponding weight space is = 0.
5.C. The isomorphism theorem
In this section we address the following question, known as the isomorphism prob-
lem for monoid rings: if RŒM| and RŒN| are isomorphic as R-algebras, are then
M and N isomorphic? The similar question for group rings (not necessarily com-
mutative) is a much studied classical problem. See the notes of this chapter for the
general background of the isomorphism problem.
The main result in this section answers the isomorphism problem for affine
monoids positively, under mild restrictions:
Theorem5.20. Let M and N be affine monoids and Ra ring. Suppose ¡ : RŒM| ÷
RŒN| is an R-algebra isomorphism. Then M Š N if one of the following conditions
is satisfied:
(a) M and N are positive and ¡ respects the monomial augmentation ideals;
(b) M is normal or homogeneous;
(c) M is positive and there exists a prime ideal n of Rsuch that char R,n = 0.
The condition in (a) means ¡ (m) ÷ R
_
N \{1]
_
for every m ÷ M\{1]; in other
words, ¡ respects the Kernels of the augmentation homomorphisms RŒM| ÷ R
and RŒN| ÷ R that map all nonunit monomials to 0. It is of course enough that
M is positive. Then N must also be positive, since U(M) Š U(RŒM|), U(R) Š
U(RŒN|), U(R) Š U(N) by Proposition 4.13. The conditions in (b) and (c) can
be refined further; see Lemma 5.28.
The first and obvious reduction step in the proof of Theorem5.20 is the passage
to a residue class ring R,n with respect to a prime ideal n and to an algebraic
closure of its field of fractions. Whenever necessary we may therefore assume that
R = k is an algebraically closed field.
Then the theorem will be reduced to the augmented case. Note that it is very
easy to find nonaugmented isomorphisms of monoid algebras: for example, con-
sider ¡ : kŒX| ÷ kŒX|, ¡ (X) = X ÷ 1. However, it will turn out that nonaug-
mented isomorphisms can only appear in the presence of polynomial variables,
roughly speaking.
After the passage to the augmented case, we will identify RŒM| and RŒN|.
Then we are dealing with two monomial structures on the same R-algebra that
have the same augmentation ideal m. Finally Borel’s theorem on maximal tori will
be applied to a suitable truncation modulo m
d
for d ;0.
5.C. The isomorphism theorem 157
Before delving into the technical details of the proof, we explain the relevance
of Borel’s theorem by proving an instructive special case of Theorem 5.20 and a
variant for residue class rings of polynomial rings modulo monomial ideals.
Two applications of Borel’s theorem. If the isomorphism RŒM| Š RŒN| does not
only respect the augmentation ideals, but even a grading, then the case is rather
simple.
Proposition 5.21. Let R be a ring and M and N be positive affine monoids, both
endowed with a positive grading. If RŒM| Š RŒN| as graded R-algebras, then
M Š N.
Proof. As pointed out above, we can replace Rby an algebraically closed field and
identify the k-algebras A = kŒM| and kŒN| along a graded isomorphism. Then
Lemma 5.7 provides us with two (rank M)-dimensional tori in the linear group
1
k
(A) (Proposition 5.12) acting rationally on A. Denote them by T
1
and T
2
. By
Theorem 5.2 there exists g ÷ 1
k
(kŒM|) such that g
-1
T
1
g = T
2
. Thus Lemma 5.11
implies that the monoids generated by the weights of T
1
and T
2
occurring in Aare
isomorphic. But the first of these monoids is isomorphic to M and the second is
isomorphic to N, as follows from Lemma 5.10.
The proof of the next theorem contains a reduction to the augmented case.
Such a reduction must be based on the ring-theoretic behavior of monoid ele-
ments, which is preserved under ring isomorphism. Here we can use the property
of being a zerodivisor.
Theorem 5.22. Let I Ç RŒX
1
. . . . . X
m
| and J Ç RŒY
1
. . . . . Y
n
| be monomial ideals
such that {X
1
. . . . . X
m
]¨I = 0, {Y
1
. . . . . Y
n
]¨J = 0 and ¡ : RŒX
1
. . . . . X
m
|,I ÷
RŒY
1
. . . . . Y
n
|,J be an R-algebra isomorphism. Then m = n and there exists a
bijective mapping 0 : {X
1
. . . . . X
m
] ÷{Y
1
. . . . . Y
m
] transforming I into J.
Proof. As in Proposition 5.21, we can assume that R = k is an algebraically closed
field.
The monomial ideal I has a unique minimal set of monomial generators
j
1
. . . . . j
q
. Let {i
1
. . . . . i
u
] be the set of indices i ÷ {1. . . . . n] such that X
i
does
not divide any j
j
, and let {j
1
. . . . . j
¡
] be the complementary set. Then
A Š
_
kŒX
j
1
. . . . . X
j
v
|,(j
1
. . . . . j
q
)
_
ŒX
i
1
. . . . . X
i
u
|.
Clearly, X
j
1
. . . . . X
j
v
are zerodivisors modulo I, and X
i
1
. . . . . X
i
u
are not.
Let m be the maximal ideal of A = kŒX
1
. . . . . X
m
|,I generated by the residue
classes x
i
of the indeterminates X
i
. Then all zerodivisors of A are contained in m,
and the same holds for the maximal ideal n of B = kŒY
1
. . . . . Y
n
|,J generated by
the residue classes y
i
of the Y
i
.
Therefore, if ¡ (x
i
) … n, then ¡ (x
i
) is not a zerodivisor in B, so x
i
is not a
zerodivisor in A, and i ÷ {i
1
. . . . . i
u
]. Since X
i
1
. . . . . X
i
u
are polynomial variables
in Aover its subalgebra A
t
= kŒX
j
1
. . . . . X
j
v
|,(j
1
. . . . . j
q
), the substitution X
i
÷
X
i
÷ ¡ (x
i
)(0), i ÷ {1. . . . . n] induces an A
t
-algebra automorphism g of A. (Here
¡ (x
i
)(0) is the constant term of ¡ (x
i
) in its monomial representation.)
158 5. Isomorphisms and Automorphisms
The k-algebra isomorphism ¡g : A ÷ B maps the maximal ideal m onto the
maximal ideal n, both being the kernel of the natural augmentation map in A and
B, respectively. Replace ¡ by ¡g.
Our next goal is to pass to an isomorphism of A and B that is even homoge-
neous with respect to the standard grading by total degree. To this end we con-
sider the associated graded rings gr
m
(A) =

o
k=0
m
k
,m
k÷1
and gr
n
(B). Since
¡ (m) = n, we obtain an induced isomorphism gr(¡ ) : gr
m
(A) ÷gr
n
(B).
The passage to the associated graded rings is always possible if a ring isomor-
phism respects the augmentation ideals. But in the present situation both alge-
bras are generated by their homogeneous degree 1 elements. Therefore the natural
k-vector space isomorphism A Š gr
m
(A) is even a k-algebra isomorphism, and
similarly for B. Identifying both rings with their associated graded rings and re-
placing ¡ by gr(¡ ), we may now assume that our rings are isomorphic under a
homogeneous isomorphism.
We identify A and B through ¡ and let 1 be the group of graded k-automor-
phisms of A. This is a linear k-group by Proposition 5.12. It contains two copies of
T
n
(k): one consisting of the automorphisms of form x
i
÷ a
i
x
i
, and the other
consisting of the automorphisms of form y
i
÷ a
i
y
i
, a
i
÷ k
+
, i = 1. . . . . n.
Denote these tori by T
1
and T
2
. Both are maximal tori in the group of linear k-
automorphisms of the graded component R
1
(generated by the linearly indepen-
dent elements x
1
. . . . . x
n
). A fortiori T
1
and T
2
are maximal in the subgroup 1 .
By Borel’s theorem we find an automorphism g with g
-1
T
1
g = T
2
. It maps
each x
i
to an element c
i
y
j(i)
, c
i
÷ k
+
, i = 1. . . . . n. After a “toric correction”
and a permutation we can assume g(x
i
) = y
i
. Then x
i
1
x
i
p
= 0 if and only if
y
i
1
y
i
p
= 0, and this implies that the monomial ideals I and J correspond to
each other under the assignment X
i
÷Y
i
.
Example 5.23. Let z be simplicial complex and k a field. The commutative k-
algebras naturally associated with z are
(a) its Stanley-Reisner ring kŒz| = kŒX
¡
: : ÷ vert(z)|,I(z) (see p. 125), and
(b) kŒz|
sqf
= kŒz|,
_
X
2
¡
: : ÷ vert(z)
_
.
The algebra in (b) is the “squarefree” version of kŒz|: only squarefree monomials
are nonzero in kŒz|
sqf
, and the nonzero monomials are in bijective correspondence
with the faces of z, so that its Hilbert function is given by the ¡ -vector of z.
Theorem 5.22 implies that zis uniquely determined by the k-algebra structure
of any of these two algebras: if kŒz
1
| Š kŒz
2
| or kŒz
1
|
sqf
Š kŒz
2
|
sqf
as k-algebras,
then z
1
Š z
2
as simplicial complexes. See Exercise 5.10 for an exterior algebra
analogue.
Reduction to the augmented case. A key argument for the reduction of Theorem
5.20 to the augmented case is the following lemma.
Lemma 5.24. Let k be a field and R =

o
i=0
R
i
a positively graded affine normal
k-domain. Then every nonunit x outside the maximal ideal m =

o
i=1
R
i
is a
product of prime elements.
5.C. The isomorphism theorem 159
Proof. Let T be the multiplicative system generated by x. The natural map
R ÷ R
m
factors through T
-1
R. by Nagata’s theorem 4.45(b) we have a se-
quence Cl(R) ÷ Cl(T
-1
R) ÷ Cl(R
m
) of surjective homomorphisms induced by
the extension of divisorial ideals. By Theorem 4.57 the composition is an isomor-
phism, and Nagata’s theorem implies that every divisorial prime ideal containing
x must be principal, i. e. generated by a prime element. Up to a unit factor, x is the
product of powers of these prime elements. (In a noetherian ring, every nonunit
x = 0 is a product of irreducible elements, and if the minimal prime overideals of
(x) are principal, the irreducible factors generate them.)
Lemma 5.25. Let M be a positive normal affine monoid, k a field, and suppose
that A = kŒM| =

o
i=0
A
i
is positively graded (independently of any grading
of M). Let n =

o
i=1
A
i
and set M
t
= M \ n. Then M
t
= x
Z
1
× × x
Z
m
is a free monoid, generated by pairwise non-associated prime elements x
1
. . . . . x
m
.
Moreover, M = M
t
×M
tt
where M
tt
= M ¨F, and F is the unique face of R
÷
M
not containing any of x
1
. . . . . x
m
.
Proof. Note that M
t
is a submonoid of M since it is the complement of a prime
ideal. (Therefore M
t
= M ¨ G for a face G of R
÷
M.) in particular M
t
is finitely
generated. It is enough to show that M
t
is generated by prime elements, since the
rest then follows from Proposition 2.20. Note that two different prime elements in
M must be coprime since M is positive.
In the previous lemma we have already shown that every nonunit x outside n
is indeed a product of prime elements, and if x is a monomial, then it is evidently
the product of monomial prime elements.
With the next proposition we have reduced Theorem 5.20 to the augmented
case if M is positive and normal.
Lemma 5.26. Let k be a field and M and N be two positive normal affine monoids.
If kŒM| and kŒN| are isomorphic as k-algebras, then they are isomorphic also as
augmented k-algebras.
Proof. Let us identify A = kŒM| and kŒN| along the given k-algebra isomor-
phism. Let ; be a positive grading on N. It defines a positive grading on A. Let
n be the maximal ideal of A generated by the homogeneous elements of positive
degree under ;. Then n is of course the augmentation ideal of the monoid algebra
A = kŒN| with respect to its monomial structure defined by N.
It remains to find an automorphism ˛ of A that maps all nonunit elements of
M into n. Since M splits as in Lemma 5.25 we have
A =
_
kŒM
tt
|
_
ŒZ
m
÷
| =
_
kŒM
tt
|
_
ŒX
1
. . . . . X
m
|
where we have replaced x
1
. . . . . x
m
by X
1
. . . . . X
m
to indicate that they are alge-
braically independent over kŒM
tt
|. Moreover, the elements of M
tt
are already con-
tained in n.
Now we define ˛ as the kŒM
tt
|-automorphism of A given by the substitution
X
i
÷ X
i
÷ ¸
i
, where ¸
i
is the constant coefficient of X
i
in its representation as a
sum of terms av, a ÷ k
+
, v ÷ N.
160 5. Isomorphisms and Automorphisms
To finish the reduction in the normal case, we only need to get rid of the units,
and this is very easy, since a normal affine monoid splits into its group of units and
a positive normal monoid (see Proposition 2.26).
Proposition 5.27. Let k be a field and M and N be two positive normal affine
monoids, M Š U(M) × M
t
, N Š U(N) × N
t
. If kŒM| and kŒN| are isomorphic
as k-algebras, then there exists a field Lsuch that LŒM
t
| and LŒN
t
| are isomorphic
also as augmented L-algebras.
Proof. Let ¡ : kŒM| ÷kŒN| be an isomorphism. Then ¡ maps the group of units
U(kŒM|) = U(M) × k
+
(see Proposi6tion 4.12) isomorphically onto U(kŒN|) =
U(N) × k
+
. Therefore ¡ maps kŒU(M)| isomorphically onto kŒU(N)|. Moreover,
taking residue classes modulo k
+
, we immediately get U(M) Š U(N).
Since kŒM| Š
_
kŒU(M)|
_
ŒM
t
| and because of the analogous decomposition
for kŒN|, we are allowed to identify the Laurent polynomial rings kŒU(M)| and
kŒU(N)| and denote this ring by R. After the identification, kŒM| Š kŒN| as R-
algebras. Inverting the nonzero elements of R (or working modulo a maximal
ideal), we obtain a field L such that LŒM
t
| Š LŒN
t
|. By Lemma 5.26 these two
algebras are even isomorphic as augmented L-algebras.
Now let M and N be positive affine monoids. If kŒM| and kŒN| are isomorphic
as k-algebras, then their normalizations kŒ
¨
M| and kŒ
¨
N| are isomorphic over k as
well. (See 4.37 for the normalization of a monoid algebra.) The monoids
¨
M and
¨
N
are also positive, and so we can apply Lemma 5.26 to showthat they are isomorphic
as augmented k-algebras. However, it is not clear that the corrected isomorphism
maps kŒM| to kŒN|.
Lemma 5.28. Let M and N be affine monoids and suppose that kŒM| and kŒN|
are isomorphic k-algebras. Furthermore, suppose that for each prime element of
¨
M
there exists q ÷ N such that x
q
÷ M and q is not divisible by the characteristic of
k. Then kŒM| and kŒN| are isomorphic as augmented k-algebras.
Proof. We choose a positive grading on N and consider kŒN| as a positively
graded k-algebra. Let ¡ be the given isomorphism from kŒM| to kŒN|. If its ex-
tension to the normalization is augmented, then we are done. Assume it is not.
Then we “correct” it as in the proof of Lemma 5.26 by changing the values of
those prime elements x of
¨
M for which ¡ (x) has a nonzero constant term ¸ in its
representation as a linear combination of monomials in N.
We claim: if ¸ = 0, then ¡ (x) ÷ kŒN|. Assume the contrary, and let v be an
element of least degree in
¨
N \N such that v ÷ supp(¡ (x)), Since x
q
÷ M, we have
¡ (x)
q
÷ kŒN|. But under the hypothesis on the characteristic of k, the monomial
v belongs also to supp(¡
q
), as the reader may check. This is a contradiction.
This shows that we have to correct the values only for such prime elements x
for which ¡ (x) ÷ kŒN|. We can do this successively for the finitely many prime
elements, and it remains to show that (¡ (x) ÷ ¸)
k
¡ (y) ÷ kŒN| for y ÷ M, y not
divisible by x. This follows immediately if ¡ (y) ÷ kŒN|.
5.C. The isomorphism theorem 161
Now it is an easy exercise to show that z ÷ kŒN| if there exists n ÷ kŒ
¨
N|
with nonzero constant term such that nz ÷ kŒN|. Apply this to z = ¡ (y) and
n = ¡ (x
q
).
The hypothesis of Lemma 5.28 are satisfied if char k = 0 or x ÷ M for all prime
elements of
¨
M, and especially if all extreme integral generators of
¨
M belong to
M. The latter is the case if M is homogeneous. Together with Proposition 5.27
this covers the reduction to the augmented case under all the hypotheses listed in
Theorem 5.20.
The augmented case. The ring-theoretic arguments in the last subsection have
put us in a position where we can again apply arguments form the theory of linear
algebraic groups that we have prepared in Section 5.A. In the graded case (Propo-
sition 5.21) we had a natural candidate for a linear algebraic group to which we
could apply Borel’s theorem on maximal tori. If M and N are positive and the
isomorphism A = kŒM| Š kŒN| respects the augmentation ideals, then, after
the identification of kŒM| and kŒN| along the given isomorphism, we can replace
G
k
(A) by
G
k
(A. m) = {ç ÷ G
k
(A) : ç(m) = m).
where m is the common augmentation ideal, generated by M \ {1] as well as by
N \ {1]. Clearly, T
k
(M). T
k
(N) Ç G
k
(A. m). They are both maximal tori in
G
k
(A. m) by Lemma 5.10(d). However, G
k
(A. m) is almost never a linear algebraic
group. Therefore it is necessary to pass to a truncation of A.
For every ç ÷ G
k
(A. m) one has ç(m
d
) = m
d
for all d. Therefore ç induces
a k-linear map ¨ ç on the vector space m,m
d
. The assignment ç ÷ ¨ ç is a group
homomorphism fromG
k
(A. m) into GL(m,m
d
). We denote it by ˚
d
.
We want to show that the images of T
k
(M) and T
k
(N) are maximal tori in
˚
d
_
G
k
(A. m)
_
if d ; 0. It is of course enough to consider one of them, since all
data so far only depend on m.
Lemma 5.29. Let M be a positive affine monoid, A = kŒM| and m the maximal
ideal generated by M\{1]. Then the (isomorphic) image of the torus T = T
k
(M) Ç
G
k
(A. m) is a maximal torus in ˚
d
_
G
k
(A. m)
_
for
d ¸ 1 ÷max{;(x) : x ÷ Hilb(M)].
where ; is a positive grading on M.
Proof. Let Hilb(M) = {x
1
. . . . . x
n
]. By definition we have d ¸ 2. Already for
d = 2 the restriction of ˚
d
to T is injective since every element t of T is uniquely
defined by the coefficients c
i
of the images t(x
i
) = c
i
x
i
, i = 1. . . . . n, and the
residue classes of x
1
. . . . . x
n
form a basis of m,m
2
. We nay now identify T and
˚
d
(T ).
Let g ÷ G
k
(A. m) and suppose that ˚
d
(g) commutes with all elements of T .
Set M
~d
= M \ ({1] L m
d
). Then the residue classes of the elements of M
~d
form a k-basis of m,m
d
, and, as far as the structure of k-vector space is concerned,
162 5. Isomorphisms and Automorphisms
we are allowed to identify the elements of M
~d
with their images. We have the
decomposition
m,m
d
=

x÷M
<d
kx
and the kx are subspaces of m,m
d
corresponding to the different characters y
x
,
x ÷ M
~d
; see Lemma 5.10(b). Since ˚
d
(g) commutes with all t ÷ T, also
_
˚
d
(g)
_
(x) is an eigenvector for the character y
x
. By Lemma 5.10(c) it follows
that
_
˚
d
(g)
_
(x) = c
x
x with c
x
÷ k
+
, for all x ÷ M
~d
.
In particular, we have
_
˚
d
(g)
_
(x
i
) = c
i
x
i
with c
i
÷ k
+
for the elements
x
1
. . . . . x
n
of the Hilbert basis. Suppose we can show that
¡ (c
1
. . . . . c
n
) = 0 (5.4)
for all binomials ¡ ÷ I, where I is defined as in Lemma 5.8. Then, by part (b) of
that lemma, there exists o ÷ T such that o(x
i
) = c
i
x
i
for all i , and it follows that
˚
d
(g) = o ÷ T, as desired.
At this point we must lift the data from m,m
d
back to A. By what has been
shown,
g(x
i
) = c
i
x
i
÷y
i
. y
i
÷ m
d
. i = 1. . . . . n.
The choice of d guarantees that c
i
x
i
is the lowest degree component of g(x
i
) with
respect to the grading defined by ;. Let ¡ = X
a
1
1
X
a
n
n
÷ X
b
1
1
X
b
n
n
÷ I.
Then ¡ (x
1
. . . . . x
n
) = 0, and so ¡ (g(x
1
). . . . . g(x
n
)) = 0, too. The lowest de-
gree component of g(x
1
)
a
1
g(x
n
)
a
n
is c
a
1
1
c
a
n
n
x
a
1
1
x
a
n
n
. Considering also
the monomial X
b
1
1
X
b
n
n
and comparing lowest degree components, we obtain
c
a
1
1
c
a
n
n
x
a
1
1
x
a
n
n
= c
b
1
1
c
b
n
n
x
b
1
1
x
b
n
n
.
It follows that c
a
1
1
c
a
n
n
= c
b
1
1
c
b
n
n
, and this is equation (5.4).
We can now conclude the proof of Theorem 5.20. Only part (a) still needs to
be proved. So suppose that kŒM| and kŒN| with M and N positive are isomorphic
as augmented k-algebras with respect to the augmentation ideals generated by the
elements = 1 in M and N, respectively. We identify the algebras along the isomor-
phism, and call the algebra A. The augmentation ideals are also identified in the
maximal ideal m. Lemma 5.29 yields that the tori T
k
(M). T
k
(N) Ç G
k
(A. m) are
both maximal tori in ˚
d
_
G
k
(A. m)
_
for d ; 0. A fortiori, they are maximal tori
in every subgroup of ˚
d
_
G
k
(A. m)
_
that contains them. Moreover, they are closed
subgroups of the linear algebraic group GL(m,m
d
): with the respect to the basis
M
~d
of m,m
d
, each element ¸ ÷ T
k
(M) is mapped to a diagonal matrix whose
entries are Laurent monomials in the entries of a matrix of ¸ with respect to a basis
of gp(M). The image of T
k
(M) under a morphism of algebraic groups is closed,
and by symmetry the same applies to T
k
(N).
Let U be the subgroup of GL(m,m
d
) generated by T
k
(M) and T
k
(N). By
Proposition 5.1 it is a linear algebraic group, and Borel’s theorem applies. It follows
that T
k
(M) and T
k
(N) are conjugate in GL(m,m
d
). By Lemma 5.11 the monoids
generated by the characters of T
k
(M) and T
k
(N) that occur in V are isomorphic.
5.D. Automorphisms 163
Since d ¸ 2, the first of these monoids is isomorphic to M, and the second is
isomorphic to N (see Lemma 5.10(d)). In fact, the characters corresponding to the
Hilbert bases are present in both cases. Altogether we conclude that M and N are
isomorphic.
5.D. Automorphisms
In this section we want to determine the groups of graded automorphisms of
polytopal and normal affine monoid algebras. So far we have discussed only the
toric automorphisms, but in general the group of graded automorphisms is strictly
larger.
Termic automorphisms. Let M be an arbitrary monoid. Then we denote the
group of automorphisms of M by ˙(M). Clearly every o ÷ ˙ induces an auto-
morphism of the monoid algebra A = kŒM| where k is an arbitrary field. There-
fore we can consider o as an element of G
k
(A).
We say that ; ÷ G
k
(A) is termic if ;(t ) is a term for every term t = ax,
a ÷ k, x ÷ M. The termic automorphisms evidently form a subgroup Tm
k
(A) of
G
k
(A). Clearly T
k
(M) and ˙(M) are subgroups of Tm
k
(M), and they determine
its structure as follows.
Proposition 5.30. For ; ÷ Tm
k
(A) define o
;
: M ÷ M by o
;
(x) = y if ;(x) ÷
ky. Then:
(a) Then o
;
÷ ˙(M), and the assignment ; ÷ o
;
defines a surjective homo-
morphism Tm
k
(M) ÷˙(M) with kernel T
k
(M).
(b) Tm
k
(M) is the semidirect product of the normal subgroup T
k
(M) and the
subgroup ˙(M).
We will need the following criterion for termic automorphisms. It uses the ideal
o Ç kŒM| generated by all monomials x ÷ int(M) (see Remark 2.4.)
Lemma 5.31. Let M be an affine monoid, A = kŒM|, and ; ÷ G
k
(A). Then the
following are equivalent:
(a) ; is termic;
(b) ; leaves the ideal o invariant;
(c) ; permutes the prime ideals p
F
where F runs through the facets of R
÷
M.
Moreover, ; is toric if and only if ;(p
F
) = p
F
for all facets F.
Proof. First suppose that ; is termic. Then ; permutes the prime ideals p
F
asso-
ciated with the facets F of C(M). In fact, it maps monomial ideals to monomial
ideals, prime ideals to prime ideals, and preserves the height of ideals. But then ;
leaves the intersection o of the ideals p
F
invariant. Conversely, if ; permutes the
ideals p
F
, then it leaves their intersection invariant.
So it remains to show (b) == (a). Suppose that ; leaves o invariant, and
consider a monomial x ÷ int(M). We claim that the ring A
x
= AŒx
-1
| is the full
group ring kŒgp(M)|. This follows from Exercise 2.9, but since we have to consider
A
;(x)
, we give a ring-theoretic argument. It is clearly enough that every monomial
164 5. Isomorphisms and Automorphisms
y ÷ M is a unit in A
x
, and this in turn follows if y is not contained in a proper ideal
of A
x
. Observe that all minimal prime ideals containing y are of the formp
F
where
F is a facet of C(M) (and p
F
contains y if and only if y … F; see Corollary 4.30).
Therefore the minimal prime ideals of A
x
containing y are of the form A
x
p
F
. But
A
x
p
F
= A
x
since x ÷ p
F
for all facets (and faces) F of C(M).
By hypothesis we have ;(x) ÷ o as well, and by the same argument it follows
that every monomial y ÷ M is invertible in A
;(x)
, too. So
A
x
= kŒgp(M)| Ç A
;(x)
.
The crucial point is to compare the groups of units of A
x
and A
;(x)
. Evidently,
these groups are isomorphic via the natural extension of ; to an automorphism
of A
x
and A
;(x)
, and ; maps the subgroup k
+
of U(A
x
) to the same subgroup of
U(A
;(x)
). Therefore, with r = rank M,
Z
r
Š U(A
x
),k
+
Š U(A
;(x)
),k
+
.
Moreover, the residue classes of the monomials in U(A
;(x)
),k
+
form a subgroup
H of rank r.
Assume that ;(x) is not a term. Then none of the powers of ;(x) is a term. In
other words, none of the multiples of the class of ;(x) is in the subgroup H. This
shows that rank
_
U(A
;(x)
_
,k
+
) > r, a contradiction.
The proof of the last statement is left to the reader (Exercise 5.13).
Column structures on lattice polytopes. The example M = Z
n
÷
, A = kŒM| =
kŒX
1
. . . . . X
n
| graded by total degree, shows that the group 1
k
(A) is much larger
than Tm
k
(M). (In general, Tm
k
(M) is not contained in 1
k
(A), however.) While
the toric automorphisms correspond to the diagonal matrices in 1
k
(A) = GL
n
(A)
and the automorphism group ˙(M) to the group of permutations of the indeter-
minates, the analogue of the elementary transformations is still missing. We will
construct them for polytopal monoids, using column vectors of the underlying
polytope.
Definition 5.32. Let P Ç R
n
be a lattice polytope (with respect to the lattice Z
n
).
We assume that the lattice points of P generate Z
n
as an affine lattice. Then an
element : ÷ Z
n
, : = 0, is a column vector (for P) if there is a facet F Ç P such
that x ÷ : ÷ P for every lattice point x ÷ P \ F. The set of column vectors is
denoted by Col(P).
For such P and : we also say that : defines a column structure on P. The
corresponding facet F is called its base facet and denoted by P
¡
.
The assumption that P ¨Z
n
generates Z
n
affinely is only for simplicity. If it is
not satisfied, then we can replace Z
n
by the lattice affinely generated by P ¨ Z
n
.
One sees easily that for a column structure on P the set of lattice points in P is
contained in the union of rays,visualized as columns, parallel to the column vector
: and with end-points in F. This is illustrated by Figure 5.1.
We further illustrate the notion of column vector by Figure 5.2: the polytope
P
1
has 4 column vectors, whereas the polytope P
2
has no column vector.
5.D. Automorphisms 165
:
Figure 5.1. A column structure
P
1
P
2
Figure 5.2. Two polytopes and their column structures
Every lattice point x of P has a unique decomposition in the formx = y ÷mx
where m ÷ Z
÷
and y ÷ P
¡
. The assumption that Z
n
is affinely generated by
P ¨ Z
n
implies that
ht
P
v
(x) = m
(Exercise 5.15). See Remark 1.78 and Section 3.B for the definition of ht; we can
omit the lattice of reference, since it is fixed to be Z
n
. It follows easily that : is part
of a basis of Z
n
. For simplicity we set ht
¡
(x) = ht
P
v
(x).
Remark 5.33. One can easily control column structures in such formations as ho-
mothetic images and direct products of lattice polytopes:
(a) The polytopes P
1
and cP
1
have the same column vectors for every c ÷ N.
(b) Let P
i
be a lattice n
i
-polytope, i = 1. 2. Then the systemof column vectors
of P
1
× P
2
is the disjoint union of those of P
1
and P
2
(embedded into
Z
n
1
÷n
2
).
(c) Actually, (a) is a special case of a more general observation on polytopes:
if every facet of P
1
is parallel to a facet of P
2
, then Col(P
2
) Ç Col(P
1
).
The proofs of these statements are left to the reader (Exercise 5.16).
Elementary automorphisms. We want to analyze the group 1
k
_
kŒM(P)|
_
where
M(P) is the polytopal monoid associated with the lattice polytope P. (As in the
previous subsection we assume that the lattice points of P generate Z
n
as an affine
lattice.) In order to simplify notation we set kŒP| = kŒM(P)| and
1
k
(P) = 1
k
_
kŒP|
_
.
It is a crucial observation that the group of graded k-algebra automorphisms
does not change if one passes form kŒP| to its normalization. The normalization
is given by kŒ
¨
P| where
¨
M(P) is the normalization of M(P). To simplify notation
further, we set

