Combinatorial Reciprocity Theorems
Enumerative Combinatorics with a Polyhedral Angle
Matthias Beck & Raman Sanyal
v
Q
w
−Q + v + w
−Q
San Francisco State University
E-mail address: [email protected]
¨ t Berlin
Freie Universita
E-mail address: [email protected]
2010 Mathematics Subject Classification. Primary 05A; Secondary 05C15,
05C21, 05C30, 05C31, 05E40, 11H06, 11P21, 11P81, 52B20, 52C07, 68R05
Key words and phrases. Combinatorial reciprocity theorem, enumerative
combinatorics, partially ordered set, M¨obius function, order polynomial,
polyhedron, cone, polytope, Euler characteristic, rational generating
function, composition, partition, quasipolynomial, subdivision, triangulation,
order cone, order polytope, P -partition, chromatic polynomial, hyperplane
arrangement, graph orientation, flow polynomial
c 2015 by the authors. All rights reserved. This work may not be
translated or copied in whole or in part without the permission of the
authors. This is a beta version of a book that will be published by the
American Mathematical Society in 2016. The most current version of this
manuscript is available at the website
http://math.sfsu.edu/beck/crt.html.
April 1, 2015
Contents
Preface
xi
Chapter 1.
Four Polynomials
1
§1.1.
Graph colorings
1
§1.2.
Flows on graphs
6
§1.3.
Order polynomials
9
§1.4.
Ehrhart polynomials
13
Notes
18
Exercises
19
Chapter 2.
Partially Ordered Sets
25
§2.1.
Order Ideals and the Incidence Algebra
25
§2.2.
The M¨
obius Function and Order Polynomial Reciprocity
29
§2.3.
Zeta Polynomials, Distributive Lattices, and Eulerian Posets
32
§2.4.
M¨
obius Inversion
34
Notes
35
Exercises
36
Chapter 3.
Polyhedral Geometry
39
§3.1.
Polyhedra, Cones, and Polytopes
40
§3.2.
Faces
45
§3.3.
The Euler Characteristic
47
§3.4.
M¨
obius Functions of Face Lattices
56
§3.5. Uniqueness of the Euler Characteristics and Zaslavsky’s
Theorem
61
vii
viii
§3.6.
Contents
The Brianchon–Gram Relation
66
Notes
68
Exercises
69
Chapter 4.
Rational Generating Functions
75
§4.1.
Matrix Powers and the Calculus of Polynomials
75
§4.2.
Compositions
82
§4.3.
Plane Partitions
83
§4.4.
Restricted Partitions
86
§4.5.
Quasipolynomials
89
§4.6.
Integer-point Transforms and Lattice Simplices
91
§4.7.
Gradings of Cones and Rational Polytopes
95
§4.8.
Stanley Reciprocity for Simplicial Cones
98
Notes
103
Exercises
105
Chapter 5.
Subdivisions
111
§5.1.
Decomposing a Polytope
111
§5.2.
M¨
obius Functions of Subdivisions
116
§5.3.
Beneath, beyond, and half-open decompositions
118
§5.4.
h∗ -vectors
120
§5.5.
Solid angles?
and unimodular triangulations
120
Exercises
120
§5.6.
Ehrhart–Macdonald Reciprocity
121
§5.7.
Solid Angles
123
§5.8.
Ehrhart–Macdonald Reciprocity Revisited [old]
125
§5.9.
Unimodular Triangulations
126
Notes
127
Exercises
128
Chapter 6.
Partially Ordered Sets, Geometrically
131
§6.1.
The Geometry of Order Cones
132
§6.2.
Order Polytopes
135
§6.3.
Generating Functions and Permutation Statistics
137
§6.4.
Order Polynomials and P -Partitions
140
§6.5.
Chromatic and Order Polynomials
145
§6.6. Applications: Extension Problems and Marked Order Polytopes 146
Contents
ix
Notes
146
Exercises
147
Chapter 7.
Hyperplane Arrangements
151
§7.1.
Inside-out Polytopes
152
§7.2.
Graph Colorings and Acyclic Orientations
155
§7.3.
Graph Flows and Totally Cyclic Orientations
159
Notes
159
Exercises
159
Bibliography
161
Index
167
Preface
Combinatorics is not a science, it’s an attitude.
Mark Haiman
A combinatorial reciprocity theorem relates two classes of combinatorial
objects via their counting functions. Consider a class X of combinatorial
objects and let f (n) be the function that counts the number of objects in
X of size n, where “size” refers to some specific quantity that is naturally
associated with the objects in X . As in canonization, it requires two miracles
for a combinatorial reciprocity to occur:
1. The function f (n) is the restriction of some reasonable function (e.g., a
polynomial) F (x) to the positive integers, and
2. the evalutation F (−n) is an integer of the same sign σ ∈ {±1} for all
n ∈ Z>0 .
In this situation it is human to ask if σ F (−n) has a combinatorial meaning,
that is, if there is a natural class X ◦ of combinatorial objects such that
σ F (−n) counts the objects of X ◦ of size n (where “size” again refers to
some specific quantity, possibly quite different from the quantity we were
measuring for X ). Combinatorial reciprocity theorems are among the most
charming results in mathematics and, in contrast to canonization, can be
found all over enumerative combinatorics and beyond.
As a first example consider the class of maps [k] → Z>0 from a finite
set [k] := {1, 2, . . . , k} into the positive integers and let f (n) = nk count
the number of maps with range [n]. Thus f (n) is the restriction of a
polynomial and (−1)k f (−n) = nk satisfies our second requirement above.
This relates the number of maps [k] → [n] to itself. This relation is a genuine
combinatorial reciprocity but the impression one is left with is that of being
xi
xii
Preface
underwhelmed rather than charmed. Later in the book it will become clear
that this example is not boring at all, but for now let’s try again.
The term combinatorial reciprocity theorem was coined by Richard Stanley
in his 1974 paper [73] of the same title, in which he developed a firm standing
of the subject. Stanley starts with a very appealing reciprocity that he
attributes to John Riordan: For a fixed integer d, let f (n) count the number
of d-subsets of an n-set, that is, the number of choices of taking d elements
from a set of n elements without repetition. The counting function is given
by the binomial coefficient
n
1
n(n − 1) · · · (n − d + 2)(n − d + 1)
(0.1)
f (n) =
=
d!
d
which is the restriction of a polynomial in n of degree d, and from the
factorization we can read off that (−1)d f (−n) is a positive integer for every
n > 0. More precisely,
1
n+d−1
(−1)d f (−n) =
(n + d − 1)(n + d − 2) · · · (n + 1)n =
d!
d
is the number of d-multisubsets of an n-set, that is, the number of picking d
elements from [n] with repetition. Now this is a combinatorial reciprocity!
In formulas it reads
n+d−1
d n
(−1)
=
.
(0.2)
d
d
This is a charming result in more than one way. It presents an intriguing
connection between subsets and multisubsets via their counting functions, and
its formal justification is completely within the realms of an undergraduate
class in combinatorics. The identity (0.1) can be found in Riordan’s book [?]
on combinatorial analysis without further comment and, charmingly, Stanley
states that his paper [73] can be considered as “further comment.” That
further comment is necessary is apparent from the fact that the formal proof
above falls short of explaining why these two sorts of objects are related
by a combinatorial reciprocity. In particular, comparing coefficients in (0.2)
cannot be the tool of choice for establishing more general reciprocity relations.
In this book we develop tools and techniques for handling combinatorial
reciprocities. However, our own perspective is firmly rooted in geometric
combinatorics and, thus, our emphasis is on the geometric nature of the
combinatorial reciprocities. That is, for every class of combinatorial objects
we associate a geometric object (such as a polytope or a polyhedral complex)
in such a way that combinatorial features, including counting functions and
reciprocity, are reflected in the geometry. In short, this book can be seen as
further comment with pictures.
The prerequisites for this book are minimal: basic knowledge of linear
algebra and combinatorics should suffice. The numerous exercises throughout
Preface
xiii
the text are designed so that the book could easily be used for a graduate
class in combinatorics.
Matthias Beck • Raman Sanyal
Chapter 1
Four Polynomials
To many, mathematics is a collection of theorems. For me, mathematics is a
collection of examples; a theorem is a statement about a collection of examples and
the purpose of proving theorems is to classify and explain the examples...
John B. Conway
In the spirit of the above quote, this chapter serves as a source of examples
and motivation for the theorems to come and the tools to be developed. Each
of the following four sections introduces a family of examples together with
a combinatorial reciprocity statement which we will prove in later chapters.
1.1. Graph colorings
Graphs and their colorings are all-time favorites in introductory classes on
discrete mathematics, and we too succumb to the temptation to start with
one of the most beautiful examples. A graph G = (V, E) is a discrete
structure composed of a finite1 set of nodes V and a collection E ⊆ V2
of unordered pairs of nodes, called edges. More precisely, this defines a
simple graph as it excludes the existence of multiple edges between nodes
and, in particular, edges with equal endpoints, i.e., loops. We will, however,
need such non-simple graphs in the sequel but we dread the formal overhead
non-simple graphs entail and will trust the reader’s discretion to make the
necessary modifications. The most charming feature of graphs is that they
are easy to visualize and their natural habitat are the margins of textbooks
or notepads. Figure 1 shows some examples.
An n-coloring of a graph G is a map c : V → [n] := {1, 2, . . . , n}. An
n-coloring c is called proper if no two nodes sharing an edge get assigned
1 Infinite graphs are interesting in their own right; however, they are no fun to color-count
and so will play no role in this book.
1
2
1. Four Polynomials
Figure 1. Various graphs.
the same color, that is,
c(u) 6= c(v)
whenever
uv ∈ E .
The name coloring comes from the natural interpretation of thinking of c(v)
as one of n possible colors that we use for the node v. A proper coloring is
one where adjacent nodes get different colors. Here is a first indication why
considering simple graphs often suffices: the existence and even the number
of n-colorings is unaffected by parallel edges, and there are simply no proper
colorings in the presence of loops.
Much of the fame of graph colorings stems from a question that was
asked around 1852 by Francis Guthrie and answered only some 124 years
later. In order to state the question in modern terms, we call a graph G
planar if G can be drawn in the plane (or scribbled in the margin) such
that edges do not cross except possibly at nodes. For example, the last row
in Figure 1 shows a planar and nonplanar embedding of the (planar) graph
K4 . Here is Guthrie’s famous conjecture, now a theorem:
Four-color Theorem. Every planar graph has a proper 4-coloring.
There were several attempts at the four-color theorem before the first
correct proof by Kenneth Appel and Wolfgang Haken. Here is one particularly interesting (but not yet successful) approach to proving the four-color
theorem, due to George Birkhoff. For a (not necessarily planar) graph G, let
χG (n) := |{c : V → [n] proper n-coloring}| .
The following observation, due to George Birkhoff and Hassler Whitney, is
that χG (n) is the restriction to Z>0 of a particularily nice function:
1.1. Graph colorings
3
Proposition 1.1. Let G = (V, E) be a loopless graph. Then χG (n) agrees
with a polynomial of degree |V | with integral coefficients.
By a slight abuse of notation, we identify χG (n) with this polynomial
and call it the chromatic polynomial of G. Nevertheless, let’s emphasize
that, so far, only the values of χG (n) at positive integral arguments have an
interpretation in terms of G.
One proof of Proposition 1.1 is interesting in its own right, as it exemplifies
deletion–contraction arguments which we will revisit in Chapter 7. For e ∈ E,
the deletion of e results in the graph G\e := (V, E \ {e}). The contraction
G/e is the graph obtained by identifying the two nodes incident to e and
removing e. An example is given in Figure 2.
v
u
Figure 2. Contracting the edge e = uv.
Proof of Proposition 1.1. We induct on |E|. For |E| = 0 there are no
coloring restrictions and χG (n) = n|V | . One step further, assume that G has
a single edge e = uv. Then we can color all nodes V \ {u} arbitrarily and,
assuming n ≥ 2, can color u with any color 6= c(v). Thus, the chromatic
polynomial is χG (n) = nd−1 (n − 1) where d = |V |.
For the induction step, let e = uv ∈ E. We claim
χG (n) = χG\e (n) − χG/e (n) .
(1.1)
Indeed, a coloring c of G \ e fails to be a coloring of G if c(u) = c(v). That
is, we are over-counting by all proper colorings that assign the same color to
u and v. These are precisely the proper n-colorings of G/e.
By (1.1) and the induction hypothesis, χG (n) is the difference of a
polynomial of degree d = |V | and a polynomial of degree d − 1, both with
integer coefficients.
The above proof and, more precisely, the deletion–contraction relation
(1.1) reveal more about chromatic polynomials, which we invite the reader
to show in Exercise 1.6:
Corollary 1.2. Let G be a loopless graph on d ≥ 1 nodes and χG (n) =
cd nd + cd−1 nd−1 + · · · + c0 its chromatic polynomial. Then
4
1. Four Polynomials
(a) the leading coefficient cd = 1;
(b) the constant coefficient c0 = 0;
(c) (−1)d χG (−n) > 0 for all integral n ≥ 1.
In particular the last property prompts the following natural question
that we alluded to in the preface and that lies at the heart of this book.
Do the evaluations (−1)|V | χG (−n) have combinatorial meaning?
This question was first asked (and beautifully answered) by Richard
Stanley in 1973. To reproduce his answer, we need the notion of orientations
on graphs. Again, to keep the formal pain level at a minimum, let’s denote
the nodes of G by v1 , v2 , . . . , vd . We define an orientation on G through a
subset ρ ⊆ E; for an edge e = vi vj ∈ E with i < j we direct
e
vi ←− vj if e ∈ ρ
and
e
vi −→ vj if e ∈
/ ρ.
We denote the oriented graph by ρ G and will sometimes write ρ G = (V, E, ρ).
Said differently, we may think of G as canonically oriented by directing edges
from small index to large, and ρ records the edges on which this orientation
is reversed; see Figure 3 for an example.
v15
v5 2
v45
v5 3
Figure 3. An orientation given by ρ = {14, 23, 24}.
A directed path is a constellation vi0 → vi1 → · · · → vis in ρ G such
that vik 6= vil for 0 ≤ k < l ≤ s, and it is called a directed cycle if vi0 = vis .
An orientation ρ of G is acyclic if there are no directed cycles in ρ G.
Here is the connection between proper colorings and acyclic orientations:
Given a proper coloring c, we define the orientation
ρ := {vi vj ∈ E : i < j, c(vi ) > c(vj )} .
That is, the edge from lower index i to higher index j is directed along its
color gradient c(vj ) − c(vi ). We call this orientation ρ induced by the
coloring c. For example, the orientation pictured in Figure 3 is induced by
the coloring shown in Figure 4.
Proposition 1.3. Let c : V → [n] be a proper coloring and ρ the induced
orientation on G. Then ρ G is acyclic.
1.1. Graph colorings
5
44
5
5
22
1
1
Figure 4. A coloring that induces the orientation in Figure 3.
Proof. Assume that vi0 → vi1 → · · · → vis → vi0 is a directed cycle in ρ G.
Then c(vi0 ) > c(vi1 ) > · · · > c(vis ) > c(vi0 ) which is a contradiction.
As there are only finitely many acyclic orientations on G,
colorings according to the acyclic orientation they induce.
ρ and an n-coloring c of G are called compatible if for
edge u → v in ρ G we have c(u) ≥ c(v). The pair (ρ, c) is
compatible if c(u) > c(v) for every oriented edge u → v.
we might count
An orientation
every oriented
called strictly
Proposition 1.4. If (ρ, c) is strictly compatible, then c is a proper coloring
and ρ is an acyclic orientation on G. In particular, χG (n) is the number of
strictly compatible pairs (ρ, c) where c is a proper n-coloring.
Proof. If (ρ, c) are strictly compatible then, since every edge is oriented,
c(u) > c(v) or c(u) < c(v) whenever uv ∈ E. Hence c is a proper coloring
and ρ is exactly the orientation induced by c. The acyclicity of ρ G now
follows from Proposition 1.3.
We are finally ready for our first combinatorial reciprocity theorem.
Theorem 1.5. Let G be a finite graph on d nodes and χG (n) its chromatic
polynomial. Then (−1)d χG (−n) is the number of compatible pairs (ρ, c)
where c is an n-coloring and ρ is an acyclic orientation. In particular,
(−1)d χG (−1) is the number of acyclic orientations of G.
As one illustration of this theorem, consider the graph G in Figure 5;
its chromatic polynomial is χG (n) = n(n − 1)(n − 2)2 , and so Theorem 1.5
suggests that G should admit 18 acyclic orientations. Indeed, there are 6
acylic orientations of the subgraph formed by v1 , v2 , and v4 , and for the
remaining two edges, one of the four possible combined orientations of v2 v3
and v3 v4 produces a cycle with v2 v4 , so there are a total of 6 · 3 = 18 acyclic
orientations.
A deletion–contraction proof of Theorem 1.5 is outlined in Exercise 1.9;
we will give a geometric proof in Section 6.5.
6
1. Four Polynomials
v14
5
v2
v42
1
v3
Figure 5. This graph has 18 acyclic orientations.
1.2. Flows on graphs
Given a graph G = (V, E) together with an orientation ρ and a finite Abelian
group Zn = Z/nZ, a Zn -flow is a map f : E → Zn that assigns a value
f (e) ∈ Zn to every edge e ∈ E such that there is conservation of flow at
every node v:
X
X
f (e) =
f (e) ,
e
→v
e
v→
that is, what “flows” into the node v is precisely what “flows” out of v. This
physical interpretation is a bit shaky as the commodity flowing along edges
are elements of Zn , and the flow conservation is with respect to the group
structure. The set
supp(f ) := {e ∈ E : f (e) 6= 0}
is the support of f , and a Zn -flow f is nowhere zero if supp(f ) = E. In
this section we will be concerned with counting nowhere-zero Zn -flows, and
so we define
ϕG (n) := f nowhere-zero Zn -flow on ρ G .
A priori, the counting function ϕG (n) depends on our chosen orientation
ρ, but our language suggests that this is not the case, which we invite the
reader to verify in Exercise 1.11:
Proposition 1.6. The flow-counting function ϕG (n) is independent on the
orientation ρ of G.
A connected component of the graph G is a maximal subgraph of G in
which any two nodes are connected by a path. A graph G is connected if it
has only one connected component.2 As the reader will discover (at the latest
when working on Exercises 1.13 and 1.14), G will not have any nowhere-zero
flow if G has a bridge, that is, an edge whose removal increases the number
of connected components of G.
To motivate why we care about counting nowhere-zero flows, let’s assume
that G is a planar bridgeless graph with a given embedding into the plane.
2These notions refer to an unoriented graph.
1.2. Flows on graphs
7
The drawing of G subdivides the plane into connected regions in which two
points lie in the same region whenever they can be joined by a path in R2
that does not meet G. Two such regions are neighboring if their topological
closures share a proper (i.e., 1-dimensional) part of their boundaries. This
induces a graph structure on the subdivision of the plane: For the given
embedding of G, we define the dual graph G∗ as the graph with nodes
corresponding to the regions and two regions C1 , C2 share an edge e∗ if an
original edge e is properly contained in both their boundaries. As we can
see in the example pictured in Figure 6, the dual graph G∗ is typically not
simple with parallel edges. If G had bridges, G∗ would have loops.
Figure 6. A graph and its dual.
Given an orientation of G, an orientation on G∗ is induced by, for example,
rotating the edge clockwise. That is, the dual edge will “point” east assuming
that the primal edge “points” north.
By carefully adding G∗ to the picture we can see that dualizing G∗ recovers
G, i.e., (G∗ )∗ = G.
Our interest in flows lies in the connection to colorings: Let c be an ncoloring of G, and for a change let’s assume that c takes on colors in Zn . After
giving G an orientation, we can record the color gradients t(uv) = c(v) − c(u)
for every oriented edge u → v and, knowing the color of a single node v0 ,
we can recover the coloring from t : E → Zn : For a node v ∈ V simply
choose an undirected path v0 = p0 p1 p2 · · · pk = v from v0 to v. Then while
walking along this path we can color every node pi by adding or subtracting
t(pi−1 pi ) to c(pi−1 ) depending whether we walked the edge pi−1 pi with or
against its orientation. This process is illustrated in Figure 7. The color
c(v) is independent of the chosen path and thus, walking along a cycle in
8
1. Four Polynomials
Figure 7. Passage from colorings to dual flows and back; in this example, the computations are done in Z6 .
G the sum of the values t(e) of edges along their orientation minus those
against their orientation has to be zero. Now, via the correspondence of
primal and dual edges, t induces a map f : E ∗ → Zn on the dual graph
G∗ . Every node of G∗ represents a region that is bounded by a cycle in G,
and the orientation on G∗ is such that walking around this cycle clockwise,
each edge traversed along its orientation corresponds to a dual edge into
the region while counter-clockwise edges dually point out of the region. The
cycle condition above then proves:
Proposition 1.7. Let G be a connected planar graph with dual G∗ . For
every n-coloring c of G, the induced map f is a Zn -flow on G∗ , and every
such flow arises this way. Moreover, the coloring c is proper if and only if f
is nowhere zero.
Conversely, for a given (nowhere-zero) flow f on G∗ one can construct
a (proper) coloring on G (see Exercise 1.12). In light of all this, we can
rephrase the four-color theorem as follows:
Corollary 1.8 (Dual four-color theorem). If G is a planar bridgeless graph,
then ϕG (4) > 0.
This perspective to colorings of planar graphs was pioneered by William
Tutte who initiated the study of ϕG (n) for all (not necessary planar) graphs.
1.3. Order polynomials
9
To see how much flows differ from colorings, we observe that there is no
universal constant n0 such that every graph has a proper n0 -coloring. The
analagous statement for flows is not so clear and, in fact, Tutte conjectured
the following:
5-flow Conjecture. Every bridgeless graph has a nowhere-zero Z5 -flow.
This sounds like a rather daring conjecture, as it is not even clear that
there is any such n such that every bridgeless graph has a nowhere-zero
Zn -flow. However, it was shown by Paul Seymour that n ≤ 6 works. In
Exercise 1.17 you will show that there exist graphs that do not admit a
nowhere-zero Z4 -flow. So, similar to the history of the four-color theorem,
the gap between the conjectured and the actual truth could not be smaller.
On the enumerative side, we have the following:
Proposition 1.9. If G is a bridgeless connected graph, then ϕG (n) agrees
with a polynomial with integer coefficients of degree |E| − |V | + 1 and leading
coefficient 1.
Again, we will abuse notation and refer to ϕG (n) as the flow polynomial
of G. The proof of the polynomiality is a deletion–contraction argument
which is deferred to Exercise 1.13.
Towards a reciprocity statement, we need a notion dual to acyclic orientations: an orientation ρ on G is totally cyclic if every edge in ρ G is
contained in a directed cycle. Let’s quickly define the cyclotomic number
of G as ξ(G) := |E| − |V | + c where c is the number of connected components
of G.
Theorem 1.10. Let G be a bridgeless graph. For every positive integer n,
the evaluation (−1)ξ(G) ϕG (−n) counts the number of pairs (f, ρ) where f is
a Zn -flow and ρ is a totally-cyclic reorientation of G/ supp(f ). In particular,
(−1)ξ(G) ϕG (0) is the number of totally-cyclic orientations of G.
We will prove this theorem in Section 7.3.
1.3. Order polynomials
A partially ordered set, or poset for short, is a set Π together with a
binary relation Π that is
reflexive: a Π a,
transitive: a Π b Π c implies a Π c, and
anti-symmetric: a Π b and b Π a implies a = b,
for all a, b, c ∈ Π. We write if the poset is clear from the context.
Partially ordered sets are ubiquitous structures in combinatorics and,
as we will amply demonstrate soon, are indispensable in enumerative and
10
1. Four Polynomials
geometric combinatorics. The essence of a poset is encoded by its cover
relations: an element a ∈ Π is covered by an element b if
[a, b] := {z ∈ Π : a z b} = {a, b} ,
in plain English: a b and there is nothing “between” a and b. From its
cover relations we can recover the poset by taking the transitive closure and
adding in the reflexive relations. The cover relations can be thought of as
a directed graph, and this gives an effective way to picture a poset: The
Hasse diagram of Π is a drawing of the directed graph of cover relations in
Π as an (undirected) graph where the node a is drawn lower than the node b
whenever a ≺ b. Here is an example: For n ∈ Z>0 we define Dn as the set
[n] = {1, 2, . . . , n} ordered by divisibilty, that is, a b if a divides b. The
Hasse diagram of D10 is given in Figure 8.
Figure 8. D10 : the set [10], partially ordered by divisibility.
This example truly is a partial order as, for example, 2 and 7 are not
comparable. A poset in which every element is comparable to every other
element is a chain. To be more precise: the poset Π is a chain if we have
either a b or b a for any two elements a, b ∈ Π. The elements of a chain
are totally or linearly ordered.
A map φ : Π → Π0 is (weakly) order preserving if for all a, b ∈ Π
a Π b
=⇒
φ(a) Π0 φ(b)
and strictly order preserving if
a ≺Π b
=⇒
φ(a) ≺Π0 φ(b) .
For example, we can label the elements of a chain Π such that
Π = {a1 ≺ a2 ≺ · · · ≺ an }
which makes Π isomorphic to [n] := {1 < 2 < · · · < n}, in the sense that
there is an order preserving bijection between Π and [n].
Order preserving maps are the natural morphisms (even in a categorical
sense) between posets, and in this section we will be concerned with counting
(strictly) order preserving maps from a poset into chains. A strictly order
preserving map φ from one chain [d] into another [n] exists only if d ≤ n and
is then determined by
1 ≤ φ(1) < φ(2) < · · · < φ(d) ≤ n .
Thus, the number of such maps equals nd , the number of d-subsets of an nset. In the case of a general poset Π, we define the strict order polynomial
1.3. Order polynomials
11
Ω◦Π (n) := |{φ : Π → [n] strictly order preserving}| .
As we have just seen, Ω◦Π (n) is indeed a polynomial when Π = [d]. We now
show that polynomiality holds for all posets Π:
Proposition 1.11. For a finite poset Π, the function Ω◦Π (n) agrees with a
polynomial of degree |Π| with rational coefficients.
Proof. Let d := |Π| and φ : Π → [n] be a strictly order preserving map.
Now φ factors uniquely
ΠB
B
φ
BB σ
BB
BB
!!
/ [n]
z=
z
ι zz
z
. zzz
φ(Π)
into a surjective map σ onto φ(Π) followed by an injection ι. (Use the
functions σ(a) := φ(a) and ι(a) := a, defined with domains and codomains
pictured above.) The image φ(Π) is a subposet of a chain and so is itself
a chain. Thus Ω◦Π (n) counts the number of pairs (σ, ι) of strictly order
preserving maps Π [r] [n] for r = 1, 2, . . . , d. For fixed r, there are only
finitely many order preserving surjections σ : Π → [r], say, sr many. As we
discussed earlier, the number of strictly order preserving maps [r] → [n] is
exactly nr , which
is a rational polynomial in n of degree r. Hence, for fixed
n
r, there are sr r many pairs (σ, ι) and we obtain
n
n
n
Ω◦Π (n) = sd
+ sd−1
+ · · · + s1
,
d
d−1
1
which finishes our proof.
◦
As an aside, the above proposition
n proves that
ΩΠ (n) is a polynomial
with integral coefficients if we use r : r ∈ Z≥0 as a basis for R[n]. That
the binomial coefficients indeed form a basis for the univariate polynomials
follows from the proposition above: If Π is an antichain on d elements, i.e.,
a poset in which no elements are related,
n
n
n
◦
d
ΩΠ (n) = n = sd
+ sd−1
+ · · · + s1
.
(1.2)
d
d−1
1
In this case, the coefficients sr = S(d, r) are the Stirling numbers of the
second kind which count the number of surjective maps [d] [r]. (The
Stirling numbers might come in handy in Exercise 1.10.)
For the case that Π is a d-chain, the reciprocity statement (0.2) says
that (−1)d Ω◦Π (−n) gives the number of d-multisubsets of an n-set, which
equals, in turn, the number of (weak) order preserving maps from a d-chain
12
1. Four Polynomials
to an n-chain. Our next combinatorial reciprocity theorem expresses this
duality between weak and strict order preserving maps from a general poset
into chains. You can already guess what is coming. We define the order
polynomial
ΩΠ (n) = |{φ : Π → [n] order preserving}| .
A slight modification (which we invite the reader to check in Exercise 1.19)
of our proof of Proposition 1.11 implies that ΩΠ (n) indeed agrees with
polynomial in n, and the following reciprocity theorem gives the relationship
between the two polynomials ΩΠ (n) and Ω◦Π (n).
Theorem 1.12. Let Π be a finite poset. Then
(−1)|Π| Ω◦Π (−n) = ΩΠ (n) .
We will prove this theorem in Chapter 2. To further motivate the study
of order polynomials, we remark that a poset Π gives rise to an oriented
graph by way of the cover relations of Π. Conversely, the binary relation
given by an oriented graph G can be completed to a partial order Π(G) by
adding the necessary transitive and reflexive relations if and only if G is
acyclic. Figure 9 shows an example, for the orientation pictured in Figure 3.
The following result will be the subject of Exercise 1.18.
Figure 9. From an acyclic orientation to a poset.
Proposition 1.13. Let ρ G = (V, E, ρ) be an acyclic graph and Π(ρ G) the
induced poset. A map c : V → [n] is strictly compatible with the orientation
ρ of G if and only if c is a strictly order preserving map Π(ρ G) → [n].
In Proposition 1.4 we identified the number of n-colorings χG (n) of G as
the number of colorings c strictly compatible with some acyclic orientation ρ
of G, and so this proves:
Corollary 1.14. The chromatic polynomial χG (n) of a graph G is the sum
of the order polynomials Ω◦Π(ρ G) (n) for all acyclic orientations ρ of G.
1.4. Ehrhart polynomials
13
1.4. Ehrhart polynomials
The formulation of (0.1) in terms of d-subsets of a n-set has a straightforward
geometric interpretation that will fuel much of what is about to come:
The d-subsets of [n] correspond precisely to the points in Rd with integral
coordinates in the set
n
o
(n + 1)∆◦d = (x1 , . . . , xd ) ∈ Rd : 0 < x1 < x2 < · · · < xd < n + 1 . (1.3)
Let’s explain the notation on the left-hand side: we define
n
o
∆◦d := (x1 , . . . , xd ) ∈ Rd : 0 < x1 < x2 < · · · < xd < 1 ,
and for a set S ⊆ Rd and a positive integer n, we define
n S := {n x : x ∈ S} ,
the n-th dilate of S. (We hope the notation in (1.3) now makes sense.) For
example, when d = 2,
∆◦2 = (x1 , x2 ) ∈ R2 : 0 < x1 < x2 < 1
is the interior of a triangle, and every integer point (x1 , x2 ) in the (n + 1)-st
dilate of ∆◦2 satisfies 0 < x1 < x2 < n + 1 or, equivalently, 1 ≤ x1 < x2 ≤ n.
We illustrate these integer points for the case n = 5 in Figure 10.
x2
x1 = x2
x2 = 6
x1
Figure 10. The integer points in 6∆◦2 .
A convex lattice polygon P ⊂ R2 is the smallest convex set containing
a given finite set of non-collinear integer points in the plane. The interior
of P is denoted by P◦ . Convex polygons are 2-dimensional instances of
convex polytopes, which live in any dimension and whose properties we
14
1. Four Polynomials
will study in detail in Chapter 3. For now, we count on the reader’s intuition
about objects such as vertices and edges of a polygon, which will be defined
rigorously later on.
For a bounded set S ⊂ R2 , we write E(S) := S ∩ Z2 for the number
of lattice points in S. Our example above motivates the definitions of the
counting functions
ehrP◦ (n) := E(nP◦ ) = nP◦ ∩ Z2 and
ehrP (n) := E(nP) = nP ∩ Z2 ,
called the Ehrhart functions of P, for historical reasons that will be given
in Chapter 4.
As we know from (0.1), the number of integer lattice points in the
(n + 1)-st dilate of ∆◦2 is given by the polynomial
n
ehr∆◦2 (n + 1) =
.
2
To make the combinatorial reciprocity statement given by (0.1) geometric,
we observe that the number of weak order preserving maps from [n] into [2]
is given by the integer points in the (n − 1)-st dilate of
∆2 = (x1 , x2 ) ∈ R2 : 0 ≤ x1 ≤ x2 ≤ 1 ,
the closure of ∆◦2 . The combinatorial reciprocity statement given by (0.1) now
reads (−1)2 −n
equals the number of integer points in (n − 1)∆2 . Unraveling
2
the parameters (and making appropriate shifts), we can rephrase this as:
(−1)2 ehr∆◦2 (−n) equals the number of integer points in n∆2 . The reciprocity
theorem featured in this section states that this holds for all convex lattice
polygons; in Section 5.6 we will prove an analogue in all dimensions.
Theorem 1.15. Let P ⊂ R2 be a lattice polygon. Then ehrP (n) agrees with
a polynomial of degree 2 with rational coefficients, and (−1)2 ehrP (−n) equals
the number of integer points in nP◦ .
In the remainder of this section we will prove this theorem. The proof
will be a series of simplifying steps that are similar in spirit to those that we
will employ for the general result in Section 5.6.
As a first step, we reduce the problem of showing polynomiality for
Ehrhart functions of arbitrary lattice polygons to that of lattice triangles.
Let P be a lattice polygon in the plane with n vertices. We can triangulate P
by cutting the polygon along sufficiently many (exactly n−3) non-intersecting
diagonals, as in Figure 11. The result is a set of n − 2 lattice triangles that
cover P. We denote by T the collection of faces of all these triangles, that is,
T consists of n zero-dimensional polytopes (vertices), 2n − 3 one-dimensional
polytopes (edges), and n − 2 two-dimensional polytopes (triangles).
1.4. Ehrhart polynomials
15
Figure 11. A triangulation of a hexagon.
Our triangulation is a well-behaved collection of polytopes in the plane
in the sense that they intersect nicely: if two elements of T intersect, then
they intersect in a common face of both. This is useful, as counting lattice
points is a valuation:3 For S, T ⊂ R2 ,
E(S ∪ T ) = E(S) + E(T ) − E(S ∩ T ) ,
(1.4)
and applying the inclusion-exclusion relation (1.4) repeatedly to the elements
in our triangulation of P yields
X
ehrP (n) =
µT (F) ehrF (n) ,
(1.5)
F∈T
where the µT (F)’s are some coefficients that correct for over-counting. If F is
a triangle, then µT (F) = 1—after all, we want to count the lattice points in
P which are covered by the triangles. For an edge F of the triangulation, we
have to make the following distinction: F is an interior edge of T if it is
contained in two triangles. In this case the lattice points in F get counted
twice, and in order to compensate for this, we set µT (F) = −1. In the case
that F is a boundary edge, i.e., F lies in only one triangle of T , there is no
over-counting and we can set µT (F) = 0. To generalize this to all faces of
T , let’s call a face F ∈ T a boundary face of T if F ⊂ ∂P and an interior
face otherwise. We can give the coefficients µT (F) explicitly as follows.
Proposition 1.16. Let T be a triangulation of a lattice polygon P ⊂ R2 .
Then the coefficients µT (F) in (1.5) are given by
(
(−1)2−dim F if F is interior,
µT (F) =
0
otherwise.
For boundary vertices F = {v}, we can check that µT (F) = 0 is correct:
the vertex is counted positively as a lattice point by every incident triangle and
negatively by every incident interior edge. As there are exactly one interior
3We’ll have more to say about valuations in Section 3.3.
16
1. Four Polynomials
edge less than incident triangles, we do not count the vertex more than once.
For an interior vertex, the number of incident triangles and incident (interior)
edges are equal and hence µT (F) = 1. (In triangulations of P obtained by
cutting along diagonals we never encounter interior vertices, however, they
will appear soon when we consider a different type of triangulation.)
The coefficient µT (F) for a triangulation of a polygon was easy to argue
and to verify in the plane. For higher-dimensional polytopes we will have
to resort to more algebraic and geometric means. The right algebraic setup
will be discussed in Chapter 2 where we will make use of the fact that a
triangulation T constitutes a partially ordered set. In the language of posets,
µT (F) is an evaluation of the M¨
obius function for the poset T . M¨obius
functions are esthetically satisfying but are in general difficult to compute.
However, we are dealing with situations with plenty of geometry involved,
and we will make use of that in Chapter 5 to give a statement analogous to
Proposition 1.16 in general dimension.
Returning to our 2-dimensional setting, showing that ehrF (n) is a polynomial whenever F is a lattice point, a lattice segment, or a lattice triangle
gives us the first half of Theorem 1.15. If F is a vertex, then ehrF (n) = 1. If
F ∈ T is an edge of one of the triangles and thus a lattice segment, verifying
that ehrF (n) is a polynomial is the content of Exercise 1.20.
The last challenge now is the reciprocity for lattice triangles. For the
rest of this section, let 4 ⊂ R2 be a fixed lattice triangle in the plane. The