¨
P| = kŒ
¨
M(P)|.
166 5. Isomorphisms and Automorphisms
although the symbol
¨
P has no meaning by itself. There is no need to add a bar to
1
k
(P):
Lemma 5.34. The natural map 1
k
(P) ÷ 1
k
(kŒ
¨
P|), given by the extension of auto-
morphisms, is a group isomorphism.
Proof. A k-algebra automorphism ; of A = kŒP| extends to a k-algebra auto-
morphism of the field QF(A) of fractions, and can then be restricted to a k-algebra
automorphism of the normalization. Moreover, the normalization is contained in
the subalgebra B of QF(A) generated by all fractions of homogeneous elements,
and since the extension of a graded automorphism to B is again graded, ; extends
to a graded automorphism of the normalization if it is graded.
Therefore it remains to show that any graded k-algebra automorphism of kŒ
¨
P|
can be restricted to an automorphism of A. In general this is impossible, but the
algebras under consideration agree in degree 1, and A is generated by its degree 1
component; see Proposition 2.27.
The monoids M(P) and
¨
M(P) are submonoids of Z
n÷1
. We identify a column
vector : ÷ Col(P) with the vector (:. 0) ÷ Z
n÷1
, since it represents the difference
of lattice points in P, which in Z
n÷1
have degree 1. Thus : has degree 0 as an
element of Z
n÷1
with respect to the grading given by the last coordinate.
Let F be a facet of P. Then ht
F
has a natural extension to Z
n÷1
by the height
with respect to the hyperplane through C(F). In the following it will also be de-
noted by ht
F
. It is of course nothing but the support form of C(P) associated with
the facet C(F). Here we prefer the term “height” because of its geometric flavor.
Remark 5.35. As elements of Z
n÷1
the vectors : ÷ Col(P) can be characterized as
follows:
(i) deg : = 0;
(ii) there exists a facet F of P such that ht
F
(:) = ÷1;
(iii) ht
G
(:) ¸ 0 for all facets G = F.
Clearly F = P
¡
.
The proof of the next lemma is now straightforward.
Lemma 5.36. Let : ÷ Col(P). Then x ÷: ÷ M(P) for all x ÷ M(P), x … C(P
¡
),
and x ÷: ÷
¨
M(P) for all x ÷
¨
M(P), x … C(P
¡
).
At this point we have to switch to multiplicative notation since the addition in
kŒP| has to be used. Thus lattice points and column vectors are identified with
the corresponding monomials in kŒZ
n÷1
|. Given : ÷ Col(P) and z ÷ k, we now
define an injective mapping fromM(P) to QF(kŒP|), the quotient field of kŒP|, by
the assignment
x ÷(1 ÷z:)
ht
v
(x)
x.
Since ht
¡
is a group homomorphismZ
n÷1
÷Z, our mapping is a homomorphism
fromM(P) to the multiplicative group of QF(kŒP|). Now it is immediate from the
definition of ht
¡
and Lemma 5.36 that the (isomorphic) image of M(P) lies actually
in kŒP|. Hence this mapping gives rise to a graded k-algebra endomorphism e
z
¡
of
5.D. Automorphisms 167
kŒP| preserving the degree of an element. Since e
z
¡
is a k-linear injective map on
each (finite-dimensional) graded component, it is an automorphism. In the same
way it defines a k-automorphism of kŒ
¨
P|, by extension from kŒP| or directly.
One can give an alternative description of e
z
¡
. By a suitable integral change of
coordinates we may assume that : = (0. . . . . . 0. ÷1) and that P
¡
lies in the sub-
space R
n-1
(thus P is in the “upper” halfspace). Now consider the standard uni-
modular n-simplex z
n
with vertices at the origin and the unit vectors. It is clear
that there is a sufficiently large natural number c, such that P is contained in a par-
allel translate of cz
n
by a vector fromZ
n-1
. Let zdenote such a parallel translate.
Then we have a graded k-algebra embedding kŒP| Ç kŒz|. Moreover, kŒz| can be
identified with the c-th Veronese subring of the polynomial ring kŒx
0
. . . . . x
n
| in
such a way that : = x
0
,x
n
. Now the automorphism of kŒx
0
. . . . . x
n
| mapping x
n
to x
n
÷zx
0
and leaving all the other variables invariant induces an automorphism
˛ of the subalgebra kŒz|, and ˛ in turn can be restricted to an automorphism of
kŒP|, which is nothing else but e
z
¡
.
It is clear from this description of e
z
¡
that it becomes an elementary matrix (e
z
0n
in our notation) in the special case when P = z
n
, after the identification 1
k
(P) =
GL
n÷1
(k). Therefore the automorphisms of type e
z
¡
will be called elementary.
We now describe the structure of the set of all elementary automorphisms with
the same base facet; G
a
(k) denotes the the additive group (k. ÷).
Lemma 5.37. Let :
1
. . . . . :
s
be pairwise different column vectors for P with the same
base facet F = P
¡
i
, i = 1. . . . . s.
(a) Then the mapping
ç : G
a
(k)
s
÷1
k
(P). (z
1
. . . . . z
s
) ÷e
z
1
¡
1
e
z
s
¡
s
.
is an embedding of groups. In particular, e
z
i
¡
i
and e
z
j
¡
j
commute for any
i. j ÷ {1. . . . . s] and the inverse of e
z
i
¡
i
is e
-z
i
¡
i
.
(b) For x ÷
¨
M(P) with ht
F
(x) = 1 one has
e
z
1
¡
1
e
z
s
¡
s
(x) = (1 ÷z
1
:
1
÷ ÷z
s
:
s
)x.
Proof. Set [ = ç(z
1
. . . . . z
s
). We define a new k-algebra automorphism 0 of
kŒP| by first setting
0(x) = (1 ÷z
1
:
1
÷ ÷z
s
:
s
)
ht
F
(x)
x.
for x ÷ M(P) and then extending 0 linearly. Arguments very similar to those
above show that 0 is a graded k-algebra automorphism of kŒP| (and kŒ
¨
P|). The
lemma is proved once we have verified that [ = 0.
Choose a lattice point x ÷
¨
M(P) such that ht
F
(x) = 1. (Such a point exists
already in P by Exercise 5.15.) We know that gp(M(P)) = Z
n÷1
is generated by
x and the lattice points in F. The lattice points in F are left unchanged by both 0
and ç, and elementary computations show that [(x) = (1÷z
1
:
1
÷ ÷z
s
:
s
)x =
0(x); hence [ = 0.
The image of the embedding ç given by Lemma 5.37 is denoted by E(F). Of
course, E(F) may consist only of the identity map of kŒP|, namely if there is no
168 5. Isomorphisms and Automorphisms
column vector with base facet F. In the case in which P is the unit simplex and
kŒP| is the polynomial ring, E(F) is the subgroup of all matrices in GL
n
(k) that
differ from the identity matrix only in the non-diagonal entries of a fixed column.
If k is algebraically closed, then by Proposition 5.1(a) E(F) Š G
a
(k)
s
is closed
subgroup of 1
k
(P).
Already in the proof of Lemma 5.31 it has become clear that the prime ideals
p
F
play an important role for the automorphism group 1
k
(P). In the next lemma
we describe the action of elementary automorphisms on them. In the following p
F
will always denote the prime ideal of the normalization kŒ
¨
P| defined by the facet
F.
Lemma 5.38. Let :
1
. . . . . :
s
be column vectors with the common base facet F = P
¡
i
,
and z
1
. . . . . z
s
÷ k. Then
e
z
1
¡
1
e
z
s
¡
s
(p
F
) = (1 ÷z
1
:
1
÷ ÷z
s
:
s
)p
F
and
e
z
1
¡
1
e
z
s
¡
s
(p
G
) = p
G
. G = F.
Proof. Note that (1 ÷z
1
:
1
÷ ÷z
s
:
s
)
ht
F
(x)-1
÷ p
F
for all monomials x ÷ p
F
.
Using the automorphism 0 from the proof of Lemma 5.37 one concludes that
e
z
1
¡
1
e
z
s
¡
s
(p
F
) Ç (1 ÷z
1
:
1
÷ ÷z
s
:
s
)p
F
.
The left hand side is a height 1 prime ideal (being an automorphic image of such)
and the right hand side is a proper divisorial ideal inside kŒ
¨
P|. Then the inclusion
is an equality since the only divisorial ideal of kŒ
¨
P| containing a height 1 prime
ideal is kŒ
¨
P| itself.
For the second assertion it is enough to treat the case s = 1, : = :
1
, z = z
1
.
One has
e
z
¡
(x) = (1 ÷z:)
ht
F
(x)
x.
and all the terms in the expansion of the right hand side belong to p
G
since
ht
G
(:) ¸ 0. As above, the inclusion e
z
¡
(p
G
) Ç p
G
implies equality.
Lemma 5.39. Let F Ç P be a facet, z
0
. . . . . z
s
÷ k
+
and :
0
. . . . . :
s
÷ Z
n÷1
be
pairwise different elements of degree 0. Then (z
0
:
0
÷ ÷z
s
:
s
)p
F
Ç kŒ
¨
P| if and
only :
i
is a column vector for P with base facet F for each i with :
i
= 0.
Proof. Both of the properties that are claimed to be equivalent are themselves
equivalent to the fact that :
i
x ÷ M(P) for all i and x ÷ M(P). Since the :
i
are of
degree 0, this means :
i
÷x ÷ P for all x ÷ P in additive notation.
The Gaussian algorithmfor polytopes. The Gaussian algorithm allows us to trans-
form a matrix to a diagonal matrix. Theorem 5.40 claims that such a diagonaliza-
tion is possible for automorphisms of polytopal algebras, and we will carry it out
by a procedure generalizing the Gaussian algorithm.
Clearly, T
n÷1
= T
k
(M(P)) Ç 1
k
(P), and since all generators of M(P) are
of degree 1, the symmetry group ˙(P) = ˙(M(P)) is contained in 1
k
(P), too.
5.D. Automorphisms 169
Moreover, the elementary automorphisms are homogeneous. As we will now see,
together they generate 1
k
(P).
The normal form in the theorem generalizes the fact that the elementary trans-
formations e
z
ij
, j fixed, in the diagonalization of an invertible matrix can be carried
out consecutively.
Theorem 5.40. Let P be a convex lattice n-polytope and k a field. Then there exists
an enumeration F
1
. . . . . F
r
of the facets of R
÷
M such that every element ; ÷ 1
k
(P)
has a (not uniquely determined) presentation
; = ˛
1
˛
2
˛
r
to.
where o ÷ ˙(P), t ÷ T
n÷1
, and ˛
i
÷ E(F
i
), i = 1. . . . ..
We isolate the crucial step of the proof in the next lemma. But first we have to
specify the enumeration of the facets. To this end we choose a number g ÷ Nsuch
that for each facet F of P there exists an element (x. h) ÷ Z
n÷1
, x ÷ hP, with
ht
F
(x) = 1 and h ¸ g. (Such g exists by Proposition 2.73.) Let N be a finitely
generated graded module over kŒ
¨
P|. Then we set
v
g
(N) =
g

i=0
dim
k
N
i
.
In the proof of the next lemma we will use the following properties of v
g
:
(a) v
g
(p
F
¨ p
G
) < v
g
(p
F
) if F and G are different facets of P;
(b) v
g
(p
(a)
F
) < v
g
(p
F
) if a ¸ 2.
The first assertion holds for all g ¸ 1 since P \ (F L G) contains fewer lattice
points than P \ F. The second holds for our choice of g since in the passage from
p
F
to p
(a)
F
we lose at least one lattice point, namely (x. h) as above. For a graded
ideal I of kŒ
¨
P| and ; ÷ 1
k
(P) one has v
g
(;(I)) = v
g
(I) since the homogeneous
automorphism ; preserves the k-dimension of the graded components.
Lemma 5.41. Let ; ÷ 1
k
(P), and choose an enumeration F
1
. . . . . F
r
of the facets
such that v
g
(p
F
1
) ¸ ¸ v
g
(p
F
r
). Then there exists a permutation ¬ of {1. . . . . r]
such that for all i and
˛
r
˛
1
;(p
F
i
) = p
t(i)
with suitable ˛
i
÷ E(F
t(j)
).
In fact, this lemma implies Theorem 5.40: the resulting automorphism ı =
˛
r
˛
1
; permutes the divisorial prime ideals p
F
. By virtue of Lemma 5.31 we
then have ı ÷ Tm
k
(kŒP|). So ı = ot with o ÷ ˙(P) and t ÷ T
n÷1
. Finally one
just replaces each ˛
i
by its inverse and each F
i
by F
t(i)
.
Proof of Lemma 5.41. By Corollary 4.54 the divisorial ideal ;(p
F
) Ç kŒ
¨
P| is
isomorphic to some monomial divisorial ideal z, i. e. there is an element × ÷
QF(kŒ
¨
P|) such that ;(p
F
) = ×z. The inclusion × ÷ (;(p
F
) : z) shows that × is a
k-linear combination of some Laurent monomials corresponding to lattice points
170 5. Isomorphisms and Automorphisms
in Z
n÷1
. We factor out one of the terms of ×, say m, and rewrite the above equality
as follows:
;(p
F
) = (m
-1
×)(mz).
Then m
-1
× is of the form1÷m
1
÷ ÷m
s
for some Laurent terms m
1
. . . . . m
s
… k,
while mzis necessarily a divisorial monomial ideal of kŒ
¨
P| (since 1 belongs to the
supporting monomial set of m
-1
×). Now ; is a graded automorphism. Hence
(1 ÷ m
1
÷ ÷ m
s
)(mz) Ç kŒ
¨
P| is a graded ideal. This implies that the terms
m
1
. . . . . m
s
are of degree 0. Thus there is always a presentation
;(p
F
) = (1 ÷m
1
÷ ÷m
s
)z.
where m
1
. . . . . m
s
… k
+
are Laurent terms of degree 0 and z Ç kŒ
¨
P| is a monomial
ideal (we do not exclude the case s = 0). A representation of this type is called
admissible.
For ; ÷ 1
k
(P) consider an admissible representation
;(p
F
1
) = (1 ÷m
1
÷ ÷m
s
)z.
One has v
g
(z) = v
g
(p
F
1
). Since z Ç kŒ
¨
P| is a divisorial monomial ideal, there
are integers a
i
¸ 0 such that
z =
r
_
i=1
p
(a
i
)
F
i
:
see Theorem 4.47. Now apply the function v
g
. By its properties (a) and (b) stated
above the lemma and because v
g
(p
F
1
) is maximal among the v
g
(p
F
i
) it follows that
exactly one exponent is 1 and all the others are 0. So z = p
G
1
for some facet G
1
.
By Lemmas 5.38 and 5.39 there exists ˛
1
÷ E(G
1
) such that
˛
1
;(p
F
1
) = p
G
1
.
Now we proceed inductively. Let 1 ¸ t < r. Assume there are facets G
1
. . . . . G
t
of P and ˛
1
÷ E(G
1
). . . . . ˛
t
÷ E(G
t
) such that
˛
t
˛
1
;(p
F
i
) = p
G
i
. i = 1. . . . . t.
(Observe that the G
i
are automatically different and that v
g
(p
G
i
) = v
g
(p
F
i
).) In
view of Lemma 5.38 we must show there are a facet G
t ÷1
Ç P, different from
G
1
. . . . . G
t
and an element ˛
t ÷1
÷ E(G
t ÷1
) such that
˛
t ÷1
˛
t
˛
1
;(p
F
tC1
) = p
G
tC1
.
For simplicity of notation we put ;
t
= ˛
t
˛
1
;. Again, consider an admissible
representation
;
t
(p
F
tC1
) = (1 ÷m
1
÷ ÷m
s
)z.
Rewriting this equality in the form
;
t
(p
F
tC1
) =
_
m
-1
j
(1 ÷m
1
÷ ÷m
s
)
_
(m
j
z).
where j ÷ {0. . . . . s] and m
0
= 1, we get another admissible representation. We
will show that by varying j we can obtain a monomial divisorial ideal m
j
z such
5.D. Automorphisms 171
that in the primary decomposition
m
j
z =
r
_
i=1
p
(a
i
)
F
i
there appears a positive power of p
G
for some facet G different from G
1
. . . . . G
t
.
Then v
g
((m
j
z) ¸ v
g
(p
G
) ¸ v
g
(p
F
tC1
) (due to our enumeration) and the inequal-
ity would be strict whenever

r
1
a
i
¸ 2. Thus m
j
z = p
G
and we can proceed as
for the ideal p
F
1
.
Assume to the contrary that in the primary decompositions of all the monomial
ideals m
j
z there appear only the prime ideals p
G
1
. . . . . p
G
t
. We have
(1 ÷m
1
÷ ÷m
s
)z Ç z÷m
1
z÷ ÷m
s
z
and
Œ(1 ÷m
1
÷ ÷m
s
)z| = Œz| = Œm
1
z| = = Œm
s
z|
in Cl(kŒ
¨
P|). Applying (;
t
)
-1
we arrive at the conclusion that p
F
tC1
is contained in
a sum of monomial divisorial ideals ˚
0
. . . . . ˚
s
, such that the primary decompo-
sition of each of them involves only p
F
1
. . . . . p
F
t
. (This follows from the fact that
(;
t
)
-1
maps p
G
i
to the monomial ideal p
F
i
for i = 1. . . . . t ; thus intersections of
symbolic powers of p
G
1
. . . . . p
G
t
are mapped to intersections of symbolic powers
of p
F
1
. . . . . p
F
t
, which are automatically monomial.) Furthermore, p
F
tC1
has the
same divisor class as each of the ˚
i
.
Now, if m = 1 is a degree 0 monomial such that p
F
tC1
= m˚ with ˚ Ç kŒ
¨
P|,
then m
-1
must be a column vector for p
F
tC1
. Choose a lattice point x ÷ P for
which ht
F
tC1
(x) is maximal. Since ht
F
tC1
(m
-1
i
y) = ht
F
tC1
(y) ÷ 1 < ht
F
tC1
(x) for
all lattice points y ÷ P, the ideal p
F
tC1
is not contained in the sumof the monomial
ideals ˚
i
, a contradiction.
The structure of the graded automorphism group. In order to give some more
information on 1
k
(P) we have to introduce a subgroup of ˙(P) defined as follows.
Assume : and ÷: are both column vectors. Then for every point x ÷ P ¨Z
n
there
is a unique y ÷ P ¨ Z
n
such that ht
¡
(x) = ht

(y) and x ÷ y is parallel to :.
The mapping x ÷ y gives rise to a monoid automorphism of M(P): it “inverts
columns” that are parallel to :. It is easy to see that these automorphisms generate
a normal subgroup of ˙(P), which we denote by ˙(P)
inv
.
Theorem 5.42. Let P be a lattice n-polytope and let k be an algebraically closed
field.
(a) The connected component of unity 1
k
(P)
0
Ç 1
k
(P) is generated by the
subgroups E(F
i
) and T
n÷1
.
(b) It consists precisely of those graded automorphisms of kŒP| which induce
the identity map on the divisor class group of the normalization of kŒP|.
(c) dim1
k
(P) = # Col(P) ÷n ÷1.
(d) One has 1
k
(P)
0
¨˙(P) = ˙(P)
inv
and 1
k
(P),1
k
(P)
0
Š ˙(P),˙(P)
inv
.
172 5. Isomorphisms and Automorphisms
Proof. (a) Since T
n÷1
and the E(F
i
) are connected groups they generate a con-
nected subgroup H of 1
k
(P) (Proposition 5.1(b)). This subgroup acts trivially on
Cl(kŒ
¨
P|) by Lemma 5.38 and the fact that the classes of the p
F
i
generate the di-
visor class group. Furthermore H has finite index in 1
k
(P) bounded by #˙(P).
Therefore H = 1
k
(P)
0
.
(b) Suppose ; ÷ 1
k
(P) acts trivially on Cl(kŒ
¨
P|). We want to show that ; ÷ H.
Let E denote the connected subgroup of 1
k
(P), generated by the elementary auto-
morphisms. Any automorphism that preserves the divisorial ideals p
F
i
is automat-
ically toric (Lemma 5.31). Therefore we have only to show that there is an element
c ÷ E such that
c;(p
F
i
) = p
F
i
. i = 1. . . . . r.
From Lemma 5.41 we know that there are c
1
÷ E and a permutation ¬ of
{1. . . . . r] such that
c
1
;(p
j
) = p
t(j)
. j = 1. . . . . r. (5.5)
Since c
1
and ; both act trivially on Cl(kŒ
¨
P|), we get
p
t(j)
= m
j
p
j
. j = 1. . . . . r.
for some monomials m
j
of degree 0.
By Lemma 5.39 we conclude that if m
j
= 1 (in additive notation, m
j
= 0), then
both m
j
and m
-1
j
are column vectors with the base facets F
j
and F
t(j)
respectively.
Observe that the automorphism
c
t(j)
= e
1
m
j
e
-1
m
1
j
e
1
m
j
÷ E
interchanges the ideals p
j
and p
t(j)
, provided m
j
= 1. Now we can complete the
proof by successively “correcting” the equations (5.5).
(c) In the enumeration of the facets used for Theorem 5.40 we have the surjec-
tive algebraic map
E(F
1
) × × E(F
r
) × T
n÷1
× ˙(P) ÷1
k
(P).
induced by composition. The left hand side has dimension # Col(P)÷n÷1. Hence
dim1
k
(P) ¸ # Col(P) ÷n ÷1.
To derive the opposite inequality we can additionally assume that P contains
an interior lattice point. Indeed, Remark 5.33(a) and Theorem 5.40 show that the
natural group homomorphism 1
k
(P) = 1
k
(kŒ
¨
P|) ÷ 1
k
(cP), induced by restric-
tion, is surjective for every c ÷ N. (The surjectivity of T
k
(M(P)) ÷ T
k
(M(cP))
follows from the fact that k is closed under taking roots). So we can work with cP,
which contains an interior point provided c is large.
Let x ÷ P be an interior lattice point and let :
1
. . . . . :
s
be different column vec-
tors. Then the supporting monomial set of e
z
i
¡
i
(x), z ÷ k
+
, is not contained in the
union of those of e
z
j
¡
j
(x), j = i (just look at the projections of x through :
i
into the
corresponding base facets). This shows that we have # Col(P) linearly independent
tangent vectors of 1
k
(P) at 1 ÷ 1
k
(P). Since the tangent vectors corresponding to
the elements of T
n÷1
clearly belong to a complementary subspace and 1
k
(P) is a
smooth variety, we are done.
5.D. Automorphisms 173
(d) Assume : and :
-1
both are column vectors. Then the automorphism
c = e
1
¡
e
-1
¡
1
e
1
¡
÷ 1
k
(P)
0
maps monomials to terms; more precisely, c “inverts up to scalars” the columns
parallel to :: every x ÷
¨
M(P) is sent either to the appropriate y ÷
¨
M(P) or to
÷y ÷ kŒ
¨
P| (Exercise 5.21). Then it is clear that there is an element t ÷ T
n÷1
such
that tc is a generator of ˙(P)
inv
. Hence ˙(P)
inv
Ç 1
k
(P)
0
.
Conversely, if o ÷ ˙(P)¨1
k
(P)
0
then o induces the identity map on Cl(kŒ
¨
P|).
Hence o(p
F
i
) = m
j
p
F
i
for some monomials m
j
, and the very same arguments we
have used in the proof of (b) show that o ÷ ˙(P)
inv
. Thus 1
k
(P),1
k
(P)
0
=
˙(P),˙(P)
inv
.
Observe that we do not claim the existence of a normal form as in Theorem
5.40 for the elements of 1
k
(P)
0
if ˙(P)
inv
is excluded from the generating set. In
fact, in general such a normal form does not exist (Exercise 5.20).
Normal affine monoids. While we have introduced then notion of column vector
only for polytopes P, the proofs of Theorems 5.40 and 5.42 uses the data of the
normalization kŒ
¨
P| of the monoid ring kŒP|. It is not difficult to see that both
theorems can be generalized from the polytopal to the normal case. (In view of
Lemma 5.34 this change is indeed a generalization.)
Let M be an arbitrary affine monoid with a positive grading denoted by deg
(and extended to gp(M). Then : ÷ gp(M), : = 0, is a column vector of M (with
respect to deg) if deg : = 0 and there exists a facet F of R
÷
M such that x÷: ÷ M
for all x ÷ M, x … M \ F. It follows immediately from Lemma 5.36 that this
general notion of column vector is consistent with the polytopal case. Moreover, if
: is a column vector for M, then it is also a column vector for
¨
M (but if M is not
polytopal, the converse does not hold in general).
Using the same formula a in the polytopal case one now defines the elementary
automorphisms e
z
¡
÷ 1
k
(M) = 1
k
(kŒM|). Lemma 5.37 remains true without
changes, and Lemma 5.39 remains true if one considers column vectors of
¨
M.
After these generalizations one obtains
Theorem 5.43. Let M be a normal affine monoid with a positive grading, and k a
field. Then there exists an enumeration F
1
. . . . . F
r
of the facets of R
÷
M such that
every element ; ÷ 1
k
(M) has a (not uniquely determined) presentation
; = ˛
1
˛
2
˛
r
to.
where o ÷ ˙(M), t ÷ T
k
(M), and ˛
i
÷ E(F
i
). Moreover, Theorem 5.42 holds
accordingly.
A determinantal application. As an application to rings and varieties outside the
class of monoid algebras we determine the groups of graded automorphisms of
the determinantal rings, a result which goes back to Frobenius [34, p. 99] and has
been re-proved many times since then. See, for instance, [98] for a group-scheme
theoretical approach for general commutative rings of coefficients, covering also
174 5. Isomorphisms and Automorphisms
the classes of generic symmetric and alternating matrices. (The generic symmetric
case can be done by the same method as the one below.)
In plain terms, Corollary 5.44 answers the following question: let k be an in-
finite field, ç : k
mn
÷ k
mn
a k-automorphism of the vector space k
mn
of m × n
matrices over k, and r an integer, 1 ¸ r < min(m. n); when is rank ç(A) ¸ r
for all A ÷ k
mn
with rank A ¸ r? This holds obviously for transformations
ç(A) = SAT
-1
with S ÷ GL
m
(k) and T ÷ GL
n
(k), and for the transposition
if m = n. Indeed, these are the only such transformations:
Corollary 5.44. Let k be an algebraically closed field, X an m × n matrix of inde-
terminates, and let R = kŒX|,I
r÷1
(X) be the residue class ring of the polynomial
ring kŒX| in the entries of X modulo the ideal generated by the (r ÷ 1)-minors of
X, 1 ¸ r < min(m. n).
(a) The connected component G
0
of unity in 1
k
(R) is the image of GL
m
(k) ×
GL
n
(k) in GL
mn
(k) under the map (S. T ) ÷ (A ÷ SAT
-1
), and is
isomorphic to GL
m
(k) × GL
n
(k),k
+
where k
+
is embedded diagonally.
(b) If m = n, the group G is connected, and if m = n, then G
0
has index 2 in
1
k
(R) and 1
k
(R) = G
0
L tG
0
where t is the transposition.
Proof. The singular locus of Spec R is given by V(p) where p = I
r
(X),I
r÷1
(X);
p is a prime ideal in R(see Bruns and Vetter [25, (2.6), (6.3)]). It follows that every
automorphism of R must map p onto itself. Thus a linear substitution on kŒX|
for which I
r÷1
(X) is stable also leaves I
r
(X) invariant and therefore induces an
automorphism of kŒX|,I
r
(X). This argument reduces the corollary to the case
r = 1.
For r = 1 one has the isomorphism
R ÷kŒY
i
Z
j
: i = 1. . . . . m. j = 1. . . . . n| Ç kŒY
1
. . . . . Y
m
. Z
1
. . . . . Z
n
|
induced by the assignment X
ij
÷ Y
i
Z
j
. Thus R is just the Segre product of
kŒY
1
. . . . . Y
m
| and kŒZ
1
. . . . . Z
n
|, or, equivalently, R Š kŒP| where P is the di-
rect product of the unit simplices z
m-1
and z
n-1
. Part (a) follows now from an
analysis of the column structures of P (see Remark 5.33(b)) and the torus actions.
For (b) one observes that Cl(R) Š Z; ideals representing the divisor classes 1
and ÷1 are given by (Y
1
Z
1
. . . . . Y
1
Z
n
) and (Y
1
Z
1
. . . . . Y
m
Z
1
) [25, 8.4]. If m = n,
these ideals have different numbers of generators; therefore every automorphism
of R acts trivially on the divisor class group. In the case m = n, the transposition
induces the map s ÷ ÷s on Cl(R). Now the rest follows again from the theorem
above. (Instead of the divisorial arguments one could also discuss the symmetry
group of z
m-1
× z
n-1
.)
Exercises
In the following k denotes an algebraically closed field whenever an exercise uses notions
that we have introduced only for such fields.
5.1. Show that the monomials in kŒX
˙
1
1
. . . . . X
˙
1
n
| are the characters of T
n
(k).
Exercises 175
5.2. Let G Ç k
+
be a finite group. Show that the character group X(G) is isomorphic to G.
5.3. Prove Corollary 5.4. Hint: if G is a closed subgroup of G
t
, then the natural morphism
O(G
t
) ÷O(G) is surjective.
5.4. Show that the assignments G ÷ X(G) and B ÷ Hom(B. k
+
) induce an equiva-
lence between the category of diagonalizable groups and the category of finitely generated
abelian groups without p-torsion.
5.5. Let D be a diagonalizable group. Show that the natural pairing D × X(D) ÷ k
+
,
induces an isomorphismD Š Hom(X(D). k
+
).
5.6. Let M be a normal monoid of monomials in R = kŒX
1
. . . . . X
n
| where k is an alge-
braically closed field of characteristic p > 0. Set M
t
= {x ÷ R : x
p
e
÷ M for some e ÷
N]. Show that kŒM
t
| is the smallest subalgebra of R that contains kŒM| and is of type R
D
for some subgroup D Ç T
n
(k).
5.7. (a) Let R = kŒX
1
. . . . . X
n
|. Represent the Veronese subalgebras kŒR
k
| as rings of
invariants of R if char k = 0. (R
k
denotes the vector space spanned by the monomials of
total degree k.)
(b) Represent kŒR
k
| as a ring of invariants of a torus action in arbitrary characteristic.
5.8. Let the torus T
1
act on kŒY
1
. . . . . Y
m
. Z
1
. . . . . Z
n
| act by Y
i
÷¸Y
i
, i = 1. . . . . m, and
Z
j
÷ ¸
-1
Z
j
, j = 1. . . . . n. Compute the ring S of invariants and all weight spaces. Find
minimal systems of generators of the weight spaces as S-modules.
5.9. Give an example of an action of T
1
on kŒX
1
. X
2
| whose weight spaces do not all have
rank 1. (This is very easy.)
5.10. Let z be an abstract simplicial complex on the vertex set V = {:
1
. . . . . :
n
]. For a
field k consider the exterior algebra E = k(e
1
. . . . . e
n
), and let J(z) be the ideal generated
by all monomials e
i
1
. .e
i
s
such that {:
i
1
. . . . . :
i
s
] … z. One calls E,J(z) the exterior
face ring of z.
Formulate and prove the analogue of the statements in Example 5.23 for the exterior face
ring. (Note that all monomials in E are squarefree so that exterior face rings cover all
monomial residue class algebras of E.)
5.11. Find the missing details in the proof of Lemma 5.28.
5.12. Let V be an affine variety over an algebraically closed field k, containing the torus
T
n
(k) as an Zariski open subset. Suppose the group structure of T
n
(k) extends to an
algebraic group action T
n
(k) × V ÷ V . Show that O(V ) Š kŒM| for an affine monoid
M of rank n.
Hint: use the fact that O(T
n
(k)) = kŒX
˙1
1
. . . . . X
˙1
n
| is a localization of the affine algebra
O(V ), and the latter is stable under the extension of the toric automorphisms of T
n
(k).
This gives an alternative description of affine monomial algebras and, more importantly,
indicates what the right non-affine analogues of the varieties associated with affine monoid
algebras should be. These global toric varieties will show up in Chapter ??.
5.13. Let M be an affine monoid, A = kŒM|, and ; ÷ G
k
(A). Show ; is toric if and only
if ;(p
F
) = p
F
for all facets F of R
÷
M.
5.14. Let M be an affine monoid and k an arbitrary field.
(a) Prove Proposition 5.30.
176 5. Isomorphisms and Automorphisms
(b) Let A = kŒZ
n
|. Show that every automorphism of A is termic and that G
k
(A) is a
semidirect product of T
n
and GL
n
(Z).
5.15. Let P Ç R
n
be a lattice polytope for which P¨Z
n
generates Z
n
affinely. Furthermore
let : be a column vector of P. If x = y÷m: for x ÷ P¨Z
d
and y ÷ P
¡
, showm = ht
¡
(x).
5.16. Prove the statements in Remark 5.33.
5.17. Let P be a lattice polytope, : ÷ Col(P), and let G be a facet with ht
G
(:) > 0. Prove:
if there exists a column vector n with base facet G and ht
F
(n) > 0, then n = ÷:.
5.18. Let A = kŒX. Y. Z| and n ÷ N, n ¸ 3. Find a grading of Aunder which dim1
k
(A) =
n.
5.19. Using the arguments in the proof of Lemma 5.41, show that a column vector : with
base facet F is invertible if v
g
(p
F
) ¸ v
g
(p
G
) for all facets G.
5.20. Let ; be the automorphism of kŒX
1
. X
2
| exchanging X
1
and X
2
. Show that ; can not
be represented by a product c
1
c
2
t where c
1
. c
2
are elementary and t is toric.
5.21. Show that the automorphism
c = e
1
¡
e
-1
¡
1
e
1
¡
÷ 1
k
(P)
0
appearing in the proof of Theorem 5.42 “inverts up to scalars” the columns parallel to ::
every x ÷
¨
M(P) is sent either to the appropriate y ÷
¨
M(P) or to ÷y ÷ kŒ
¨
P|. Proceed as
follows:
(a) Reduce the claim to a single “column”, i. e. to the polytopal algebra k
_
Œ0. m|
_
of a lattice
line segment.
(b) Show that the algebra in (a) is isomorphic to kŒX
m
. X
m-1
Y. . . . . XY
m-1
. Y
m
|.
(c) Now reduce the claim to kŒX. Y | and finish the proof.
5.22. Let k be a field and 1
k
(R
n
÷
) be the group of graded automorphisms of the infinite
polytopal k-algebra kŒR
n
÷
| whose underlying “polytope” is the infinite polyhedron R
n
÷
, de-
fined in a natural way (work out the details!). Let O denote the origin of R
n
÷
and e
i
denote
the i th standard basis element. Prove:
(a) For every element ¸ = (¸
0
. ¸
1
. . . . . ¸
n
) ÷ T
n
(k) the assignments
(O. 1) ÷¸
0
(O. 1).
(e
i
. 1) ÷¸
i
(e
i
. 1). i = 1. . . . . n
extend in a unique way to an element of 1
k
(R
n
÷
).
(b) The action of T
n÷1
(k) on kŒR
n
÷
| introduced in (a) defines a group embedding T
n
(k) ÷
1
k
(R
n
÷
). Show that the image is a maximal torus.
(c) Identifying k
+
with the image in 1
k
(R
n
÷
) of the first coordinate subgroup of T
n
(k),
show that 1
k
(R
n
÷
),k
+
is isomorphic to the group of augmented k-algebra automorphisms
of the polynomial algebra kŒX
1
. . . . . X
n
|.
(d) Introduce the notion of a column vector for R
n
÷
in exact analogy to Definition 5.32 with
the following extra condition: if a column vector : is associated with a facet F Ç R
n
÷
, then
for every lattice point x ÷ R
n
÷
\ F there is a natural number k for which c ÷ :k ÷ F.
(This is condition is automatic for polytopes!). Then a pair of elements : ÷ Col(R
n
÷
) and
z ÷ k defines an element e
z
¡
÷ 1
k
(R
n
÷
), called an elementary automorphism. Moreover,
Exercises 177
the exact analogue of Lemma 5.37 is true. (Hint: use the same argument as in the polytopal
case.)
(e) The subgroup of 1
k
(R
n
÷
), generated by the elementary automorphisms, intersects k
+
trivially (as introduced in (c)). As a consequence, this subgroup is isomorphic to a sub-
group of the group of augmented k-automorphisms of kŒX
1
. . . . . X
n
| (i. e. those leaving
the ideal (X
1
. . . . . X
n
) invariant) and we get the notion of an elementary automorphism of
kŒX
1
. . . . . X
n
|.
(f) Any automorphism of kŒX
1
. . . . . X
n
| of the form
X
j
÷X
j
. j ,= i.
X
i
÷X
i
÷g
i-1
. g
i-1
÷ kŒX
1
. . . . . X
i-1
|. g
i-1
(0. . . . . 0) = 0.
is a composition of elementary automorphisms.
(g) The subgroup of the group of augmented k-automorphisms of kŒX
1
. . . . . X
n
|, gener-
ated by the elementary automorphisms, is exactly the group of tame automorphisms of
kŒX
,
. . . . X
n
|. Here “tame” means to be a composition of linear changes of the variables
and automorphisms of the form
X
1
÷X
1
.
X
2
÷X
2
÷¡
1
.
.
.
.
X
n
÷X
n-1
÷¡
n-1
where ¡
i
÷ kŒX
1
. . . . . X
i
|. The latter are called triangular automorphisms. (Hint: show
that any triangular automorphism is a composite of automorphisms of the form in (f).
Then show that every elementary automorphism is tame.)
However, the analogue of Theorem 5.40 fails for 1
k
(R
n
÷
). In other words, there are (aug-
mented) automorphisms of kŒX
1
. . . . . X
n
| that are not tame, contrary to what was known
as a the tame generation conjecture. This conjecture was finally disproved in [76] using
Nagata’s old candidate for a wild automorphism.
Chapter 6
Homological properties and enumerative com-
binatorics
179
Chapter 7
Gröbner bases, triangulations, and Koszul alge-
bras
181
Part III
K-theory of monoid algebras
Chapter 8
Projective modules over monoid rings
8.A. Projective modules
Recall that (for us) a module P over a (commutative) ring R is called projective if
it is a direct summand of a free R-module of finite rank. The class of projective
R-modules is denoted by P(R).
Moreover, for a ring homomorphism ¡ : R ÷ R
t
and an R
t
-module M
t
we
say that M
t
is extended from R if there exists an R-module M such that M
t
Š
R
t
˝
R
M as R
t
-modules. Occasionally we use the notation ¡
+
(M) = R
t
˝
R
M.
The restriction to finitely generated modules is justified since, by a theorem of
Bass [3], nonfinitely generated projective modules over a noetherian domain are
free.
Recall that a module P is called finitely presented if there exists an exact se-
quence R
m
÷ R
n
÷ P ÷ 0 of R-modules with m. n ÷ N. The following basic
property of finitely presented modules, which is an easy exercise on the right ex-
actness of the functor R˝÷, will frequently be used later on.
Lemma 8.1. Let I be a filtered partially ordered set and let {¡
ij
: R
i
÷R
j
: i. j ÷
I. i < j ] be a family of ring homomorphisms such that ¡
jk
ı ¡
ij
= ¡
ik
whenever
i < j < k. Let R denote the direct limit lim
÷÷
R
i
. Then for every finitely presented
R-module P there exist an index i ÷ I and a finitely presented R
i
-module P
i
such
that P Š R˝
R
i
P
i
.
If P is projective, then one can choose i ÷ I such that P
i
is projective.
Of course, over a noetherian ring all our modules are finitely presented. But
later on it will be crucial that we work with nonnoetherian rings as well. By defini-
tion, a projective module is a direct summand of a free module of finite rank and,
therefore, is finitely presented whether the ground ring is noetherian or not.
This is not the place to develop the general theory of projective modules. Rather
we confine ourselves to stating some important basic properties. For proofs see [4,
Chapter III, §7] or [10, Ch. 2, §5].
Lemma 8.2. Let Rbe a ring and P be an R-module.
185
186 8. Projective modules over monoid rings
(a) P is projective if and only if P is a finitely presented R-module and P
m
is
free over R
m
for every maximal ideal m Ç R.
(b) If P is projective, then the map rank
P
: Spec(R) ÷ Z, given by p ÷
rank
R
p
(P
p
), is continuous with respect to the Zariski topology on Spec(R)
and the discrete topology on Z.
If ¡ : R ÷ R
t
is a ring retraction and an R
t
-module M
t
is extended from
R then the module M such that M
t
= R
t
˝
R
M is unique up isomorphism:
M Š R˝
R
0 M
t
. Obviously, M inherits both projectivity and finite presentability
fromM
t
(and conversely).
Example 8.3. We have already constructed nonfree projective modules of rank 1 in
Section 4.F. A very interesting, completely different example of a nonfree projec-
tive module is given by the following construction due to Kaplansky [59, Chapter
1, §4]. Consider the R-algebra R = RŒx
1
. x
2
. x
3
| generated by three generators
x
1
. x
2
. x
3
, subject to the relation x
2
1
÷x
2
2
÷x
2
3
= 1; we consider R as the ring of
the real valued polynomial functions on the standard unit sphere S
2
Ç R
3
. The
R-module P is defined by the split exact sequence
0 P R
3
(
R 0
. ¡ (ç. j. 0) = x
1
ç ÷x
2
j ÷x
3
0.
It is a projective module because P ˚R Š R
3
. In other words, P is even a stably
free R-module; see Section 8.I.
We claim that P is not free. Assume to the contrary that P is free. Then there
exists an R-automorphism of ˛ : R
3
÷ R
3
fitting into the commutative diagram
with exact rows
0 P R
3
˛
(
R
1
R
0
0 R
2
t
R
3
t
R 0
where |(ç. j) = (ç. j. 0) and ¬(ç. j. 0) = 0. This is equivalent to the existence of
an invertible 3 × 3 matrix over Rhaving the form
_
_
x
1
x
2
x
3
y
1
y
2
y
3
z
1
z
2
z
3
_
_
.
In particular, for no point z = (z
1
. z
2
. z
3
) ÷ S
2
the vector
F(z) =
_
y
1
(z). y
2
(z). y
3
(z)
_
÷ R
3
is parallel to z. We obtain a continuous vector field on S
2
if we assign to each point
z ÷ S
2
the orthogonal projection of F(z) onto the tangent plane of S
2
at z. As just
seen, this vector field vanishes nowhere, in contradiction to the theorem that every
such vector field must have a zero. This shows that the P is not free.
Actually, R has strong ring theoretical properties: it is regular and a unique
factorization domain (see [59, Chapter 1, §4] for the details). Therefore P does not
split into a direct sum of two nonzero modules since each of the two summands
8.B. The main theorem and the plan of the proof 187
would be a rank 1 projective module, and over a unique factorization domain such
modules are free.
8.B. The main theorem and the plan of the proof
In this chapter we prove
Theorem 8.4. Let R be a principal ideal domain (PID) and M be a seminormal
monoid. Then all finitely generated projective RŒM|-modules are free.
Note that we do not require that the monoid M is finitely generated or without
nontrivial units. (However, a seminormal monoid is cancellative and torsionfree
by definition.) That the condition of seminormality is inevitable has already been
shown in Corollary 4.66. In Theorem 8.27 we will obtain an even stronger converse
result.
Theorem 8.4 generalizes the classical Quillen-Suslin theorem that solved a
problem posed by Serre:
Theorem 8.5 (Quillen-Suslin). Let R be a PID and m a nonnegative integer. Then
all finitely generated projective modules over the polynomial ring RŒZ
m
÷
| are free.
Various extensions and applications of Theorem 8.4 will be discussed in Sec-
tions 8.I and 10.D. The proof of the theoremis given in Sections 8.D–8.G, while Sec-
tion 8.C contains the classical results on projective modules on which the Quillen-
Suslin theorem is based, to be extended and adapted to monoid rings in general.
Before delving into the technical details we sketch the course of the proof of 8.4.
The key case. When rank M = 0 there is nothing to prove – all (i. e. finitely gen-
erated) torsion free modules over a PID are free. The case of rank M = 1 is easily
reduced to the Quillen-Suslin theorem; see the discussion following Lemma 8.17.
Therefore, without loss of generality, we can assume that the following conditions
hold:
(a) rank M ¸ 2;
(b) Theorem 8.4 has been shown for monoids N such that rank N < rank M.
Later on we will see that the key case of Theorem 8.4 is the one in which the follow-
ing additional conditions hold:
(c) Ris a local PID;
(d) M is an affine, positive and normal monoid.
The reduction to this case (see Corollary 8.19) is based on classical results on pro-
jective modules to be recapitulated in Section 8.C. We mention the inductive hy-
pothesis (b) explicitly since some of the reduction steps rely on it.
In the rest of this section we assume that the conditions (a)–(d) all hold.
Pyramidal descent. As proved in Proposition 1.26, every pointed cone has a cross-
section: if ˛ is a linear form in the (absolute) interior of the polar cone C
+
, then
the intersection of the hyperplane H with equation ˛(x) = 1 is a polytope.
188 8. Projective modules over monoid rings
For a positive affine monoid M we choose ˛ as a rational linear form. Then we
obtain a rational cross-section of R
÷
M. For a subset N of RM we set
˚(N) = conv
_
x
˛(x)
: x ÷ N. ˛(x) = 0
_
.
This notation will be used frequently in this chapter and the next. Clearly ˚(M) is
the cross-section of R
÷
M defined by ˛.
In the following it will be crucial that we can shrink the monoid M in a way
that is controlled by a (rational) subpolytope of ˚(M). Therefore we introduce
the following notation that can be applied to every submonoid M of R
d
and every
subset W Ç R
d
: we set
M[W = M ¨ R
÷
W.
If M is an affine monoid and W is a rational polytope, then M[W is again affine.
This follows from Lemma 2.7 and Corollary 2.9.
The K-theoretical heart of the proof of Theorem 8.4 is
Theorem 8.6 (Pyramidal descent). Let : ÷ ˚(M) be a vertex and ˚(M) = zL1
be a decomposition of ˚(M) into two rational polytopes z and 1 such that z is a
pyramid of dimension d ÷ 1 with apex : and 1 meets z in the facet opposite to :.
Then every projective RŒM|-module is extended fromRŒM[1 |.
1
z
˚(M)
1
Figure 8.1. Pyramidal descent
If a polytope P decomposes in the formzL1 like ˚(M) in Theorem8.6 we say
that P is a pyramidal extension of 1 and that z¨1 is a pyramidal decomposition
of P. We do not exclude the case in which P = z and 1 is the facet of z opposite
of :.
We remark that in the proof of Theorem 8.6 we will make heavy use of pro-
jective modules over nonnoetherian rings of type RŒM
+
| (see Remark 2.4 for the
definition of M
+
), and here we leave the realm of noetherian rings – M
+
is never
a finitely generated monoid unless d = 1 (which is a trivial case in the context of
projective modules).
Shrinking the cross-section. The rest of the proof of Theorem 8.4 is a purely geo-
metric argument.
Theorem 8.7. Let P be a (rational) polytope and z ÷ int(P) a rational point. Then
there exists a sequence (P
i
)
i÷N
of (rational) polytopes with the following properties:
(a) for all i ÷ None has
(i) P
i
Ç P;
(ii) P
i÷1
Ç P
i
or P
i÷1
¸ P
i
;
8.C. Projective modules over polynomial rings 189
(iii) if P
i÷1
Ç P
i
, then P
i
is a pyramidal extension of P
i÷1
.
(b) for every c > 0 there exists an i ÷ Nsuch that P
i
Ç U
t
(z) ¨ P.
This theorem will be proved in Section 8.G. Let us show here that Theorem
8.6 and Theorem 8.7 reduce Theorem 8.4 to the Quillen-Suslin theorem. The cone
R
÷
M has a unimodular triangulation by Theorem 2.67, and we choose a rational
point z in the interior of an intersection ˚(M) ¨ D where D is a d-dimensional
unimodular cone of the triangulation. Next we choose a sequence of rational poly-
topes P
i
Ç P = ˚(M) as in Theorem 8.7. Then there exists c > 0 such that
D¨P contains U
t
¨P. Figure 8.2 illustrates this construction. It shows members
Figure 8.2. Extending modules from shrunk polytopes
in the sequence (P
i
) that are homothetic images of P with respect to the center z,
and the inclusion of such images in the sequence will be a crucial argument in its
construction.
Let Q be a projective RŒM|-module. The extension P
1
Ç P is pyramidal, and
therefore there exists a projective RŒM[P
1
|-module Q
1
such that Q is extended
from Q
1
. Now we can define a projective RŒM[P
i÷1
|-module Q
i÷1
, i ¸ 1, re-
cursively by (i) Q
i÷1
= Q
i
˝ RŒM[P
i÷1
| if P
i÷1
¸ P
i
, or (ii) as a projective
RŒM[P
i÷1
|-module that extends to Q
i
if P
i÷1
Ç P
i
. Clearly, this construction is
only possible because of Theorem 8.6. Note that Q is extended from each of the
Q
i
.
For c as above there exists j ÷ Nsuch that P
j
Ç U
t
(z)¨P, and we can consider
the extension Q
t
= Q
j
˝RŒD¨Z
d
|. Since D is unimodular, D¨Z
d
Š Z
d
÷
, and
so Q
t
is a free module over RŒD ¨ Z
d
| by the Quillen-Suslin theorem. But Q is
extended fromQ
t
, and so Qis free.
8.C. Projective modules over polynomial rings
In this section we recall the required results on Serre’s problem on projective mod-
ules over polynomial rings.
190 8. Projective modules over monoid rings
Milnor patching. A diagram of commutative rings
A A
1
(
1
A
2
(
2
A
t
(8.1)
is said to have the Milnor patching property if the following condition is satisfied:
whenever we are given finitely generated projective modules P
1
and P
2
over A
1
and A
2
and an isomorphismA
t
˝
A
1
P
1
Š A
t
˝
A
2
P
2
of A
t
-modules, the pullback
P of the diagram
P
1
P
2
A
t
˝P
1
Š A
t
˝P
2
(i. e. the set of pairs (x. y) ÷ P
1
×P
2
such that x and y map to the same element in
A
t
˝P
2
) is a finitely generated projective A-module and the natural maps A
i
˝
A
P ÷P
i
are isomorphisms.
We will need the following special cases in which (8.1) has the Milnor patching
property:
(A) Milnor squares. The diagram (8.1) is cartesian, i. e. A = {(x. y) ÷ A
1
×
A
2
: ¡
1
(x) = ¡
2
(y)], and A
1
÷A
t
is surjective.
(B) Karoubi squares. The diagram has the form
A
(
B
S
-1
A
¡ (S)
-1
B
where S Ç A is regular on A and B, and A,sA ÷ B,¡ (s)B is an iso-
morphism for every s ÷ S. (Actually, it suffices if we only require that
A,sA ÷B,¡ (s)B is a surjective homomorphism.)
(C) Localization squares. The diagram has the form
A A
s
A
t
A
st
where As ÷At = A.
Case (A) is proved in [65, §2], (B) is considered in [90, Appendix A], and case
(C) is just standard sheaf patching on Spec(A) (see, for instance, [59, Chapter 1,
Corollary 3.12]).
The following lemma is an immediate consequence of the definition of Milnor
patching.
8.C. Projective modules over polynomial rings 191
Lemma 8.8. Suppose the diagram (8.1) has the Milnor patching property and P ÷
P(A
1
). If A
t
˝P is extended fromA
2
, then P is extended fromA.
The following lemma is an application of localization squares.
Lemma 8.9. Let P be a projective RŒX. X
-1
|-module. Let ¡ ÷ RŒX
-1
| be monic in
X
-1
. If P
(
is extended fromR, then P is extended fromRŒX|.
Proof. Write ¡ = X
-n
g where g ÷ RŒX| and g(0) = 1. Then
RŒX| RŒX. X
-1
|
RŒX|
g RŒX. X
-1
|
g
= RŒX. X
-1
|
(
is a localization square. Since P
(
is extended from R, it is extended from RŒX|
g
and we are done by Milnor patching.
Quillen’s local-global principle. Quillen’s proof [71] of the Quillen-Suslin theorem
is based in the following local-global principle that we will use without proof in the
following.
Theorem 8.10. Let R be a commutative ring and P be a finitely presented RŒX|-
module. Then P is extended fromRif and only if P
m
is extended fromR
m
for every
maximal ideal m Ç R.
We need the extension of this theorem to general graded rings:
Theorem 8.11. Let R = R
0
˚R
1
˚ be a graded ring and P a finitely presented
R-module. Suppose P
m
is extended from (R
0
)
m
for every maximal ideal m Ç R
0
.
Then P is extended fromR
0
.
Here P
m
denotes the localization of P with respect to the multiplicative subset
R
0
\ m Ç R.
Although it is possible to simply adapt Quillen’s original argument to the more
general situation of graded rings, we explain how Theorem 8.11 follows from the
special case of polynomial rings, using the so-called Swan-Weibel homotopy trick.
Proof of Theorem 8.11. Consider the following ring homomorphisms:
(i) the inclusion maps i : R
0
÷ R and j : R ÷ RŒX|; moreover, for max-
imal ideals m Ç R
0
and n Ç R we will also use i and j for the inclusion
maps (R
0
)
m
÷R
m
and R
n
÷R
n
ŒX|;
(ii) the map n : R ÷RŒX| defined by r ÷rX
n
for r ÷ R
n
;
(iii) the augmentation c : R ÷R
0
sending R
n
to 0 for n > 0;
(iv) e
k
: RŒX| ÷Rfor k = 0. 1 ÷ Rby sending X to k.
Then e
0
j = e
1
j = e
1
n = 1
R
and e
0
n = i c (in this proof we suppress the symbol
ı for composition). Now let W = n
+
(P). If m Ç R
0
is a maximal ideal then P
m
is
extended from (R
0
)
m
and so has the form P
m
= i
+
(V ) for some finitely presented
(R
0
)
m
-module V . Therefore W
m
Š n
+
(P
m
) Š n
+
i
+
(V ) Š j
+
i
+
(V ). If n Ç R is a
maximal ideal choose m ¸ n¨R
0
a maximal ideal of R
0
. Then W
n
is a localization
192 8. Projective modules over monoid rings
of W
m
and so is extended via j . By the original Quillen local-global principle, W is
also extended fromRvia j : W Š j
+
(Q) for some Q, and this is the crucial point.
So
P Š e
1+
(W) Š e
1+
j
+
(Q) = e
0+
j
+
(Q) Š e
0+
(W) Š e
0+
n
+
(P) Š i
+
c
+
(P).
showing that P is extended from R
0
.
Inverting monic polynomials. For a ring R and a variable X we let R(X) denote
the localization of the polynomial ring RŒX| with respect of the multiplicative set
of all monic polynomials, i. e. those with invertible leading coefficient.
Although based on substantially different approaches, both Quillen’s and
Suslin’s solutions to Serre’s problem [71, 84] use the following criterion for a pro-
jective module over a polynomial ring to be extended from the ring of coefficients.
Theorem 8.12. Let R be a ring and P be a projective RŒX|-module. Then P is
extended from Rif and only if R(X) ˝
RŒXj
P is extended fromR.
Below we will include a proof of this theorem. Next we explain how it implies
the Quillen-Suslin theorem 8.5. Actually, we do so for its extension to Laurent
polynomial rings, due to Swan [91]:
Theorem 8.13. Let R be a PID and m. n nonnegative integers. Then all finitely
generated projective modules over the Laurent polynomial ring RŒZ
m
÷
˚ Z
n
| are
free.
We need the following two lemmas.
Lemma 8.14. If Ris a PID then so is R(X).
This is proved in [59, Chapter 4, Corollary 1.3].
Lemma 8.15. Let P be a projective RŒX. X
-1
|-module. If R(X)˝P and R(X
-1