idea that we will use is to triangulate the dilates n4 for n ≥ 1, but the
triangulation will change with n. Figure 12 gives the picture for n = 1, 2, 3.
Figure 12. Special triangulations of dilates of a lattice triangle.
We trust that the reader can imagine the triangulation for all values of n.
The special property of this triangulation is that up to lattice translations,
there are only few different pieces. In fact, there are only two different
lattice triangles used in the triangulation of n4: there is 4 itself and (lattice
translates of) the reflection of 4 in the origin, which we will denote by .
As for edges, we have three different kinds of edges, namely the edges , —,
and . Up to lattice translation, there is only one vertex •.
—
—
4
1.4. Ehrhart polynomials
17
Now let’s count how many copies of each tile occur in these special
triangulations; let t(P, n) denote the number of times P appears in our
triangulation of n4. For triangles, we count
n+1
n
t(4, n) =
and t( , n) =
.
2
2
4
For the interior edges, we observe that each interior edge is incident to a
unique upside-down triangle
and consequently
n
t( , n) = t(—, n) = t( , n) =
.
2
—
—
4
Similarly, for interior vertices,
t(•, n) =
n−1
.
2
Thus with (1.5), the Ehrhart function for the triangle 4 is
n+1
n
ehr4 (n) =
E(4) +
E( )
2
2
n
−
E( ) + E(—) + E( )
2
n−1
+
E(•) .
2
—
—
4
(1.6)
This proves that ehr4 (n) agrees with a polynomial of degree 2, and together
with (1.5) this establishes the first half of Theorem 1.15.
To prove the combinatorial reciprocity of Ehrhart polynomials in the
plane, let’s make the following useful observation.
Proposition 1.17. If for every lattice polygon P ⊂ R2 we have that ehrP (−1)
equals the number of lattice points in the interior of P, then ehrP (−n) =
E(nP◦ ) for all n ≥ 1.
Proof. For fixed n ≥ 1, let’s denote by Q the lattice polygon nP. We see
that ehrQ (m) = E(m(nP)) for all m ≥ 1. Hence the Ehrhart polynomial of
Q is given by ehrP (mn) and for m = −1 we conclude
ehrP (−n) = ehrQ (−1) = E(Q◦ ) = E(nP◦ )
which finishes our proof.
To establish the combinatorial reciprocity of Theorem 1.15 for triangles,
we can simply substitute n = −1 into (1.6) and use (0.2) to obtain
—
) − E(—) − E(
—
ehr4 (−1) = E( ) − E(
) + 3E(•) ,
4
which equals the number of interior lattice points of . Observing that 4
and
have the same number of lattice points finishes the argument.
4
4
18
1. Four Polynomials
For the general case, Exercise 1.20 gives
X
ehrP (−1) =
E(F◦ ) = E(P◦ )
F∈T
and this (finally!) concludes our proof of Theorem 1.15.
Exercises 1.20 and 1.22 also answer the question why we carefully cut P
along diagonals (as opposed to cutting it up arbitrarily to obtain triangles):
Theorem 1.15 is only true for lattice polygons. There are versions for polygons with rational and irrational coordinates but they become increasingly
complicated. By cutting along diagonals we can decompose a lattice polygon
into lattice segments and lattice triangles. This part becomes nontrivial
already in dimension 3, and we will worry about this in Chapter 5.
In Exercise 1.24 we will look into the question as to what the coefficients
of ehrP (n), for a lattice polygon P, tell us. Let’s finish this chapter by
considering the constant coeffient c0 = ehrP (0). This is the most tricky
one, as we could argue that ehrP (0) = E(0P) and since 0P is just a single
point, we get c0 = 1. This argument is flawed: we defined ehrP (n) only for
n ≥ 1. To see that this argument is, in fact, plainly wrong, let’s consider
S = P1 ∪ P2 ⊂ R2 , where P1 and P2 are disjoint lattice polygons. Since they
are disjoint, ehrS (n) = ehrP1 (n) + ehrP2 (n). Now 0S is also just a point and
therefore
1 = ehrS (0) = ehrP1 (0) + ehrP2 (0) = 2 .
It turns out that c0 = 1 is still correct but the justification will occupy most
of Section 5.2. In Exercise 1.25, you will prove a more general version for
Theorem 1.15 that dispenses of convexity.
Notes
Graph-coloring problems started as map-coloring problems, and so the fact
that the chromatic polynomial is indeed a polynomial was proved for maps
(in 1912 by George Birkhoff [16]) before Hassler Whitney proved it for graphs
in 1932 [83]. As we mentioned, the first proof of the four-color theorem is
due to Kenneth Appel and Wolfgang Haken [5, 6]. Theorem 1.5 is due to
Richard Stanley [72]. We will give a proof from a geometric point of view in
Section 6.5.
As already mentioned, the approach of studying colorings of planar graphs
through flows on their duals was pioneered by William Tutte [82], who also
conceived the 5-flow Conjecture. This conjecture becomes a theorem when
“5” is replaced by “6”, due to Paul Seymour [66]; the 8-flow theorem had
previously been shown by Fran¸cois Jaeger [?, 44]. Theorem 1.10 is due to
Felix Breuer and Raman Sanyal [18]. We will give a proof in Section 7.3.
Order polynomials were introduced by Richard Stanley [71, 76] as
‘chromatic-like polynomials for posets’ (this is reflected in Corollary 1.14);
Exercises
19
Theorem 1.12 is due to him. We will study order polynomials in depth in
Chapters 2 and 6.
Theorem 1.15 is essentially due to Georg Pick [57], whose famous formula
is the subject of Exercise 1.24. In some sense, this formula marks the
beginning of the study of integer-point enumeration in polytopes. Our
phrasing of Theorem 1.15 suggests that it has an analogue in higher dimension,
and we will study this analogue in Chapters 4 and 5.
Herbert Wilf [84] raised the question of characterizing which polynomials
can occur as chromatic polynomials of graphs. This question has spawned a
lot of work in the algebraic combinatorics. For example, a recent theorem of
June Huh [42] says that the absolute values of the coefficients of any chromatic
polynomial form a unimodal sequence, that is, the sequence increases up to
some point, after which it decreases. Huh’s theorem had been conjectured by
Ronald Read [60] almost 50 years earlier. In fact, Huh proved much more.
In Chapter 7 we will study arrangements of hyperplanes and their associated
characteristic polynomials. Huh and later Huh and Eric Katz [43] proved
that, up to sign, the coefficients of characteristic polynomials of hyperplane
arrangements (defined over any field) form a log-concave sequence. We
will see the relation between chromatic and characteristic polynomials in
Chapter 7.
We finish by mentioning one general problem, which fits the “big picture”
point-of-view of this introductory chapter: classify chromatic, flow, order,
and Ehrhart polynomials. What we mean is: give conditions on real numbers
a0 , a1 , . . . , ad that allow us to detect whether or not a given polynomial
ad nd + ad−1 nd−1 + · · · + a0 is, say, a chromatic polynomial.
Exercises
1.1 Let f (t) = ad td + ad−1 td−1 + · · · + a0 ∈ R[t] be a polynomial such
that f (n) is an integer for every integer n > 0. Give a proof or a
counterexample for the following statements.
(a) All coefficients ai are integers.
(b) f (n) is an integer for all n ∈ Z.
(c) If (−1)k f (−n) ≥ 0 for all n > 0, then k = deg(f ).
1.2 Two graphs G1 = (V1 , E1 ) and G2 = (V2 , E2 ) are isomorphic if there
is a bijection φ : V1 → V2 such that for all u, v ∈ V1
uv ∈ E1
if and only if
φ(u) φ(v) ∈ E2 .
20
1. Four Polynomials
Let G be a planar graph and let G1 and G2 be the dual graphs for
two distinct planar embeddings of G. Is it true that G1 and G2 are
isomorphic?
If not, can you give a sufficient condition on G such that the
above claim is true? (A precise characterization is rather difficult, but
for a sufficient condition you might want to contemplate Steinitz’s
theorem [88, Ch. 4].)
1.3 Find two simple non-isomorphic graphs G and H with χG (n) = χH (n).
Can you find many (polynomial, exponential) such examples in the
number of nodes? Can you make your examples arbitrarily high connected?
1.4 Find the chromatic polynomials of
(a) the wheel with d spokes (and d + 1 nodes); for example, the wheel
with 6 spokes is this:
(b) the cycle on d nodes;
(c) the path on d nodes.
1.5 Show that if G has c connected components, then nc divides the polynomial χG (n).
1.6 Complete the proof of Corollary 1.2: Let G be a loopless nonempty
graph on d nodes and χG (n) = cd nd + cd−1 nd−1 + · · · + c0 its chromatic
polynomial. Then
(a) the leading coefficient cd = 1;
(b) the constant coefficient c0 = 0;
(c) (−1)d χG (−n) > 0.
1.7 Prove that every complete graph Kd (a graph with d nodes and all
possible edges between them) has exactly d! acyclic orientations.
1.8 Find and prove a deletion–contraction formula for the number of acyclic
orientations of a given graph.
1.9 In this exercise you will give a deletion–contraction proof of Theorem 1.5.
(a) Verify that the deletion–contraction relation (1.1) implies for the
function χG (n) := (−1)n χG (−n) that
χG (n) = χG\e (n) + χG/e (n).
(b) Define XG (n) as the number of compatible pairs acyclic orientation ρ and n-coloring c. Show XG (n) satisfies the same deletion–
contraction relation as χG (n).
(c) Infer that χG (n) = XG (n) by induction on |E|.
Exercises
21
1.10 The complete bipartite graph Kr,s is the graph on the vertex set
V = {1, 2, . . . , r, 10 , 20 , . . . , s0 } and edges E = {ij 0 : 1 ≤ i ≤ r, 1 ≤
j ≤ s}. Determine the chromatic polynomial χKr,s (n) for m, k ≥ 1.
(Hint: Proper n-colorings of Kr,s correspond to pairs (f, g) of maps
f : [r] → [n] and g : [s] → [n] with disjoint ranges.)
1.11 Prove Proposition 1.6: The flow-counting function ϕG (n) is independent
on the orientation of G.
1.12 Let G be a connected planar graph with dual G∗ . By reversing the steps
in our proof before Proposition 1.7, show that every (nowhere-zero)
Zn -flow f on G∗ naturally gives rise to n different (proper) n-colorings
on G.
1.13 Prove Proposition 1.9: If G is a bridgeless connected graph, then ϕG (n)
agrees with a monic polynomial of degree |E| − |V | + 1 with integer
coefficients.
1.14 Let G = (V, E) be a graph, and let n be a positive integer.
(a) Prove that
ϕG (n) 6= 0
ϕG (n + 1) 6= 0 .
implies
(Hint: use Tutte’s Lemma [?].)
(b) Even stronger, prove that
ϕG (n) ≤ ϕG (n + 1) .
(This is nontrivial. But you will easily prove this after having read
Chapter 7.)
1.15 Let ρ G = (V, E, ρ) be an oriented graph and n ≥ 2.
(a) Let f : E → Zn be a nowhere-zero Zn -flow and let e ∈ E. Show
that
f : E \ {e} → Zn
is a nowhere-zero Zn -flow on the contraction ρ G/e.
(b) For S ⊆ V let E in (S) be the in-coming edges, i.e., u → v with
v ∈ S and u ∈ V \ S and E out (S) the out-going edges. Show that
f : E → Zn is a nowhere-zero Zn -flow if and only if
X
X
f (e) =
f (e)
e∈E in (S)
e∈E out (S)
for all S ⊆ V . (Hint: For the sufficiency contract all edges in S and
V \ S.)
(c) Infer that ϕG ≡ 0 if G has a bridge.
1.16 Discover the notion of tensions.
1.17 Consider the Petersen graph G pictured in Figure 13.
(a) Show that ϕG (4) = 0.
22
1. Four Polynomials
Figure 13. The Petersen graph.
(b) Show that the polynomial ϕG (n) has nonreal roots.
(c) Construct a planar4 graph whose flow polynomial has nonreal roots.
(Hint: think of the dual coloring question.)
1.18 Prove Proposition 1.13: Let ρ G = (V, E, ρ) be an acyclic graph and
Π = Π(ρ G) the induced poset. A map c : V → [n] is strictly compatible
with the orientation ρ of G if and only if c is a strictly order preserving
map Π → [n].
1.19 Show that ΩΠ (n) is a polynomial in n.
with a1 , a2 , b1 , b2 ∈ Z, a lattice segment.
1.20 Let S = conv ab11 , ab22
Show that
ehrS (n) = L n + 1
where L = gcd(a2 − a1 , b2 − b1 ), the lattice length of S. Conclude
further that − ehrS (−n) equals the number of lattice points of nS other
than the endpoints, in other words,
(−1)dim S ehrS (−n) = ehrS ◦ (n) .
Can you find an explicit formula for ehrS (n) when S is a segment
with rational endpoints?
1.21 Let O be a closed polygonal lattice path, i.e., the union of lattice
segments, such that any vertex on O lies on precisely two such segments,
and that topologically O is a closed curve. Show that
ehrO (n) = L n
where L is the sum of the lattice lengths of the lattice segments that
make up O or, equivalently, the number of lattice points on O.
1.22 Let v1 , v2 ∈ Z2 , and let Q be the half-open parallelogram
Q := {λv1 + µv2 : 0 ≤ λ, µ < 1}
Show (for example, by tiling the plane by translates of Q) that
ehrQ (n) = A n2
4The Petersen graph is a (famous) example of a nonplanar graph.
Exercises
23
a c where A = det
.
b d 1.23 A lattice triangle conv{v1 , v2 , v3 } is unimodular if v2 −v1 and v3 −v1
form a lattice basis of Z2 .
(a) Prove that a lattice triangle is unimodular if and only if it has
area 12 .
(b) Conclude that for any two unimodular triangles ∆1 and ∆2 , there
exist T ∈ GL2 (Z) and x ∈ Z2 such that ∆2 = T (∆1 ) + x.
(c) Compute the Ehrhart polynomials of all unimodular triangles.
(d) Show that any lattice polygon can be triangulated into unimodular
triangles.
(e) Use the above facts to give an alternative proof of Theorem 1.15.
1.24 Let P ⊂ R2 be a lattice polygon, denote the area of P by A, the number
of integer points inside the polygon P by I, and the number of integer
points on the boundary of P by B. Prove that
A = I + 12 B − 1
(a famous formula due to Georg Alexander Pick). Deduce from this
formulas for the coefficients of the Ehrhart polynomial of P.
1.25 Let P, Q ⊂ R2 be lattice polygons, such that Q is contained in the
interior of P. Generalize Exercise 1.24 (i.e., both a version of Pick’s
theorem and the accompanying Ehrhart polynomial) to the “polygon
with a hole” P − Q. Generalize your formulas to a lattice polygon with
n “holes” (instead of one).
Chapter 2
Partially Ordered Sets
The mathematical phenomenon always develops out of simple arithmetic, so useful
in everyday life, out of numbers, those weapons of the gods: the gods are there,
behind the wall, at play with numbers.
Le Corbusier
Partially ordered sets, posets for short, made an appearance twice so far.
First (in Section 1.3) as a class of interesting combinatorial objects with a rich
counting theory that is intimately related to graph colorings and, second (in
Section 1.4), as a natural book-keeping structure for geometric subdivisions
of polygons. In particular, the stage for the principle of overcounting-andcorrecting, more commonly referred to as inclusion–exclusion, is naturally
set in the theory of posets. Our agenda in this chapter is twofold: we need to
introduce machinery that will be crucial tools in later chapters, but we will
also prove our first combinatorial reciprocity theorems in a general setting,
from first principles—that is, without the geometric intuition that will appear
in later chapters (where we will revisit the theorems from this chapter in
light of the geometry). Let’s recall that for us, a poset Π is a finite set with
a binary relation Π that is reflexive, transitive, and anti-symmetric.
2.1. Order Ideals and the Incidence Algebra
Let’s go back to the problem of counting (via ΩΠ (n)) order preserving maps
φ : Π → [n] which satisfy
a Π b
=⇒
φ(a) ≤ φ(b)
for all a, b ∈ Π. The preimages φ−1 (j), for j = 1, 2, . . . , n, partition Π and
uniquely identify φ, but from a poset point of view they do not have enough
structure. A better perspective comes from the following observation: Let
25
26
2. Partially Ordered Sets
φ : Π → [2] be an order preserving map into the 2-chain, and let I := φ−1 (1).
Now
y ∈ I and x Π y =⇒ x ∈ I .
A subset I ⊆ Π with this property is called an order ideal of Π. Conversely,
if I ⊆ Π is an order ideal, then φ : Π → [2] with φ−1 (1) = I defines an order
preserving map. Thus, order preserving maps φ : Π → [2] are in bijection
with the order ideals of Π, of which there are exactly ΩΠ (2) many. Dually,
the complement F = Π \ I of an order ideal I is characterized by the property
that x y ∈ F implies x ∈ F . Such a set is called a dual order ideal or
filter.
To characterize general order preserving maps into chains in terms of Π,
we note that every order ideal of [n] is principal, that is, every order ideal
I ⊆ [n] is of the form
I = {j ∈ [n] : j ≤ k} = [k]
for some k. In particular, if φ : Π → [n] is order preserving, then the preimage
φ−1 ([k]) of an order ideal [k] ⊆ [n] is an order ideal of Π, and this gives us
the following bijection.
Proposition 2.1. Order preserving maps φ : Π → [n] are in bijection with
multichains1 of order ideals
∅ = I0 ⊆ I1 ⊆ I2 ⊆ · · · ⊆ In = Π
of length n. The map φ is strictly order preserving if and only if Ij \ Ij−1 is
an antichain for all j = 1, 2, . . . , n.
Proof. We only need to argue the second part. We observe that φ is strictly
order preserving if and only if there are no elements x ≺ y with φ(x) = φ(y).
Hence, φ is strictly order preserving if and only if φ−1 (j) = Ij \ Ij−1 does
not contain a pair of comparable elements.
The collection J (Π) of order ideals of Π is itself a poset under set inclusion,
which we call the lattice of order ideals or the Birkhoff lattice2 of Π.
What we just showed is that ΩΠ (n) counts the number of multichains in
J (Π) \ {∅, Π}. The next problem we address is counting multichains of
length n in general posets. To that end, we introduce an algebraic gadget:
The incidence algebra I(Π) is a C-vector space spanned by those functions
α : Π × Π → C that satisfy
α(x, y) = 0
whenever
x 6 y .
1A multichain is a sequence of comparable elements, where we allow repetition.
2The reason for this terminology will become clear shortly.
2.1. Order Ideals and the Incidence Algebra
27
We define the product of α, β : Π × Π → C as
X
α(r, s)β(s, t) ,
(α ∗ β)(r, t) :=
rst
and together with δ ∈ I(Π) defined by
(
0
δ(x, y) :=
1
if x 6= y,
if x = y,
(2.1)
this gives I(Π) the structure of an associative C-algebra with unit δ. (Those
readers for whom this is starting to feel like linear algebra are on the right
track.) A distinguished role is played by the zeta function ζ ∈ I(Π) defined
by
(
1 if x y,
ζ(x, y) :=
0 otherwise.
For the time being, the power of zeta functions lies in their powers.
Proposition 2.2. Let Π be a finite poset and x, y ∈ Π. Then ζ n (x, y) equals
the number of multichains
x = x0 x1 · · · xn = y
of length n.
Proof. For n = 1, we have that ζ(x, y) = 1 if and only if x = x0 x1 = y.
Arguing by induction, we can assume that ζ n−1 (x, y) is the number of
multichains of length n − 1 for all x, y ∈ Π, and we calculate
X
ζ n (x, z) = (ζ n−1 ∗ ζ)(x, z) =
ζ n−1 (x, y) ζ(y, z) .
xyz
Every summand on the right is the number of multichains of length n − 1
ending in y that can be extended to z.
x5
x2
x3
x4
x1
Figure 1. A sample poset.
28
2. Partially Ordered Sets
As an example, the zeta function for the poset in Figure 1 is given in
matrix form as