P are extended fromR, then P is extended fromR.
Proof. There exists a monic polynomial ¡ in X
-1
such that P
(
is extended from
R(since R(X
-1
) is a filtered union of extensions RŒX. X
-1
|
(
). By Lemma 8.9 P is
extended from RŒX|: P Š RŒX. X
-1
| ˝ Q for some projective RŒX|-module Q.
Then R(X)˝Q = R(X)˝P is extended fromR. By Theorem 8.12 Qis extended
fromRand, hence, so is P.
Proof of Theorem 8.13. We have to show that projective modules over the Lau-
rent polynomial ring RŒZ
m
÷
˚Z
n
| are free if Ris a PID. We use induction on m÷n.
The case m÷n = 0 is obvious.
Let P be a projective RŒZ
m
÷
˚ Z
n
|-module. First assume that m > 0. By the
induction hypothesis and Lemma 8.14 the module R(X)ŒZ
m-1
÷
˚Z
n
| ˝P is free
over R(X)ŒZ
n-1
÷
˚ Z
n
|, where we make the identification RŒZ
m-1
÷
˚ Z
n
|ŒX| =
RŒZ
m
÷
˚Z
n
|. Since RŒZ
m-1
÷
˚Z
n
|(X) is a further localization of R(X)ŒZ
m-1
÷
˚Z
n
|,
we are done by Theorem 8.12.
Now assume that n > 0. By Lemma 8.14 and the induction hypothesis the
modules R(X)ŒZ
m
÷
˚ Z
n-1
| ˝ P and R(X
-1
)ŒZ
m
÷
˚ Z
n-1
| ˝ P are free over
8.C. Projective modules over polynomial rings 193
R(X)ŒZ
m
÷
˚Z
n-1
| and R(X
-1
)ŒZ
m
÷
˚Z
n-1
| respectively. Here we make the identifi-
cation RŒZ
m
÷
˚Z
n-1
|ŒX. X
-1
| = RŒZ
m
÷
˚Z
n
|. Since RŒZ
m
÷
˚Z
n-1
|(X) and RŒZ
m
÷
˚
Z
n-1
|(X
-1
) are further localizations of R(X)ŒZ
m
÷
˚ Z
n-1
| and R(X
-1
)ŒZ
m
÷
˚
Z
n-1
|, respectively, we are done be Lemma 8.15.
Roberts’ theorem. The next theorem, due to P. Roberts, will play a crucial rôle for
the pyramidal descent. In the theorem and its proof we use the following notation:
J(B) is the Jacobson radical of a ring B, i. e. the intersection of all its maximal
ideals, and E
n
(B) is the group generated by all elementary n × n matrices, i. e.
those matrices that have the entry 1 on the diagonal and exactly one off-diagonal
element nonzero.
Note that a surjective ring homomorphism R ÷ S induces a surjective group
homomorphism E
n
(R) ÷ E
n
(S). Therefore, if E
n
(S) = SL
n
(S), the natural
homomorphism SL
n
(R) ÷ SL
n
(S) is surjective, too. By the Gauß algorithm, one
has E
n
(S) = SL
n
(S) if S is a local ring.
Theorem8.16. Let (L. m) be a local ring, Aan L-algebra, P a finitely generated A-
module, and S Ç A a multiplicative subset which is regular on A and P. Suppose
the following conditions hold for the natural number n:
(a) A,¡Ais a finitely generated L-module for all ¡ ÷ S;
(b) every matrix ˛ ÷ GL
n
(S
-1
A) can be decomposed as a product ˇ; with
ˇ ÷ GL
n
(S
-1
A) and ; ÷ GL
n
(
¨
A), where
¨
A = A,mA and S
-1
A =
S
-1
A,mS
-1
A;
(c) there is an L-subalgebra B Ç S
-1
Awith S
-1
A = A÷B and mB Ç J(B);
(d) S
-1
P Š
_
S
-1
A
_
n
and
¨
P Š
¨
A
n
.
Then P Š A
n
.
This is proved in [59, Chapter 4, §4] with only one difference in the hypothe-
ses: instead of the condition (b) the stronger condition of surjectivity of the map
GL
n
(S
-1
A) ÷ GL
n
(S
-1
A) is required. But the weaker condition (b) above is
enough to ensure the existence of bases of S
-1
P and
¨
P mapping to the same basis
of S
-1
P, and this is all that is needed for the proof in [59, Chapter 4, §4] .
We now explain how Theorem 8.12 follows from Theorems 8.10 and 8.16. The
reader may check that Theorem 8.10 and Theorem 8.16 are the only results on
projective modules that we have not proved, apart from Milnor patching.
Proof of Theorem 8.12. By Theorem 8.10 we can assume that Ris local. Let m Ç
R be the maximal ideal. We apply Theorem 8.16 to L = R and A = RŒX|. Let
S be the set of monic polynomials. Conditions (a) and (d) are clearly satisfied
because S
-1
A = R(X) and
¨
A = kŒX| where k = R,m.
Furthermore S
-1
A = k(X), so SL
n
(S
-1
A) = E
n
(S
-1
A) and SL
n
(S
-1
A) ÷
SL
n
(S
-1
A) is surjective. To prove (b) we observe that U(k(X)) is generated by the
images of S and U(k). Condition (b) follows, since every element ˛ of GL(S
-1
A)
can now be decomposed into a product ˛
t
ˇ; where ˛
t
÷ SL
n
(S
-1
A), ˇ is an
invertible diagonal matrix over S
-1
A, and ; is such a matrix over k(X).
194 8. Projective modules over monoid rings
For (c) we choose B to be the set of all ¡,g where g is monic and deg(¡ ) ¸
deg(g). Then S
-1
A = A ÷ B follows easily if we divide ¡ by g when deg(¡ ) >
deg(g). Finally, if m ÷ m, then 1 ÷ m¡,g = (g ÷ m¡ ),g is a unit of B and this
implies mB Ç J(B).
8.D. Reduction to the interior
The first reduction in the proof of Theorem 8.4 is given by
Lemma 8.17. It is sufficient to prove Theorem 8.4 for the case in which R is a local
PID and M is an affine positive seminormal monoid.
Proof. Let N = (M \ U(M)) L {1] (writing the monoid operation multiplica-
tively). It is clear that N is also a seminormal monoid. We have the cartesian
square of R-algebras
RŒN| RŒM|
R
RŒU(M)|
in which the vertical maps send the elements of N \ {1] to 0 ÷ R. This square is
of Milnor type (A). Since U(M) is the filtered union of finitely generated torsion-
free and, hence, free abelian groups, all projective RŒU(M)|-modules are free by
Lemma 8.1 and Theorem 8.5. Now it suffices to have all projective RŒN|-modules
free.
The monoid N is the filtered union of positive affine monoids M
i
. The semi-
normalizations of the M
i
are also affine (see p. 63), so that we can replace each M
i
by its seminormalization. By Lemma 8.1 it is enough that the projective RŒM
i
|-
modules are free for all i .
For the lemma we still have to show that R can be replaced by its localizations
if M is a positive affine seminormal monoid. By Proposition 2.15 the ring RŒM|
admits a grading RŒM| = R˚R
1
˚ in which the elements of M are homoge-
neous. But then Theorem 8.10 implies that the projective RŒM|-modules are free
if the projective R
m
ŒN|-modules are free for all maximal ideals m of R.
Lemma 8.17 finishes the proof of Theorem8.4 for rank 1 monoids: the monoids
M
i
are all seminormal of rank 1, and therefore isomorphic to Z
÷
, as is easily seen.
So the Quillen-Suslin theorem shows that all projective RŒM
i
|-modules are free.
The next lemma shows two facts simultaneously. Namely, it further reduces the
general case to that of normal monoids and, moreover, it reduces the problem to
the interior of M.
Lemma 8.18. Let M be an affine positive seminormal monoid of rank d and R
a PID. Suppose that the projective RŒN|-modules are free for every affine positive
normal monoid of rank < d. Then every projective RŒM|-module is extended from
RŒM
+
|.
8.E. Graded “Weierstraß Preparation” 195
Proof. Let F
1
. . . . . F
n
be the faces of R
÷
M, labelled in such a way that F
i
¸ F
j
implies i ¸ j . In particular, F
1
= R
÷
M and F
n
= {0]. Set M
i
= M ¨ F
i
and
U
i
= int(M
i
). Then the M
i
are the extreme submonoids of M, and the U
i
form a
partition of M (see Theorem 1.10(d)). Let W
i
= U
1
L L U
i
and W
0
= 0.
We claim that
MW
i
Ç W
i
. (8.2)
In fact, let x ÷ M and y ÷ W
i
. Then y ÷ U
j
for some j ¸ i . Since the U
k
from a
partition of M, we have xy ÷ U
q
Ç M
q
Ç F
q
for some q. It follows that y ÷ M
q
.
So U
j
¨ M
q
= 0. Since int(F
j
) ¨ F
q
= 0, we get M
j
Ç M
q
by Theorem 1.10(d).
Therefore, q ¸ j and xy ÷ U
q
Ç W
i
.
Nowlet J
i
be the R-submodule of RŒM| generated by W
i
. It is an ideal of RŒM|
by (8.2). Set A
i
= RŒM|,J
i
. The kernel of the natural homomorphismA
i-1
÷A
i
is the free R-module with basis U
i
. So for i < n we have the Milnor squares
RŒ(M
i
)
+
| A
i-1
R A
i
.
(8.3)
Since (M
i
)
+
is the filtered union of affine positive normal monoids by Proposition
2.38, all finitely generated projective RŒ(M
i
)
+
|-modules, i > 0, are free by our
inductive hypothesis. Since A
n-1
= R, descending induction on i , in conjunction
with Lemma 8.8, shows that all projective A
i
-modules are extended from R for
i < n. The lemma follows because A
0
= RŒM|.
Let us sum up what reduction we have reached in proving Theorem 8.4:
Corollary 8.19. It is enough to prove Theorem 8.4 under the following additional
assumptions:
(a) M is affine, positive, and normal;
(b) rank M ¸ 2;
(c) Ris local;
(d) (Induction hypothesis on rank) projective R
t
ŒN|-modules are free for every
PID R
t
and all positive normal monoids N with rank N < rank M;
(e) all projective RŒM|-modules are extended fromRŒM
+
|.
We only need to explain why we can restrict ourselves to normal monoids in
(a): if N is just seminormal, then N
+
is the filtered union of normal affine monoids,
and so we reach the conclusion for N via Lemma 8.18.
8.E. Graded “Weierstraß Preparation”
In the next section we will need a graded variant of the formal Weierstraß prepa-
ration theorem.
196 8. Projective modules over monoid rings
Theorem 8.20. Let A = A
0
˚ A
1
˚ be a graded ring an let : ÷ A
d
. Let
M = M
0
˚M
1
˚ be graded A-module satisfying the condition
: : M
i
÷M
i÷d
. x ÷:x. is an isomorphism for i ¸ 0. (8.4)
Let ¡ ÷ Awith ¡ ÷ a
0
÷a
1
÷ ÷a
nd-1
÷:
n
mod (nil(A)). Then every z ÷ M
can be written as z = ¡q ÷r with q ÷ M and r ÷ M
0
÷ ÷M
nd-1
. Moreover,
q and r are unique.
Proof. Let ¡ = ¡
0
÷ ¡
m
. Then ¡
nd÷1
. . . . . ¡
m
and ¡
nd
÷ :
n
are nilpotent and
therefore generate a homogeneous nilpotent ideal J. Suppose J
h
= 0. We use
induction on h. If h = 0, the usual division algorithm clearly applies. Let
N =
˚
x ÷ M : :
k
x ÷ J
h-1
M for some k ¸ 0
_
=
_
k¸0
_
J
h-1
: :
k
_
.
Then N satisfies (8.4) and so does M,N. Moreover, M,N is a module over
A,J
h-1
and N is a module over A,J since x ÷ N implies :
k
Jx = 0 and : is
regular on M by (8.4). By induction, the conclusion of the theorem applies to the
residue classes of ¡ in A,J
h-1
and A,J and the modules M,N and to N respec-
tively. Writing ¨ z = ¡ ¨ q ÷ ¨ r in M,N and lifting it to M, we get z = ¡q ÷r ÷n
where n ÷ N. An application of the induction hypothesis to N yields n = ¡q
t
÷r
t
so that z = ¡ (q ÷q
t
)÷(r ÷r
t
). For the uniqueness it is enough to consider z = 0.
Then we must have ¨ q = ¨ r = 0 for by induction applied to M,N. Hence q and r
lie in N, and so q = r = 0 by induction applied to N.
Corollary 8.21. Let A = A
0
˚A
1
˚ be a graded ring and let : be an element of
A
d
satisfying the condition
: : A
i
÷A
i÷d
. x ÷:x. is an isomorphism for i ¸ 0. (8.5)
Let ¡ ÷ A with ¡ ÷ a
0
÷ a
1
÷ ÷ a
nd-1
÷ :
n
mod (nil(A)). Then ¡ =
(1 ÷j)(:
n
÷ r) where j ÷ nil(A) and r ÷ A
0
÷ ÷A
nd-1
.
Proof. Apply the theorem with M = A and z = :
n
to get :
n
= ¡q ÷ r. The
theorem also applies to M = A, nil(A). The surjectivity in (8.4) for M follows
from that in (8.5) for A while the injectivity follows from the fact that : is regular
on A. Therefore x is nilpotent if :x is. In M = A, nil(A) we have
¨ :
n
=
¨
¡ ÷ ( ¨ a
0
÷ ¨ a
1
÷ ÷ ¨ a
nd-1
) = ¨ q
¨
¡ ÷ ¨ r
by the theorem. Thus the uniqueness of ¨ q implies ¨ q = 1, and so q = 1 ÷j with j
nilpotent. Therefore, q is invertible and q
-1
= 1 ÷j with j ÷ nil(A).
8.F. Pyramidal descent
In this section we prove Theorem 8.6. During the entire section M is supposed
to be an affine, positive, normal monoid of rank d ¸ 2. As in the statement of
Theorem 8.6, we assume that gp(M) = Z
d
and fix a rational hyperplane H Ç R
d
that defines a cross-section ˚(M) of the cone R
÷
M Ç R
d
.
Local descent. The first step is a local version of pyramidal descent:
8.F. Pyramidal descent 197
Theorem 8.22. Let (R. m) be a local PID, and M Ç Z
d
a normal monoid with
gp(M) = Z
d
. Assume that projective R
t
ŒH|-modules are free for every positive
normal monoid H with rank(H) < d and every PID R
t
. Suppose ˚(M) = zL 1
is a pyramidal decomposition of ˚(M) with pyramid z, and N = M[1 . Let
M= (m. N
÷
) Ç RŒN| be the maximal ideal generated by m and the noninvertible
elements of N.
Then for every projective RŒM|-module Qthe localized module Q
M
is free over
RŒM|
M
= (RŒN| \ M)
-1
RŒM|.
The proof mimics the derivation of Theorem 8.12 from Theorem 8.16. How-
ever, the situation is more complicated now.
Let us first define and examine all the data that depend only on the monoid
structures. The polytopes z and 1 meet in a common face of dimension d ÷ 2. It
spans a rational vector subspace H
t
of dimension d ÷1, and we choose the unique
primitive integral form ı vanishing on H
t
and having ı(x) ¸ 0 for all x ÷ z. This
linear form will serve as a degree function. Note that it separates N from M[z,
having values ¸ 0 on N and values ¸ 0 on M[z.
Let : be the vertex of ˚(M) that serves as the apex of the pyramid z. We
choose the extreme integral generator t of R
÷
M in the extreme ray R
÷
: of R
÷
M.
Lemma 8.23. Let x ÷ M with ı(x) ¸ aı(t ), a ÷ Z
÷
. Then x ÷ at ÷ M.
Proof. Since ı(x) > 0, we have x ÷ M[z. Since z is a pyramid with apex :, all
support linear forms of zvanish on t , except ı, which defines the facet opposite to
:. Therefore we only need to check that ı(x ÷ at ) ¸ 0, and this is obvious.
Since R
÷
N ÷ R
÷
t = R
÷
M, the monoid M is integral over N
t
= N ÷
Z
÷
t . In particular it is finitely generated as an N
t
-module, say by m
1
= 0 and
m
2
. . . . . m
n
÷ M[z. Then, as an N-module, it is generated by all sums kt ÷m
i
,
k ÷ Z
÷
, i = 1. . . . . n. Note that, given b ÷ Z
÷
, only finitely many of these monoid
elements have value ¸ b under ı.
From now on we switch to multiplicative notation for monoids. The ideal of
RŒM| generated by N
÷
is the free R-module with basis X = {yz : y ÷ N
÷
. z ÷
M]. Therefore, C = RŒM|,(N
÷
)) is the free R-module with basis Y = M \ X;
he multiplication on C is given by the R-bilinear extension of
y z =
_
yz if yz ÷ Y.
0 otherwise.
The linear form ı defines a grading on C. Namely,let Y
k
= {x ÷ Y : ı(x) = k]
and let C
k
= RY
k
. We have C
k
= 0 for l < 0 because N
÷
Ç X. So C is positively
graded, C = C
0
˚C
1
˚ .
Lemma 8.24. The R-linear map C
i
÷ C
i÷ı(t )
, c ÷ t c, is an isomorphism for all
i ¸ 0.
Proof. It is enough to show that the multiplication by t maps Y
i
bijectively onto
Y
i÷ı(t )
. We first show that the image lies in Y
i÷ı(t )
. Let y ÷ Y
i
and assume that
ty … Y
i÷ı(t )
. Then ty = xz with x ÷ N
÷
, z ÷ M. Now ı(y) > 0 and ı(x) ¸ 0
198 8. Projective modules over monoid rings
so ı(z) > ı(t ). By Lemma 8.23 z = st for some s ÷ M. Therefore y = xs,
contradicting the fact that y ÷ Y
i
.
The map is clearly injective, and it is easily seen that it is surjective, too: if
y ÷ Y
i÷ı(t )
, then y = tx for some x ÷ M by Lemma8.23 and x must lie in Y
i
,
since y ÷ X if x ÷ X.
We prove Theorem 8.22 by applying Roberts’ theorem 8.16 to the rings L =
RŒN|
M
and A = RŒM|
M
. We have to define a multiplicative subset S Ç RŒM|
m
and an RŒN|
M
-subalgebra B Ç S
-1
RŒM|
M
so that the conditions of Theorem
8.16 are satisfied.
We choose S Ç A as follows. First we set deg(m) = ı(m) for the elements of
M. Next we write deg(¡ ) ¸ c for an element ¡ ÷ A if ¡ = r
1
n
1
÷ ÷ r
p
n
p
where r
i
÷ L and n
i
÷ M such that ı(n
i
) < c for all i . Then we call an element
g ÷ Amonic if g = t
a
÷¡ with deg(¡ ) < a deg(t ) and set deg(g) = a deg(t ). Now
we choose S as the set of monic elements. Evidently S is a multiplicatively closed
set. (The reader may worry whether the degree is well-defined – we will justify the
definition below.)
One can imitate the usual division algorithm since A is generated as an L-
module by the elements of M:
Lemma 8.25. Let ¡ ÷ A and g ÷ S. Then we can write ¡ = gq ÷ r in A with
deg(r) < deg(g).
Finally, we put
B =
˚
˛ ÷(¡,g) : ˛ ÷ L. g ÷ S. ¡ ÷ A. deg(¡ ) < deg(g)
_
.
This is easily seen to be a subring of S
-1
A.
Now we check the conditions of Theorem 8.16. Let P = Q
M
. Since A is a
domain, it is clear that S is regular on Aand P.
(a) If ¡ ÷ S, then A,¡Ais finite over L.
By Lemma 8.25 A,¡A is generated as an L-module by the monomials m
i
t
k
where
m
i
runs through the generating set of M over N, k ÷ Z
÷
, and deg(m
i
t
k
) <
deg(¡ ). As already remarked, there are only finitely many such monomials.
(b) Every matrix ˛ ÷ GL
n
(S
-1
A) can be decomposed as a product ˇ; with ˇ ÷
GL
n
(S
-1
A) and ; ÷ GL
n
(
¨
A), where
¨
A = A,MAand S
-1
A = S
-1
A,MS
-1
A.
Since L,M= R,m = k is a field, we see that
¨
A = A,MA = RŒM|,(m. N
÷
) = kŒM|,(N
÷
).
If x ÷ M, x = 1, then x
m
= t
b
y with b ¸ 0 and y ÷ N. So either x ÷ t
Z
C
or x
is nilpotent modulo the ideal generated by N
÷
. Since t ÷ M generates an extreme
submonoid, M \ t
Z
C
is an ideal of M (see p. 56), and RŒM|,
_
M \ t
Z
C
_
Š RŒt |.
Therefore,
RŒM|,
_
N
÷
_
red
= RŒt |
and so
¨
A
red
= kŒt |.
8.F. Pyramidal descent 199
Let us examine the image of S in kŒt |. Let g = t
a
÷

s
i
m
i
with m
i
÷ M,
deg(m
i
) < a deg(t ) and s
i
÷ L. Since (M \ t
Z
C
) goes to 0, only terms s
i
m
i
with m
i
= t
j
for some j ÷ Z
÷
can survive in kŒt |. Since deg(m
i
) < a deg(t ),
it follows that j < a. Furthermore the coefficient s
i
goes to its residue class in
k. Therefore the residue class of g is indeed a monic polynomial in kŒt |. This
observation justifies the notion of degree introduced above. Moreover, the image
of S is exactly the set of monic polynomials, as every monic polynomial in kŒt | can
clearly be lifted to an element of S.
Consequently (S
-1
A)
red
= k(t ). It follows that S
-1
A is a local ring (having
exactly one prime ideal), so that SL
n
(S
-1
A) = E
n
(S
-1
A), and SL
n
(S
-1
A) ÷
SL
n
(S
-1
A) is surjective. Therefore, we only need to prove that U(S
-1
A) ×
U(
¨
A) ÷U(S
-1
A) is surjective.
Now we use the graded structure of RŒM|,(N
÷
) that we have introduced
above. It follows that
¨
A = C,MC =
¨
A
0
˚
¨
A
1
˚ where
¨
A
i
= C
i
,MC
i
and
Lemma 8.24 implies that the multiplication by t is an isomorphism
¨
A
i
÷
¨
A
i÷e
,
e = deg(t ), for all i ¸ 0. A unit of S
-1
A has the form ¡,s where s is in the
image of S Ç U(S
-1
A) and ¡ ÷
¨
A divides some element of the image of S.
Since
¨
A
red
= kŒt | we see that, up to a factor from U(k), ¡ maps to a monic
polynomial in kŒt |. By Corollary 8.21, ¡ = (1 ÷ j)g where j is nilpotent and
g = t
m
÷ a
me-1
÷ ÷ a
0
with a
i
÷
¨
A
i
. Since 1 ÷ j ÷ U(
¨
A), it will suf-
fice to show that g lies in the image of S. We can lift g to h in A, choosing
h = t
m
÷b
me-1
÷ ÷b
0
with b
i
a linear combination over L of element of Y
i
.
Since the elements of Y
i
have smaller degree than t
m
for i < me, we have indeed
lifted g to a monic element.
(c) S
-1
A = A÷B and MB Ç J(B).
Every element of S
-1
S has the form¡,g with g monic. Writing ¡ = gq ÷r as in
Lemma 8.25 we see that ¡,g = q ÷r,g ÷ A÷B so S
-1
A = A÷B.
If m
i
÷ Mand ˛
i
÷¡
i
,g ÷ B, then u = 1÷

m
i

i
÷¡
i
,g) = c÷¡,g where
c = 1 ÷

m
i
˛
i
is a unit in L and ¡ =

m
i
¡
i
. Therefore, u = c(g ÷c
-1
¡ ),g
which is a unit of S
-1
Asince g ÷c
-1
¡ is monic. This shows that MB Ç J(B).
(d) S
-1
P Š (S
-1
A)
n
and
¨
P Š (
¨
A)
n
.
Since
¨
A
red
= kŒt |, the projective module
¨
P,nil(
¨
A)
¨
P is free over
¨
A
red
and hence
¨
P
is free over
¨
A, say
¨
P Š (
¨
A)
n
. Suppose P
t
is free over A
t
. Since t ÷ S, S
-1
P Š
(S
-1
A)
n
0
for some n
t
. Both rings S
-1
Aand
¨
Amap to the same ring S
-1
A(which
is a nilpotent extension of k(t )). This proves n = n
t
.
So it only remains to show that P
t
is free over A
t
. In its turn, this would follow
if Q
t
were free over RŒM|
t
. By Proposition 2.31 we have RŒM|
t
Š RŒN|ŒZ| Š
RŒN|ŒX. X
-1
where N is an affine positive normal monoid of rank d ÷ 1, and
RŒN|(X) is a localization of R(X)ŒN|. Therefore, Lemma 8.14 and Theorem 8.12,
together with the induction assumption on rank in Theorem 8.22, give us the de-
sired freeness of Q
t
.
200 8. Projective modules over monoid rings
Global descent. The derivation of Theorem 8.6 from Theorem 8.22, to be given
now, is not a local-to-global passage in the sense of patching local data. Rather
we will apply Theorem 8.22 to an infinite family of rings distinct from the original
RŒM| to derive the analogue of Theorem 8.6 for M
+
. The final step to M is taken
with the help of Lemma 8.18. We keep the notation introduced in Theorem 8.22.
Proof of Theorem 8.6. By Corollary 8.19 we can additionally assume that (i)
R = (R. m) is a local PID and (ii) projective RŒM|-modules are extended from
RŒM
+
|.
We want to prove that for every projective RŒM|-module P there exists a pro-
jective RŒM[1 |-module P
T
such that
P Š RŒM| ˝
RŒM[T j
P
T
. (8.6)
By Lemma 8.18 we find P
+
÷ P(RŒM
+
|) such that
P Š RŒM| ˝
RŒM

j
P
+
Fix such P
+
and consider the maximal ideal M
+
=
_
m. int(M[1 )
_
Ç RŒM
+
|. (Note
that int(M[1 ) is the ideal of noninvertible elements in (M[1 )
+
.)
First we prove that
(P
+
)
M

is free over RŒM
+
|
M

. (8.7)
There are rational equidimensional polytopes 1
i
Ç int(1 ) and rational points :
i
÷
int(z), i ÷ Nsuch that the following conditions are satisfied:
(i) :
i
÷: as i ÷o,
(ii) 1
1
Ç 1
2
Ç ,
(iii) int(1 ) =
_
i÷N
1
i
,
(iv) for every i ÷ Nexactly one facet of 1
i
is visible from :
i
,
This approximation of 1 is illustrated by Figure 8.3. For example, it can be real-
:
:
2
:
1
1
1
1
2
1
Figure 8.3. The approximation of 1
ized as follows: we choose a rational point in the interior of 1 , and consider the
homotheties 0
i
with center z and factor i,(i ÷ 1); then we set :
i
= 0
i
(:) and
8.F. Pyramidal descent 201
1
i
= 0
i
(1 ). (However, in a similar situation for higher K-groups the choice of the
polytopes 1
i
and the points :
i
must be made much more carefully; see [45, §6].)
Consider the affine positive normal submonoids M
i
= M[ conv(1
i
. :
i
). By
Lemma 8.1 there exists i such that P
+
Š RŒM
+
| ˝
RŒM
i
j
P
i
. Then
(P
+
)
M

= RŒM
+
|
M

˝
RŒ(M
i
)
M
i
j
(P
i
)
M
i
where M
i
=
_
m. int(M
i
[1
i
)
_
Ç RŒM
i
|. Therefore, it is enough to show that (P
i
)
M
i
is free over RŒ(M
i
)
M
i
|. But now we are exactly in the situation of Theorem 8.22,
the rôles of M and Qbeing played by M
i
and P
i
. This proves (8.7).
Now we prove that P
+
is extended from RŒM[1
+
|. Then it follows that P is
extended from RŒM[1 |, and we are done. By Lemma 8.8 and (8.7) it suffices to
show that the commutative diagram
RŒM[1
+
| RŒM
+
|
RŒM[1
+
|
M

RŒM
+
|
M

is a Karoubi square. Therefore, it is enough to show that for all elements s ÷
RŒM[1
+
| \ M
+
the map
RŒM[1
+
|,s ÷RŒM
+
|,s
is an isomorphism. This can be checked locally on RŒM[1
+
|, i. e. it suffices to show
that
˛
N
: (RŒM[1
+
|,s)
N
÷(RŒM
+
|,s)
N
(8.8)
is an isomorphism for every maximal ideal N ÷ max(RŒM[1
+
|).
Suppose first that N ¨ M[1
+
= 0. Then M[1
+
Ç RŒM[1
+
| \ N and ˛
N
is a
further localization of (M[1
+
)
-1
RŒM[1
+
| ÷ (M[1
+
)
-1
RŒM
+
|. But by Corollary
2.25 we have
gp(M[1 ) = gp(M[1
+
) Ç (M[1
+
)
-1
M
+
Ç gp(M
+
) = gp(M) = gp(M[1 )
and ˛
N
turns out to be a localization of an isomorphism.
Now suppose that N¨M[1
+
= 0. We will see that N = M
+
. This implies that
the source and target of ˛
N
are both the zero module.
In fact, N¨ M[1
+
is a prime ideal of the monoid M[1
+
. But there is no other
prime ideal in M[1
+
except int(M[1 )
1
: This follows as in the proof of Proposition
2.19: every x ÷ int(M[1 ) is nilpotent over the ideal (in the monoid M
+
) gen-
erated by an arbitrary y ÷ int(M[1 ). Moreover, since R is local, the inclusion
int(M[1 ) Ç N implies that m Ç N and, thus, N = M
+
.
1
It is here where we need the fact that the cone R
÷
M[1
+
Ç R
d
is open. This explains the
necessity of the passage to interior submonoids.
202 8. Projective modules over monoid rings
8.G. How to skin a polytope
In this section we prove Theorem 8.7. We first show that a pyramidal decompo-
sition of a face can be extended to the whole polytope, an argument obviously
necessary for an inductive attack.
Lemma 8.26. Let P Ç R
d
be a rational polytope and F Ç P be a proper face
(not necessarily a facet). Let z ÷ int(P) be a rational point. Suppose G Ç F is a
pyramidal extension. Then there exists a pyramidal extension P
t
Ç P such that
P
t
¨ F = G and z ÷ int(P
t
).
Proof. By induction we can assume that F is a facet of P. Let : ÷ F be the vertex
not in G. We claim that there exists a rational affine function ˛ : aff(P) ÷ R
satisfying the following conditions
(i) ˛(x) ¸ 0 for all x ÷ G,
(ii) ˛(:) > 0,
(iii) ˛(n) < 0 for every vertex n ÷ P different from :,
(iv) there exists a vertex n ÷ P such that ˛(n) < 0.
Then the extension of polytopes
Q = {z ÷ P : ˛(z) ¸ 0] Ç P
is pyramidal. In fact, P decomposes into Q and z = {z ÷ P : ˛(z) ¸ 0], and
G
P
F
Figure 8.4. Extending a pyramidal decomposition
the hyperplane H
˛
separating Q and z intersects the interior of P. Moreover, z
is a pyramid. Namely, the ray Rfrom: through any point x ÷ zleaves P through
a facet not containing :, and at the last point y in P the function ˛ must have a
negative value. Since ˛(:) > 0, the line segment Œ:. y| meets H
˛
. Thus R leaves z
through H
˛
.
In order to construct ˛ we choose ˇ as an affine form vanishing on the facet
F and having negative values on the rest of P. Furthermore we choose ; as an
extension of an affine form on aff(F) that defines the given pyramidal decompo-
sition of F (and has positive value in :). We then set ˛ = ˇ ÷ c; for a rational
c > 0 small enough such that ˛(n) < 0 for all vertices n of P different from :,
and furthermore ˛(z) < 0.
8.G. How to skin a polytope 203
Let P be a rational polytope. We say that a (finite or infinite) sequence (P
i
) of
rational subpolytopes is admissible if the following conditions are satisfied for all
i :
(i) P
i
Ç P;
(ii) P
i÷1
Ç P
i
or P
i÷1
¸ P
i
;
(iii) if P
i÷1
Ç P
i
, then P
i
is a pyramidal extension of P
i÷1
.
As we have seen in Section 8.B the proof of Theorem 8.4 will be complete once
we have shown Theorem 8.7: given a rational point z ÷ int(P), there exists an
admissible sequence (P
i
) such that P
i
Ç U
t
(z) ¨ P for every c > 0 and i ;
0. If this condition is satisfied, we say that (P
i
) contracts to z. Observe that an
admissible sequence (P
i
) that contracts to z is necessarily infinite.
Proof of Theorem 8.7. We use induction on dimP. It is enough to prove that
there is a finite admissible sequence P = P
1
. . . . . P
n
such that z ÷ int(P
i
) for all i
and
P
n
Ç int(P). (8.9)
Then we can find a rational number z ÷ (0. 1) such that 0
z
(P) Ç P
n
where 0
z
is
the homothety with center z and factor z. Now we can replace P
n
by 0
z
(P), and
choose P
n÷i
= 0
z
(P
i
), i = 1. . . . . n. Then P
2n
Ç 0
z
2(P). By iteration we obtain
an admissible sequence of polytopes contracting to the point z, and are done.
Now we want to show the existence of an admissible sequence satisfying (8.9).
We can assume that dimP ¸ 1 for otherwise P is a point and, thus, coincides with
its own interior. The case dimP = 1 is obvious because we can chose P
1
to be P
2
with a small (rational) subsegment cut off at one end-point of P and P
3
to be P
2
with a small rational subsegment cut off at the other end-point of P.
Let d = dimP ¸ 2. Assume that for every d-polytope P and all facets Q Ç P
we can find an admissible sequence P
1
. . . . . P
n
such that
P
n
¨ Q = 0 (8.10)
and z ÷ int(P
n
). Then we apply the same argument to P
n
and “skin off ” Q
t
¨ P
n
where Q
t
is another facet of P. Iterating the process we remove all facets of the
original polytope P.
By the previous step all we need to show is that for every facet Q Ç P there
exists an admissible sequence P = P
1
. . . . .P
n
satisfying (8.10).
Let z Ç Q be a rational simplex and z
t
÷ int(z) a rational point. By the
inductive hypothesis we can find an admissible sequence (Q
i
) in Q that contracts
to z
t
. In particular, Q
n
Ç z for some n. We can extend (Q
i
) to an admissible
sequence in P: whenever we pass fromQ
i
to Q
i÷1
by removing a pyramid, Lemma
8.26, applied to P
i
and the extension Q
i÷1
Ç Q
i
,yields P
i÷1
. If Q
i÷1
¸ Q
i
, we
choose P
i÷1
= conv(P
i
L Q
i÷1
).
Now replace P
n
by conv(P
n
Lz). The intersection of the new P
n
with the facet
Qof the original polytope P is the simplex z. Choose a vertex : of z. Then zis a
pyramidal extension of its facet z
t
opposite to :. Applying Lemma 8.26 once more,
we can cut off a pyramid from P
n
in such a way that all points of z, except those
in z
t
, have been removed. Again we have reached a polytope whose intersection
204 8. Projective modules over monoid rings
with Qis a simplex, but now of smaller dimension. Iterating the procedure we can
cut off z completely.
The reader may have noticed that it would actually not be necessary to take
care of the point z in Lemma 8.26, since it is contained in 0
z
n(P) for all n.
8.H. Converse results
In this section we relax our standard convention on monoids: they are only assumed
to be commutative, but not necessarily cancellative or torsionfree.
Now we prove the following converse to Theorem 8.4. Its part (b) is a special
case of Corollary 4.66, and only mentioned here because we want to discuss its
validity under a relaxation of the conditions on M.
Theorem 8.27. (a) Let M be a cancellative, but not necessarily torsion free monoid.
Suppose Pic(RŒM|) = 0 for every PID R of characteristic 0. Then M is torsion free
and seminormal.
(b) If M is cancellative and torsion free and Pic(kŒM|) = 0 for some field k,
then M is seminormal.
The condition that M be cancellative cannot be dropped, and we cannot re-
strict ourselves in Theorem8.27(a) only to fields. This is explained by the following
examples.
Example 8.28. Consider the commutative noncancellative monoid M generated by
three elements x. y. z subject to the relations x
2
= xy = y
2
. Then all projective
RŒM|-modules are free for every PID R. In fact, RŒM|
red
Š RŒX. Y |,(X ÷Y ) is a
polynomial ring over R.
Example 8.29. Let M = (Z
÷
× Z
2
) \ {(0. 1)]. Then all projective kŒM|-modules
are free for every field k. The reader may check that kŒM| Š kŒX. Y |,(X
2
÷ Y
2
).
If char k = 2, then kŒM| = kŒX. Y |,(X ÷ Y )(X ÷ Y ) Š kŒX. Y |,(XY ), and
the freeness of projective kŒM|-modules follows from Theorem 8.49 below. But if
char k = 2, then kŒM|
red
Š kŒX|.
The proof of Theorem 8.27(a) requires some preparation.
Proposition 8.30. Let G be an abelian group. Then the group ring RŒG| is seminor-
mal for every PID Rof characteristic 0 if and only if G is torsion free.
Before we prove the proposition we derive Theorem 8.27(a) from it.
Proof of Theorem 8.27(a). Note that RŒM| is reduced because char R = 0 im-
plies kŒgp(M)| is reduced for the quotient field k of R; see Theorem 4.11.
By Lemma 8.14 R(X) is a PID if and only if R is, so Pic(R(X)ŒM|) = 0 for all
PID’s Rof characteristic 0. Therefore, the composite map
Pic
_
RŒX|ŒM|
_
÷Pic
_
R(X)ŒM|
_
÷Pic
_
RŒM|(X)
_
is zero. Since Pic(AŒX|) ÷ Pic(A(X)) is always injective by Theorem 8.12, we
get Pic
_
RŒM|ŒX|
_
= Pic
_
RŒM|
_
= 0. By the theorem of Traverso-Swan RŒM| is
8.H. Converse results 205
seminormal. Since RŒgp(M)| is a localization of RŒM| by a multiplicatively closed
set of non-zerodivisors, RŒgp(M)| is seminormal. So Proposition 8.30 implies that
M is torsion free.
The proof of Proposition 8.30 is based on a series of auxiliary results.
Lemma 8.31. Let R be a commutative ring, let G be an abelian group, and suppose
that RŒG| is seminormal. Let T be the torsion subgroup of G. Then RŒT | is also
seminormal.
Proof. The subring RŒT | of RŒG| is itself reduced. Suppose x
3
= y
2
in RŒT |.
Then there is z ÷ RŒG| such that z
2
= x and z
3
= y. We can find a subgroup
H Ç G, containing T, such that z ÷ RŒH| and H,T is finitely generated. Since
H,T is torsion free, it is free, and we can write H = T ˚F. Since H retracts to
T, the group ring RŒH| retracts onto RŒT |, sending z to n ÷ RŒT | with n
2
= x
and n
3
= y.
Lemma 8.32. Let T be an abelian torsion group and assume that RŒT | is seminor-
mal for a PID R with an infinite number of distinct residue characteristics. Let H
be a finite subgroup of T . Then RŒH| is also seminormal.
Proof. Suppose that x
3
= y
2
in RŒH|. Then there is z ÷ RŒT | such that z
2
= x
and z
3
= y. We can find a finite subgroup H
t
Ç T containing H such that z ÷
RŒH
t
|. Choose a maximal ideal m Ç R such that the characteristic of k = R,m is
prime to the order of H
t
. Consider the diagram
RŒH| RŒH
t
|
kŒH| kŒH
t
|
and let ¨ x. ¨ y. ¨ z denote the images of x. y. z in kŒH
t
|. Since gcd(char k. [H
t
[) =
1, kŒH| and kŒH
t
| are products of copies of k by Maschke’s theorem, and hence
reduced and seminormal. Since z
2
. z
3
÷ kŒH| it follows that ¨ z ÷ kŒH|. Therefore,
z ÷ RŒH| ÷ mRŒH
t
|. Now RŒH
t
| = RŒH| ˚ F as an R-module where F is
the free R-module on H
t
\ H. This show that z ÷ RŒH| ˚ mF. Since there are
an infinite number of choices for m, we have
_
mF = 0. This show that z ÷
_
(RŒH| ˚mF) = RŒH| ˚0.
Lemma 8.33. Let R be a domain of characteristic 0 and let H be a finite group of
order n. If RŒH| is seminormal, then R,nRis reduced.
Proof. Let N ÷ RŒH| be the sum of elements of H and let S = RH,(N). Since
kŒH| = k × kŒH|,(N) where k = QF(R), we see easily that S is reduced and the
diagram
RŒH|
S
R
R,nR
is cartesian. So we are done by the following lemma.
206 8. Projective modules over monoid rings
Lemma 8.34. Let
A B
C D
be a cartesian diagram with B and C reduced and B ÷ D surjective. If A is
seminormal, then D is also reduced.
Proof. We think of A as the set of pairs (b. c) ÷ B × C such that b and c have
the same image in D. Clearly, A is reduced. If D were not reduced, we could find
t ÷ B mapping to
¨
t with
¨
t = 0, but
¨
t
2
= 0. Let x = (t
2
. 0), y = (t
3
. 0) in A.
Then x
3
= y
2
, but there is no z ÷ A with z
2
= x and z
3
= y. If there were such
a z = (b. c), then b = t and c = 0 since B and C are reduced, and it would follow
that
¨
t = 0.
Proof of Proposition 8.30. If the torsion subgroup T Ç G is not trivial we
choose a finite nontrivial subgroup H of T. Let Abe the ring of integers of Q(
_
n)
where n = [H[, and let p
1
. . . . . p
m
be the prime ideals of Acontaining n. Consider
B = S
-1
Rwhere S = A\(p
1
L Lp
m
). Then B is a semilocal Dedekind domain,
and therefore a PID [10, Ch. VII, §2, Prop. 1]. Since the class group of A is finite,
we can find prime ideals q
1
. . . . . q
p
, all different from p
1
. . . . . p
m
, that generate the
class group of A. Now we choose an element s ÷ (q
1
¨ ¨ q
p
) \ (p
1
L L p
m
).
It follows that R = AŒs
-1
| is a PID with infinitely many residue characteristics.
By Lemma 8.31 RŒT | is seminormal. But, since R,nR is not reduced, RŒH| is not
seminormal by Lemma 8.33 – a contradiction with Lemma 8.32.
8.I. Generalizations
The Grothendieck group K
0
. First we give an overview of the basic facts on the
Grothendieck functor K
0
. For the general theory the reader is referred to the clas-
sical books [4] and [87].
We start with some observations on an arbitrary functor F : Rings ÷
AbGroups from the category of rings to that of abelian groups. (Recall that we
consider only commutative rings.)
(i) Let ¡ : R ÷ R
t
be a ring retraction, i. e. there exists a ring homomorphism
g : R
t
÷ R such that g¡ = 1
R
. Then F(R) is a direct summand of F(R
t
), and we
identify F(R) with its isomorphic image in F(R
t
).
For us, the most relevant examples of ring retractions are the identity embed-
dings R = R
0
÷ R
0
˚ R
1
˚ = R
t
where the graded ring R
t
is a monoid
algebra RŒM| of an affine positive monoid M. However, R is always a retract
of RŒM|, whether the monoid M is affine and positive or not: the surjective R-
homomorphism : c
1
: RŒM| ÷ R, M ÷ 1, splits the identity embedding
R ÷RŒM| (see Section 4.A).
8.I. Generalizations 207
(ii) To a commutative diagram of rings
A
(
1
(
2
B
g
1
C
g
2
D
we usually associate the sequence of group homomorphisms
F(A)
;
F(B) ˚F(C)
ˇ
F(D).
where
;(x) =
_
F(¡
1
)(x). F(¡
2
)(y)
_
. x ÷ F(A).
ˇ(y. z) = F(g
1
)(y) ÷ F(g
2
)(z). (y. z) ÷ F(B) ˚F(C).
The commutativity of the diagram implies that ˇ; = 0.
(iii) Let F : Rings ÷AbGroups be a functor. If a ring homomorphism¡ : A ÷B
induces an isomorphism F(¡ ) : F(A) ÷ F(B) we simply write F(A) = F(B),
assuming the ring homomorphism ¡ is clear from the context.
The functor K
0
: Rings ÷AbGroups assigns to a ring Rthe group
K
0
(R) = F
X
,G
where F
X
is the free abelian group on the set X of isomorphism classes (P) of
projective R-modules and G is the subgroup generated by the elements of type
(P ˚ Q) ÷ (P) ÷ (Q). For a ring homomorphism ¡ : R ÷ R
t
the functorial
homomorphismK
0
(¡ ) is given by ŒP| ÷ŒR
t
˝
R
P|.
Two projective R-modules P and Q define the same element ŒP| = ŒQ| in
K
0
(R) if and only if P and Q are stably isomorphic: P ˚R
n
Š Q˚R
n
for some
n. In particular, K
0
(R) = Z means that every projective R-module P is stably
free, i. e. P ˚ R
m
Š R
n
for some m. n ÷ N. If M is seminormal and R is a PID
then K
0
(RŒM|) = Z by Theorem 8.4. A generalization of this equality for higher
dimensional regular rings is given by Theorem 8.37 below.
The tensor product of two projective modules induces a commutative ring
structure on K
0
(R), and we have a natural embedding Pic(R) ÷ U(K
0
(R)). The
injectivity of this homomorphism follows from the observation that if ŒP| = ŒQ| in
K
0
(R) for some P. Q ÷ Pic(R) then P Š Q: in fact, if P˚R
n
Š Q˚R
n
for some
n ÷ N then, taking the (n ÷1)st exterior powers, we get P Š
_
n÷1
(P ˚R
n
) Š
_
n÷1
(Q ˚ R
n
) Š Q. In general, Pic(R) = U(K
0
(R)); see Example 9.26 in
Chapter 9.
Consider the ring H
0
(R) of continuous functions Spec(R) ÷ Z where
Spec(R) carries the Zariski topology and Z is a discrete set. It follows easily
from Lemma 8.2(b) that the assignment ŒP| ÷ rank
P
induces a well defined ring
homomorphismK
0
(R) ÷H
0
(R).
208 8. Projective modules over monoid rings
Next we define a group homomorphism
_
rank
: K
0
(R) ÷ Pic(R). Consider
a module P ÷ P(R). Because the affine spectrum Spec(R) is quasicompact (ev-
ery open cover contains a finite subcover) the image of rank
P
÷ H
0
(R) is a fi-
nite subset of Z
÷
. Then (rank
P
)
-1
(N) Ç Spec(R) is the disjoint union of certain
subsets that are simultaneously open and closed. This decomposition of Spec(R)
induces a decomposition of the form R = e
1
R × × e
n
R where e
1
. . . . . e
n
are
mutually orthogonal idempotents, and e
i
P ÷ P(e
i
R) is of constant rank for every
i = 1. . . . . n. (Here we view e
i
R as a ring whose unit element is e
i
.) We have
_
rank(e
i
P)
(e
i
P) ÷ Pic(e
i
R), i = 1. . . . . n, where the exterior powers are taken over
e
i
R. Then the direct sum of the rank one modules over the rings e
i
Ris a rank one
module over R which we denote by
_
rank
(P). It is easily checked that the assign-
ment ŒP| ÷
_
rank
(P) gives rise to a well defined group homomorphism, denoted
by
_
rank
: K
0
(R) ÷Pic(R).
The diagram below with exact rows and columns serves as the definition of
further groups associated to K
0
(R):
0
SK
0
(R)
0
¯
K
0
(R)
K
0
(R)
_
rank
rank
H
0
(R)
0
Pic(R)
1
Pic(R)
0 0
When Spec(R) is connected (i. e. Rhas no nontrivial idempotents) then H
0
(R) =
Z. In this case we have SK
0
(R) = 0 if and only if every projective R-module stably
has the form “free ˚rank one”.
For every Milnor square of rings
A B
(
C D
we have the exact K
0
-Mayer-Vietoris sequence
K
0
(A) ÷K
0
(B) ˚K
0
(C) ÷K
0
(D). (8.11)
see [65, §3]. It is easily observed that we also have a similar exact sequence for H
0
.
Therefore, the cartesian square above yields the following commutative diagram
with exact rows and columns – the right half of the diagram in [4, Chapter IX,
8.I. Generalizations 209
Corollary 5.12]:
0 0 0
SK
0
(A) SK
0
(B) ˚SK
0
(C) SK
0
(D)
¯
K
0
(A)
¯
K
0
(B) ˚
¯
K
0
(C)
¯
K
0
(D)
Pic(A) Pic(B) ˚Pic(C) Pic(D)
0 0 0.
(8.12)
Finally, we have the following
Lemma 8.35. If I Ç A is an ideal contained in the Jacobson radical J(A) then
K
0
(A) ÷ K
0
(A,I) is a monomorphism and Pic(A) ÷ Pic(A,I) is an isomor-
phism.
In particular, the homomorphisms
¯
K
0
(A) ÷
¯
K
0
(A,I) and SK
0
(A) ÷SK
0
(A,I)
are monomorphisms.
The proof is based on the following consequence of the Nakayama Lemma:
modules P. Q ÷ P(A) are isomorphic if and only if their reductions mod I are
isomorphic over A,I See [4, Chapter IX, Proposition 1.3].
Homotopy invariance. One of the first results in algebraic K-theory is Grothen-
dieck’s theorem [4, Chapter XII, §3] (originally published in [75]) on the K
0
-
homotopy invariance of a regular ring of arbitrary Krull dimension:
Theorem8.36. For every regular ring Rthe identity embedding into the polynomial
and Laurent polynomial rings R ÷ RŒX| ÷ RŒX. X
-1
| induces natural isomor-
phisms K
0
(R) = K
0
(RŒX|) = K
0
(RŒX. X
-1
|).
2
By iterated use of Theorem 8.36 we get the equality K
0
(R) = K
0
(RŒZ
m
÷
×Z
n
÷
|)
for every regular ring R and all m. n ÷ Z
÷
. In view of Theorem 8.4 it is natural to
ask to what extent the same equality holds for general monoid rings. The complete
answer to this question is given by the following
Theorem 8.37. Let Rbe a regular ring and M be a monoid. Then
(a) the following conditions are equivalent:
(i) K
0
(R) = K
0
(RŒM|),
(ii) Pic(R) = Pic(RŒM|),
(iii) M is seminormal;
(b) SK
0
(R) = SK
0
(RŒM|).
2
Grothendieck’s theorem is true for regular schemes which are not necessarily affine, where
the group K
0
is defined in terms of locally trivial sheaves, see Chapter 10.C.
210 8. Projective modules over monoid rings
In the proof of Theorem 8.37 we will use the following stable version of Theo-
rem 8.11 for projective modules:
Theorem 8.38. Let R = R
0
˚ R
1
˚ be a graded ring and P be a projective
R-module. Then P is stably extended from R
0
, i. e. ŒP| ÷ Im
_
K
0
(R
0
) ÷ K
0
(R)
_
,
if and only if P
m
is stably free for every maximal ideal m Ç R
0
.
In the special case when R = R
0
ŒX| this is proved in [96], along with the same
equality for all higher K-groups. Then the general case of graded rings is derived
in the same way as Theorem 8.11 was derived from Theorem 8.10; see Section 8.C.
The functorial version of the Swan-Weibel homotopy trick (see the proof of Theo-
rem 8.11) to be used here is the following
Lemma 8.39. Let F : Rings ÷ AbGroups be a functor and A = A
0
˚A
1
˚ be
a graded ring. If F(AŒX|) = F(A) then F(A) = F(A
0
).
The proof is based on the same ideas as Theorem8.11 and we leave it as Exercise
8.1.
One more result we will use in the proof of Theorem 8.37 is the following:
Theorem8.40 (Bass’ cancellation). Let Abe a noetherian ring of finite Krull dimen-
sion dimA. Then two modules P. Q ÷ P(A) of rank > d (i. e. rank
P
(p) > d for all
p ÷ Spec(A)) are stably isomorphic if and only if they are isomorphic.
See [4, Chapter IV, Corollary 3.5].
Proof of Theorem 8.37(a). (i) == (ii) The identity embedding R ÷ RŒM|
is split by the surjective R-algebra homomorphism RŒM| ÷ R, M ÷ 1. By
functoriality, the map Pic(R) ÷ Pic(RŒM|) is injective. On the other hand,
since K
0
(R) ÷ K
0
(RŒM|) is an isomorphism, the two invertible RŒM|-modules
RŒM| ˝
RŒMj
˝
R
P and P are stably isomorphic. But we have already observed
that stably isomorphic invertible modules are necessarily isomorphic.
(ii) == (iii) If R is an integral domain, then this implication is part of Corol-
lary 4.66. However, R
m
is an integral domain for every maximal ideal m of R, and
the construction of the critical RŒM|-module I commutes with localization.
(iii) == (i) This is similar to the proof of Theorem 8.4. We describe the steps
that correspond to each other in more detail.
(1) Since the homomorphism K
0
(R) ÷ K
0
(RŒM|) is split injective, it is enough
to show its surjectivity.
(2) The functor K
0
commutes with filtered direct limits, and so the general case
reduces to the situation when M is finitely generated.
(3) By Theorem 8.11 we can further assume that R is a regular local ring. Then all
projective R-modules are free. We want to prove that all projective RŒM|-modules
are stably free.
For any module P ÷ P(RŒM|) we have the equivalence
ŒP| ÷ Im
_
K
0
(R) ÷K
0
(RŒM|)
_
” ŒP ˚RŒM|
dimR÷rank M÷1
| ÷ Im
_
K
0
(R) ÷K
0
(RŒM|)
_
.
8.I. Generalizations 211
Therefore, it is enough to prove the following stronger claim: all projective RŒM|-
modules of rank > dimR÷rank M are free.
This will be done by induction on rank M, the case rank M = 1 following from
Theorems 8.36 and 8.40.
(4) Using the inductive hypothesis on rank M and the same cartesian squares as in
Section 8.D, we can assume that M is an affine positive normal monoid and that it
is enough to showthat all projective RŒM
+
|-modules of rank > dimR÷rank M are
free. Observe that the monoid ring RŒM| is a domain since R is local and regular.
In particular, the projective modules, involved in the Milnor patching associated
to any of the cartesian squares mentioned above, have the same constant rank.
(5) Next we go on with exactly the same argument as in the proof of Theorem 8.6
– the pyramidal descent. The only difference that we encounter in this process
is the verification of the last condition in Roberts’ theorem for the corresponding
objects, namely that S
-1
P is free over S
-1
A, see p. 199: we have to show that
all projective RŒN|ŒZ|-modules of rank > dimR ÷ rank M free where N is an
affine positive normal monoid with rank N = rank M ÷ 1. By Lemma 8.15 it is
enough to show that all projective modules over R(X)ŒN| and R(X
-1
)ŒN|, having
rank > dimR÷rank M, are extended correspondingly from R(X) and R(X
-1
).
Since dimR = dimR(X) = dimR(X
-1
) [59, Chapter 4, Corollary 1.3], the result
follows from Theorem 8.11 and induction on the monoid rank.
(6) Using exactly the same argument as in the subsection Global descent of Sec-
tion 8.F, we conclude that all projective RŒM
+
|-modules are extended from an R-
algebra isomorphic to RŒZ
rank M
÷
|. But then Theorems 8.36 and 8.40 give the desired
freeness.
Proof of Theorem 8.37(b). Theorem 8.36 implies SK
0
(R) = SK
0
(RŒZ
n
|) for n ÷
N. Moreover, the functor SK
0
, like K
0
, commutes with inductive limits. Therefore
the Mayer-Vietoris SK
0
-sequence (the upper row of the diagram (8.12)), associated
to the cartesian square of rings
RŒN| RŒM|
R
RŒU(M)|
with N = (M \ U(M)) L {1] reduces the general case to the situation in which
M is an affine positive monoid.
Applying the Mayer-Vietoris SK
0
-sequences to cartesian squares of the form
(8.3), we see that it is enough to show that SK
0
(R) = SK
0
(RŒM
+
|) where M is an
affine positive monoid. (Observe that here we consider squares of the form (8.3)
for arbitrary, not necessarily normal monoids.)
By Corollary 2.33 we have
R =
_
RŒM
+
|,Rm
¨
M
_
red
=
_

¨
M
+
|,Rm
¨
M
_
red
. (8.13)
212 8. Projective modules over monoid rings
With N = m
¨
M L {1] we have the cartesian squares
RŒN| RŒM
+
|
R RŒM
+
|,Rm
¨
M
RŒN|

¨
M
+
|
R RŒ
¨
M
+
|,Rm
¨
M
(8.14)
where the vertical arrows represent the reduction modulo the ideal Rm
¨
M Ç
RŒM| and the horizontal arrows represent the identity embeddings. For simplicity
of notation we set I = Rm
¨
M in the following.
By (8.13), the bottom homomorphisms in the diagrams (8.14) are split nilpo-
tent extensions. By Lemma 8.35 we have
SK
0
(R) = SK
0
(RŒM
+
|,I) = SK
0
(RŒ
¨
M
+
|,I).
Suppose we have shown that the embedding R ÷ RŒN| induces an isomor-
phism of SK
0
-groups. Then we apply the Mayer-Vietoris sequence associated to
the left diagram in (8.14) to get an exact sequence
SK
0
(R) ÷SK
0
(R) ˚SK
0
(RŒM
+
|) ÷SK
0
(R)
In this sequence the map on the left maps SK
0
(R) diagonally to SK
0
(R)˚SK
0
(R) Ç
SK
0
(R)˚SK
0
(RŒM
+
|), whereas the map on the right, when restricted to SK
0
(R)˚
SK
0
(R) is the codiagonal map (u. u
t
) ÷u÷u
t
. It follows easily that SK
0
(RŒM
+
|) =
SK
0
(R), as desired.
In order to show the crucial equation SK
0
(R) = SK
0
(RŒN|) we have to borrow
the following Mayer-Vietoris K
1
-K
0
-sequence from Section 9.A: it extends the se-
quence for SK
0
to the left. In particular, the right square in (8.14) gives rise to the
following exact sequence
SK
1
(R) ˚SK
1
(RŒ
¨
M
+
|) ÷SK
1
(RŒ
¨
M
+
|,I) ÷SK
0
(RŒN|)
÷SK
0
(R) ˚SK
0
(RŒ
¨
M
+
|) ÷SK
0
(RŒ
¨
M
+
|,I). (8.15)
Theorem 8.4 implies that SK
0
(R) = SK
0
(RŒ
¨
M
+
|). We also have SK
1
(R) =
SK
1
(RŒ
¨
M
+
|,I) because R ÷ RŒ
¨
M
+
|,I is a nilpotent extension (see Section 9.A).
More precisely, the restriction of the map SK
1
(R)˚SK
1
(RŒ
¨
M
+
|) ÷SK
1
(RŒ
¨
M
+
|,I)
is an isomorphism SK
1
(R) ÷ SK
1
(RŒ
¨
M
+
|,I). Therefore, (8.15) yields the exact
sequence
0 ÷SK
0
(RŒN|) ÷SK
0
(R) ˚SK
0
(R) ÷SK
0
(R) ÷0.
By the same arguments on diagonal and codiagonal maps as above we conclude
SK
0
(R) = SK
0
(RŒRN|), and are done.
Remark 8.41. Using the same technique as above, but applying cancellation re-
sults for Laurent polynomial rings over Noetherian coefficient rings [91] that
are stronger than Theorem 8.40, Swan [92, Corollary 1.4] deduces the following
stronger version of Theorem 8.37: for a regular ring R of finite Krull dimension
dimR and an arbitrary monoid M all projective RŒM|-modules of rank > dimR
are extended fromR. Combining this technique with the work of Bhatwadekar and
8.I. Generalizations 213
Rao [6] Swan furthermore shows that for every localization R of a regular affine
algebra over a field and every monoid M without nontrivial invertible elements all
projective RŒM|-modules are actually extended from R. See Theorem 1.2 and the
discussion following Conjecture Q
n
in Section 3 of [92].
Picard groups. For positive affine monoids M one can describe the Picard group
Pic(RŒM|). In combination In with Theorem 8.37 the following result then gives a
complete characterization of the additive structure of K
0
(R) when R is a regular
ring and Q Ç R. (The restriction to reduced rings of coefficients is irrelevant since
Pic is stable under the passage to a residue class ring modulo a nilpotent ideal.)
Theorem 8.42. Let Rbe a seminormal reduced ring such that Q Ç Rand M be an
affine positive monoid. Then
Pic(RŒM|) Š Pic(R) ˚
_

sn(M)\M
R
_
.
We need two lemmas on units.
Lemma 8.43. Let Q Ç Abe commutative ring and n Ç Abe a nilpotent ideal. Then
the multiplicative group 1÷n = {1÷x : x ÷ n] is isomorphic to the additive group
n.
Proof. A pair of mutually inverse homomorphisms is given by
exp : n ÷1 ÷n. x ÷

n¸0
x
n
n!
log : 1 ÷n ÷n. 1 ÷x ÷÷

n>0
(÷x)
n
n

Lemma 8.44. Let A = A
0
˚A
1
˚ be a graded ring. Then
U(A) = {u ÷a
1
÷ ÷a
n
: u ÷ U(A
0
). n ÷ N. a
i
÷ A
i
is nilpotent.
i = 1. . . . . n]
The proof if straightforward.
Proof of Theorem 8.42. Note that sn(M) is an affine monoid (see the discus-
sion immediately after Definition 2.36). Therefore, sn(M) is also an affine positive
monoid and by Proposition 2.15(f) we can fix a grading RŒsn(M)| = R˚R
1
˚
making the monomials homogeneous elements. Then we have induced gradings
on RŒM
t
| for any submonoid M
t
Ç sn(M) as well as on monomial quotients of
RŒM
t
|.
First we consider the special case when M = sn(M). By Theorem 4.69
RŒsn(M)| is seminormal. Therefore, by Theorem 4.67(b), we have Pic(RŒM|) =
Pic(RŒM|ŒX|). Because of the grading on RŒM| Lemma 8.39 implies the equality
Pic(R) = Pic(RŒM|).
Now suppose that M = sn(M). Then there exists x ÷ sn(M) \ M such that
x
2
. x
3
÷ M. Therefore, I = RŒM|x
2
÷ RŒM|x
3
is an ideal in both rings RŒM|
214 8. Projective modules over monoid rings
and RŒM
t
| where we have set M
t
= Mx
Z
C
(we are using multiplicative notation).
In particular, we have the Milnor square
RŒM| RŒM
t
|
RŒM|,I RŒM
t
|,I.
(8.16)
The U-Pic Mayer-Vietoris sequence, associated to (8.16) (see [4, Chapter IX, Corol-
lary 5.12])
3
, contains the exact sequence
1 ÷U(RŒM|) ÷U(RŒM|,I) ˚U(RŒM
t
|)
(
÷÷U(RŒM
t
|,I) ÷Pic(RŒM|)
÷Pic(RŒM|,I) ˚Pic(RŒM
t
|)
g
÷ ÷Pic(RŒM
t
|,I). (8.17)
By Proposition 4.12 one has U(RŒM
t
|) = U(R). Therefore, by Lemmas 8.43
and 8.44 we get
Coker ¡ Š
nil(RŒM
t
|,I)
nil(RŒM|,I)
Š R(M
t
\ M). (8.18)
The isomorphism on the right is only an isomorphism of R-modules, and seen
as follows. Both algebras RŒM
t
|,I and RŒM|,I are gp(M)-graded, and therefore
their nilradicals are graded ideals (Exercise 4.8). But a graded ideal in one of these
rings is a free module over R on the basis formed by the (residue classes of) the
monomials it contains. So the quotient is free on the difference of the bases, and
that is formed by the set M
t
\ M. It cannot be larger, and is not smaller since all
elements in M
t
\ M are nilpotent mod I.
The last observation implies (RŒM|,I)
red
= (RŒM
t
|,I)
red
, and we use it in
order ro compute Ker g. By Lemma 8.35, the homomorphism Pic(RŒM|,I) ÷
Pic(RŒM
t
|,I) is an isomorphism. Since g is the codiagonal of the two homomor-
phisms
Pic(RŒM|,I) ÷Pic(RŒM
t
|,I) and Pic(RŒM
t
|) ÷Pic(RŒM
t
|,I.
we arrive at the isomorphism
Ker g Š Pic(RŒM
t
|). (8.19)
Therefore, (8.17), (8.18) and (8.19) yield a short exact sequence of the form
0 ÷R(M
t
\ M) ÷Pic(RŒM|) ÷Pic(RŒM
t
|) ÷0.
Since Q Ç R(used for the second time!), the additive group of the free R-module
R(M
t
\ M) is injective. In particular, the short exact sequence above splits, and
we have Pic(RŒM|) Š Pic(RŒM
t
|) ˚R(M
t
\ M).
3
This sequence is simply obtained by combination of the lowest horizontal rows of the corre-
sponding diagrams (8.12) and (9.2).
8.I. Generalizations 215
Note that sn(M) is reached by a finite number of extensions as above (Exercise
2.13). Thus, by accumulation along the chain of extensions, we obtain
Pic(RŒM|) = Pic
_
RŒsn(M)|
_
˚
_

sn(M)\M
R
_
.
But we have already shown that Pic(RŒsn(M)|) = Pic(R).
Remark 8.45. (a) One can drop the condition that M be finitely generated. In fact,
if M is a seminormal monoid without nontrivial units then it is a filtered union
of its affine positive seminormal submonoids (see the beginning of the proof of
Lemma 8.17.) So one can use the fact that the functor Pic : Rings ÷ AbGroups
commutes with filtered direct limits (Exercise 8.2).
(b) Using a different technique, Singh and Roberts [72] obtained a very gen-
eral result, containing Theorem 8.42 as a special case. Our proof is “monoid ring
friendly”, so to speak.
(c) The theorem of Singh and Roberts holds for Q-algebras, and not even The-
orem 8.17 can be generalized to positive characteristics; see Dayton [29, 3.9] for a
counterexample. However, Singh [77] has found a modification that is valid in all
characteristics.
(d) Theorem 8.17 hints at the existence of a natural R-module structure on
Pic(RŒM|), Pic(R) under more general conditions. It does indeed exist; see [29,
72]. The source of such R-module structures for K-theoretical functors is a natural
action of the ring of big Witt vectors over R, containing the ghost copy of R. See
Section 10.C for details. They will play an important rôle in Section 10.C.
(e) The situation is much more complicated if M is not positive since the group
of units grows in the passage fromRto RŒM|. The Picard group of RŒX
˙1
| where
R is an arbitrary commutative ring has been computed by Weibel [100]. For a
seminormal monoid M and an integral domain of coefficients the computation if
Pic(RŒM|) can be reduced to Weibel’s result: see Anderson [1].
Monoid rings over Dedekind domains. The local-global principle (Theorem 8.11)
and the arguments in the proof of Lemma 8.17 readily imply the following exten-
sion of Theorem8.4: for a Dedekind domain Rand a seminormal monoid M every
projective RŒM|-module is extended from R. Since every projective R-module is
of the form free ˚rank 1, one has SK
0
(R) = 0.
By Theorem 8.37(b) we know that SK
0
(RŒM|) = 0 when M is a monoid and
R is a Dedekind domain. In other words, projective RŒM|-modules stably have
the form free ˚rank one for such R and M. Motivated by this observation, Swan
proved [92, Theorem 1.5] the following stronger nonstable result, which answers a
question of Murthy in the positive:
Theorem 8.46. For a Dedekind domain Rand a monoid M every projective RŒM|-
module is of the form free ˚rank one.
Swan’s proof is based on Theorem 8.4 and a result on projective modules of
general interest. Since the argument is beyond the scope of interaction of discrete
216 8. Projective modules over monoid rings
geometry and K-theory, we confine ourselves to stating the theorem. (See page
p. 141) for the definition of subintegral extensions.)
Theorem 8.47 (Swan). Let A Ç B be a subintegral extension of rings. Let P be a
projective A-module. Then:
(a) If B ˝P Š Q
1
˚ ˚Q
n
, then P Š P
1
˚ ˚P
n
with Q
i
Š B ˝P
i
.
(b) If B ˝P has the form free ˚rank one, then the same is true of P.
Theorem 8.46 follows from Theorem 8.47. In fact, if M is a monoid with semi-
normalization M
t
, then for any ring Rthe monoid ring extension RŒM| Ç RŒM
t
|
is subintegral while, as already observed, projective RŒM
t
|-modules are extended
fromR, provided Ris a Dedekind domain.
Monomial Quotients. The following theorem of Vorst [97] extends the Quillen-
Suslin theorem to monomial quotients of polynomial algebras.
Theorem8.48. Let Rbe a ring such that the projective modules over the polynomial
ring RŒX
1
. . . . . X
n
| are free. Let I Ç RŒX
1
. . . . . X
n
| an ideal generated by monomi-
als. Then the projective RŒX
1
. . . . . X
n
|,I-modules are also free.
The R-algebras mentioned in Theorem 8.48 have sometimes been called dis-
crete Hodge algebras [27]. Now we show that this theorem can be generalized to
arbitrary monoid rings:
Theorem 8.49. Let R be a ring and M be an affine monoid such that the projective
RŒM|-modules are extended from R. Then the projective RŒM|,RI-modules are
extended from Rfor every proper ideal I Ç M.
Remark 8.50. (a) Theorem 8.48 extends to arbitrary monoids of finite rank [42,
§3.2]. But here we prefer to avoid the involved technical subtleties. In [92] Swan
presents the following version which strengthens both the hypothesis and the con-
clusion: if the projective RŒM|-modules are extended from R for all seminormal
monoids M then the same is true for the projective RŒM|,RI-modules. These
results are based on the same idea, used also in the proof of Lemma 8.18.
(b) Although Theorem 8.49 and its proof do not refer to seminormality, it fol-
lows from Corollary 4.66 that M has to be seminormal.
We need the following
Lemma 8.51. Let Rbe a ring and
A A
1
(
A
2 A
t
be a cartesian diagramof R-algebras. Suppose ¡ is a split surjective homomorphism
of R-algebras and R ÷ A
t
is a split injective homomorphism of R-algebras. Let P
be a projective A-module such that A
1
˝
A
P and A
2
˝
A
P are extended from R.
Then P is extended fromR.
Exercises 217
Proof. Let A
i
˝
A
P Š P
i
Š A
i
˝
R
Q
i
for some projective R-modules Q
i
, i =
1. 2. Since the composite R ÷ A
i
÷ A
t
÷ R is 1
R
and A
t
˝ P
1
Š A
t
˝ P
2
we have A
t
˝
R
Q
1
Š A
t
˝
R
Q
2
. Therefore we can assume Q
1
= Q
2
= Q. Let
Q
t
= A
t
˝
R
Q. Then P is obtained by Milnor patching of P
1
and P
2
along some
automorphism˛ of Q
t
(Section 8.C), while A˝
R
Qis obtained by patching P
1
and
P
2
along 1
Q
0 . Since the splitting map A
t
÷ A
1
is an R-algebra homomorphism,
P
1
Š A
1
Š
A
0 Q
t
and we can lift ˛ to an automorphism 1 ˝˛ of A
1
˝
A
0 Q
t
. This
shows that the two patching diagrams are isomorphic and so P Š A˝
R
Q.
Proof of Theorem 8.49. By the Nakayama Lemma, projective modules are iso-
morphic if they are so modulo an ideal contained in the Jacobson radical. Since the
kernel of the natural homomorphismRŒN|,RI ÷RŒN|,R
_
I is nilpotent, there
is no loss of generality in assuming that I is a radical ideal. Then I = P
1
¨ ¨P
n
is the intersection of monomial prime ideals; see Proposition 2.19. We will use
induction on n.
The case n = 1 we have I = M \ F where F is a face of R
÷
M (Proposition
2.19). Therefore the embedding N = M ¨ F ÷ M induces an isomorphism
RŒN| Š RŒM|,RI. On the other hand RŒN| is an R-retract of RŒM| (Corollary
4.29). So if P is a projective RŒN|-module then we have
P Š RŒN| ˝
RŒMj
RŒM| ˝
RŒNj
P Š RŒN| ˝RŒM| ˝
R
P
0
Š RŒN| ˝
R
P
0
where P
0
= R˝
RŒMj
P with respect to the augmentation RŒM| ÷Rdetermined
by M \ U(M) ÷0, U(M) ÷1.
Now assume n > 1. Consider the ideals P = P
1
and J = P
2
¨ ¨ P
n
. Then
I = P ¨ J and L = P L J is a proper ideal of M since P. J Ç M \ U(M). The
diagram
RŒM|,RI RŒM|,RJ
(
RŒM|,RP RŒM|,RL
is cartesian. Moreover, the homomorphism¡ is a split R-algebra homomorphism.
In fact, let K = M \ P and S = K ¨ L = K ¨ J. Since M = K L L,
RŒK|,RS Š RŒM|,RL. Now the inclusion K ÷ M takes S into J and induces
a map RŒK|,RS ÷RŒM|,RJ splitting ¡ . So Lemma 8.51 applies.
Exercises
8.1. Prove Lemma 8.39.
8.2. Let I be a filtered partially ordered set and let {¡
ij
: R
i
÷ R
j
: i. j ÷ I. i < j ] be
a family of ring homomorphisms such that ¡
jk
ı ¡
ij
= ¡
ik
whenever i < j < k. Then
Pic(lim
÷÷I
R
i
) = lim
÷÷I
Pic(R
i
).
8.3. This problem is referred to in the next exercise.
Call a pair of functors (F. G) admissible if F. G : Rings ÷ AbGroups are covariant func-
tors, satisfying the conditions:
218 8. Projective modules over monoid rings
(i) F commutes with the direct limits of filtered diagrams,
(ii) for a Milnor square of rings
A B
C D
the associated sequence F(A) ÷F(B) ˚F(C) ÷F(D) is exact,
(iii) for a subintegral ring extension A Ç B the homomorphism F(A) ÷ F(B) is injec-
tive,
(iv) there exists a natural transformation of functoris F ÷ G such that for any ring R the
homomrophismF(R) ÷G(R) is injective.
(v) for a ring extension A Ç B, making B a free A module of finite rank n, there exists a
homomorphism T
BA
: G(B) ÷ G(A) such that the composite map G(A) ÷ G(B)
T
BA
÷÷÷÷
G(A) is multiplication by n.
(Functors F satisfying the conditions (i) and (ii) are correspondingly called continuous
and semiexact.)
Show that the pair (SK
0
. K
0
) is an admissible pair of functors.
Outline of the proof: for an arbitrary ring R show the following claims and then use them
together with Theorem 8.47.
(1) The homomorphism rank : K
0
(R) ÷ H
0
(R) has a right inverse ŒR
-
| : H
0
(R) ÷
K
0
(R), defined as follows. Any element o ÷ H
0
(R) is a difference o
1
÷ o
2
for some
o
1
. o
2
÷ H
0
(R) with nonnegative values. Put
ŒR
o
| = ŒR
o
1
| ÷ ŒR
o
2
|. ¡ ÷ H
0
(R)
where the module R
o
1
÷ P(R) is determined by the condition that if o
1
has a constant
value r on Spec(eR) Ç Spec(R), e ÷ R an idempotent, then e (R
o
1
) = (eR)
r
, and
similarly for R
o
2
.
(2) Any element of
¯
K
0
(R) is of the formŒP| ÷ŒQ| ÷ŒR
rank(P)
| ÷ŒR
rank(Q)
| for some P. Q ÷
K
0
(R). Moreover, ŒP| ÷ŒQ| ÷ŒR
rank(P)
| ÷ŒR
rank(Q)
| = 0 ÷
¯
K
0
(R) iff there exists a module
L ÷ P(R) for which P ˚Land Q˚Lare both free R-modules.
(3) If modules P. Q. L ÷ P(¹) are such that P ˚L and Q˚L are both of the form free
˚rank one then
ŒP| ÷ ŒQ| ÷ ŒR
rank(P)
| ÷ŒR
rank(Q)
| ÷ SK
0
(R) ==
ŒP| ÷ ŒQ| ÷ ŒR
rank(P)
| ÷ŒR
rank(Q)
| = 0.
8.4. Let (F. G) be an admissible pair of functoris in the sense of Exercise 8.3. For any
natural number n and any affine simplicial monoid M Ç Z
n
show the implication:
F(R) = F(RŒX
1
. . . . . X
n
|) ==F(R) = F(RŒM|).
(This gives an almost categorial proof of Theorem 8.37(b).)
Outline of the proof:
(1) By condition (iii) in Exercise 8.3 without loss of generality we can assume that M is
seminormal.
(2) Using Theorem 2.67 and Proposition 2.37, show that there exists a free basis B =
{x
1
. . . . . x
n
], n = rank M, of gp(M) such that B Ç int(M).
Exercises 219
(3) Let c ¸ 2 be a natural number and consider the affine positive submonoid
M
t
= M ÷
n