1
1
1
1
1
 0
1
0
0
1 


 0
.
0
1
0
1


 0
0
0
1
1 
0
0
0
0
1
We encourage the reader to see Proposition 2.2 in action by computing
powers of this matrix.
As a first milestone, Proposition 2.2 implies the following representation
of the order polynomial of Π which we introduced in Section 1.3.
Corollary 2.3. For a finite poset Π, let ζ be the zeta function of J (Π), the
lattice of order ideals in Π. The order polynomial associated with Π is given
by
ΩΠ (n) = ζ n (∅, Π) .
Identifying ΩΠ (n) with the evaluation of a power of ζ does not suggest
that ΩΠ (n) is the restriction of a polynomial (which we know to be true from
Exercise 1.19) but this impression is misleading: Let η ∈ I(Π) be defined by
(
1 if x ≺ y,
η(x, y) :=
(2.2)
0 otherwise.
Then ζ = δ + η and hence, using the binomial theorem (Exercise 2.1),
n X
n k
ζ n (x, y) = (δ + η)n (x, y) =
η (x, y) .
(2.3)
k
k=0
Exercise 2.3 asserts that the sum on the right stops at the index k = |Π| and
is thus a polynomial in n of degree ≤ |Π|.
The arguments in the preceding paragraph are not restricted to posets
formed by order ideals, but hold more generally for any poset Π that has
a minimum ˆ
0 and a maximum ˆ1, i.e., ˆ0 and ˆ1 are elements in Π that
ˆ
satisfy 0 x ˆ
1 for all x ∈ Π. (For example, the Birkhoff lattice J (Π) has
minimum ∅ and maximum Π.) Let’s record this:
Proposition 2.4. Let Π be a finite poset with minimum ˆ0, maximum ˆ1, and
zeta function ζ. Then ζ n (ˆ
0, ˆ1) is a polynomial in n.
To establish the reciprocity theorem for ΩΠ (n), we’d like to evaluate
ζ n (∅, Π) at negative integers n, so we first need to understand when an
element α ∈ I(Π) is invertible. To this end, let’s pause and make the
incidence algebra a bit more tangible.
Choose a linear extension of Π, that is, we label the d = |Π| elements
of Π by p1 , p2 , . . . , pd such that pi pj implies i ≤ j. (That such a labeling
2.2. The M¨
obius Function and Order Polynomial Reciprocity
29
exists is the content of Exercise 2.2.) This allows us to identify I(Π) with a
subalgebra of the upper triangular (d × d)-matrices by setting
α = (α(pi , pj ))1≤i,j≤d .
For example, for the poset D10 given in Figure 8, a linear extension is given by
(p1 , p2 , . . . , p10 ) = (1, 5, 2, 3, 7, 10, 4, 6, 9, 8) and the incidence algebra consists
of matrices of the form

1
5

2

3

7

10 

4

6

9
8
1
∗
5 2
∗ ∗
∗
∗
3
∗
∗
7
∗
10 4 6 9 8

∗ ∗ ∗ ∗ ∗

∗

∗ ∗ ∗
∗


∗ ∗


∗


∗

∗
∗


∗


∗
∗
where the stars are the possible non-zero entries for the elements in I(Π). This
linear-algebra perspective affords a simple criterion for when α is invertible.
Proposition 2.5. A transformation α ∈ I(Π) is invertible if and only if
α(x, x) 6= 0 for all x ∈ Π.
2.2. The M¨
obius Function and Order Polynomial
Reciprocity
We now return to the stage set up by Corollary 2.3, namely,
ΩΠ (n) = ζJn (Π) (∅, Π) .
We’d like to use this setup to compute ΩΠ (−n); thus we need to invert the
zeta function of J (Π). Such an inverse exists by Proposition 2.5, and we call
µ := ζ −1 the M¨
obius function. For example, the M¨obius function of the
poset in Figure 1 is given in matrix form as


1 −1 −1 −1
2
 0
1
0
0 −1 


 0
0
1
0 −1 

.
 0
0
0
1 −1 
0
0
0
0
1
It is apparent that one can compute the M¨obius function recursively, and
in fact, unravelling the condition that (µ ∗ ζ)(x, z) = δ(x, z) for all x, z ∈ Π
30
2. Partially Ordered Sets
gives
X
µ(x, z) = −
X
µ(y, z) = −
µ(x, y)
for x ≺ z and
xy≺z
x≺yz
(2.4)
µ(x, x) = 1 .
As a notational remark, the functions ζ, δ, µ, and η depend on the
underlying poset Π, so we will sometimes write ζΠ , δΠ , etc., to make this
dependence clear. For an example, let’s consider the M¨obius function of the
Boolean lattice Bd , the partially ordered set of all subsets of [d] ordered
by inclusion. For two subsets S ⊆ T ⊆ [d], we have that µBd (S, T ) = 1
whenever S = T and µBr (S, T ) = −1 whenever |T \ S| = 1. Although this is
little data, we venture that
µBd (S, T ) = (−1)|T \S| .
(2.5)
We dare the reader to prove this from first principles, or to appeal to the
results in Exercise 2.4 after realizing that Bd is the d-fold product of a
2-chain.
Towards proving the combinatorial reciprocity theorem for order polynomials (Theorem 1.12) we note the following.
Proposition 2.6.
ΩΠ (−n) = ζJ−n (∅, Π) = µnJ (∅, Π) ,
where J = J (Π) is the Birkhoff lattice of Π.
This proposition is strongly suggested by our notation but nevertheless
requires a proof.
Proof. Recall that ζ = δ + η. Hence
ζ −1 = (δ + η)−1 = δ − η + η 2 − · · · + (−1)d η d ,
by the geometric series and Exercise 2.3. If we now take powers of ζ −1 and
again appeal to the nilpotency of η, we calculate
d
X
−n
k n+k−1
ζ
=
(−1)
ηk .
k
k=0
ζn
Thus the expression of
as a polynomial given in (2.3), together with the
fundamental combinatorial reciprocity for binomial coefficients (0.2) given in
the very beginning of the book, proves the claim.
By Proposition 2.2, the right-hand side of the identity in Proposition 2.6
is
µnJ (∅, Π) =
X
µJ (I0 , I1 ) µJ (I1 , I2 ) · · · µJ (In−1 , In )
(2.6)
2.2. The M¨
obius Function and Order Polynomial Reciprocity
31
where the sum is over all multichains of order ideals
∅ = I0 ⊆ I1 ⊆ · · · ⊆ In = Π
of length n. Our next goal is thus to understand the evaluation µJ (K, M )
where K ⊆ M ⊆ Π are order ideals. This evaluation depends only on
[K, M ] := {L ∈ J : K ⊆ L ⊆ M } ,
the interval from K to M in the Birkhoff lattice J .
Theorem 2.7. Let Π be a finite poset and K ⊆ M order ideals in J = J (Π).
Then
(
(−1)|M \K| if M \ K is an antichain,
µJ (K, M ) =
0
otherwise.
Proof. Let’s first consider the (easier) case that M \ K is an antichain. In
this case K ∪ A is an order ideal for all A ⊆ M \ K. In other words, the
interval [K, M ] is isomorphic3 to the Boolean lattice Br for r = |M \ K| and
hence, with (2.5), we conclude µJ (K, M ) = (−1)r .
The case that M \ K contains comparable elements is a bit more tricky.
We can argue by induction on the length of the interval [K, M ]. The base
case is given by the situation that M \ K consists of exactly two comparable
elements m ≺ M and hence
µJ (K, M ) = − µJ (K, K) − µJ (K, K ∪ {m}) = −1 − (−1) = 0
since K ∪ {m} is an order ideal that covers K in J .
For the induction step, we use (2.4), i.e.,
X
µJ (K, M ) = −
µJ (K, L)
where the sum is over all order ideals L such that K ⊆ L ⊂ M . By induction
hypothesis, µJ (K, L) is zero unless L \ K is an antichain and thus
X
K ⊆ L ⊂ M order ideal,
|L\K|
µJ (K, M ) = −
(−1)
:
,
L \ K is an antichain
where we have used the already-proven part of the theorem. Now let m ∈
M \ K be a minimal element. The order ideals L in the above sum can be
partitioned into those containing m and those who do not. Both parts of this
partition have the same size: if m 6∈ L, then L ∪ {m} is also an order ideal;
if m ∈ L, then L \ {m} is an admissible order ideal as well. (You should
check this.) Hence, the positive and negative terms cancel each other and
µJ (K, M ) = 0.
With this we can give a (purely combinatorial) proof of Theorem 1.12.
3Two posets Π and Θ are isomorphic if there is a bijection f : Π → Θ that satisfies
x Π y ⇐⇒ f (x) Θ f (y).
32
2. Partially Ordered Sets
Proof of Theorem 1.12. With Theorem 2.7, the right-hand side of (2.6)
becomes (−1)|Π| times the number of multichains ∅ = I0 ⊆ I1 ⊆ · · · ⊆ In =
Π of order ideals such that Ij \ Ij−1 is an antichain for j = 1, . . . , n. By
Proposition 2.1 this is exactly (−1)|Π| Ω◦Π (n), and this proves Theorem 1.12.
Our proof also gives us some structural insights into ΩΠ (n).
Corollary 2.8. Let Π be a poset and 1 ≤ m ≤ |Π|. Then Ω(−k) = 0 for all
0 < k < m if and only if Π contains an m-chain.
2.3. Zeta Polynomials, Distributive Lattices, and Eulerian
Posets
We now take a breath and see how far we can generalize (the assumptions
in) Theorem 1.12. Our starting point is Proposition 2.4: for a poset Π that
has a minimum ˆ
0 and maximum ˆ1, the evaluation
ZΠ (n) := ζ n (ˆ0, ˆ1)
is a polynomial in n, the zeta polynomial of Π. For example, if we augment
the poset D10 in Figure 8 by a maximal element (think of the number 0, which
is divisible by all positive integers), Exercise 2.7 gives the accompanying zeta
polynomial as
ZΠ (n) =
1 4
24 n
+
11 3
12 n
+
35 2
24 n
−
17
12 n
+ 1.
(2.7)
In analogy with the combinatorial reciprocity theorem for order polynomials
(Theorem 1.12)—which are, after all, zeta polynomials of posets formed by
order ideals—we now seek interpretations for evaluations of zeta polynomials
at negative integers. Analogous to Proposition 2.6,
ZΠ (−n) = ζ −n (ˆ0, ˆ1) = µn (ˆ0, ˆ1)
where µ is the M¨obius function of Π. Our above example zeta function
(2.7) illustrates that the quest for interpretations at negative evaluations is
nontrivial: here we compute
ZΠ (−2) = 3
and
ZΠ (−3) = −3
and so any hope of a simple counting interpretation of ZΠ (−n) or −ZΠ (−n)
is shattered. On a more optimistic note, we can repeat the argument
behind (2.6) for a general poset Π:
X
ZΠ (−n) = µn (ˆ
0, ˆ
1) =
µ(x0 , x1 ) µ(x1 , x2 ) · · · µ(xn−1 , xn )
(2.8)
where the sum is over all multichains
ˆ
0 = x0 x1 · · · xn = ˆ1
2.3. Zeta Polynomials, Distributive Lattices, and Eulerian Posets
33
of length n. The key property that put (2.6) to work in our proof of Theorem
2.7 (and subsequently, our proof of Theorem 1.12) was that every summand
on the right-hand side of (2.6) was either 0 or (the same) constant. We thus
seek a class of posets where a similar property holds in (2.8).
For two elements x and y in a poset Π, consider all least upper bounds
of x and y, i.e., all z ∈ Π such that x z and y z and there is no w ≺ z
with the same property. If such a least upper bound of x and y exists and is
unique, we call it the join of x and y and denote it by x ∨ y. Dually, if a
greatest lower bound of x and y exits and is unique, we call it the meet of
x and y and denote it by x ∧ y.
A lattice4 is a poset in which meets and joins exist for any pair of
elements. Note that any finite lattice will necessarily have a minimum ˆ0 and
a maximum element ˆ
1. A lattice Π is distributive if meets and joins satisfy
the distributive laws
(x ∧ y) ∨ z = (x ∨ z) ∧ (y ∨ z)
and
(x ∨ y) ∧ z = (x ∧ z) ∨ (y ∧ z)
for all x, y, z ∈ Π. The reason we are interested in distributive lattices is the
following famous result, whose proof is subject to Exercise 2.8.
Theorem 2.9. Every finite distributive lattice is isomorphic to the poset of
order ideals of some poset.
Let’s experience the consequences of this theorem. Given a finite distributive lattice Π, we now know that the M¨obius-function values on the
right-hand side of (2.8) can be interpreted as stemming from the poset of
order ideals of some other poset. But this means that we can apply Theorem
2.7 in precisely the same way we applied it in our proof of Theorem 1.12: the
right-hand side of (2.8) becomes (−1)|Π| times the number of multichains
ˆ0 = x0 x1 · · · xn = ˆ1 such that the corresponding differences of order
ideals are all antichains. A moment’s thought reveals that this last condition
is equivalent to the fact that each interval [xj , xj+1 ] is a Boolean lattice.
What we have just proved is a combinatorial reciprocity theorem which, in a
sense, generalizes that of order polynomials (Theorem 1.12):
Theorem 2.10. Let Π be a finite distributive lattice. Then (−1)|Π| ZΠ (−n)
equals the number of multichains ˆ0 = x0 x1 · · · xn = ˆ1 such that each
interval [xj , xj+1 ] is a Boolean lattice.
There is another class of posets that comes with a combinatorial reciprocity theorem stemming from (2.8). To introduce it, we need a few more
definitions. A finite poset Π is graded if every maximal chain in Π has the
4This lattice is not to be confused with the integer lattice Z2 that made an appearance in
Section 1.4 and whose higher-dimensional cousins will play a central role in later chapters. Both
meanings of lattice are well furnished in the mathematical literature; we hope that they will not
be confused in this book.
34
2. Partially Ordered Sets
same length, which we call the rank of Π. The length l[x, y] of an interval
[x, y] in a poset Π is the length of a maximal chain in [x, y]. A graded poset
that has a minimal and a maximal element is Eulerian if its M¨
obius function
is
µ(x, y) = (−1)l[x,y] .
We have seen examples of Eulerian posets earlier, for instance, Boolean
lattices. Another important class of Eulerian posets are formed by faces of
polyhedra, which we will study in the next chapter.
What happens with (2.8) when the underlying poset Π is Eulerian? In
this case, the M¨obius-function values on the right-hand side are determined
by the interval length, and so each summand on the right is simply (−1)r
where r is the rank of Π. But then (2.8) says that ZΠ (−n) equals (−1)r
times the number of multichains of length n, which is ζ n (ˆ0, ˆ1) = ZΠ (n). This
argument yields a reciprocity theorem that relates the zeta polynomial of Π
to itself:
Theorem 2.11. Let Π be an Eulerian poset of rank r. Then
ZΠ (−n) = (−1)r ZΠ (n) .
We will come back to this result in connection with the combinatorial
structure of polytopes.
2.4. M¨
obius Inversion
Our use of the M¨obius function to prove Theorem 1.12 is by far not the only
instance where this function is useful—zeta and M¨obius functions make a
prominent appearance in practically all counting problems in which posets
play a structural role. The setting is the following: For a fixed poset Π, we
want to know a certain function f : Π → C but all we know is the function
X
g(y) =
f (x) c(x, y)
(2.9)
xy
where we can think of the c(x, y) as some coefficients distorting the given
function f . In such a situation, can we infer f from g? The typical situation
is given by
X
f (y) :=
f (x).
xy
The right (algebraic) setting in which to address such a question is the
incidence algebra: Every α ∈ I(Π) defines a linear transformation on CΠ via
X
(αf )(y) :=
f (x) α(x, y) .
(2.10)
x∈Π
In the warm-up Exercise 2.11 you are asked to verify that, with the above
definition, I(Π) yields a right action on the vector space CΠ . Thus, in (2.9),
Notes
35
f can be recovered from g whenever c is invertible. For the zeta function, the
procedure of recovering f from f goes by the name of M¨
obius inversion.
obius function, and
Theorem 2.12. Let Π be a poset, µ its associated M¨
f : Π → C. Then
X
X
f (y) =
f (x)
⇐⇒
f (y) =
f (x) µ(x, y)
xy
xy
and, likewise,
f (y) =
X
⇐⇒
f (x)
f (y) =
xy
X
µ(y, x) f (x) .
xy
Proof. The statement of the theorem is simply
µf = (µ ∗ ζ)f = f
where we used that µ = ζ −1 . However, it is instructive to do the yoga of
M¨
obius inversion
X
XX
(µf )(z) =
f (y) µ(y, z) =
f (x) µ(y, z)
yz
=
X
yz xy
f (x)
xz
X
µ(y, z) =
xyz
X
f (x) δ(x, z)
xz
= f (z) .
The second statement is verified the same way.
In a nutshell M¨obius inversion is what we implicitly used in our treatment
of Ehrhart theory for lattice polygons in Section 1.4. The subdivision of a
lattice polygon P into triangles, edges, and vertices is a genuine poset under
inclusion. The function f (x) is the number of lattice points in x ⊆ R2 and
we were interested in the evaluation of f (P) = (µf⊆ )(P).
Notes
Posets and lattices originated in the nineteenth century and became subjects
of their own rights with the work of Garrett Birkhoff, who proved Theorem
2.9 [15], and Philip Hall [36].
The oldest type of M¨obius function is the one studied in number theory,
which is the M¨obius function (in a combinatorial sense) of the divisor lattice
(see Exercise 2.5c). The systematic study of M¨
obius functions of general
posets was initiated by Gian–Carlo Rota’s famous paper [63] which arguably
started modern combinatorics. Rota’s paper also put the idea of incidence
algebras on firm ground, but it can be traced back much further to Richard
Dedekind and Eric Temple Bell [78, Chapter 3]. Theorem 2.7 is known as
Rota’s crosscut theorem.
36
2. Partially Ordered Sets
As we already mentioned in Chapter 1, order polynomials were introduced by Richard Stanley [71, 76] as ‘chromatic-like polynomials for posets’.
Stanley’s paper [76] also introduced order polytopes, which form the geometric face of order polynomials, as we will see in Chapter 6. Stanley introduced
the zeta polynomial of a poset in [73], the paper that inspired the title of
our book. Stanley also initiated the study of Eulerian posets in [75], though,
in his own words, “they had certainly been considered earlier.”
For (much) more on posets, lattices, and M¨
obius functions, we recommend
[70] and [78, Chapter 3], which contains numerous open problems; we mention
one representative:
Let Πn be the set all partitions (whose definition is given in (4.20) below)
of a fixed positive integer n. We order the elements of Πn by refinement, i.e.,
given two partitions (a1 , a2 , . . . , aj ) and (b1 , b2 , . . . , bk ) of n, we say that
(a1 , a2 , . . . , aj ) (b1 , b2 , . . . , bk )
if the parts a1 , a2 , . . . , aj can be partitioned into blocks whose sums are
b1 , b2 , . . . , bk . Find the M¨
obius function of Πn .
Exercises
2.1 Let (R, +, ·) be a ring with unit 1. For any r ∈ R and d ≥ 0 verify the
binomial theorem
d X
d j
d
(1 + r) =
r .
j
j=0
Show how this, in particular, implies (2.3).
2.2 Show that every finite poset Π has a linear extension. (Hint: you can
argue graphically by reading the Hasse diagram or, more formally, by
induction on |Π|.)
2.3 Let Π be a finite poset and recall that ζ = δ + η where δ and η are
defined by (2.1) and (2.2), respectively.
(a) Show that for x y,
η k (x, y) = |{x = x0 ≺ x1 ≺ x2 ≺ · · · ≺ xk−1 ≺ xk = y}| ,
the number of strict chains of length k in the interval [x, y].
(b) Infer that η is nilpotent, that is, η k ≡ 0 for k > |Π|.
(c) For x, y ∈ Π, do you know what (2δ − ζ)−1 (x, y) counts?
(d) Show that ηJn (Π) (∅, Π) equals the number of surjective order preserving maps Π → [n].
Exercises
37
2.4 For posets (Π1 , 1 ) and (Π2 , 2 ), we define their (direct) product
with underlying set Π1 × Π2 and partial order
(x1 , x2 ) (y1 , y2 )
:⇐⇒
x1 1 y1 and x2 2 y2 .
(a) Show that every interval [(x1 , x2 ), (y1 , y2 )] of Π1 × Π2 is of the form
[x1 , y1 ] × [x2 , y2 ].
(b) Show that µΠ1 ×Π2 ((x1 , x2 ), (y1 , y2 )) = µΠ1 (x1 , y1 ) µΠ2 (x2 , y2 ).
(c) Show that the Boolean lattice Bn is isomorphic to the n-fold product
of the chain [2], and conclude that for S ⊆ T ⊆ [n]
µBn (S, T ) = (−1)|T \S| .
2.5 (a) Let Π = [d], the d-chain. Show


1
µ[d] (i, j) =
−1


0
that for 1 ≤ i < j ≤ d
if i = j,
if i + 1 = j,
otherwise.
(b) Write out the statement that M¨
obius inversion gives in this explicit
case and interpret it along the lines of the Fundamental Theorem
of Calculus.
(c) The M¨obius function in number theory is the function µ : Z>0 →
Z defined for n ∈ Z>0 through µ(1) = 1, µ(n) = 0 if n is not
squarefree, that is, if n is divisible by a proper prime power, and
µ(n) = (−1)r if n is the product of r distinct primes. Show that
for given n ∈ Z>0 , the partially ordered set Dn of divisors of n is
isomorphic to a direct product of chains and use Exercise 2.4 to
verify that µ(n) = µDn (1, n).
2.6 For fixed k, n ∈ Z>0 consider the map g : Bk → Z>0 given by
g(T ) = |T |n .
Show that
k! S(n, k) = (g µBk )([k])
where S(n, k) is the Stirling number of the second kind. (Hint:
k! S(n, k) counts surjective maps [n] → [k].)
2.7 Compute the zeta polynomial of the poset D10 in Figure 8, appended
by a maximal element.
2.8 Prove Theorem 2.9: Every finite distributive lattice is isomorphic to a
poset of order ideals of some poset. (Hint: Given a distributive lattice
Π, consider the subposet Θ consisting of all join irreducible elements,
i.e., those elements that cannot be written as the join of some other
elements. Show that Θ is isomorphic to J (Π).)
2.9 State and prove a result analogous to Corollary 2.8 for distributive
lattices.
38
2. Partially Ordered Sets
2.10 Let Π be a finite graded poset that has a minimal and a maximal
element, and define the rank of x ∈ Π as the length of a maximal chain
ending in x. Prove that Π is Eulerian if and only if for all x ≺ y the
interval [x, y] has as many elements of even rank as of odd rank.
2.11 Show that (2.10) defines a right action of I(Π) on CΠ . That is, I(Π)
gives rise to a vector space of linear transformations on CΠ and (α∗β)f =
β(αf ), for any α, β ∈ I(Π).
2.12 Let A be a finite set and S1 , . . . , Sn ⊆ A subsets. The intersection
poset is the set
(
)
\
Π = SI :=
Si : I ⊆ [n]
i∈I
partially ordered by containment. The minimum of Π is ˆ0 = S1 ∩· · ·∩Sn
and the maximum is ˆ1 = S∅ = A. More precisely, Π is a sublattice of
the Boolean lattice consisting of all subsets of A.
(a) For a map f : Π → C verify the inclusion–exclusion principle
X
f (S1 ∪ · · · ∪ Sn ) =
(−1)|I|−1 f (SI ).
I⊆[n]
(Hint: You get an induced map fˆ : Bn → C by fˆ(I) = f (SI ).)
(b) For A, B ∈ Π with A ⊆ B, show that
o
Xn
µΠ (A, B) =
(−1)|I| : ∅ 6= I ⊆ [n], A ⊆ SI = B .
(Hint: Show that the right-hand side satisfies (2.4).)
(c) Can you find a more explicit formula for µΠ (A, B)? (Hint: Observe
that Π is a distributive lattice.)
Chapter 3
Polyhedral Geometry
One geometry cannot be more true than another; it can only be more convenient.
Jules Henri Poincar´e
In this chapter we define the most convenient geometry for the combinatorial
objects from Chapter 1. To give a first impression of how geometry naturally
enters our combinatorial picture, let’s return to the problem of counting
multisubsets of size d of [n]. Every such multiset corresponds to a dtuple
(m1 , m2 , . . . , md ) ∈ Zd such that
1 ≤ m1 ≤ m2 ≤ · · · ≤ md ≤ n .
Forgetting about the integrality of the mj , we obtain a genuine geometric
object containing the solutions to this system of d + 1 linear inequalities:
n ∆d = {x ∈ Rd : 1 ≤ x1 ≤ x2 ≤ · · · ≤ xd ≤ n} .
The d-multisubsets correspond exactly to the integer lattice points n ∆d ∩ Zd .
The set n ∆d is a polyhedron: a set defined by finitely many linear inequalities.
Polyhedra constitute a rich class of geometric objects, rich enough to capture
much of the enumerative combinatorics that we pursue in this book.
Besides introducing machinery to handle polyhedra, our main emphasis in
this chapter is on the faces of a given polyhedron. They form a poset that is
naturally graded by dimension, and counting the faces in each dimension gives
rise to the famous Euler–Poincar´e formula. This identity is at play (often
behind the scenes) in practically every combinatorial reciprocity theorem
that we will encounter in later chapters.
39
40
3. Polyhedral Geometry
3.1. Polyhedra, Cones, and Polytopes
The building blocks for our geometric objects are halfspaces. To this end,
we define an affine hyperplane as a set of the form
n
o
H := x ∈ Rd : ha, xi = b
for some normal a ∈ Rd \ {0} and displacement b ∈ R. Here h , i denotes the
standard inner product in Rd . We call H a linear hyperplane if 0 ∈ H or,
equivalently, b = 0. Every hyperplane H subdivides the ambient space Rd
into the two closed halfspaces
n
o
H ≥ := x ∈ Rd : ha, xi ≥ b
n
o
(3.1)
H ≤ := x ∈ Rd : ha, xi ≤ b .
Starting with such halfspaces we can manufacture more complex objects that
are bounded by hyperplanes: a polyhedron Q ⊆ Rd is the intersection of
finitely many halfspaces, such as the one shown in Figure 1. We remark that,
trivially, all affine subspaces, including Rd and ∅, are polyhedra.
Figure 1. A bounded polyhedron in the plane.
The term polyhedron appears in many parts of mathematics, unfortunately with different connotations. To be more careful, we just defined a
convex polyhedron. A set S ⊆ Rd is convex if for every p, q ∈ S, the
line segment
[p, q] := {(1 − λ)p + λq : 0 ≤ λ ≤ 1}
with endpoints p and q is contained in S. Since hyperplanes and halfspaces
are convex, a finite intersection of halfspaces produces a convex object (and
so convex polyhedra are indeed convex). Nevertheless, we will drop the
adjective convex and simply refer to Q as a polyhedron.
3.1. Polyhedra, Cones, and Polytopes
41
So, Q ⊆ Rd is a polyhedron if there are hyperplanes Hi = {x ∈ Rd :
hai , xi = bi } for i = 1, 2, . . . , k, such that
Q =
k
\
Hi≤ =
n
o
x ∈ Rd : hai , xi ≤ bi for all i = 1, 2, . . . , k .
(3.2)
i=1
T
We call a halfspace Hj≤ irredundant if i6=j Hi≤ =
6 Q. We might as well
only use irredundant halfspaces to describe a given polyhedron. By arranging
the normals to the hyperplanes as the rows of a matrix A ∈ Rk×d and letting
b := (b1 , b2 , . . . , bk ), we can compactly write
n
o
Q = x ∈ Rd : A x ≤ b .
For example,


1
1
0




 −1 −1
0





−1
0
0
x ∈ R3 : 
 0 −1
0





 0
0
−1



0
0
1



 
 x1



  x2  ≤ 


 x3



1
−1
0
0
0
1
















(3.3)
is a square embedded in R3 which is pictured in Figure 2.
Figure 2. The square defined by (3.3).
As in all geometric disciplines, a fundamental notion is that of dimension.
For polyhedra, this turns out to be pretty straightforward. We define the
affine hull aff(Q) of a polyhedron Q ⊆ Rd as the smallest affine subspace of
Rd that contains Q. The reader might want to check (in Exercise 3.1) that
o
\n
aff(Q) =
H hyperplane in Rd : Q ⊆ H .
(3.4)
42
3. Polyhedral Geometry
The dimension of a polyhedron is the dimension of its affine hull. When
dim Q = n, we call Q an n-polyhedron. For example, the square defined by
(3.3) has affine hull {x ∈ R3 : x1 + x2 = 1} and so (surprise!) its dimension
is 2.
The (topological) interior of a polyhedron Q given in the form (3.2) is
n
o
x ∈ Rd : hai , xi < bi for all i = 1, 2, . . . , k .
(3.5)
However, this notion of interior is not intrinsic to Q but makes reference
to the ambient space. For example, a triangle might or might not have
an interior depending on whether we embed it in R2 or R3 . Luckily, every
polyhedron comes with a canonical embedding into its affine hull and we
can define the relative interior of Q as the set of points of Q that are in
the interior of Q relative to its embedding into aff(Q). We will denote the
relative interior of Q by Q◦ .1 When Q is full dimensional, Q◦ is given by
(3.5). In the case that Q is not full dimensional, we have to be a bit more
careful (the details are the content of Exercises 3.4 and 3.5): Assuming Q
is given in the form (3.2), let I(Q) := {i ∈ [k] : hai , xi = bi for all x ∈ Q}.
Then
Q◦ = {x ∈ Q : hai , xi < bi for all i 6∈ I(Q)} .
(3.6)
For instance, our running example, the square defined by (3.3), has relative
interior



 