i=1
Z
÷
x
i
c
Ç Q˝gp(M).
Show that RŒM
t
| is a free RŒM| module of rank c
n
, with a basis (using multiplicative
notation):
{x
a
1
{c
1
. . . x
a
n
{c
n
[ 0 ¸ a
1
. . . . . a
n
¸ c ÷ 1].
(4) Using the conditions (iv, v), show that the group Ker
_
F(RŒM|) ÷ F(RŒM
t
|)
_
is of
c
n
-torsion.
(5) Using again the condition (iii) in Exercise 8.3, show Ker
_
F(RŒM|) ÷ F(RŒsn(M
t
)|)
_
is also of c
n
-torsion.
(6) Let ZŒ1,c|M
+
be the c-divisible hull of M
+
, i. e. ZŒ1,c|M
+
= {m,c
a
[ m ÷ M
+
. a ÷
Z
÷
] (writing additively). Show that sn(M
t
) = M ÷ZŒ1,c|M
+
. In particular we have the
following Milnor square of R-algebras:
RŒZŒ1,c|M
+
| RŒsn(M
t
)|
R RŒsn(M
t
)|,R
_
ZŒ1,c| int(M)
_
(7) Show that ZŒ1,c|M
+
is a filtered union of rank n free monoids.
(8) Using the continuity and semiexactness of F (conditions (i, ii) in Exercise 8.3), show
the implication:
F(R) = RŒsn(M
t
)|,R
_
ZŒ1,c| int(M)
_
==F(RŒM|),F(R) is of c
n
-torsion.
(9) Apply an inductive argument, similar to the one in the proof of Lemma 8.18, involving
the monoids ZŒ1,c|(M ¨ F)
+
where F runs through the faces of R
÷
M, and arrive at the
conclusion that F(RŒM|),F(R) is of c
n
-torsion.
(10) Complete the proof by consideration of another natural number c
t
¸ 2 which is co-
prime with c.
8.5. Let M Ç Q
n
be an integrally closed positive submonoid such that the convex subset
R
÷
M Ç R
n
is a cone. Let F : Rings ÷AbGroups be a continuous semiexact functor. For
any ring Rsuch that F(R) = F(RŒX
1
. . . . . X
n
|), n = rank M, show the equivalence
F(RŒM|),F(R) is a finitely generated group ”F(R) = F(RŒM|).
Outline of the proof:
(1) Fix rational hyperplane H Ç R
n
such that R
÷
M is spanned over O by the intersection
˚(M) = R
÷
M ¨ H. Using the induction on dimension and the Milnor squares similar
to those in the proof of Lemma 8.18, show that the identity embedding RŒR
÷
int(˚(M)) ¨
Q
n
| ÷ RŒM| induces the surjective homomorphism F
_
RŒR
÷
int(˚(M)) ¨ Q
n
|
_
÷
F(RŒM|). Informally, this can be phrased as ‘all elements of F(RŒM|) are extended from
the interior of the rational polytope ˚(M)’.
(2) Fix any rational point z ÷ int ˚(M). For a rational point 0 < z < 1 let ˚(M)
z
denote
the homothetic image of ˚(M) with factor z and centered at z. Showthat if z is sufficiently
close to 1 then identity embedding RŒR
÷
˚(M)
z
¨ Q
n
| ÷ RŒM| induces the surjective
220 8. Projective modules over monoid rings
homomorphism F
_
RŒR
÷
˚(M)
z
¨ Q
n
|
_
÷ F(RŒM|). On the other hand the monoids
M and R
÷
˚(M)
z
¨ Q
n
are isomorphic.
(3) Iterate the process and use the idea of a “sandwiched simplex” from the proof of Theo-
rem 8.4.
8.6. Let M Ç Q
3
be a integrally closed positive submonoid such that R
÷
M is a cone with
four extremal rays. Let H Ç R
3
be a rational hyperplane such that R
÷
M is spanned over
O by the intersection ˚(M) = R
÷
M ¨ H. Denote by A. B. C. D the mid-points of the
edges of the quadrangle ˚(M) = R
÷
M ¨ H.
Show the following implication for for any functor F : Rings ÷ AbGroups and any ring
Rthe following implication holds:
F(R) = F(RŒX
1
. X
2
. X
3
. X
4
|) ==
Im
_
F(RŒR
÷
conv(A. B. C. D) ¨ Q
3
|) ÷F(RŒM|)
_
= F(R).
Hint: think of the embedding of quadrangles conv(A. B. C. D) ÷ ˚(M) as a parallel
projection of a suitable high dimensional picture.
8.7. Let R be a ring and ˘ = {p. ¬
p
: C
p
÷ p] be a conical complex, i. e. a polyhedral
complex (see Definition 1.40) ˘ where all polyhedral P
p
are cones C
p
Ç R
n
p
.
Assume M
˘
= {M
p
Ç C
p
¨ Z
n
p
] is a system of additive submonoids, satisfying the
conditions:
(i) R
÷
M
p
= C
p
(full dimensionality),
(ii) for all p. q ÷ ˘ the map ¬
-1
q
ı ¬
p
restricts to a monoid isomorphism between M
p
¨
¬
-1
p
(p ¨ q) and M
q
¨ ¬
-1
q
(p ¨ q) (compatibility).
Such a systemM
˘
will be called monoidal complex, supported by ˘.
For a monoidal complex M
˘
one defines the ring RŒM
˘
| as the inverse limit of the R-
algebras RŒM
p
|:
lim
÷÷
_
¡
pq
: RŒM
p
| ÷RŒM
píq
| [ ¡
pq
the face projection
_
.
This construction extends the notion of a polyhedral algebra as well as Stanley’s ‘face’ rings
associated to fans, both discussed in Section 4.C.
Let R be a ring, ˘ a conical complex, and M
˘
a monoidal complex supported by ˘. The
set of monomials in RŒM
˘
| is defined as follows:
[M
˘
[ = lim
_
¬
-1
q
ı ¬
p
: M
p
÷M
q
[ p Ç q a face
_
.
Show that the results of Chapter 8.4 extend to RŒM
˘
| as follows:
(a) If R is a PID and the monoids M
p
are seminormal for all p ÷ ˘, then all projective
RŒM
˘
|-modules are free.
(b) If Ris regular, then SK
0
(R) = SK
0
(RŒM
˘
|).
(c) The submonoids sn(M
p
) Ç Z
n
p
, p ÷ ˘, form a monoidal complex, which we denote
by sn(M)
˘
. On the other hand, the submonoids
¨
M
p
Ç Z
n
p
, p ÷ ˘, may fail to form a
monoidal complex.
(d) If R is reduced seminormal and all monoids M
p
, p ÷ ˘, are seminormal, then
Pic(R) = Pic(RŒM
˘
|).
Exercises 221
(e) If Ris reduced seminormal, then
Pic(RŒM
˘
|), Pic(R) Š Pic(R) ˚
_

[ sn(M
˘
)[\[M
˘
[
R
_
.
Hint: in (a), (b), and (d) use the corresponding statement in Chapter 8 in combination with
Milnor squares of the form:
Rint(M
p
) RŒM
˘
|
R
RŒM
˘
|,Rint(M
p
)
where p ÷ ˘ are maximal elements with respect to inclusion: after successive applications
of these squares, where p runs over the maximal elemengts of ˘, the claim in question
reduces to the corresponding claim for the monoidal subcomplex of M
˘
, supported by
the codimension 1 skeleton of ˘. In particular, the induction on max(dim(C
p
) : p ÷ ˘)
applies.
Chapter 9
Bass-Whitehead groups of monoid rings
9.A. The functors K
1
and K
2
The functor K
1
. For a ring R the general linear group of order n is the group
GL
n
(R) of invertible n × n-matrices over R. In particular, GL
1
(R) = U(R). The
stable general linear group GL(R) is the inductive limit of the diagram
GL
1
(R) ÷ ÷GL
n
(R) ÷GL
n÷1
(R) ÷ . + ÷
_
+ 0
0 1
_
.
The elements of GL(R) can be thought of as invertible infinite square matrices with
only finitely many entries of the main diagonal = 1 and only finitely many non-
diagonal entries = 0. The nonstable and stable special linear groups SL
n
(R) and
SL(R) are defined similarly, using only matrices of determinant 1. The determi-
nant maps det : GL
n
(R) ÷ U(R) are compatible with the diagram above, and we
get a map GL(R) ÷U(R), denoted again by det. It is a surjective homomorphism,
split by the embedding U(R) = GL
1
(R) ÷GL(R).
The subgroup of elementary matrices E
n
(R) Ç SL
n
(R) is generated by the
standard elementary matrices, i. e. matrices of the form e
a
ij
where a ÷ R is the
entry on the position (i. j ), i = j , the main diagonal entries are equal to 1, and
all other entries are zero. The embeddings GL
n
(R) ÷ GL
n÷1
(R) respect the sub-
groups of elementary matrices and we get the stable subgroup of elementary matri-
ces E(R) Ç SL(R). The easily checked relation
_
e
a
ij
. e
b
kl
_
= e
ab
il
for j = k, i = l ,
and the Whitehead Lemma [4, Ch. V, Prop. 1.7] imply
E(R) = ŒE(R). E(R)| = ŒGL(R). GL(R)| (9.1)
where ŒG. G| denotes the commutator subgroup of a group G. In particular, E(R)
is a normal subgroup of GL(R) and GL(R), E(R) is the abelianization of GL(R).
Since the determinant of an elementary matrix is 1, we have the induced homo-
morphism det : GL(R), E(R) ÷U(R).
The Bass functor K
1
: Rings ÷ AbGroups is defined by R ÷ GL(R), E(R)
(and in the obvious way for ring homomorphisms). The group K
1
(R) is also called
the Bass-Whitehead group of R.
223
224 9. Bass-Whitehead groups of monoid rings
The determinant homomorphism det : GL(R) ÷ U(R) gives rise to the func-
torial splitting
K
1
(R) = U(R) ˚SK
1
(R)
where SK
1
(R) = SL(R), E(R). This defines another functor SK
1
: Rings ÷
AbGroups.
Consider a Milnor square of rings
A B
(
C D.
The stable groups and homomorphisms introduced above assemble in a natural
way into the following commutative diagram with exact rows and columns, called
the (K
1
-) Mayer-Vietoris sequence [4, Ch. IX, Cor. 5.12]:
0 0 0
SK
1
(A) SK
1
(B) ˚SK
1
(C) SK
1
(D)
K
1
(A)
det
K
1
(B) ˚K
1
(C)
det
K
1
(D)
det
0
U(A) U(B) ˚U(C) U(D)
0 0 0.
(9.2)
Moreover, there is a connecting homomorphismd : K
1
(D) ÷
¯
K
0
(A) that connects
the diagrams (8.12) and (9.2) above, yielding the K
1
-K
0
-Mayer-Vietoris sequence
(of size 3 × 6) whose middle row is
K
1
(A) ÷K
1
(B) ˚K
1
(C) ÷K
1
(D) ÷
¯
K
0
(A) ÷
¯
K
0
(B) ˚
¯
K
0
(C) ÷
¯
K
0
(D).
One defines d as follows: for an element z ÷ K
1
(D) pick a representative ˛ ÷
GL
n
(D), n ; 0, and apply Milnor patching to the following diagram of free mod-
ules
B
n
C
n
D
n
Š
˛
D
n
to obtain a projective A-module P; then set d(z) = ŒP| ÷ ŒA
n
| ÷
¯
K
0
(A). That d
is well defined and possesses the properties stated above is easy to verify. For the
details see [4, Ch. IX, §5].
For a Milnor square as above we also have the exact sequence
K
1
(A) ÷K
1
(B) ˚K
1
(C) ÷K
1
(D) ÷K
0
(A) ÷K
0
(B) ˚K
0
(C) ÷K
0
(D)
9.A. The functors K
1
and K
2
225
whose connecting map K
1
(D) ÷ K
0
(A) is the composite of d and the identity
embedding
¯
K
0
(A) ÷K
0
(A).
Here is an SK
1
-analogue of Lemma 8.35.
Lemma 9.1. For any ideal I Ç R, contained in the Jacobson radical J(R), we have
SK
1
(R) = SK
1
(R,I).
Proof. A matrix ˛ ÷ SL(R) belongs to E(R) whenever ˛ ÷ 1 mod I – a con-
sequence of the standard Gauss-Jordan reduction. This shows that the preimage
of E(R,I) in SL(R) is E(R). In other words, the homomorphism SK
1
(R) ÷
SK
1
(R,I) is a monomorphism. But it is also an epimorphism because, firstly, any
element ˇ ÷ SL(R,I) lifts to some element ; ÷ GL(R) – a consequence of the fact
that the preimage of U(R,I) in Rcoincides with U(R), and, secondly, dividing the
first row of ; by det(;) we get a preimage of ˇ in SL(R).
The functor K
2
. Let R be a ring and a. b ÷ R. Then we have the following
Steinberg relations:
e
a
ij
e
b
ij
= e
a÷b
ij
. i = j .
and
_
e
a
ij
. e
b
kl
_
=
_
e
ab
il
if j = k. i = l.
1 if j = k and i = l.
The Steinberg group St(R) is defined as the group generated by ‘symbols’ x
a
ij
, i. j ÷
N, i = j , a ÷ R, subject to the relations
x
a
ij
x
b
ij
= x
a÷b
ij
. i = j .
and
_
x
a
ij
. x
b
kl
_
=
_
x
ab
il
if j = k. i = l.
1 if j = k and i = l.
Recall that a surjective group homomorphism¡ : H ÷G is called a universal
central extension of G if it satisfies the following conditions:
(i) Ker(¡ ) is contained in the center of H;
(ii) for every surjective homomorphism¡
t
: H
t
÷G with Ker(¡
t
) contained
in the center of H
t
, there exists a unique homomorphism h : H ÷ H
t
such that ¡
t
h = ¡ .
Theorem 9.2 (Milnor [65, §5]). Let R be a ring. The surjective group homomor-
phism [ : St(R) ÷ E(R), [
_
x
a
ij
_
= e
a
ij
, is a universal central extension of E(R).
Moreover, Ker([) coincides with the center of St(R).
The Milnor group K
2
(R) is defined to be Ker([), with notation as in Theorem
9.2. Clearly, we actually get a functor K
2
: Rings ÷AbGroups.
226 9. Bass-Whitehead groups of monoid rings
Theorem 9.3 (Milnor [65, §6]). Let
A B
C D
be a cartesian square of rings in which all homomorphisms are surjective. Then
there is a natural connecting homomorphism K
2
(D) ÷ K
1
(A) that gives rise to
the exact sequence
K
2
(A) ÷K
2
(B) ˚K
2
(C) ÷K
2
(D) ÷
÷K
1
(A) ÷K
1
(B) ˚K
1
(C) ÷K
1
(D) ÷
÷K
0
(A) ÷K
0
(B) ˚K
0
(C) ÷K
0
(D).
Relative groups. Let I be an ideal in a ring A. Then we have the cartesian square
of rings

A{I
A
t
1
t
2
A
A
A,I.
Following Milnor [65], we define the relative K-groups by
K
i
(A. I) = Ker
_

2
)
+
: K
i
(A×
A{I
A) ÷K
i
(A)
_
. i = 0. 1. 2.
The 9-term Mayer-Vietoris exact sequence from Theorem 9.3 then yields the fol-
lowing exact sequence
K
2
(A. I)
(t
1
)

÷÷÷÷K
2
(A) ÷÷÷÷K
2
(A,I) ÷÷÷÷
÷÷÷÷K
1
(A. I)
(t
1
)

÷÷÷÷K
1
(A) ÷÷÷÷K
1
(A,I) ÷÷÷÷ (9.3)
÷÷÷÷K
0
(A. I)
(t
1
)

÷÷÷÷K
0
(A) ÷÷÷÷K
0
(A,I).
The relative group K
1
(A. I) can alternatively be defined by
K
1
(A. I) = GL(A. I), E(A. I) (9.4)
where GL(A. I) Ç GL(A) is the subgroup of the matrices congruent to the identity
matrix modulo I, and E(A. I) Ç E(A) is the minimal normal subgroup containing
the matrices of the form e
a
ij
, i = j , a ÷ I. In this way the relative K
1
-group is
defined in [4, Ch. V, §2].
We also have the relative groups
SL(A. I) = SL(A) ¨ GL(A. I). SK
1
(A. I) = SL(A. I), E(A. I).
9.A. The functors K
1
and K
2
227
Remark 9.4. (a) It is known that K
0
(A. I) does not depend on the ambient ring,
i. e. K
0
satisfies has the excision property: for a commutative diagram of the form
I A
(
J B
where ¡ : A ÷ B is a ring homomorphism mapping the ideal I Ç A bijectively
onto the ideal J Ç B, the resulting homomorphism K
0
(A. I) ÷ K
0
(B. J) is an
isomorphism [4, Ch. IX, Ex. 6.5].
(b) Swan has shown [88] that, firstly, already K
1
fails to satisfy excision and,
secondly, there is no functor K
3
: Rings ÷ AbGroups that would extend the 9-
term exact sequence in Theorem 9.3 to the corresponding 12-term exact sequence.
The nontrivial elements in the relative groups constructed by Swan will play an
important rôle later in this chapter.
(c) The higher relative groups K
i
(A. I) are defined as the homotopy groups
of the homotopy fiber of the map BGL(A)
÷
÷ BGL(A,I)
÷
between certain K-
theoretical spaces, whose homotopy groups, introduced by Quillen [69] are the
higher K-groups of Aand A,I.
Since higher K-groups do not satisfy excision, one could ask for which nonuni-
tal rings I the relative groups K
i
(A. I) are independent of the ring A containing
I as an ideal.This so-called excision problem was solved by Suslin and Wodzicki
[86]. As already remarked, the main result of this chapter shows that a direct K
1
-
analogue of Theorem 8.4 is impossible. However, a weaker version of Theorem 8.4
is proved for all higher groups in [46] (see Section 9.C), and the results of [86] are
among the key ingredients of the techniques used there.
The fundamental theorem. The K
1
-analogue of Theorem 8.36 was established in
[5]: K
1
(R) = K
1
(RŒX|) for all regular rings R. This was extended to all higher
K-groups at once by Quillen in his fundamental work [70]. This is the so-called
K-homotopy invariance of regular rings. Below we state the general result, known
as the Fundamental Theorem of K-theory, containing the homotopy invariance of
regular rings.
Let R be a ring (not necessarily regular) and F : Rings ÷ AbGroups be a
functor. Then we have the group homomorphisms
F(R) F(RŒX|)
F(RŒX
-1
|) F(RŒX. X
-1
|)
induced by the identity ring embeddings. The group F(R) functorially splits off
from the other three groups in this diagram, and we can consider the decomposi-
tion F(RŒX|) = F(R) ˚NF(R).
Theorem 9.5 (Fundamental Theorem of K-theory). Let Rbe a ring and i ÷ Z.
228 9. Bass-Whitehead groups of monoid rings
(a) The homomorphisms K
i
(RŒX|) ÷ K
i
(RŒX. X
-1
|) and K
i
(RŒX
-1
|) ÷
K
i
(RŒX. X
-1
|) are split monomorphisms. Moreover, we have the functorial
splitting
K
i
(RŒX. X
-1
|) = K
i
(R) ˚NK
÷
i
(R) ˚NK
-
i
(R) ˚K
i-1
(R)
where K
i
(R)˚NK
÷
i
(R) is the isomorphic image of K
i
(RŒX|) and K
i
(R)˚
NK
-
i
(R) is that of K
i
(RŒX
-1
|).
(b) If Ris regular, then NK
i
(R) = 0, and so K
i
(R) = K
i
(RŒX|).
Remark 9.6. (a) Theorem 9.5 is stated for all integer indices i , while the group
K
i
(A) were defined above only for i = 0. 1. 2. In this book we are not going to
work out higher K-groups in detail, although they will be mentioned occasionally.
The groups K
i
(A) are higher homotopy groups of certain huge CW-complexes
BQP(A), built up from the category of projective A-modules, see [70]. That the
alternative definition of higher K-groups, mentioned in Remark 9.4(c), coincides
with this one is also due to Quillen [38].
(b) Theorem 9.5 holds for schemes that are not necessarily affine.
(c) Theorem9.5(a) is actually the definition of the groups K
i
for i < 0. Namely,
one applies the following formula iteratively, starting with i = 0:
K
i-1
(R) = Coker
_
K
i
(RŒX|) ˚K
i
(RŒX
-1
|) ÷K
i
(RŒX. X
-1
|)
_
.
see [4, Ch. XII, §7].
(d) Since K
i-1
(R) is a natural direct summand of K
i
(RŒX. X
-1
|), the higher
K-groups can be considered as invariants of rings that are finer than the lower K-
groups. In particular, it is natural to expect that higher analogues of the results in
Chapter 8.4 are more difficult to obtain, and as we will see in this chapter, this is
indeed the case.
9.B. The nontriviality of SK
1
(RŒM|)
The mainresult. In Section 8.I the homotopy invariance of K
0
(Theorem8.36) has
been generalized as far as possible to a statement about monoid rings (Theorem
8.37). In view of Theorem 9.5 it is natural to expect that
K
i
(R) = K
i
(RŒM|). i ¸ 0.
for all regular rings R and all positive seminormal monoids M. It is clear that
we have to require the positivity of M: for nonpositive monoids the equation
K
i
(R) = K
i
(RŒM|) fails already for i = 1 because U(R) is a proper direct sum-
mand of U(RŒM|).
However, we will show below that the group K
i
(RŒM|),K
i
(R), i ¸ 1, dramat-
ically fail to be trivial for essentially all nonfree affine positive monoids. In other
words, a monoid generalization of the Fundamental Theorem 9.5(b) is impossible
for higher K-groups.
The main result of this section is
9.B. The nontriviality of SK
1
(RŒM|) 229
Theorem 9.7. Let R be a regular ring and M be an affine simplicial monoid M.
Then
(a) K
1
(RŒM|) = K
1
(RŒM|ŒX|) if and only if M is free;
(b) if M is seminormal, then K
1
(R) = K
1
(RŒM|) if and only if M is free;
(c) if ˝
1
R
= 0 then K
1
(R) = K
1
(RŒM|) if and only if M is free.
Remark 9.8. By Theorem 9.5(a) and Theorem 9.7, the affine simplicial monoid M
is free if K
i
(RŒM|) = K
i
(RŒM|ŒX|) for some i ¸ 1. The same holds if ˝
1
R
= 0
and K
i
(R) = K
i
(RŒM|).
Combinatorial preparations. We need some auxiliary results about affine mono-
ids. The first concerns rank 2 affine monoids.
Lemma 9.9. Let M be a nonfree affine positive normal monoid of rank 2. Then M
is isomorphic to a submonoid L Ç Z
2
÷
which is normal in Z
2
÷
and satisfies the
conditions (0. 1) ÷ L, (1. 1) ÷ L
+
, (1. 0) … L.
Proof. We identify gp(M) with Z
2
. Each of the two extreme integral generators of
M can be extended to a basis of Z
2
, and therefore we can assume that x = (1. 0) is
one of them. Moreover we can assume that M is contained in the upper halfplane.
Since M is nonfree, its second extreme integral generator (m. n) has n > 1, and,
moreover, mand n are coprime. We choose k ÷ Zsuch that
k <
m
n
< k ÷1
and set z = (k. 1). Then x. z is a basis of Z
2
. After a change of bases keeping x
fixed and sending z to (0. 1) we have reached the desired embedding.
The proof of the next lemma is similar to that of Lemma 3.32.
Lemma 9.10. Let M be an affine positive monoid and x be an extreme generator of
M. Then there exists an affine normal submonoid L Ç M of rank r ÷ 1 such that
L
+
Ç M
+
and Zx ÷M = Zx ÷L.
Proof. By Proposition 2.31 we have Zx ÷M Š Z × M
0
for some affine positive
normal monoid of rank r ÷ 1. Moreover, the isomorphism maps Zx to Z. Taking
the isomorphic image of M
0
in Zx ÷M we can assume that M
0
Ç Zx ÷M. By
Proposition 2.15(e) there is a free intermediate monoid F with M
0
Ç F Ç gp(M
0
)
such that gp(F) = gp(M
0
). Let F = Z
÷
x
1
÷. . . Z
÷
x
r-1
.
If c is a sufficiently large natural number then the monoid
L = M ¨
_
Z
÷
(x
1
÷cx) ÷ ÷Z
÷
(x
r-1
÷cx)
_
satisfies the desired condition. In fact, if c is big enough then elementary geometric
considerations show that
int(R
÷
M) ¨ int(R
÷
L) = int(R
÷
M) ¨ int
_

R
÷
(x
i
÷cx)
_
= 0.
This implies rank(L) = dimaff(L) = r ÷ 1.
230 9. Bass-Whitehead groups of monoid rings
On the other hand, since gp(F) = Zx ÷Z(x
1
÷cx) ÷ ÷Z(x
r-1
÷cx) =
gp(M) and M is normal, we also have that L is integrally closed in Z
÷
(x
1
÷
cx) ÷ ÷ Z
÷
(x
r-1
÷ cx). By Corollary 2.25 we have gp(L) = Z(x
1
÷ cx) ÷
÷Z(x
r-1
÷cx). In particular, gp(M) = Zx ÷gp(L).
Finally, the equality of the two normal affine monoids Zx ÷ M and Zx ÷ L
follows since their groups of differences coincide as well as the cones they span in
RM.
Finally we formulate a very simple criterion for the freeness of simplicial affine
monoids:
Lemma 9.11. Let M be a simplicial affine monoid and let : ÷ ˚(M) be a vertex.
Then M is free if the following conditions hold:
(i) the extreme submonoids M
t
Ç M of rank 2 with : ÷ ˚(M
t
) are free;
(ii) there exists a free submonoid F Ç M such that gp(F) = gp(M), : is a
vertex of ˚(F), and the cones spanned by the polytopes ˚(M) and ˚(F)
at their vertex : are the same.
Proof. Let t
1
. . . . . t
r
denote the extreme integral generators of M, labelled in such
a way that : ÷ R
÷
t
1
, and consider the submonoid L = Z
÷
t
1
÷ ÷ Z
÷
t
r
. We
claim that L =
¨
M. Then L = M since L Ç M.
Since ˚(L) = ˚(
¨
M) and L and
¨
M are normal, we only need to prove that
gp(L) = gp(M) (= gp(
¨
M)). Let F = Z
÷
s
1
÷ ÷Z
÷
s
r
with : ÷ R
÷
s
1
. Then
s
1
= t
1
. Moreover the hypotheses imply s
i
÷ Z
÷
t
1
÷ Z
÷
t
i
for i = 2. . . . . r. In
particular, gp(M) = gp(F) Ç gp(L).
Swan’s elements in K
2
(RŒX. Y |,(XY )). For a ring Aand elements u. : ÷ Asuch
that 1 ÷ u: ÷ U(A) Dennis and Stein [81] defined an element (u. :) ÷ K
2
(A).
In the special case when A = RŒX. Y |,(XY ) for some ring R and u and : are
respectively the residue classes of X and Y , the element (u. :) coincides with
_
x
u
12
. x
¡
21
_
÷ K
2
(A), used by Swan in [88]. Here the x
+
ij
denote the standard gen-
erators of the Steinberg group (Section 9.A). The inclusion
_
x
u
12
. x
¡
21
_
÷ K
2
(A)
follows from the equation
_
e
u
12
. e
¡
21
_
= 1 and Theorem 9.2.
Lemma 9.12. Let A, u and : be as above. Then:
(a) (u. 0) = (0. :) = 0;
(b) (u
a
. :
b
) = 0 for all natural numbers a and b with max(a. b) ¸ 2.
The claims (b) and (c) follow immediately from the general relations between
the Dennis-Stein elements [81].
The following theorem is exactly Corollary 4.8 in [31], stated there for a bigger
class of coefficient rings R. It answers a question of Swan [88].
Theorem 9.13. Let R be a regular ring and RŒu. :| = RŒX. Y |,(XY ) as above.
Then K
2
(RŒu. :|) = K
2
(R) ˚(u. :)R where (u. :)R is the rank 1 free R-module
over the basis {(u. :)].
9.B. The nontriviality of SK
1
(RŒM|) 231
Remark 9.14. The R-module structure on K
2
(RŒu. :|),K
2
(R) Š (u. :)R is de-
fined by r (u. :) = (ru. :). More generally, for any graded ring A = A
0
˚A
1
˚
and all indices i the relative groups K
i
(A),K
i
(A
0
) are modules over the Witt vec-
tors W(A
0
). The module structures are given by the Bloch-Stienstra-Weibel opera-
tions on nil-K-theory [8, 82, 99], see Section 10.C.
As usual, we make the identification RŒZ
2
÷
| = RŒX. Y |. Let RŒu. :| be as above
and consider the group homomorphisms
St(RŒX. Y |)
)
E(RŒX. Y |)
St(RŒu. :|).
(9.5)
The key technical fact used in the proof of Theorem 9.7 is contained in
Proposition 9.15. For every convex neighborhood W of the point ˚(XY ) ÷ ˚(Z
2
÷
)
there exists a preimage z ÷ St(RŒX. Y |) of (u. :) such that
[(z) ÷ E(RŒX. Y |) ¨ SL(RŒZ
2
÷
[W|)
and, simultaneously, [(z) reduces to the unit element of SL(R) modulo (XY ).
Proof. Consider the following system of preimages of (u. :) in St(RŒX. Y |):
z
1
= Œx
12
(X). x
21
(Y )| .
z
i÷1
=
_
x
12
(X
i÷1
Y
i
). x
21
(÷X
i
Y
i÷1
)
_
. i ¸ 1.
In this formulas we have used “functional” notation for Steinberg symbols instead
of the usual exponential one. Then
[(z
i
) =
_
_
¡
i
÷X
i÷1
Y
i
X
i
Y
i÷1
1 ÷ XY
_
_
÷ SL(RŒX. Y |).
where ¡
i
=