−1
0
0
0 




 0 −1
 x1
 0 
0
3






x2 <   .
x ∈ R : x1 + x2 = 1, 
0
0 −1 
0 



x3


0
0
1
1
Complementary to the affine hull of a polyhedron Q, we define the
lineality space lineal(Q) ⊆ Rd to be the inclusion-maximal linear subspace
of Rd such that there exists q ∈ Rd for which q + lineal(Q) ⊆ Q. It follows
from convexity that p + lineal(Q) ⊆ Q for all p ∈ Q; see Exercises 3.11 and
3.12. For example,




x1


1 −1 −1 
0
x2  ≤
x ∈ R3 :
−1 −1
1
0 

x3
is a wedge with lineality space {x ∈ R3 : x1 = x3 , x2 = 0}, which is a line.
(A picture of this, which we encourage the reader to draw, should look a bit
like the left side of Figure 3.) If lineal(Q) = {0} we call Q pointed or line
free.
1 A note on terminology: A polyhedron is, by definition, closed. However, we will sometimes
talk about an open polyhedron, by which we mean the relative interior of a polyhedron. In a few
instances we will simultaneously deal with polyhedra and open polyhedra, in which case we may
use the superfluous term closed polyhedron to distinguish one from the other.
3.1. Polyhedra, Cones, and Polytopes
43
Figure 3. Two polyhedral cones, one of which is pointed.
A set C ⊆ Rd is a convex cone if C is a convex set such that µ C ⊆ C for
all µ ≥ 0. In particular, every linear subspace is a convex cone and, stronger,
any intersection of linear halfspaces is a cone. Intersections of finitely many
linear halfspaces are called polyhedral cones; see Figure 3 for two examples
and Exercise 3.13 for more. Not all cones are polyhedral: for example,
C = (x, y, z) ∈ R3 : z ≥ x2 + y 2
is a cone but not polyhedral (Exercise 3.14). To ease notation, we will
typically drop the annotation, as henceforth all the cones to be encountered
will be polyhedral. We remark (Exercise 3.15) that a cone C is pointed if
and only if p, −p ∈ C implies p = 0.
A connection between convex sets and convex cones is the following: For
a convex set K ⊂ Rd we define the homogenization of K as
n
o
hom(K) := (x, λ) ∈ Rd+1 : λ ≥ 0, x ∈ λK .
In particular, the homogenization of the polyhedron Q = {x ∈ Rd : A x ≤ b}
is the polyhedral cone
n
o
hom(Q) = (x, λ) ∈ Rd+1 : λ ≥ 0, A x ≤ λb .
We can recover our polyhedron Q from its homogenization as the set of those
points x ∈ hom(Q) for which xd+1 = 1. Moreover, hom(Q) is pointed if and
only if Q is. Homogenization seems like a simple construction but it will
come in quite handy in this and later chapters. The homogenization of a
pentagon is shown on the right in Figure 3.
As the class of convex sets is closed under taking intersections, there is
always a unique inclusion-minimal convex set conv(S) containing a given
set S ⊆ Rd . This set is thus the intersection of all convex sets containing S
and is called the convex hull of S. For finite S, the resulting objects will
44
3. Polyhedral Geometry
play an important role: A (convex) polytope P ⊂ Rd is the convex hull of
finitely many points in Rd . Thus the polytope defined as the convex hull of
S = {p1 , p2 , . . . , pm } is
(
)
m
X
conv(S) = λ1 p1 + λ2 p2 + · · · + λm pm :
λi = 1, λ1 , . . . λm ≥ 0 .
i=1
This construction is the essence of “closed under taking line segments”; see
Exercise 3.9. Similar to our definition of irredundant halfspaces, we call
v ∈ S a vertex of P := conv(S) if P 6= conv(S \ {v}), and we let vert(P) ⊆ S
denote the collection of vertices of P. This is the unique inclusion-minimal
subset T ⊆ S such that P = conv(T ); see Exercise 3.8. For example, the
vertices of our square defined by (3.3) are (1, 0, 0), (0, 1, 0), (1, 0, 1), and
(0, 1, 1). This square is an example of a lattice polytope—a polytope in
Rd whose vertices are in Zd . Similarly, a rational polytope is one with
vertices in Qd .
A special class of polytopes consists of the simplices: A d-simplex ∆ is
the convex hull of d + 1 affinely independent points in Rn . Simplices are the
natural generalizations of line segments, triangles, and tetrahedra to higher
dimensions.
Convex cones are also closed under taking intersections, and so we define
the conical hull cone(S) of a set S ⊆ Rd as the smallest convex cone that
contains S. It turns out (this is at the heart of Theorem 3.1 below) that
polyhedral cones are precisely the conical hull of finite sets. We call a minimal
set S such that a given cone C equals cone(S) a set of generators of C. The
homogenization of a polytope P is thus given by (Exercise 3.18)
hom(P) = cone(P × {1}) = cone {(v, 1) : v ∈ vert(P)}
(3.7)
(and so {(v, 1) : v ∈ vert(P)} is a set of generators for hom(P)) and, as
above, the polytope P ⊂ Rd can be recovered by intersecting hom(P) ⊂ Rd+1
with the hyperplane H = {x ∈ Rd+1 : xd+1 = 1}. This relation holds in
somewhat greater generality—see Exercise 3.19. By analogy with the notion
of a lattice/rational polytope, we call a cone in Rd rational if we can choose
its generators in Qd (in which case we might as well multiply them to lie
in Zd ).
Polyhedra and convex/conical hulls of finite sets are two sides of the
same coin. To make this more precise, we need the following notion: The
Minkowski sum of two convex sets K1 , K2 ⊂ Rd is
K1 + K2 := {p + q : p ∈ K1 , q ∈ K2 } .
An example is depicted in Figure 4. That K1 + K2 is again convex is
the content of Exercise 3.21. Minkowski sums are key to the following
fundamental theorem of polyhedral geometry.
3.2. Faces
45
+
=
Figure 4. A Minkowski sum.
Theorem 3.1 (Minkowski–Weyl). A set Q ⊆ Rd is a polyhedron if and only
if there exist a polytope P and a polyhedral cone C such that
Q = P + C.
This theorem highlights the special role of polyhedra among all convex
bodies. It states that polyhedra possess a discrete intrinsic description in
terms of finitely many vertices and generators of P and C, respectively, as
well as a discrete extrinsic description in the form of finitely many linear
inequalities. The benefit of switching between different presentations is
apparent.
A proof of Theorem 3.1 would need a chapter on its own and, while
important, is not at the core of what we will develop in this book. So we get
around proving it by pointing the reader instead to [88, Lecture 1]. At any
rate, we will need the content of Theorem 3.1. As a first application we get
the following nontrivial operations on polyhedra, polytopes, and cones.
Corollary 3.2. Let Q ⊂ Rd be a polyhedron and φ(x) = A x + b an affine
projection Rd → Re . Then φ(Q) is a polyhedron. If P ⊂ Rd is a polytope,
then P ∩ Q is a polytope.
3.2. Faces
A hyperplane H := {x ∈ Rd : ha, xi = b} is a supporting hyperplane of
the polyhedron Q if
Q ∩ H 6= ∅ and either Q ⊆ H ≤ or Q ⊆ H ≥ ,
in words, Q is entirely contained in one of the halfspaces bounded by H. A
face of Q is a set of the form Q ∩ H, where H is a supporting hyperplane
of Q; we always include Q itself (and sometimes ∅) in the list of faces of
Q. A face F 6= Q is called a proper face. The 0-dimensional faces of
Q are precisely the vertices of Q (Exercise 3.16). The faces of dimension
1 and d − 1 of a d-dimensional polyhedron Q are called edges and facets,
respectively. It follows almost by definition that a face of a polyhedron is a
polyhedron (Exercise 3.24). An equally sensible but somewhat less trivial
46
3. Polyhedral Geometry
fact is that every face of a polyhedron Q is the intersection of some facets of
Q (Exercise 3.25).
The faces of a given d-polyhedron Q (including ∅) are naturally ordered
by set containment, which gives rise to a poset (in fact, a lattice), the face
lattice Φ(Q). Figure 5 gives an example, the face lattice of a square pyramid.
To emphasize the poset structure of the set of faces of a polyhedron, we
write F G whenever F and G are faces of Q with F ⊆ G; in this notation,
F ≺ G means that F is a proper face of G.
Figure 5. The face lattice of a square pyramid.
In a sense, much of what remains in this chapter is devoted to face
lattices of polyhedra. For starters, we note that Φ(Q) is naturally graded by
dimension, which is one of many motivations to introduce the face numbers
fk = fk (Q) := number of faces of Q of dimension k.
The face numbers are often recorded in the f -vector
f (Q) := (f0 , f1 , . . . , fd−1 )
It is not hard to see (Exercise 3.17) that the j-faces of a d-polytope P are in
bijection with the (j + 1)-faces of hom(P) for all j = 0, 1, . . . , d, and so, in
particular,
f (hom(P)) = (1, f (P)) .
The faces of a polyhedron Q give rise to the following natural decomposition of Q.
Lemma 3.3. Let Q be a polyhedron. For every point p ∈ Q there is a unique
face F of Q such that p ∈ F ◦ . In particular, we have the disjoint union2
]
Q=
F ◦.
F Q
2We use the symbol ] to denote disjoint unions.
3.3. The Euler Characteristic
47
Proof. Suppose Q is given as the intersection of irredundant halfspaces
Q=
k
\
Hj≤ ,
j=1
and p ∈ Q. After possibly renumbering the halfspaces, we may assume
=
p ∈ H1= , . . . , Hm
<
p ∈ Hm+1
, . . . , Hk<
and
where 0 ≤ m < k. Thus
F :=
m
\
Hj= ∩
j=1
k
\
Hj≤
j=m+1
is a face of Q whose interior (see (3.6))
F
◦
=
m
\
Hj=
j=1
∩
k
\
Hj<
j=m+1
contains p. The uniqueness of F follows from Exercise 3.7 and the fact that
F is, by construction, inclusion minimal.
We will soon see that one often wants to “mod things out” by the
lineality space, in the following sense: For a polyhedron Q with lineality
space L = lineal(Q), we write F/L for the orthogonal projection of the face
F of Q to L⊥ .
Lemma 3.4. The map Φ(Q) → Φ(Q/L) given by F 7→ F/L is an inclusionpreserving bijection.
• a polyhedron Q is proper if Q is not an affine subspace.
6 aff(P). Then conv(P ∪ v) is pyramid over P
• P polytope and v ∈
with apex v.
3.3. The Euler Characteristic
We now come to an important notion that will allow us to relate geometry
to combinatorics, the Euler characteristic. Our approach to the Euler
characteristics of convex polyhedra is by way of sets built up from polyhedra.
A set S ⊆ Rd is polyconvex if it is the union of finitely many relatively
open polyhedra:
S = P1 ∪ P2 ∪ · · · ∪ Pk
where P1 , . . . , Pk ⊆ Rd are relatively open polyhedra. For example, a polyhedron is polyconvex: according to Lemma 3.3, we can write it as the (disjoint)
union of its relatively open faces. Note, however, that our definition entails
that polyconvex sets are not necessarily convex, not necessarily connected,
and not necessarily closed. As we will see, they form a nice bag of sets to
48
3. Polyhedral Geometry
draw from, but not every reasonable set—e.g., the unit disc in the plane
(Exercise 3.26)—is a polyconvex set.
Let’s denote by PCd the collection of polyconvex sets in Rd . This is an
(infinite) poset under inclusion with minimal and maximal elements ∅ and
Rd , respectively. The intersection and the union of finitely many polyconvex
sets are polyconvex, which renders PCd a distributive lattice.
A map φ from PCd to some Abelian group is a valuation if φ(∅) = 0
and
φ(S ∪ T ) = φ(S) + φ(T ) − φ(S ∩ T )
(3.8)
for all S, T ∈ PCd . Here’s what we’re after:
Theorem 3.5. There exists a valuation χ : PCd → Z such that χ(∅) = 0
and χ(P) = 1 for every nonempty closed polytope P ⊂ Rd .
This is a nontrivial statement, as we cannot simply define χ(S) = 1
whenever S 6= ∅. Indeed, if P ⊂ Rd is a d-polytope and H = {x ∈ Rd :
ha, xi = b} is a hyperplane such that H ∩ P 6= ∅, then
P1 := {x ∈ P : ha, xi < b}
and
P2 := {x ∈ P : ha, xi ≥ b}
are polyconvex sets such that P1 ∩ P2 = ∅ and thus
χ(P) = χ(P1 ) + χ(P2 ) .
Therefore, if χ is a valuation satisfying the conditions of Theorem 3.5, then
necessarily χ(P1 ) = 0.
The goal of this section is to construct a valuation χ satisfying the
properties dictated by Theorem 3.5. The valuation property (3.8) will be the
key to simplifying the computation of χ(S) for arbitrary polyconvex sets: If
S = P1 ∪ P2 ∪ · · · ∪ Pk , where each Pi ⊆ Rd is a relatively open polyhedron,
then by iterating (3.8) we obtain the inclusion–exclusion formula
X
X
χ(S) =
χ(Pi ) −
χ(Pi ∩ Pj ) + · · ·
i
=
X
i<j
(−1)|I|−1 χ(PI ) ,
(3.9)
∅6=I⊆[k]
where PI := ∩i∈I Pi . In particular, the value of χ(S) does not depend on the
presentation of S as a union of relatively open polyhedra.
Here is a way to construct polyconvex sets. Let H = {H1 , H2 , . . . , Hn } be
an arrangement (i.e., a finite set) of hyperplanes Hi = {x ∈ Rd : hai , xi =
bi } in Rd . An example of an arrangement of six hyperplanes (here: lines) in
the plane is shown in Figure 6.
Continuing our definitions in (3.1), for a hyperplane Hi we denote by
Hi> := {x : hai , xi > bi }
3.3. The Euler Characteristic
49
Figure 6. An arrangement of six lines in the plane.
the open positive halfspace bounded by H. We analogously define Hi<
and Hi= := Hi . For every σ ∈ {<, =, >}n , we obtain a (possibly empty)
relatively open polyhedron
Hσ := H1σ1 ∩ H2σ2 ∩ · · · ∩ Hnσn ,
(3.10)
Rd .
and these relatively open polyhedra partition
For a point p ∈ Rd , let
σ(p) ∈ {<, =, >}n record the position of p relative to the n hyperplanes; that
is, Hσ(p) is the unique relatively open polyhedron among the Hσ ’s containing
p. For example, the line arrangement in Figure 6 decomposes R2 into 57
relatively open polyhedra: 19 of dimension two, 28 of dimension one, and 10
of dimension zero.
For a fixed hyperplane arrangement H, we define the class of H-polyconvex sets PC(H) ⊂ PCd to consist those sets that are finite unions of relatively
open polyhedra of the form Hσ given by (3.10). That is, every S ∈ PC(H)
has a representation
S = Hσ1 ] Hσ2 ] · · · ] Hσk
σ1, σ2, . . . , σk
(3.11)
{<, =, >}n
for some
∈
such that Hσj 6= ∅ for all j = 1, 2, . . . , k.
Note that the relatively open polyhedra Hσj are disjoint and thus the
representation of S given in (3.11) is unique. For S ∈ PC(H) we define
χ(H, S) :=
k
X
(−1)dim(Hσj ) .
(3.12)
j=1
The next result, whose proof we leave as Exercise 3.27, states that this
function is a valuation.
Proposition 3.6. The function χ(H, ·) : PC(H) → Z is a valuation.
50
3. Polyhedral Geometry
We can consider χ(H, S) as a function in two arguments, the arrangement
H and the set S ⊆ Rd ; note that a set S is typically polyconvex with respect
to various arrangements. It is a priori not clear how the value of χ(H, S)
changes when we change the arrangement. The power of our above definition
is that it doesn’t:
Lemma 3.7. Let H1 , H2 be two hyperplane arrangements in Rd and let
S ∈ PC(H1 ) ∩ PC(H2 ). Then
χ(H1 , S) = χ(H2 , S) .
Proof. It is sufficient to show that
χ(H1 , S) = χ(H1 ∪ {H}, S)
where
H ∈ H1 \ H2 .
(3.13)
Iterating this yields χ(H1 , S) = χ(H1 ∪ H2 , S) and, similarly, χ(H2 , S) =
χ(H1 ∪ H2 , S), which proves the claim.
As a next simplifying measure, we observe that it suffices to show (3.13)
for S = Hσ for some σ. Indeed, from the representation in (3.11) and the
inclusion–exclusion formula (3.9), we then obtain
χ (H1 , S) = χ (H1 , Hσ1 ) + χ (H2 , Hσ2 ) + · · · + χ (Hk , Hσk ) .
Thus suppose that S = Hσ ∈ PC(H1 ) and H ∈ H1 \ H2 . There are three
possibilities how S can lie relative to H. The easiest cases are S ∩ H = S
and S ∩ H = ∅. In both cases S is genuinely a polyconvex set for H1 ∪ {H}
and
χ(H1 ∪ {H}, S) = (−1)dim S = χ(H1 , S) .
The only interesting case is ∅ 6= S ∩ H =
6 S. Since S is relatively open,
S < := S ∩ H < and S > := S ∩ H > are both nonempty, relatively open