2i
j=0
(XY )
j
. Indeed,
[(z
1
) =
_
_
1 ÷XY ÷X
2
Y
2
÷X
2
Y
XY
2
1 ÷ XY
_
_
and
[(z
i
) =
_
_
¡
i-1
÷X
2i-1
Y
2i-1
÷X
2i
Y
2i
÷X
i÷1
Y
i
X
i
Y
i÷1
1 ÷ XY
_
_
. i ¸ 2.
so the induction process applies.
Let W
i
denote the subsegment
_
˚(X
i÷1
Y
i
). ˚(X
i
Y
i÷1
)
_
Ç ˚(Z
2
÷
). It is clear
that W
i
is a neighborhood of ˚(XY ) in ˚(Z
2
÷
), W
1
¸ W
2
¸ W
3
¸ , and
_
i
W
i
= ˚(XY ). Since [(z
i
) ÷ SL(RŒZ
2
÷
[W
i
|) and every [(z
i
) reduces to the
identity matrix modulo (XY ) we are done.
232 9. Bass-Whitehead groups of monoid rings
The case of rank 2 seminormal monoids. For seminormal rank 2 monoids we can
prove a stronger version of Theorem9.7. We need this stronger version in the proof
of Proposition 9.17.
Proposition 9.16. Let R be a regular ring and M is a nonfree affine positive semi-
normal monoid of rank 2. Then SK
1
(R) is a proper subgroup of the image of
SK
1
(RŒM
+
|) in SK
1
(RŒM|).
Proof. We use the notation introduced in Proposition 9.15.
(1) First we consider the case in which M is normal. By Lemma 9.9 we can
assume that M Ç Z
2
÷
, Y ÷ M, XY ÷ M
+
and X … M
+
. In this situation
˚(M
+
) is a convex neighborhood of ˚(XY ) in ˚(Z
2
+
). By Proposition 9.15 there
is a preimage z ÷ St(RŒX. Y |) of (u. :) ÷ K
2
(RŒu. :|) such that
[(z) ÷ E(RŒX. Y |) ¨ SL
_
RŒX. Y |. (XY )
_
¨ SL(RŒM
+
|).
Let ç denote the class of [(z) in SK
1
(RŒM|). We claim that ç … SK
1
(R).
Suppose ç ÷ SK
1
(R). Since [(z) reduces to the identity matrix modulo (XY )
we have Œ[(z)| = 0 in SK
1
(R), or equivalently [(z) ÷ E(RŒM|). Let s be a preim-
age of [(z) in St(RŒM|), s
1
be the image of s in St(RŒX. Y |), and s
2
be the im-
age of s in St(RŒu. :|). Since both s
1
and z map to [(z) ÷ E(RŒX. Y |) we have
s
1
z
-1
÷ K
2
(RŒX. Y |). By Theorem 9.5(b) K
2
(RŒX. Y |) = K
2
(R). Therefore,
s
2
(u. :)
-1
÷ K
2
(R) Ç K
2
(RŒu. :|). Modifying the choice of s by this element of
K
2
(R), we can assume that s
2
= (u. :)
-1
. On the other hand, the image of RŒM|
in RŒu. :| under the composite map RŒM| ÷ RŒX. Y | ÷ RŒu. :| is RŒ:| because
of our special embedding of M into Z
2
÷
. We get (u. :) ÷ St(RŒ:|) Ç St(RŒu. :|).
But then the commutative diagram
St(RŒu. :|)
u|-0
¡|-¡
St(RŒu. :|)
St(RŒ:|)
implies (u. :) = (0. :) = 0 (Lemma 9.12(a)) – in contradiction with Theorem9.13.
(2) Now we treat the case in which M is seminormal and its normalization
¨
M
is not free. By Corollary 2.25 and Proposition 2.37 we have (
¨
M)
+
= M
+
. Consider
the commutative diagram
SK
1
(RŒM|)
SK
1
(RŒ
¨
M|).
SK
1
(RŒM
+
|)
By the previous case there exists ç ÷ SK
1
(RŒM
+
|), whose image in SK
1
(RŒ
¨
M|)
does not belong to SK
1
(R). Then the diagram above implies that the image of ç in
SK
1
(RŒM|) does not belong to SK
1
(R).
(3) It remains to consider the case when M is seminormal and
¨
M = Z
÷
.
Then Proposition 2.37 implies that there exist natural numbers a and b satisfying
the following conditions:
9.B. The nontriviality of SK
1
(RŒM|) 233
(i) max(a. b) ¸ 2:
(ii) M = Z
2
÷
\ {X
i
: a i ] L {Y
j
: b j ].
Then, as in (1), we can find a preimage z ÷ St(RŒX. Y |) of (u. :) ÷ K
2
(RŒu. :|)
such that
[(z) ÷ E(RŒX. Y |) ¨ SL
_
RŒX. Y |. (XY )
_
¨ SL(RŒM
+
|).
Let ç denote the corresponding element of SK
1
(RŒM|). We claim that ç … SK
1
(R).
Assume to the contrary that ç ÷ SK
1
(R). Then, as in (1), z ÷ E(RŒM|). Let s
be a lifting of z in St(RŒM|). Consider the commutative diagram of rings
RŒM| RŒX. Y |
RŒu
a
. :
b
| RŒu. :|
where the vertical maps represent the reduction modulo int(Z
2
÷
)R. Let s
0
denote
the image of s in St(RŒu
a
. :
b
|), s
1
denote the image of s in St(RŒX. Y |), and s
2
de-
note the image of s
1
in St(RŒu. :|). Both s
1
and z map to z ÷ E(RŒM|). Therefore,
s
1
z
-1
÷ K
2
(RŒX. Y |) = K
2
(R). In particular, s
2
(u. :)
-1
÷ K
2
(R). Modifying s
as in (1), we can assume s
1
= z and s
2
= (u. :). On the other hand, s
2
is the image
of s
0
under the map St(RŒu
a
. :
b
|) ÷St(RŒu. :|) and, simultaneously, the image of
s
0
in E(RŒu
a
. :
b
|) is 1. Therefore, s
0
÷ K
2
(RŒu
a
. :
b
|). By Theorem 9.13, applied to
the ring RŒu
a
. :
b
| (Š RŒu. :|), we have s
0
= p ˚ r(u
a
. :
b
) for some p ÷ K
2
(R)
and r ÷ R. (Here we have used the notation (u
a
. :
b
) =
_
x
u
a
12
. x
¡
b
21
_
.) By Lemma
9.12(b) s
0
maps to p under the map K
2
(RŒu
a
. :
b
|) ÷K
2
(RŒu. :|). Thus we obtain
(u. :) ÷ K
2
(R), a contradiction to Theorem 9.13.
The case of free extremal submonoids.
Proposition9.17. Let Rbe regular ring and M be a nonfree, affine, simplicial, semi-
normal monoid. Assume that for every facet z Ç ˚(M) the extremal submonoid
M[z Ç M is free. Then SK
1
(R) = SK
1
(RŒM|).
We need an auxiliary result.
Lemma 9.18. Let R be a regular ring and M be an affine positive seminormal
monoid all of whose proper extremal submonoids are free. Then the following con-
ditions are equivalent:
(a) SK
1
(R) = SK
1
(RŒM|);
(b) SK
1
(R) = Im
_
SK
1
(RŒM
+
|) ÷SK
1
(RŒM|)
_
.
Proof. First we prove the equality
SK
1
(R) = SK
1
_
RŒM|,(Rint(M)
_
. (9.6)
The proof is based on the ideas in Section 8.D, making use of the Mayer-Vietoris
sequence (9.2).
234 9. Bass-Whitehead groups of monoid rings
As in the proof of Lemma 8.18, we let M
1
. . . . . M
n
be the extremal submonoids,
labelled in such a way that M
i
¸ M
j
implies i ¸ j . In particular, M
1
= M and
M
n
= 0. In the proof of Lemma 8.18 we have constructed Milnor squares
RŒ(M
i
)
+
| A
i-1
R A
i
. i = 1. 2. . . . . n. (9.7)
where A
0
= RŒM|, A
1
= RŒM|,(Rint(M)), and A
n-1
= R. Observe that in
these squares the map RŒ(M
i
)
+
| ÷ A
i-1
factors through the subring RŒM
i
| Ç
A
i-1
for every index i , 1 ¸ i ¸ n.
In our situation the rings RŒM
i
|, i = 2. . . . . n are polynomial rings. So
by the Bass-Heller-Swan theorem (Theorem 9.5 for i = 1) we have SK
1
(R) =
SK
1
(RŒM
i
|), i ¸ 2. In particular, the Mayer-Vietoris exact sequences, associated
to the squares (9.7), yield the embeddings
SK
1
(A
1
) ÷SK
1
(A
2
) ÷ ÷SK
1
(A
n-1
) = SK
1
(R).
But SK
1
(R) is naturally a subgroup of SK
1
(A
1
) and, hence, (9.6) follows.
Now we turn to the proof of the equivalence (a) ” (b). Clearly, we only
need to show the implication (a) == (b). Assume to the contrary that SK
1
(R) =
Im
_
SK
1
(RŒM
+
|) ÷SK
1
(RŒM|)
_
. Then the Mayer-Vietoris sequence, associated to
(9.7) for i = 1, shows that we have an embedding
SK
1
(RŒM|) = SK
1
(A
0
) ÷SK
1
(A
1
) = SK
1
_
RŒM|,(Rint(M))
_
.
In view of (9.6) this implies SK
1
(R) = SK
1
(RŒM|), contradicting (a).
Proof of Proposition 9.17. We use induction on r = rank(M). For r = 1 there
is nothing to prove since a rank 1 seminormal monoid is free. The case r = 2
follows from Proposition 9.16. So we can assume r ¸ 3 and that Proposition 9.17
has been proved for monoids of rank r ÷ 1. Assume to the contrary SK
1
(R) =
SK
1
(RŒM|).
(1) Let :
1
. . . . . :
r
be the vertices of ˚(M) and let t
i
denote the generator of the
monoid M[:
i
Š Z
÷
.
By Lemma 9.10 there exists a submonoid L Ç
¨
M satisfying the following con-
ditions (in multiplicative notation):
v Lis integrally closed in
¨
M;
v L
+
Ç
¨
M
+
= M
+
(Proposition 2.37);
v (
¨
M[:
1
)
-1
¨
M = gp(
¨
M(:
1
)) × L.
We have the sequence of four injective monoid homomorphisms followed by the
projection onto L:
L
+
÷M
+
÷M ÷
¨
M ÷gp(M(:
1
)) × L ÷L.
Notice that the composite map coincides with the embedding L
+
÷Land, more-
over, the image of M
+
in L is L
+
. In particular, there is the commutative diagram
9.B. The nontriviality of SK
1
(RŒM|) 235
of monoids
M
+
L
+
M L
(9.8)
whose vertical arrows are the identity embeddings.
(2) Let L
t
be the image of M in L. It is an affine simplicial monoid for which
˚(L
t
) = ˚(L). By (9.8) L
+
Ç L
t
.
Consider a facet ı Ç ˚(L). Let us showthat the extremal submonoid L
t
[ı Ç L
t
is free. In view of the inclusion L
+
Ç L
t
and Proposition 2.37, this would imply
that L
t
is seminormal.
Let z Ç ˚(M) be the uniquely determined facet that contains ı and :
1
. There
exists j , 2 ¸ j ¸ r, such that z = conv(:
i
: i = j ). By the condition on M the
monoid M[z is free. It is generated by {t
i
: i = j ]. But then L
t
[ı is generated
by the images of t
i
, i = 1. j . Since L
t
[ı is generated by rank L
t
[ı elements, the t
i
,
i = 1. j , must generate the monoid L
t
[ı freely.
Next we show that L
t
= L. Suppose L
t
= L. Then, in view of the equali-
ties ˚(L
t
) = ˚(L) and gp(L
t
) = gp(L), the monoid L
t
cannot be normal. In
particular, L
t
is not free. So L
t
satisfies all conditions in Proposition 9.17. By the
induction hypothesis on rank we have SK
1
(R) = SK
1
(RŒL
t
|). By Lemma 9.18 this
is equivalent to SK
1
(R) = Im(SK
1
(RŒM
+
|) ÷SK
1
(RŒM|)).
On the other hand, the commutative diagram
SK
1
(RŒL
+
|) SK
1
(RŒL
t
|)
SK
1
(RŒM|)
in conjunction with the assumption SK
1
(R) = SK
1
(RŒM|) implies that SK
1
(R) =
Im
_
SK
1
(RŒL
+
|) ÷SK
1
(RŒL
t
|)
_
. This contradiction shows that L
t
= Land, more-
over, Lis free.
(3) Here we prove that M is free, in contradiction with the hypothesis that M
is not free. Then our assumption SK
1
(R) = SK
1
(RŒM|) has proved to be wrong.
Let s
i
denote the generator of
¨
M[:
i
Š Z
÷
, i = 1. . . . . r. For every index i we
have t
i
= s
a
i
i
for some a
i
÷ N. First we show that a
i
= 1 for all i .
By (2) L is a free monoid. Let t
i
, 2 ¸ i ¸ r ÷ 1, denote the free generators of
L so that ˚(t
i
) ÷ conv
_
˚(t
1
). ˚(t
i
)
_
. The equality gp(M) = gp(
¨
M[:
1
) × gp(L)
and the freeness of L imply that s
1
and t
i
generate gp
_
¨
M[˚(s
1
. s
i
)
_
for every i =
2. . . . . r. In particular, there exist representations
s
i
= t
x
i
s
-y
i
1
. x
i
÷ N. y
i
÷ Z
÷
. i = 2. . . . . r.
The image of t
i
= s
a
i
i
in Lunder the map M ÷Lis t
a
i
x
i
for i ¸ 2, and it is 1 for
i = 1. By (2) we know that these images generate the free monoid L. In particular,
a
i
x
i
= 1, and so a
i
= 1 for i = 2. . . . . r, as claimed.
Applying all these arguments to any other vertex :
i
÷ ˚(M), different form:
1
we get t
1
= s
1
as well.
236 9. Bass-Whitehead groups of monoid rings
Now we are in the situation of Lemma 9.11 with respect to the monoid M, the
vertex :
1
÷ ˚(M), and the free submonoid of M generated by t
1
and L. Hence
the desired freeness of M.
The general case. Before proving Theorem 9.7 we explain the relationship with
the module of absolute differentials.
Lemma 9.19. For every commutative ring R there is a surjective homomorphism
K
2
_
RŒX|,(X
2
)
_
,K
2
(R) ÷˝
1
R
.
Proof. We only indicate the argument as it involves techniques beyond the scope
of our book.
It is shown in [94] that for a commutative ring A and its nilpotent ideal I the
relative group K
2
(A. I) has a presentation by Dennis-Stein symbols [81]. In the
special case A = RŒX|,(X
2
) and I = Rc, where c is the residue class of X modulo
(X
2
), if one factors out the extra relation (rc. sc) = 0, r. s ÷ R, the Dennis-
Stein presentation becomes precisely the definition of ˝
1
R
. So there is always a
surjection K
2
(RŒX|,(X
2
)),K
2
(R) ÷ ˝
1
R
whose kernel is generated by Dennis-
Stein symbols (rc. sc).
Remark 9.20. The map K
2
(RŒX|,(X
2
)),K
2
(R) ÷ ˝
1
R
is the first glimpse of the
Dennis trace maps from higher K-theory to Hochschild and eventually to cyclic
homology groups – Chern classes – [54, 61]. Deep facts from this theory (such
as Goodwillie’s theorem [37] and Cortinas’ proof of the KABI conjecture [28]) play
crucial rôle in [46] which presents a higher analogue of Theorem 8.4; see Section
9.C.
Lemma 9.21. Let R be a regular ring and X a variable. Then there is a surjective
homomorphism SK
1
(RŒX
2
. X
3
|), SK
1
(R) ÷˝
1
R
.
In the special case when R is a field this has been proved by Krusemeyer [56].
However, his argument implies the general case, as we now show.
Proof. Being regular, Ris a reduced ring. In particular, U(R) = U(RŒX|). It then
follows easily that K
1
_
RŒX|. (X
2
)
_
= SK
1
_
RŒX|. (X
2
)
_
. Thus by (9.3) we have the
exact sequence
K
2
(RŒX|) ÷K
2
(RŒX|,(X
2
)) ÷SK
1
(RŒX|. (X
2
)) ÷K
1
(RŒX|).
By Theorem 9.5(b) K
2
(R) = K
2
(RŒX|) and K
1
(R) = K
1
(RŒX|). On the other
hand SK
1
_
RŒX|. (X
2
)
_
÷K
1
(R) is the zero map. So we get
SK
1
_
RŒX|. (X
2
)
_
= K
2
_
RŒX|. (X
2
)
_
,K
2
(R).
Now we are done by Lemma 9.19 once we notice that SK
1
(RŒX
2
. X
3
|), SK
1
(R)
maps naturally onto SK
1
_
RŒX|. (X
2
)
_
– an immediate consequence of the defini-
tion of relative K
1
-groups in (9.4).
Proof of Theorem 9.7. First observe that the “if ” part of the theorem follows
from Theorem 9.5(b). Therefore, we only need to prove the following three im-
plications (restated here for the reader’s convenience):
9.B. The nontriviality of SK
1
(RŒM|) 237
(a) if K
1
(RŒM|) = K
1
(RŒM|ŒX|), then M is free;
(b) if M is seminormal and K
1
(R) = K
1
(RŒM|), then M is free;
(c) if ˝
1
R
= 0 and K
1
(R) = K
1
(RŒM|), then M is free.
Let us show that (b) == (a). Assume K
1
(RŒM|) = K
1
(RŒM|ŒX|). For an
arbitrary ring Athe following implication is known [96]:
K
i
(A) = K
i
(AŒX|) == K
i-1
(A) = K
i-1
(AŒX|). i ÷ N.
On the other hand if K
0
(A) = K
0
(AŒX|) then Pic(A) = Pic(AŒX|): exactly the
same argument as in the proof of the implication (i) == (ii) in Theorem 8.37
applies. Therefore, Theorem 4.67 and the equality K
1
(RŒM|) = K
1
(RŒM|ŒX|)
imply that the ring RŒM| seminormal. By Theorem 4.69 M is seminormal. But we
also have K
1
(R) = K
1
(RŒM|), because the monoid ring RŒM| admits a grading
of the form RŒM| = R ˚ R
1
˚ and Lemma 8.39 applies. Thus we are in the
situation of (b). In particular, M is free.
Now we prove (b). The case r = rank(M) = 1 is trivial because already the
seminormality guarantees that M Š Z
÷
. Suppose r ¸ 2. Let N Ç M be any
proper extreme submonoid. Since the identity embedding RŒN| Ç RŒM| splits the
surjective R-homomorphism RŒM| ÷ RŒN|, m ÷ m for all m ÷ N and m ÷ 0
for m ÷ M \ N, the equality K
1
(R) = K
1
(RŒM|) implies K
1
(R) = K
1
(RŒN|).
Because of the functorial splitting K
1
= U˚SK
1
we get SK
1
(R) = SK
1
(RŒN|).
By the induction hypothesis on rank we conclude that N is free. That is, all proper
extremal submonoids of M are free. By Proposition 9.17 M is itself free.
To prove (c) we first notice that, as mentioned above, K
1
(R) = K
1
(RŒM|)
implies SK
1
(R) = SK
1
(RŒM|).
Consider the case r = rank(M) = 1. The monoid operation will be written
additively. Without loss of generality we can assume M Ç Z
÷
and gp(M) = Z.
By Proposition 2.32 there exists m ÷ M such that m÷Z
÷
Ç M. We want to show
that M = Z
÷
. Assume to the contrary that M = Z
÷
. Then M Ç {0. 2. 3. . . . ].
Both RŒM| and RŒX
2
. X
3
| contain the ideal I = (X
m
. X
m÷1
. X
m÷2
. . . . ). (Here
we make the identification RŒZ
÷
| = RŒX|.) Since RŒX
2
. X
3
|,I is a nilpotent ex-
tension of R, any element of SK
1
(RŒX
2
. X
3
|) reduces modulo I to an element of
SK
1
(R); see Lemma 9.1. Therefore, every element of SK
1
(RŒX
2
. X
3
|) has a rep-
resentative in SL(RŒX
2
. X
3
|. I) Ç SL(RŒM|). In particular, the homomorphism
SK
1
(RŒM|) ÷ SK
1
(RŒX
2
. X
3
|) is surjective. We are done because the target of
this map is larger than SK
1
(R) by Lemma 9.21.
Assume r ¸ 2, SK
1
(R) = SK
1
(RŒM|), and that (c) is proved for affine sim-
plicial monoids of rank r ÷ 1. As in the proof of (b), all proper extremal sub-
monoids of M are free. By Proposition 2.37, M[z = sn(M)[zfor any proper face
z Ç ˚(M). (Recall, sn(M) is the seminormalization of M, p. 63.)
We claim that SK
1
(R) = SK
1
(RŒsn(M)|). By Proposition 2.32 there exists an
element m ÷ int(M) such that m ÷
¨
M Ç M (writing the monoid operation
additively). Let I = m
¨
MR (switching back to multiplicative notation). Then I is
an ideal of both RŒM| and RŒsn(M)|. Moreover, we have
(RŒM|,I)
red
= (RŒsn(M)|,I)
red
= RŒF|,(int(F))
238 9. Bass-Whitehead groups of monoid rings
for the free submonoid F of M generated by M[:
1
L LM[:
r
where :
1
. . . . . :
r
are the vertices of the simplex ˚(M). Equation (9.6) in the proof of Lemma 9.18
implies SK
1
(R) = SK
1
_
RŒF|,(int(F))
_
. Then the same argument we have used in
the case r = 1 shows that the homomorphism SK
1
(RŒM|) ÷ SK
1
(RŒsn(M)|) is
surjective. Hence the equality SK
1
(R) = SK
1
(RŒsn(M)|).
Proposition 9.17 implies that sn(M) is free. Now the equations sn(M)[z =
F[z, z running through the proper faces of ˚(M), imply that sn(M) = F. We
are done because F Ç M Ç sn(M) Ç F.
Remark 9.22. (a) It follows fromthe W(R)-module structure on K
1
(RŒM|),K
1
(R),
mentioned in Remark 9.14 (for general graded rings) that if Rcontains Qthen the
group K
1
(RŒM|),K
1
(R) is not even finitely generated once it is nontrivial. We
will make an essential use of this very fact in Section 10.C.
(b) For a natural number c and a monoid M we have the monoid endomor-
phism M ÷ M, m ÷ m
c
. This gives rise to a group endomorphism c
+
:
K
1
(RŒM|),K
1
(R) ÷ K
1
(RŒM|),K
1
(R). It follows easily from Lemma 9.12(b)
that all nonzero elements of K
1
(RŒM|),K
1
(R), produced in the proof of Theo-
rem 9.7, are trivialized by an endomorphism of type c
+
for c > 1. We do not
know whether the endomorphisms c
+
: K
1
(RŒM|),K
1
(R) ÷K
1
(RŒM|),K
1
(R),
c > 1, are all zero maps for R a regular ring and M a simplicial positive monoid.
A “nilpotent” version is given by Theorem 9.30(a) below.
(c) Let R be a regular ring. It can be shown that the proof of Theorem
9.7 actually produces many nonsimplicial affine positive monoids M for which
K
1
(RŒM|),K
1
(R) = 0; for instance see [43, Theorem 9.1(b)]). However, we do
not know whether Theorem 9.7 can be generalized to all affine positive monoids.
It is interesting to remark that in the special case of normal monoids it is enough
to consider only rank 3 monoids:
K
1
(RŒM|),K
1
(R) = 0 for all affine normal positive nonfree monoids M
”K
1
(RŒM|),K
1
(R) = 0 for all monoids M of rank ¸ 3 in the same class.
The proof of this equivalence is based on the fact that simple and simplicial
1
poly-
topes in dimension ¸ 3 are simplices. In more detail, assume M is an affine normal
positive nonfree monoid. If M is simplicial, then Theorem 9.7 applies. If M is not
simplicial, then either there exists a nonsimplicial extreme submonoid N Ç M
or there is a nonsimple vertex : ÷ ˚(M) and we have a nonsimplicial monoid
L Ç M as in Lemma 9.10. In the first case RŒN| is an R-retract of RŒM| and in
the second case RŒL| is an R-retract of RŒM|, so that the induction on rank goes
through.
Examples and applications.
Example 9.23 (Krusemeyer’s ring). The argument in the proof of Lemma 9.19 shows
that in the special case when ˝
1
R
= 0 and the Dennis-Stein symbols are trivial
1
A polytope P is simple if at any vertex : ÷ P exactly dim(P) facets of P meet, and it is
simplicial if all proper faces of P are simplices.
9.B. The nontriviality of SK
1
(RŒM|) 239
there are nonfree monoids M with K
1
(R) = K
1
(RŒM|). In particular, we recover
Krusemeyer’s example: k
+
= K
1
(k) = K
1
(kŒX
2
. X
3
|) for any algebraic extension
of fields Q Ç k. Notice, though, K
1
(kŒX
2
. X
3
|) = K
1
(kŒX
2
. X
3
|ŒY |) by Theorem
9.7(a).
Example 9.24 (Srinivas’ element). Let R be a regular ring. Consider the algebra
RŒX
2
. XY. Y
2
|. It is isomorphic to the monomial algebra RŒU. UV. UV
2
| (X
2
÷
U, XY ÷ UV , Y
2
÷ UV
2
). The embedding of the multiplicative monoid
of the monomials of the second algebra into the free monoid of all monomi-
als in U and V satisfies the condition in Lemma 9.9. Let ˛ ÷ SL
2
(RŒU. UV.
UV
2
|) be the same matrix as [(z
1
) in the proof of Lemma 9.15. Then step
(1) in the proof of Proposition 9.16 shows that Œ[(z
1
)| … SK
1
(RŒU. UV. UV
2
|).
Let ˛ ÷ SL
2
(RŒU. UV. UV
2
|) be the matrix obtained from [(z
1
) by permuting
rows and columns. Then Œ˛| = Œ[(z
1
)| ÷ SK
1
(RŒU. UV. UV
2
|). In particular,
Œ˛| … SK
1
(RŒU. UV. UV
2
|).
For a commutative ring A and two comaximal elements a. b ÷ A there is the
Mennicke symbol Œa. b| ÷ SK
1
(A) defined as follows [4, Ch. VI, §1]. Choose c. d ÷
Aso that ad ÷ bc = 1. Let
Œa. b| = class of
_
a b
c d
_
÷ SK
1
(A).
It can be shown that Œa. b| is independent of the choice of c and d [4].
Returning to our example we obtain that
Œ1 ÷ UV. UV
2
| ÷ SK
1
(RŒU. UV. UV
2
|) \ SK
1
(R).
or equivalently
Œ1 ÷ XY. Y
2
| ÷ SK
1
(RŒX
2
. XY. Y
2
|) \ SK
1
(R).
In the particular case R = Cwe recover Srinivas’ example in [78].
Example 9.25 (Toric degeneration). Let k be a field and - be a term order on the
polynomial ring kŒX
1
. . . . . X
n
| (See Section 4.A). Let A Ç kŒX
1
. . . . . X
n
| be an
affine subring. If the initial algebra
in
-
(A) = {in(¡ )
-
: ¡ ÷ A] Ç kŒX
1
. . . . . X
n
|
finitely generated over k then many ring theoretic properties of in
-
(A) transfer to
A; for instance see [16]. However, this is not the case for the property “all projective
modules are free”. Here is an example.
Let A = kŒX. Y. Z
2
. Z
3
÷ XYZ| and - be the lexicographic term order in-
duced by Z - Y - X. Then SK
0
(A) = 0 (i. e. there are projective A-modules
which are not even stably free), while in
-
(A) is finitely generated and all projective
in
-
(A)-modules are free.
240 9. Bass-Whitehead groups of monoid rings
To prove the claim above observe that (Z
2
÷XY )kŒX. Y. Z| Ç A, yielding the
Milnor square
A
kŒX. Y. Z|
kŒU. V | kŒS
2
. ST. T
2
|
where the horizontal arrows represent the identity embeddings, U. V. S. T de-
note variables, and X ÷ S
2
, Y ÷ T
2
, Z ÷ ST , U ÷ S
2
, V ÷ T
2
. It is
rather straightforward to check that in
-
(A) = kŒX. Y. Z
2
. XYZ| – a seminormal
monoid k-algebra: in view of the representation
kŒX. Y. Z
2
. XYZ| = kŒX. Y | ÷kŒX. Z
2
| ÷kŒY. Z
2
| ÷kŒint(Z
3
÷
)|
the seminormality of in
-
(A) follows from Proposition 2.37. So by Theorem 8.4
projective modules over in
-
(A) are free. On the other hand, the Mayer-Vietoris
sequence, associated to the diagram above, implies
SK
0
(A) = SK
1
(kŒS
2
. ST. T
2
|) = 0.
the inequality being a consequence of Theorem 9.7.
Example 9.26 (Pic(A) can be smaller than U(K
0
(A))). Let Abe as in Example 9.25
above. Then the six-term U-Pic Mayer-Vietoris sequence, i. e. the lower rows in
(8.12) and (9.2), associated to the Milnor square in Example 9.25, implies Pic(A) =
0.
Above we have shown that SK
0
(A) ,= 0. In particular,
¯
K
0
(A) = 0. Now
the desired nontrivial elements in U(K
0
(A)) are given by 1 ÷ x ÷ K
0
(A) where
x ÷
¯
K
0
(A), x = 0. In fact, it is known that
¯
K
0
(B) is a nilpotent ideal in the ring
K
0
(B) for any commutative ring B [4, Ch. IX, Prop. 4.6].
9.C. Further results: a survey
Nonstable results. The Quillen-Suslin theorem (Theorem 8.5) contains more
information than K
0
(kŒX
1
. . . . . X
n
|) = Z: it says that projective modules over
kŒX
1
. . . . . X
n
| are free, and not just stably free. Can one strengthen the equal-
ity K
1
(k) = K
1
(kŒX
1
. . . . . X
n
|) in the same way, namely to the assertion that
SL
r
(kŒX
1
. . . . . X
n
|) = E
r
(kŒX
1
. . . . . X
n
|)? The following theorem of Suslin [85]
shows that this is indeed the case, at least if r is large enough:
Theorem 9.27. Let R be a noetherian ring of finite Krull dimension dimR. Then
the natural map
GL
r
(RŒX
1
. . . . . X
n
|), E
r
(RŒX
1
. . . . . X
n
|) ÷K
1
(RŒX
1
. . . . . X
n
)
is surjective for r ¸ max(dimR÷1. 2) and bijective for r ¸ max(dimR÷2. 3).
Remark 9.28. (a) For r ¸ 3 the source of the map in Theorem 9.27 is not merely
the set of cosets: Suslin [85] has shown that E
r
(A) is a normal subgroup of GL
r
(A)
for any ring Awhenever r ¸ 3.
9.C. Further results: a survey 241
(b) In the special case M = 0 Theorem 9.31 is essentially equivalent to the
Bass-Vaserstein surjective K
1
-stabilization estimate [4, 95] for noetherian rings.
Combining Theorems 9.5(b) and 9.27, we get
Corollary 9.29. For every field k and all n ÷ Nwe have
SL
n
(kŒX
1
. . . . . X
r
|) = E
r
(kŒX
1
. . . . . X
n
|). r ¸ 3.
That the results above are sharp follows fromCohn’s observation [26]: for every
field k
_
1 ÷XY X
2
÷Y
2
1 ÷ XY
_
… E
2
(kŒX. Y |).
It follows from Example 9.24 that the matrix above does not belong to the stable
group E(kŒX
2
. XY. Y
2
|).
The proof of Theorem 9.27 in [85] is a K
1
-analogue of Quillen’s solution [71] to
the Serre problem, sketched in Section 8.C. It involves several results on invertible
matrices over polynomial rings that are of independent interest, like matrix ana-
logues of Quillen’s local-global principle (8.10) and of the affine Murthy-Horrocks
theorem (8.12). The proof is constructive and, like that in [71], can be converted
into an algorithm for factoring elements of SL
r
(kŒX
1
. . . . . X
n
|) into elementary
matrices [66]. Such algorithms have applications in signal processing, see [67].
Suslin’s techniques can be generalized to monoid rings. This opens up a possi-
bility for results like Theorems 9.30 and 9.31 below.
c-divisible monoids. Let c ÷ N. We say that a monoid M is c-divisible if for every
element x ÷ M there is an element y ÷ M such that x = cy, writing the monoid
operation additively. The typical examples of c-divisible monoids are intersections
of cones C Ç R
r
with the subgroup ZŒ1,c|
r
Ç Q
r
. (Here ZŒ1,c| is understood as
a ring extension.)
Observe that a c-divisible monoid is seminormal if c ¸ 2:
x ÷ gp(M). 2x. 3x ÷ M == cx ÷ M == x =
1
c
(cx) ÷ M.
However, no c-divisible monoid is finitely generated except M = 0.
The following K
1
-analogue of Theorem 8.4 is proved in [39].
Theorem 9.30. Let M be a c-divisible monoid. Then:
(a) if R is an Euclidean domain (such as a field, or Z), then SL
r
(RŒM|) =
E
r
(RŒM|) for r ¸ 3;
(b) if Ris a regular ring and M is positive, then K
1
(R) = K
1
(RŒM|).
Elementary actions on unimodular rows. To state the next result we introduce a
special class of cones. A cone C Ç R
r
is called of simplicial growth if there is a
sequence of cones
{0] = C
0
Ç C
1
Ç Ç C
k
= C
such that the closures of the sets C
i
\ C
i-1
, i = 1. . . . . k, are simplicial cones. It is
easily seen that all cones of dimension ¸ 3 are of simplicial growth, while starting
242 9. Bass-Whitehead groups of monoid rings
Figure 9.1. 3-cones are of simplicial growth
from dimension 4 this is no longer the case. For instance, the 4-cone C Ç R
4
over
the platonic octahedron in
_
R
3
. 1
_
,
C = R
÷
(1. 0. 0. 1) ÷R
÷
(÷1. 0. 0. 1) ÷R
÷
(0. 1. 0. 1)÷
R
÷
(0. ÷1. 0. 1) ÷R
÷
(0. 0. 1. 1) ÷R(0. 0. ÷1. 1).
is not of simplicial growth, for otherwise at least one of its vertices would be simple.
Recall, for a commutative ring A an r-row (a
1
. . . . . a
r
) ÷ A
r
is unimodular if
Aa
1
÷ ÷ Aa
r
= A. The set of unimodular r-rows over A will is denoted by
Um
r
(A). The general linear group GL
r
(A) acts on Um
r
(A) by right multiplication,
and we can restrict the action to the subgroup E
r
(A).
The following extension of Theorem 9.27 has been obtained in [40, 41].
Theorem 9.31. Let R be a noetherian ring of finite Krull dimension and M Ç Z
d
be an affine positive monoid for which the cone C(M) Ç R
d
is of simplicial growth.
Then the group of elementary matrices E
r
(RŒM|) acts transitively on the set of uni-
modular rows Um
r
(RŒM|) whenever r ¸ max(dimR÷2. 3).
Remark 9.32. (a) Theorem9.31 is nontrivial already in the special case of simplicial
affine monoids.
(b) It follows from Theorem 9.31 that the map
GL
r
(RŒM|), E
r
(RŒM|) ÷K
1
(RŒM|)
is surjective for r ¸ max(dimR ÷ 1. 2). In the special case M = Z
r
÷
we recover
the ‘surjective’ part of Theorem 9.27. This may indicate that general monoid rings
and polynomial rings have similar stabilizations for all higher K-groups. However,
currently not even the validity of Theorem 9.31 for general monoids is known. For
more information on this topic see [44].
Elementary actions and subintegral extensions. Concluding this section we give
an application of Theorem 9.31 which goes beyond the class of monoid rings. It is
a K
1
-stabilization analogue of Swan’s theorem on projective modules over subin-
tegral extensions (Theorem 8.47).
9.C. Further results: a survey 243
Corollary 9.33. Let A Ç B be a subintegral extension and ¨ a = (a
1
. . . . . a
r
) ÷
Um
r
(A) for some r ¸ 3. Then we have the equivalence
¨ a -
E
r
(A)
(1. 0. . . . . 0) ” ¨ a -
E
r
(B)
(1. 0. . . . . 0).
An auxiliary result we need is the following Milnor patching for unimodular
rows.
Lemma 9.34. Let
A
˛
ˇ
B
ı
C
;
D
be a Milnor square. Assume the image
¨
d ÷ Um
r
(D) of
¨
b ÷ Um
r
(B) is in the E
r
(D)-
orbit of (1. 0. . . . . 0). Then there is an element ¨ a ÷ Um
r
(A) such that ˛( ¨ a) -
E
r
(B)
¨
b.
Proof. Let c
D
÷ E
r
(D) be such that
¨
dc
D
= (1. 0. . . . 0).
(1) ; is surjective. In this situation E
r
(C) ÷ E
r
(D) is a surjective homo-
morphism. Let c
C
÷ E
r
(C) be a preimage of c
-1
D
. Then the first row of c
C
, say
¨ c ÷ C
r
, maps to
¨
d under ; and is unimodular. Because the diagram is carte-
sian, there exists a (unique) row ¨ a = (a
1
. . . . . a
r
) ÷ A
r
mapping respectively to
¨
b and ¨ c. We only need to prove that ¨ a ÷ Um
r
(A). Since in our situation ˛ is
also surjective, the ideal of A generated by the components of ¨ a contains an ele-
ment of the from 1 ÷ x for some x ÷ Ker(˛). If we knew that this ideal would
contain also x ÷ y for some y ÷ Ker(ˇ), then the desired unimodularity would
follow because 1 ÷ x ÷ x
_
(1 ÷ x) ÷ (x ÷ y)
_
= 1 ÷ xy = 1. (Here we use
that Ker(˛) ¨ Ker(ˇ) = 0.) Now the existence of such element y is established as
follows. There are z
1
. . . . . z
r
÷ C such that
_
ˇ(x)z
1
_
c
1
÷ ÷
_
ˇ(x)z
r
_
c
r
= ˇ(x)
where ¨ c = (c
1
. . . . . c
r
). Since
ˇ(x)z
1
. . . . . ˇ(x)z
r
÷ Ker(;)
there are elements a
t
1
. . . . . a
t
r
÷ A mapping to ˇ(x)z
1
. . . . . ˇ(x)z
r
, respectively.
Then a
1
a
t
1
÷ ÷a
r
a
t
1
= x ÷y for some y ÷ Ker(ˇ).
(2) ı is surjective. In this situation c
D
has a preimage c
B
÷ E
r
(B). The ele-
ments
¨
bc
B
and (1. 0. . . . . 0) ÷ Um
r
(C) have the same image in Um
r
(D). Then the
argument in (1) shows that
¨
bc
B
has a preimage in Um
r
(A), and this completes the
proof.
Proof of Corollary 9.33. We use Theorem 9.31 with R = Z: for an affine sim-
plicial monoid M the group of elementary matrices E
r
(ZŒM|) acts transitively on
Um
r
(ZŒM|) if r ¸ 3.
By induction it is enough to treat the case of an elementary subintegral exten-
sion B = AŒx| with x
2
. x
3
÷ A, x … A. Let S = A L {x] and consider the
polynomial algebra ZŒt
s
: s ÷ S|. The ring homomorphism
g : ZŒt
s
: s ÷ S| ÷B. t
s
÷s. s ÷ S.
244 9. Bass-Whitehead groups of monoid rings
is surjective and we have the Milnor square
D A
ZŒt
s
: s ÷ S|
g
B.
Since A ÷ B is injective, D can be identified with the preimage of A in ZŒt
s
: s ÷
S|. By Lemma 9.34 it is enough to show that E
r
(D) acts transitively on Um
r
(D)
for r ¸ 3.
Let N be the multiplicative monoid consisting of 1 and those monomials in the
indeterminates t
s
which are divisible by t
2
x
. Then N is the filtered union of affine
simplicial monoids. In fact, for a finite subset T Ç {t
s
: s ÷ A] let M
T
be the
monoid of those monomials that involve only indeterminates from T L {t
x
] and,
in addition, are divisible by t
2
x
. In the following we identify monomials in M
T
with
their exponent vectors in Z
N÷1
, N = #T.
Let M
Tk
, k ÷ N, be the integral closure in M
T
of the submonoid of M
T
gener-
ated by the monomials t
2
x
and t
2
x
t
k
s
, s ÷ T . Then the M
Tk
form an ascending chain
of affine simplicial monoids and their union is M
T
. (We leave the easy proof of
this claim to the reader.) Thus M
T
is the filtered union of simplicial submonoids.
Now N, too, is the filtered union of affine simplicial monoids because it is the fil-
tered union of the M
T
. In particular, E
r
(ZŒN|) acts transitively on Um
r
(ZŒN|) for
r ¸ 3.
The image of the ideal I = Z
_
N \ {1]
_
Ç ZŒt
s
: s ÷ S| under g is contained
in A since any natural number ¸ 2 is a nonnegative integral combination of 2 and
3. Therefore, I is an ideal of the subring D Ç ZŒt
s
: s ÷ S| as well. Consider the
Milnor square
ZŒN|
D
Z
D,I.
If the action of E
r
(D,I) on Um
r
(D,I) were transitive then, in view of the transi-
tivity of E
r
(ZŒN|) on Um
r
(ZŒN|), Milnor patching would imply the desired tran-
sitivity of E
r
(D) on Um
r
(D). But (D,I)
red
= ZŒt
s
: s ÷ A|, and the property we
are interested in is invariant modulo nilpotents. So it is enough to apply Theorem
9.31 again, this time to ZŒt
s
: s ÷ A|.
Nilpotence of higher K-theory of monoid rings. Somewhat unexpectedly, one can
prove that the K
0
-version of Theorem 9.30 implies Theorem 8.37(a). More pre-
cisely, the following implication is shown in [45] for an arbitrary ring R:
K
0
(R) = K
0
(RŒM|) for all c-divisible monoids M ==
K
0
(R) = K
0
(RŒM|) for all seminormal monoids.
Exercises 245
(See Exercise 9.3.) In view of Theorems 9.7 and 9.30, no K
1
-version of this impli-
cation is possible. On the other hand, one can deduce from [96] that
K
i
(R) = K
i
(RŒM|) for all c-divisible positive monoids M ==
K
i-1
(R) = K
i-1
(RŒM|) for all c-divisible positive monoids M.
We also have the following equivalence for a ring Rand i ÷ Z
÷
:
K
i
(R) = K
i
(RŒZ
r
÷
|) ”
_
K
i
(R) = K
i
_
R
_
ZŒ1,c|
r
÷
_
_
for coprime natural numbers c = c
t
. c
tt
¸ 2.
(Here ZŒ1,c|
÷
= ZŒ1,c| ¨ Q
÷
.) In fact, the endomorphism RŒ÷
c
| : RŒZ
r
÷
| ÷
RŒZ
r
÷
| makes RŒZ
r
÷
| a free module of rank c
r
over itself. Thus, using transfer
maps, we see that for every element x ÷ K
i
(RŒZ
r
÷
|) there exists j ÷ N such that
(c
t
)
j
x ÷ K
i
(R). Since the same is true for the number c
tt
and gcd(c
t
. c
tt
) = 1
we get x ÷ K
i
(R). That there is no way to use the same trick for general finitely
generated monoids is a consequence of Theorem 9.7. Moreover, it is shown in ??,
for an affine monoid M the R-endomorphism RŒM| ÷ RŒM|, m ÷ m
c
, makes
RŒM| an RŒM|-module of finite projective dimension if and only if M Š Z
s
÷
for
some s ÷ Z
÷
(or c = 1).
The implications listed above have motivated the following theorem [46]. Its
proof involves a series of important techniques, designed for studying higher K-
groups, which are well beyond the scope of the K-theoretical tools used in this
book.
Theorem 9.35. Let k be a field of characteristic 0. Let i. c ÷ N, c ¸ 2, and M be a
c-divisible positive monoid. Then K
i
(k) = K
i
(kŒM|).
Based on Mushkudiani’s earlier work the following stronger version for K
2
is
obtained in [47].
Theorem 9.36. Let M be a positive c-divisible monoid for some natural number
c ¸ 2 and R be a regular ring. Then we have K
2
(R) = K
2,r
(R) = K
2,r
(RŒM|)
for all r ¸ max(5. dimR÷3).
In the formulation of Theorem9.36 we have used the rth unstable Milnor group.
For any ring Athe K
2,r
(A) it is defined to be Ker
_
St
r
(A) ÷E
r
(A)
_
where St
r
(A)
is generated by the Steinberg symbols x
z
ij
, 1 ¸ i. j ¸ r, subject to the usual Stein-
berg relations.
Exercises
9.1. Give an example of a commutative noncancellative monoid M
1
and a commu-
tative non-torsion-free monoid M
2
such that such that SK
1
(R) = SK
1
(RŒM
1
|) =
SK
1
(RŒM
2
|) for every regular ring R.
9.2. Let ˘ be a conical complex such that all cones C
p
Ç R
n
p
, p ÷ ˘, are simplicial and
the monoids M
p
= C
p
¨ Z
n
p
form a monoidal complex in the sense of Exercise 8.7. Let
RŒ˘| denote the ring associated to this monoidal complex.
246 9. Bass-Whitehead groups of monoid rings
Show that the following conditions are equivalent for any regular ring R:
(a) K
1
(R) = K
1
(RŒ˘|),
(b) RŒ˘| is K
1
-regular,
(c) every cone C
p
, p ÷ ˘, is unimodular.
9.3. Let (F. G) be an admissible pair of functors in the sense of Exercise 8.3 and c. c
t
¸ 2
be coprime natural numbers.
(a) Show the following implications
F(R) = F(RŒM|) for all positive c-divisible and c
t
-divisible monoids M
==F(R) = F(RŒM|) for all positive monoids.
Hint: use the same approach as suggested in Exercise 8.3.
(b) Show the same implication as in (a) without the restriction to positive monoids.
9.4. For a regular ring Rgive a detailed proof of the equivalence
K
1
(RŒM|),K
1
(R) = 0 for all affine normal positive nonfree monoids M
”K
1
(RŒM|),K
1
(R) = 0 for all monoids M of rank ¸ 3 in the same class.
which has been mentioned in Remark 9.22.
9.5. Prove the special case of Theorem 9.30 when M Ç Q
n
is a c-divisible monoid for
some c ¸ 2 and R
÷
M Ç R
d
is a simplicial cone.
Hint: apply claim (7) in the hint for Exercise 8.4 to Milnor squares similar to the ones in
the proof of Lemma 8.18.
9.6. For a natural number r ¸ 2 and a ring Rshow the equivalence
(a
1
. a
2
. . . . . a
n
)
E
r
(R)
- (1. 0. . . . . 0) ”( ¨ a
1
. ¨ a
2
. . . . . ¨ a
n
)
E
r
(R{ J(R))
- (
¨
1.
¨
0. . . . .
¨
0)
where J(R) is the Jacobson ideal of R.
9.7. For a natural number r ¸ 2, a ring R, an affine monoid M, and a proper ideal I Ç M
show the implication:
E
r
(RŒM|) acts transitively on Um
r
(RŒM|)
==E
r
(RŒM|,RI) acts transitively on Um
r
(RŒM|,RI).
Hint: use the same strategy as in the proof of Theorem 8.49. This is possible in view of
Lemma 9.34 and Exercise 9.6.
9.8. Let Rbe a noetherian ring of finite Krull dimension, r ¸ max(dimR÷2. 3) a natural
number, M an affine positive monoid, and I Ç M a nonempty proper ideal. Suppose that
every proper face of R
÷
M is simplicial if rank M > 4. Using Theorem 9.31, prove that
E
r
(RŒM|,RI) acts transitively on Um
r
(RŒM|,RI).
Chapter 10
Global varieties
10.A. Toric varieties
10.B. Intersection theory for toric varieties
10.C. Toric varieties with huge Grothendieck group
10.D. The equivariant Serre Problem for abelian groups
247
Bibliography
1. D. F. Anderson. The Picard group of a monoid domain. II. Arch. Math. 55, 2 (1990), 143–145.
2. V. Barucci, D. E. Dobbs, and M. Fontana. Maximality properties in numerical semigroups
and applications to one-dimensional analytically irreducible local domains. MAMS 598, 1997.
3. H. Bass. Big projective modules are free. Illinois J. Math. 7 (1963), 24–31.
4. H. Bass. Algebraic K-Theory. W. A. Benjamin, Inc., 1968.
5. H. Bass, A. Heller, and R. Swan. The Whitehead group of a polynomial extension. Publ.
Math. I.H.E.S. 22 (1964), 61–79.
6. S. M. Bhatwadekar and R. A. Rao. On a question of quillen. Trans. Amer. Math. Soc. 279
(1983), 801–810.
7. A. M. Bigatti, R. L. Scala, and L. Robbiano. Computing toric ideals. J. Symb. Comp. 27
(1999), 351–365.
8. S. Bloch. Some formulas pertaining to the K-theory of commmutative groupschemes. J. Al-
gebra 53 (1978), 304–326.
9. A. Borel. Linear algebraic groups. Springer, 1991.
10. N. Bourbaki. Algèbre commutative. Hermann, 1967.
11. N. Bourbaki. Lie groups and Lie algebras, Ch. 4–6. Springer, 2002.
12. C. Bouvier and G. Gonzalez-Sprinberg. Système generateurs minimal, diviseurs essen-
tiels et G-désingularizations de variétés toriques. Tôhoku Math. J. 46 (1994), 125–149.
13. J. W. Brewer, D. L. Costa, and K. McCrimmon. Seminormality and root closure in polyno-
mial rings and algebraic curves. J. Algebra 58 (1979), 217–226.
14. M. Brun, W. Bruns, and T. Römer. Cohomology of partially ordered sets. Adv. in Math. (to
appear).
15. M. Brun and T. Römer. On algebras associated to partially ordered sets. Preprint.
16. W. Bruns and A. Conca. Gröbner bases and determinantal ideals. In Commutative algebra,
singularities and computer algebra, J. Herzog and V.Vuletescu, Eds., Kluwer, 2003, pp. 9–66.
17. W. Bruns and J. Gubeladze. Normality and covering properties of affine semigroups. J.
Reine Angew. Math. 510 (1999), 151–178.
18. W. Bruns and J. Gubeladze. Rectangular simplicial semigroups. In Commutative algebra,
algebraic geometry, and computational methods, D. Eisenbud, Ed., Springer Singapore, 1999,
pp. 201–214.
19. W. Bruns and J. Gubeladze. Polyhedral algebras, arrangements of toric varieties, and their
groups. In Computational Commutative Algebra and Combinatorics, T. Hibi, Ed., Advanced
Studies in Pure Mathematics 33, Math. Soc. of Japan, 2002, pp. 1–51.
20. W. Bruns and J. Gubeladze. Semigroup algebras and discrete geometry. In Toric geometry,
L. Bonavero and M. Brion, Eds., Séminaires et Congrès 6, SMF, 2002, pp. 43–127.
21. W. Bruns and J. Gubeladze. Unimodular covers of multiples of polytopes. Doc. Math., J.
DMV 7 (2002), 463–480.
22. W. Bruns and J. Gubeladze. Divisorial linear algebra of normal semigroup rings. Algebr.
Represent. Theory 6 (2003), 139–168.
249
250 Bibliography
23. W. Bruns, J. Gubeladze, M. Henk, A. Martin, and R. Weismantel. A counterexample to
an integer analogue of Carathéodory’s theorem. J. Reine Angew. Math. 510 (1999), 179–185.
24. W. Bruns and J. Herzog. Cohen–Macaulay rings. Cambridge University Press, 1998.
25. W. Bruns and U. Vetter. Determinantal rings. Lect. Notes Math. 1327, Springer, 1988.
26. P. M. Cohn. The similarity reduction of matrices over a skew field. Math. Z. 132 (1973), 151–
163.
27. C. D. Concini, D. Eisenbud, and C. Procesi. Hodge Algebras. Astérisque 91, Soc. Math. de
France, 1982.
28. G. Cortinas. The obstruction to excision in K-theory and in cyclic homology. INV (to ap-
pear).
29. B. H. Dayton. The Picard group of a reduced G-algebra. J. Pure Appl. Algebra 59, 3 (1989),
237–253.
30. J. A. De Loera, J. Rambau, and F. Santos. Triangulations of point sets: Applications, Struc-
tures, Algorithms. In preparation.
31. R. K. Dennis and M. Krusemeyer. K
2
(AŒX. Y |,(XY )), a problem of Swan and related com-
putations. J. Pure Applied Algebra 15 (1979), 125–148.
32. D. Eisenbud. Commutative algebra with a view toward algebraic geometry. Springer, 1995.
33. R. Fossum. The divisor class group of a Krull domain. Springer, 1973.
34. F. G. Frobenius. Gesammelte Abhandlungen, Band 3. Springer, 1968.
35. I. M. Gelfand, M. M. Kapranov, and A. V. Zelevinsky. Discriminants, resultants, and mul-
tidimensional determinants. Birkhäuser, 1994.
36. R. Gilmer. Commutative semigroup rings. University of Chicago Press, 1984.
37. T. G. Goodwillie. Relative algebraic K-theory and cyclic homology. Ann. of Math. 124 (1986),
347–402.
38. D. Grayson. Higher algebraic K-theory: II. In Algebr. K-Theory, Proc. Conf. Evanston 1976,
vol. 551 of Lect. Notes Math. 1976, pp. 217– 240.
39. J. Gubeladze. Classical algebraic k-theory of monoid algebras. In K-theory and homological
algebra, Proc. Semin., Tbilisi 1987-88, Lect. Notes Math., Springer, 1990, pp. 36–94.
40. J. Gubeladze. The elementary action on unimodular rows over a monoid ring. J. Algebra 148
(1992), 135–161.
41. J. Gubeladze. The elementary action on unimodular rows over a monoid ring II. J. Algebra
155 (1993), 171–194.
42. J. Gubeladze. Geometric and algebraic representations of commutative cancellative monoids.
Proc. A. Razmadze Math. Inst., Georg. Acad. Sci. 113 (1995), 31–81.
43. J. Gubeladze. Nontriviality of SK
1
(RŒM|). J. Pure Applied Algebra 104 (1995), 169–190.
44. J. Gubeladze. K-theory of affine toric varieties. Homology Homotopy Appl. 1 (1999), 135–145.
45. J. Gubeladze. Higher K-theory of toric varieties. K-Theory 28 (2003), 285–327.
46. J. Gubeladze. The nilpotence conjecture in K-theory of toric varieties. Invent. Math. (2005).
47. J. Gubeladze. The Steinberg group of a monoid ring, nilpotence, and algorithms. J. Algebra
(to appear).
48. C. Haase and G. M. Ziegler. On the maximal width of empty lattice simplices. Eur. J. Comb.
21 (2000), 111–119.
49. F. Halter-Koch. Finitely generated monoids, finitely primary monoids and factorization
properties of integral domains. In Factorization in integral domains, D. D. A. (Ed.), Ed., M.
Dekker, 1997, pp. 31–72.
50. E. Hamann. The R-invariance of RŒX|. J. Algebra 35 (1975), 1–16.
51. J. E. Humphreys. Linear algebraic groups. Springer, 1975.
52. H. Jarchow. Locally convex spaces. Teubner, 1981.
53. J. M. Kantor and K. S. Sarkaria. On primitive subdivisions of tetrahedra. Pac. J. Math. (to
appear).
54. M. Karoubi. Homologie cyclique et K-théorie. Astérisque, 149 (1987), 147.
Bibliography 251
55. G. Kempf, F. Knudsen, D. Mumford, and B. Saint-Donat. Toroidal embeddings I. Lect.
Notes Math. 339, Springer, 1973.
56. M. Krusemeyer. Fundamental groups, algebraic K-theory, and a problem of Abhyankar. In-
vent. Math. 19 (1973), 15–47.
57. E. Kunz. Introduction to commutative algebra and algebraic geometry. Birkhäuser, 1985.
58. J. C. Lagarias and G. M. Ziegler. Private communication.
59. T. Y. Lam. Serre’s Conjecture. Lect. Notes Math. 635, Springer, 1978.
60. S. Lang. Algebra. Springer, 2002.
61. J.-L. Loday. Cyclic Homology, Second edition. Springer, 1998.
62. D. Maclagan, S. Hosten, and B.Sturmfels. Supernormal vector configurations. J. Algebr.
Comb. (to appear).
63. H. Matsumura. Commutative ring theory. Cambridge University Press, 1989.
64. E. Miller and B. Sturmfels. Combinatorial commutative algebra. Springer, 2005.
65. J. Milnor. Introduction to algebraic K-theory. Annals of Mathematics Studies 72, Princeton
University Press, 1971.
66. H. Park andC. Woodburn. An algorithmic proof of Suslin’s stability theoremfor polynomial
rings. J. Algebra 178 (1995), 277–298.
67. H. Park andC. Woodburn. An algorithmic proof of Suslin’s stability theoremfor polynomial
rings. J. Algebra 178, 1 (1995), 277–298.
68. D. S. Passman. The algebraic structure of group rings. Wiley-Interscience, 1977.
69. D. Quillen. On the cohomology and K-theory of the general linear groups over a finite field.
Ann. of Math. 96 (1972), 552–586.
70. D. Quillen. Higher algebraic K-theory: I. In Algebr. K-Theory I, Proc. Conf. Battelle Inst.
1972, Lect. Notes Math. 341, Springer, 1973, pp. 85–147.
71. D. Quillen. Projective modules over polynomial rings. Invent. Math. 36 (1976), 167–171.
72. L. G. Roberts and B. Singh. Subintegrality, invertible modules and the Picard group. Com-
pos. Math. 85, 3 (1993), 249–279.
73. A. Seb˝ o. An introduction to empty lattice simplices. Lect. Notes Comput. Sci. 1610 (1999), 400–
414.
74. A. Seidenberg. A note on the dimension theory of rings. Pac. J. Math. 3 (1953), 505–512.
75. J. P. Serre. Modules projectifs et espaces fibrés á fibre vectorielle. Semin. Dubreil 23 (1957-58).
76. I. P. Shestakov and U. U. Umirbaev. The tame and the wild automorphisms of polynomial
rings in three variables. J. Amer. Math. Soc. 17 (2004), 197–227.
77. B. Singh. The Picard group and subintegrality in positive characteristic. Compos. Math. 95, 3
(1995), 309–321.
78. V. Srinivas. K
1
of the cone over a curve. J. Reine Angew. Math. 381 (1987), 37–50.
79. R. Stanley. Generalized h-vectors, intersection cohomology of toric varieties, and related
results. Commutative algebra and combinatorics, Adv. Stud. Pure Math. 11 (1987), 187–213.
80. R. P. Stanley. Combinatorics and commutative algebra. Birkhäuser, second edition 1996.
81. M. R. Stein and R. K. Dennis. K
2
of radical ideals and semi-local rings revisited. In Algebr.
K-Theory II, Proc. Conf. Battelle Inst. 1972, Lect. Notes Math. 342, Springer, 1973, pp. 281–303.
82. J. Stienstra. Operations in the higher K-theory of endomorphisms. In Current trends in alge-
braic topology, Semin. London/Ont. 1981, CMS Conf. Proc. 2, 1, Amer. Math. Soc., 1982, pp. 59–
115.
83. B. Sturmfels. Gröbner bases and convex polytopes. Amer. Math. Soc., 1996.
84. A. A. Suslin. Projective modules over a polynomial ring are free. Sov. Mat. Dokl. 17 (1976),
1160–1164.
85. A. A. Suslin. On the structure of the special linear group over polynomial rings. Math. USSR-
Izv. 11 (1977), 221–238.
86. A. A. Suslin and M. Wodzicki. Excision in algebraic K-theory. Ann. of Math. 136, Ser. II
(1992), 51–122.
87. R. G. Swan. Algebraic K-theory. Lect. Notes Math. 76, Springer, 1968.
252 Bibliography
88. R. G. Swan. Excision in algebraic K-theory. J. Pure Applied Algebra 1 (1971), 221–252.
89. R. G. Swan. On seminormality. J. Algebra 67 (1980), 210–229.
90. R. G. Swan. Projective modules over binary polyhedral groups. J. Reine Angew. Math. 342
(1983), 66–172.
91. R. G. Swan. Projective modules over laurent polynomial rings. Trans. Amer. Math. Soc. 237
(1987), 111–120.
92. R. G. Swan. Gubeladze’s proof of Anderson’s conjecture. Contemp. Math. 124 (1992), 215–250.
93. C. Traverso. Seminormality and Picard group. Ann. Scuola Norm. Sup. Pisa (3) 24 (1970),
585–595.
94. W. van der Kallen, H. Maazen, and J. Stienstra. A presentation for some K
2
(n. R). Bull.
Amer. Math. Soc. 81 (1975), 934–936.
95. L. Vaserstein. On stabilization of the general linear group over a ring. Math. USSR-Sb. 79
(1969), 405–424.
96. T. Vorst. Localization of the K-theory of polynomial extensions. Math. Ann. 244 (1979), 33–
53.
97. T. Vorst. The Serre problem for discrete Hodge algebras. Math. Z. 184 (1983), 425–433.
98. W. Waterhouse. Automorphisms of det(x
ij
) : the group scheme approach. Adv. in Math. 65
(1987), 171–203.
99. C. A. Weibel. Module structures on the K-theory of graded rings. J. Algebra 105 (1987), 465–
483.
100. C. A. Weibel. Pic is a contracted functor. Invent. Math. 103, 2 (1991), 351–377.
101. G. K. White. Lattice tetrahedra. Canad. J. Math. 16 (1964), 389–396.
102. O. Zariski and P. Samuel. Commutative algebra, Vols. I and II. Van Nostrand (new edn.
Springer 1975), 1958, 1960.
103. G. M. Ziegler. Lectures on polytopes. Springer, 1998.

Sign up to vote on this title
UsefulNot useful

Master Your Semester with Scribd & The New York Times

Special offer for students: Only $4.99/month.

Master Your Semester with a Special Offer from Scribd & The New York Times

Cancel anytime.