polyhedra of dimension dim S, and S = := S ∩ H = is relatively open of
dimension dim S − 1; see Exercise 3.6. Therefore,
S = S< ] S= ] S>
is a presentation of S as an element of PC(H1 ∪ {H}), and
χ(H1 ∪ {H}, S)
= χ(H1 ∪ {H}, S < ) + χ(H1 ∪ {H}, S = ) + χ(H1 ∪ {H}, S > )
= (−1)dim S + (−1)dim S−1 + (−1)dim S
= (−1)dim S = χ(H1 , S) ,
which proves the claim.
The argument used in the above proof is typical when working with
valuations. The valuation property (3.8) allows us to refine polyconvex sets
by cutting them with hyperplanes and halfspaces. Clearly, there is no finite
set of hyperplanes H such that PCd = PC(H), but as long as we only worry
3.3. The Euler Characteristic
51
about finitely many polyconvex sets at a time, we can restrict ourselves to
PC(H) for some H.
Proposition 3.8. Let S ∈ PCd be a polyconvex set. Then there is a hyperplane arrangement H such that S ∈ PC(H).
Proof. This should be intuitively clear. We can write S = P1 ∪ P2 ∪ · · · ∪ Pk
for some relatively open polyhedra Pi . Now for every Pi thereTis a finite
σ
set of hyperplanes Hi := {H1 , H2 , . . . , Hm } such that Pi = j Hj j for
some σ ∈ {<, =, >}m . Thus Pi ∈ PC(Hi ), and by refining we conclude
S ∈ PC(H1 ∪ H2 ∪ · · · ∪ Hk ).
We can express the content of Proposition 3.8 more conceptually. For
two hyperplane arrangements H1 and H2 ,
H1 ⊆ H2
=⇒
PC(H1 ) ⊆ PC(H2 ) .
In more abstract terms then, PCd is the limit of PC(H) over all arrangements H.
Lemma 3.7 and Proposition 3.8 get us closer to Theorem 3.5:
Proposition 3.9. There is unique valuation χ : PCd → Z such that for
S ∈ PCd ,
χ(S) = χ(H, S)
for all hyperplane arrangements H for which S ∈ PC(H).
Proof. For a given S ∈ PCd , Lemma 3.7 implies that the definition
χ(S) := χ(H, S) ,
(3.14)
for any H such that S ∈ PC(H), is sound, and Proposition 3.8 ensures that
there is such an H. For uniqueness, we look back at (3.12) and conclude
immediately that χ(P) = (−1)dim P for any relatively open polyhedron P.
But then uniqueness follows for any polyconvex set, as we can write it as a
disjoint union of finitely many relatively open polyhedra.
We emphasize one part of the above proof for future reference:
Corollary 3.10. If P is a nonempty relatively open polyhedron, then
χ(P) = (−1)dim P .
The valuation χ in Proposition 3.9 is the Euler characteristic, and for
the rest of this book, we mean the Euler characteristic of P when we write
χ(P).
What is left to show to finish our proof of Theorem 3.5 is that χ(P) = 1
whenever P is a (nonempty) polytope. Let’s first note that (3.12) gives us
52
3. Polyhedral Geometry
an effective way to compute the Euler characteristic of a polyhedron via face
numbers: if Q is a polyhedron, then Lemma 3.3 states
]
Q =
F ◦,
F Q
where we recall that our notation F Q means F is a face of Q. The
inclusion–exclusion formula (3.9) gives
χ(Q) =
X
dim F
(−1)
∅≺F Q
=
dim
XQ
(−1)i fi (Q) .
(3.15)
i=0
This is called the Euler–Poincar´
e formula.
d
Now let P ⊂ R be a nonempty polytope. To compute χ(P ), we may
assume that the origin is contained in the relative interior of P. (Any other
point in the relative interior would work, but the origin is just too convenient.)
For a nonempty face F ⊆ P, let
n
o
[
C0 (F ) :=
tF ◦ = p ∈ Rd : 1t p ∈ F ◦ for some t > 0
t>0
and C0 (∅) := {0}. Figures 7 and 8 illustrates the following straightforward
facts whose proofs we leave as Exercise 3.28.
Figure 7. Cones over two faces of a polytope.
Proposition 3.11. Let P be a nonempty polytope with 0 ∈ P◦ . For any
proper face F ≺ P, the set C0 (F ) is a relatively open polyhedral cone of
dimension dim F + 1. Furthermore,
]
C0 (P) = aff(P) =
C0 (F ) .
∅≺F ≺P
We can now finally complete the proof of Theorem 3.5.
3.3. The Euler Characteristic
53
Figure 8. The decomposition of Proposition 3.11.
Proof of Theorem 3.5. We will show that the Euler characteristic χ satisfies the properties stated in Theorem 3.5. Let P be a nonempty closed
polytope of dimension d. By the Euler–Poincar´e formula (3.15), the Euler
characteristic is invariant under translation and thus we may assume that 0
is in the relative interior of P.
Proposition 3.11 yields two representations of the affine subspace aff(P),
and computing the Euler characteristic using the two different representations
gives
X
(−1)d = χ(C0 (P)) = χ(aff(P)) =
χ(C0 (F ))
F ≺P
=1−
X
dim F
(−1)
∅≺F ≺P
where the last equality follows from Proposition 3.11(a) and (c). Both P◦
and C0 (P) = aff(P) are relatively open polyhedra of the same dimension and
hence χ(C0 (P)) = χ(P◦ ) = (−1)d , by Corollary 3.10. Thus
X
1=
(−1)dim F = χ(P) .
∅≺F P
We don’t know yet the Euler characteristic of an unbounded polyhedron
Q. It turns out that this depends on whether Q is pointed or not. We start
with the easier case.
Corollary 3.12. If Q is a polyhedron with lineality space L = lineal(Q),
then
χ(Q) = (−1)dim L χ(Q/L) .
This is pretty straightforward considering the relationship between faces
and their dimensions of Q and Q/L given by Lemma 3.4; we leave the details
54
3. Polyhedral Geometry
to Exercise 3.29. In preparation for the case of a general pointed unbounded
polyhedron, we first treat pointed cones.
Proposition 3.13. If C ⊂ Rd is a pointed cone, then χ(C) = 0.
Proof. Let C = cone(u1 , . . . , um ) for some u1 , . . . , um ∈ Rd \ {0}. Since C
is pointed, there is a supporting hyperplane
H0 := {x ∈ Rd : ha, xi = 0}
such that C ⊆ H0≥ and C ∩ H0 = {0}. In particular, this means that
ha, ui i > 0 for all i and by rescaling, if necessary, we may assume that
ha, ui i = δ for some δ > 0. Let Hδ := {x ∈ Rd : ha, xi = δ} and consider
C := C ∩ Hδ≤ and
C∞ := C ∩ Hδ .
(See Figure 9 for an illustration.) By construction, C is a polytope and C∞
Figure 9. The polytopes C and C∞ .
is a face of C (and thus also a polytope). Moreover, every unbounded face of
C (here this means every nonempty face F 6= {0}) meets Hδ . Thus, every
k-face F of C gives rise to a k-face of C and, if F is unbounded, a (k − 1)-face
of C∞ . These are all the faces of C and C∞ . We thus compute
X
X
X
χ(C) =
(−1)dim F =
(−1)dim F −
(−1)dim F = χ(C)−χ(C∞ ) = 0
F C
F C
F C∞
where the last equality uses the fact that the Euler characteristic satisfies
the properties in Theorem 3.5.
3.3. The Euler Characteristic
55
Interestingly, the general case of a pointed unbounded polyhedron is not
much different from that of pointed cones. Recall from Theorem 3.1 that
every polyhedron Q is of the form Q = P + C where P is a polytope and C is a
polyhedral cone. In Exercise 3.23 you will show that for every nonempty face
F Q there are unique faces F 0 P and F 00 C such that F = F 0 + F 00 . In
particular, F is an unbounded face of Q if and only if F 00 is an unbounded
face of C.
Corollary 3.14. If Q is a pointed unbounded polyhedron then χ(Q) = 0.
Proof. We extend the idea of the proof of Proposition 3.13: We will find
a hyperplane H such that H is disjoint from any bounded face but the
intersection of H with an unbounded face F ⊆ Q yields a polytope of
dimension dim F − 1.
Let Q = P + C where P is a polytope and C a pointed cone. Let
Hδ = {x : ha, xi = δ} be the hyperplane constructed for the cone C in the
proof of Proposition 3.13. This time we choose δ > 0 sufficiently large so
that P ∩ Hδ = ∅. Thus, if F ≺ Q is a bounded face, then F = F 0 + {0}, for
some face F 0 of P, and F ∩ Hδ = ∅. See Figure 10 for a simple illustration.
Figure 10. “Compactifying” a polyhedron
Let F = F 0 + F 00 be an unbounded face of Q, for some F 0 P and
C, and let F∞ := F ∩ Hδ . We claim that F∞ is bounded. If it is not
bounded, then there exist p ∈ F∞ and u 6= 0 such that p + tu ∈ F∞ ⊆ F
for all t ≥ 0. This implies that u ∈ F 00 . But by construction ha, ui > 0 and
thus there is only one t for which p + tu ∈ Hδ .
As for the dimension of F∞ , we only have to show the impossibility of
dim F∞ < dim F − 1. This can only happen if Hδ meets F in the boundary.
But for any p ∈ F 0 and u ∈ F 00 \ {0} we have that p + 0 and p + tu are
points in F , but for t > 0 sufficiently large they lie on different sides of Hδ .
F 00
56
3. Polyhedral Geometry
Hence, Q := Q ∩ Hδ≤ and Q∞ := Q ∩ Hδ are both polytopes such that
every k-face of Q yields a k-face of Q and a (k − 1)-face of Q∞ , provided it
was unbounded. The same computation as in the proof of Proposition 3.13
then gives
X
X
X
χ(Q) =
(−1)dim F =
(−1)dim F = χ(Q) − χ(Q∞ )
(−1)dim F −
F Q
F Q
F Q∞
= 0.
3.4. M¨
obius Functions of Face Lattices
The Euler characteristic is a fundamental concept throughout mathematics.
In the context of geometric combinatorics it ties together the combinatorics
and the geometry of polyhedra in an elegant way. A first evidence of this is
provided by the central result of this section: The M¨obius function of the face
lattice of a polyhedron can be computed in terms of the Euler characteristic.
Theorem 3.15. Let Q be a polyhedron with face lattice Φ = Φ(Q). For any
two faces F, G ∈ Φ with ∅ ≺ F G,
µΦ (F, G) = (−1)dim G−dim F ,
and µΦ (∅, G) = (−1)dim G+1 χ(G) for G 6= ∅.
Towards a proof of this result, let ψ : Φ × Φ → Z be the map stipulated
in Theorem 3.15, i.e.,
(
(−1)dim G−dim F
if ∅ ≺ F G ,
ψ(F, G) :=
(−1)dim G+1 χ(G) if F = ∅ ≺ G .
Recall from Section 2.2 that the M¨
obius function µΦ is the inverse of the zeta
function ζΦ and hence unique. Thus, to prove the claim in Theorem 3.15,
namely, that µΦ (F, G) = ψ(F, G), it is sufficient to show that ψ satisfies the
defining relations (2.4) for the M¨obius function. That is, we have to show
that ψ(F, F ) = 1 and, for F ≺ G,
X
ψ(K, G) = 0 .
(3.16)
F KG
That ψ(F, F ) = 1 is clear from the definition, so the meat lies in (3.16). Here
is a first calculation that shows that we are on the right track by thinking of
Euler characteristics: for F = ∅, we compute
X
X
ψ(K, G) = ψ(∅, G) +
(−1)dim G−dim K
∅KG
∅≺KG
= ψ(∅, G) + (−1)dim G χ(G)
= 0.
3.4. M¨
obius Functions of Face Lattices
57
For the general case ∅ ≺ F G Q, we would like to make the same
argument but unfortunately we don’t know if the interval [F, G] ⊆ Φ is
isomorphic to the face lattice of a polyhedron (see, however, Exercise 3.31).
We will do something else instead and take a route that emphasizes the
general geometric idea of modelling geometric objects locally by simpler ones.
Namely, we will associate to each face F a polyhedral cone that captures the
structure around F . We already used this idea in the proof of Theorem 3.5.
Let Q ⊆ Rd be a polyhedron and q ∈ Q. The tangent cone of Q at q
is defined by
Tq (Q) := {q + u : q + u ∈ Q for all > 0 sufficiently small} .
See Figure 11 for examples.
Figure 11. Sample tangent cones of a quadrilateral.
The following result says that Tq (Q) is the translate of a full-dimensional
polyhedral cone which justifies the name tangent cone.
Proposition 3.16. Let Q = {x ∈ Rd : hai , xi ≤ bi , i ∈ [n]} be a polyhedron.
Then the tangent cone of Q at q ∈ Q is
Tq (Q) = {x ∈ Rd : hai , xi ≤ bi for all i with hai , qi = bi }.
(3.17)
In particular, if p, q ∈ Q are both contained in the relative interior of a face
F , then Tq (Q) = Tp (Q).
Proof. We observe that Tq (Q) = Tq−r (Q − r) for all r ∈ Rd , and so we
may assume that q = 0. Let I := {i ∈ [n] : 0 = hai , 0i = bi }. By definition,
a point u is contained in T0 (Q) if and only if there is an 0 > 0 such that
u ∈ Q for all 0 < < 0 . That is,
hai , ui ≤ bi
58
3. Polyhedral Geometry
for all i = 1, . . . , n. For i 6∈ I, we have 0 < bi , and thus we can find such an
0 for any u. So, the only restrictions are hai , ui ≤ 0 for all i ∈ I, which are
independent of > 0. This proves the first claim.
For the second claim, we note from Lemma 3.3 that p ∈ Q is contained
in the relative interior of the same face as q = 0 if and only if I = {i ∈ [n] :
hai , pi = bi }. Hence (3.17) gives Tp (Q) = Tq (Q).
Proposition 3.16 prompts the definition of tangent cones of faces: for a
nonempty face F Q we define the tangent cone of Q at F as
TF (Q) := Tq (Q)
for any point q ∈ F ◦ . The next result solidifies our claim that the tangent
cone models the facial structure of Q around F .
Lemma 3.17. Let Q be a polyhedron and F Q a nonempty face. The
tangent cone TF (Q) is the translate of a polyhedral cone of dimension dim Q
with lineality space parallel to TF (F ) = aff(F ). The faces of TF (Q) are in
bijection with the faces of Q that contain F via
F G 7→ TF (G) .
Proof. The claims follow from the representation (3.17) and what we learned
about polyhedra in Section 3.1. We may assume that
n
o
Q = x ∈ Rd : hai , xi ≤ bi for i = 1, . . . , n
is a full-dimensional polyhedron. Let p ∈ F ◦ and I = {i ∈ [n] : hai , pi = bi }.
Then from (3.17), it follows that TF (Q) is given by a subset of the inequalities
defining Q and hence that Q ⊆ TF (Q). This shows that TF (Q) is fulldimensional. When we translate TF (Q) by −p, we obtain
n
o
TF −p (Q − p) = x ∈ Rd : hai , xi ≤ 0 for all i ∈ I ,
a polyhedral cone with lineality space
L = {x : hai , xi = 0 for all i ∈ I} = aff(F ) − p .
For the last claim, we use Exercise 3.3, which says that for every face
G Q that contains F , there is an inclusion-maximal subset J ⊆ I with
G = {x ∈ Q : hai , xi = bi for all i ∈ J}. In particular, J defines a face of
TF (Q), namely,
{x ∈ TF (Q) : hai , xi = bi for all i ∈ J}
which, by Proposition 3.16, is exactly TF (G). To see that the map G 7→
ˆ TF (Q)
TF (G) is a bijection, we observe that the map that sends faces G
ˆ ∩ Q is an inverse.
to G
3.4. M¨
obius Functions of Face Lattices
59
The gist of Lemma 3.17 is that for a nonempty face F ≺ Q, the posets
[F, Q] (considered as an interval in Φ(Q)) and Φ(TF (Q)) \ {∅} are isomorphic.
With these preparations at hand, we can prove Theorem 3.15.
Proof of Theorem 3.15. Let F G be two faces of Q. We already treated
the case F = ∅, so we may assume that F is nonempty. To show that ψ
satisfies (3.16), we use Lemma 3.17 to replace the sum in (3.16) over faces
in the interval [F, G] of Φ(Q) by the nonempty faces of TF (G). For F ≺ G,
we thus compute
X
X
ψ(F, K) =
(−1)dim K−dim F .
(3.18)
F KG
TF (F )TF (K)TF (G)
Now by Lemma 3.17, the tangent cone TF (G) is the translate of a polyhedral
cone with lineality space L = aff(F ) − p for p ∈ F ◦ . Hence, we can continue
our computation by noting that the right-hand side of (3.18) equals the
Euler characteristic of the line-free polyhedral cone TF (G)/L, and so, by
Corollary 3.12,
X
(−1)dim K−dim F = (−1)dim F χ(TF (G)/L) = 0 .
TF (F )TF (K)TF (G)
To justify the last equality, we have to rule out that TF (G)/L is bounded.
This happens exactly when TF (G) = aff(F ) = TF (F ). However, using again
Lemma 3.17, this would imply F = G which would contradict our assumption
F ≺ G.
We will use tangent cones again in Chapter 5 to compute the M¨obius
function of (seemingly) more complicated objects. Yet another application
of tangent cones is given in Section 3.6.
We finish this section by returning to a theme of Section 2.3, from which
we recall the notion of an Eulerian poset, i.e., one that comes with M¨obius
function
µ(x, y) = (−1)λ[x,y]
where λ[x, y] is the length of the interval [x, y]. The length of an interval
[F, G] in a face lattice is dim G − dim F , and so with Theorem 3.15 we
conclude:
Corollary 3.18. The face lattice Φ(Q) of a polyhedron Q is Eulerian.
This allows us to derive an important corollary to Theorem 2.11, the
reciprocity theorem for zeta polynomials of Eulerian posets, which we will
apply to face lattices of a special class of polytopes. Namely, the polytope
P is simplicial if all of its proper faces are simplices (equivalently, if all of
its facets are simplices). We will compute the zeta polynomial ZΦ(P) (n) of
60
3. Polyhedral Geometry
the face lattice of a simplicial d-polytope P via Proposition 2.2, for which we
have to count multichains
∅ = F0 F1 · · · Fn = P
(3.19)
formed by faces of P. Because P is simplicial, these multichains are of a
special form. Namely, let Fj be the largest proper face of P in (3.19). Then
(3.19) looks like
∅ = F0 F1 · · · Fj ≺ P = · · · = P
(3.20)
and the multichain ∅ = F0 F1 · · · Fj consists entirely of simplices.
The variables here are j (which ranges from 0 to n − 1) and the dimension
of Fj (which ranges from 0 to d − 1, not taking into account ∅). The latter
naturally gives rise to face numbers, and so we count
ZΦ(P) (n) = # multichains of the form (3.20)
= n+
d−1
X
k=0
fk
n−1
X
ZΦ(∆k ) (j) ,
j=0
where ∆k denotes a k-simplex. (The first term on the right-hand side counts
the possible multichains of the form ∅ = · · · = ∅ ≺ P = · · · = P.) With
Exercise 3.33, this simplifies to
ZΦ(P) (n) = n +
d−1
X
fk
k=0
n−1
X
j k+1 =
j=0
d−1
X
k=−1
fk
n−1
X
j k+1 ,
j=0
where we set f−1 := 1, accounting for ∅. The sums over j are polynomials
in n, closely related to the Bernoulli polynomials Bk (n) defined through
X Bk (n)
z enz
=
zk .
ez − 1
k!
(3.21)
k≥0
The Bernoulli numbers are Bk := Bk (0). The reason we are interested in
these constructs is the identity (Exercise 3.34)
n−1
X
j=0
j k−1 =
1
(Bk (n) − Bk ) ,
k
3.5. Uniqueness of the Euler Characteristics and Zaslavsky’s Theorem
61
whence
ZΦ(P) (n) =
=
=
=
d
X
fk−1
(Bk+1 (n) − Bk+1 )
k+1
k=0
d
k+1 X
fk−1 X k + 1
Bk+1−m nm
k+1
m
m=1
k=0
d
d X
X
k + 1 fk−1
nm+1
Bk−m
m+1 k+1
m=0
k=m
d d
m+1
X
X n
k
Bk−m fk−1 .
m
m+1
m=0
(3.22)
k=m
Here the second equation follows from Exercise 3.34. By Theorem 2.11 and
Corollary 3.18, every other coefficient of this polynomial is zero:
Theorem 3.19. If P is a simplicial d-polytope then
d X
k
Bk−m fk−1 = 0
m
k=m
for m = d − 1, d − 3, . . . .
These face-number identities for simplicial polytopes are called the Dehn–
Sommerville relations; they are usually stated in the form
d−1
X
k k+1
(−1)
fk = (−1)d−1 fm
(3.23)
m+1
k=m
for m = −1, 1, 3, . . . The fact that these are equivalent to Theorem 3.19 is
the subject of Exercise 3.35. We remark that the case m = −1 in (3.23) is
the Euler–Poincar´e formula (3.15).
3.5. Uniqueness of the Euler Characteristics and Zaslavsky’s
Theorem
It is a valid question if the Euler characteristic is the unique valuation on
PCd with the properties of Theorem 3.5, i.e.,
χ(∅) = 0
and
χ(P) = 1
(3.24)
for any nonempty closed polytope P ⊂ Rd . The answer is no: as we will
see shortly, the conditions (3.24) do not determine χ(Q) for an unbounded
polyhedron Q.
Here is the situation on the real line: the building blocks for polyconvex
sets are points {p} with p ∈ R and open intervals of the form (a, b) with
62
3. Polyhedral Geometry
a, b ∈ R ∪ {±∞}. Let χ be a valuation on polyconvex sets in R that satisfies
the properties of Theorem 3.5. We need to define χ({p}) = 1 and from
1 = χ([a, b]) = χ((a, b)) + χ({a}) + χ({b})
for finite a and b, we infer that χ((a, b)) = −1. The interesting part now
comes from unbounded intervals. For example, we can set χ((a, ∞)) =
χ((−∞, a)) = 0 and check that this indeed defines a valuation: the value of
χ on a closed unbounded interval [a, ∞) is
χ([a, ∞)) = χ({a}) + χ((a, ∞)) = 1 ,
in contrast to χ([a, ∞)) = 0 (by Proposition 3.13). What about the value
on R = (−∞, ∞)? It is easy to see (Exercise 3.36) that χ(R) = 1. This was
a proof (on hands and knees) for the case d = 1 of the following important
result:
Theorem 3.20. There is a unique valuation χ : PCd → Z of polyconvex sets
in Rd such that χ(∅) = 0 and χ(Q) = 1 for any closed polyhedron Q 6= ∅.
We can prove this result using the same arguments as in Section 3.3 by
way of H-polyconvex sets and, in particular, Lemma 3.7. However, as in
Section 3.3, we will need to make a choice for the value of χ on relatively
open polyhedra. On the other hand, if χ is unique, then there isn’t really a
choice for χ(Q◦ )...
Proposition 3.21. Suppose that χ is a valuation such that χ(Q) = 1 for all
nonempty closed polyhedra Q. If Q is a closed polyhedron with lineality space
L = lineal(Q), then
(
(−1)dim Q if Q/L is bounded,
◦
χ(Q ) =
0
otherwise.
Proof. Let Q be a nonempty polyhedron. We can think of χ as a function
on the face lattice Φ = Φ(Q) given by
X
χ(G) =
χ(F ◦ )
(3.25)
F G
and χ(∅) = 0. We apply M¨obius inversion (Theorem 2.12) to χ to obtain
X
X
χ(G◦ ) =
χ(F ) µΦ (F, G) =
µΦ (F, G) = −µΦ (∅, G).
F G
∅≺F G
The result now follows from Theorem 3.15.
Proof of Theorem 3.20. To show that χ is a valuation on PCd , we revisit the argumentation of Section 3.3 by way of H-polyconvex sets. The
value of χ on relatively open polyhedra is given by Proposition 3.21. We
explicitly give the crucial step (analogous to Lemma 3.7), namely: let Q
3.5. Uniqueness of the Euler Characteristics and Zaslavsky’s Theorem
63
be a relatively open polyhedron and H a hyperplane such that Q ∩ H 6= ∅.
Define the three nonempty polyhedra Q< , Q= , Q> as the intersection of Q
with H < , H = , H > , respectively. If Q is bounded, then we are exactly in the
situation of Lemma 3.7. Now let Q be unbounded. The closure of Q= is a
face of the closure of both Q< and Q> . Hence, if one of them is bounded,
then so is Q= . Thus
χ(Q) = χ(Q< ) + χ(Q= ) + χ(Q> ) .
That χ(Q) = 1 for all closed polyhedra Q follows from (3.25). This
also shows uniqueness: By Proposition 3.21, the value on relatively open
polyhedra is uniquely determined. By Proposition 3.8, for every polyconvex
set S there is a hyperplane arrangement H such that S is H-polyconvex and
hence can be represented as a union of disjoint relatively open polyhedra. There is merit in having several “Euler characteristics.” We will illustrate
this in the remainder of this section with a theorem that is one of the gems
of geometric combinatorics—Zaslavsky’s theorem, Theorem 3.23 below. To
state it, we need the notion of characteristic polynomials.
Let Π be a graded poset with minimal element ˆ0. The characteristic
polynomial of Π is
X
χΠ (n) :=
µΠ (ˆ0, x) nrank(x)
x∈Π
0, x], the rank of x ∈ Π. In many situations the characterwhere rank(x) = l[ˆ
istic polynomial captures interesting combinatorial information about the
poset. Here is a very simple example.
Proposition 3.22. Let P be a polytope. Let χΦ (n) be the characteristic
polynomial of the face lattice Φ = Φ(P). Then χΦ (−1) is the number of faces
of P and χΦ (1) = χ(P) − 1 = 0.
Proof. Recall from Theorem 3.15 that the M¨
obius function of Φ is given by
µΦ (F, G) = (−1)dim G−dim F whenever F G are faces of P. The rank of a
face G is given by rank(G) = dim(G) + 1. Hence
X
χΦ (n) =
(−1)dim F +1 ndim F +1 .
(3.26)
∅F P
Thus, the evaluation of χΦ (n) at n = −1 simply counts the number of faces
and for n = 1, equation (3.26) reduces to 1 − χ(P) via the Euler–Poincar´e
formula (3.15).
Let H = {H1 , H2 , . . . , Hk } be an arrangement of hyperplanes in Rd . The
lattice of flats L(H) of H is the collection of affine subspaces
(
)
\
L(H) =
Hi 6= ∅ : I ⊆ [k]
i∈I
64
3. Polyhedral Geometry
Figure 12. The poset of flats for the line arrangement in Figure 6.
ordered by reverse inclusion. That is, for two flats F, G ∈ L(H), we have
F G if G ⊆ F . (Figure 12 shows the poset of flats of the arrangement of
six lines in Figure 6.) The minimal element is the empty intersection given
by ˆ
0 = Rd . There is a maximal element precisely if all hyperplanes have a
point in common, in which case we may assume that all hyperplanes pass
through the origin and call H central. An example of a central arrangement
Figure 13. An arrangement of three coordinate hyperplanes and its
intersection poset.
(consisting of the three coordinate hyperplanes x1 = 0, x2 = 0, and x3 = 0)
is given in Figure 13. An arrangement is essential if there is no affine line
that is parallel to all hyperplanes in H. The characteristic polynomial of
H is the characteristic polynomial for its lattice of flats χH (n) = χL(H) (n).
For example, the characteristic polynomial of the arrangement of six lines in
Figure 6 is χH (n) = n2 − 6n + 12.
3.5. Uniqueness of the Euler Characteristics and Zaslavsky’s Theorem
65
As mentioned in Section 3.3, the hyperplane arrangement H decomposes
Rd into relatively open polyhedra: For every point p ∈ Rd there is a unique
σ ∈ {<, =, >}k such that
p ∈ Hσ = H1σ1 ∩ H2σ2 ∩ · · · ∩ Hkσk .
A (closed) region of H is (the closure of) a full-dimensional open polyhedron
in this decomposition. In particular, a region can be bounded or unbounded.
We define r(H) to be the number of all regions and b(H) to be the number
of bounded regions of H. For example, the arrangement in Figure 6 has
19 regions and seven bounded regions. Exercises 3.40–3.43 give a few more
examples. Our proof of the following famous theorem makes use of the fact
that we have two “Euler characteristics” at our disposal.
Theorem 3.23. Let H be an essential hyperplane arrangement in Rd and
χH (n) its characteristic polynomial. Then
r(H) = (−1)d χH (−1)
and
b(H) = (−1)d χH (1) .
This theorem also holds true for non-essential arrangements if we replace
the term bounded by relatively bounded ; see Exercise 3.44.
S
Proof. WeSdenote by H = H1 ∪ · · · ∪ Hk the union of all hyperplanes.
Then Rd \ H = R1 ] R2 ] · · · ] Rm where R1 , . . . , Rm are the regions of H.
Since every region is a d-dimensional open polyhedron,
[ r(B) = (−1)d χ Rd \ H and
[ b(B) = (−1)d χ Rd \ H ,
by Corollary 3.10 and Proposition 3.21. Let φ : PCd → R be any valuation.
Then
[ X
φ Rd \ H = φ Rd − φ (H1 ∪ H2 · · · ∪ Hk ) =
(−1)|I| φ (HI )
I⊆[k]
where
the last equation is the inclusion–exclusion formula (3.9) with HI :=
T
i∈I Hi . Now by Exercise 2.12,
[ X
X
φ Rd \ H =
(−1)|I| φ(HI ) =
µL(H) (ˆ0, F ) φ(F ) . (3.27)
I⊆[k]
F ∈L(H)
Since every F ∈ L(H) is an affine subspace and hence a closed polyhedron,
we have χ(F ) = 1, and so for φ = χ in (3.27) we get the evaluation χH (1).
Simiarly, for φ = χ, we have χ(F ) = (−1)dim F and hence (3.27) for φ = χ
yields χH (−1).
66
3. Polyhedral Geometry
3.6. The Brianchon–Gram Relation
In this final section we want to once again allure to Euler characteristics to
prove an elegant result in geometric combinatorics, the Brianchon–Gram
relation—Theorem 3.24 below. This is a simple geometric equation that
can be thought of as a close relative to the Euler–Poincar´e formula (3.15).
It is best stated in terms of indicator functions: For a subset S ⊆ Rd , the
indicator function of S is the function [S] : Rd → Z defined by
(
1 if p ∈ S,
[S](p) :=
0 otherwise.
Theorem 3.24. Let P ⊂ Rd be a polytope. Then
X
[P ] =
(−1)dim F [TF (P)] .
(3.28)
∅≺F P
Recall that TF (P) is the tangent cone of P at a face F , a notion that
we popularized in Section 3.4 in connection with the M¨obius function of a
face lattice. The Brianchon–Gram relation is a simple truth which we will
see in action in Section 5.7. We devote the remainder of this section to its
verification.
First note that Lemma 3.17 implies P ⊆ TF (P) for all F 6= ∅. Hence,
for a point p ∈ P, we compute
X
X
(−1)dim F [TF (P)](p) =
(−1)dim F = χ(P) = 1 ,
∅≺F P
∅≺F P
by Theorem 3.5 and the Euler–Poincar´e formula. This leaves us to prove
that the right-hand side of (3.28) evaluates to 0 for all p 6∈ P. This is clear
when p 6∈ aff(P), and so we may assume that P is full-dimensional.
We will use the following terminology: we call a point p ∈ Rd beneath
a face F P if p ∈ TF (P) and beyond F otherwise. Let
[
|Bp (P)| :=
{F ◦ : p is beyond F ⊂ P} .
Since we assume that P is full-dimensional, there is no point beyond P and
thus |Bp (P)| is a subset of the boundary. More importantly, |Bp (P)| is a
proper subset of ∂P by Exercise 3.38. Now |Bp (P)| is a polyconvex set and
for p ∈ Rd \ P the right-hand side of (3.28) equals
X
X
X
(−1)dim F =
(−1)dim F −
(−1)dim F
F ≺P
p beneath F
F ≺P
F ≺P
p beyond F
= χ(P) − χ(|Bp (P)|) .
Hence, we want to show that |Bp (P)| has Euler characteristic 1 whenever
p 6∈ P. By translating both the point p and the polytope P by −p, we may
assume that p = 0. Let C := cone(P); see Figure 14 for a sketch. Since
3.6. The Brianchon–Gram Relation
67
Figure 14. The cone C = cone(P) and the faces beyond 0.
0 6∈ P, the cone C is pointed. The following construction is reminiscent of
our proof of Theorem 3.5 in Section 3.3; it is illustrated in Figure 15.
Figure 15. The decomposition of Proposition 3.25.
Proposition 3.25. Let P be a full-dimensional polytope and C := cone(P).
Then
]
C \ {0} =
{cone(F )◦ : 0 is beyond F ≺ P} .
Proof. A point u ∈ Rd is contained in C \ {0} if and only if λu ∈ P for some
λ > 0. Since P is compact, there are a minimal such λ and a unique face F
such that λu ∈ F ◦ , by Lemma 3.3. Checking the definition shows that 0 is
beyond F .
Now if F ≺ P is a face such that 0 is beyond F , then cone(F )◦ is
a relatively open polyhedral cone of dimension dim F + 1. Hence, with
68
3. Polyhedral Geometry
Proposition 3.25,
χ(C) = χ({0}) +
X
(−1)dim F +1 = 1 − χ(|B0 (P)|) .
F ≺P
06∈TF (P)
Since C is pointed, we conclude, with the help of Proposition 3.13, that
χ(|B0 (P)|) = 1, which was the missing piece to finish the proof of Theorem 3.24. We will encounter the set |Bp (P)| again in Chapter 5.3, where also
the notation will make more sense.
Notes
Since polyhedra are simply solution sets to a system of linear inequalities and
equations, they are ubiquitous in mathematics. The 3-dimensional case of the
Euler–Poincar´e formula (3.15) was proved by Leonard Euler in 1752 [28, 29].
The full (i.e., higher-dimensional) version of (3.15) was discovered by Ludwig
Schl¨afli in 1852 (though published only in 1902 [65]), but Schl¨
afli’s proof
implicitly assumed that every polytope is shellable (as did numerous proofs
of (3.15) that followed Schl¨afli’s), a fact that was established only in 1971
by Heinz Bruggesser and Peter Mani [20]. The first airtight proof of (3.15)
(in 1893, using tools from algebraic topology) is due to Henri Poincar´e [58]
(see also, e.g., [37, Theorem 2.44]).
The Dehn–Sommerville relations (Theorem 3.19 and, equivalently, (3.23))
are named after Max Dehn, who proved their 5-dimensional instance in
1905 [24], and D. M. Y. Sommerville, who establised the general case in
1927 [69]. That the Dehn–Sommerville relations follow from the reciprocity
theorem for the zeta polynomial of the face lattice of a simplicial polytope
was realized by Richard Stanley in [73] at the inception of zeta polynomials‘.
As already mentioned in Chapter 1, classifying face numbers is a major
research problem; in dimension 3 this question is answered by Steinitz’s
theorem [81] (see also [88, Lecture 4]). The classification question is open in
dimension 4. One can ask a similar question for certain classes of polyhedra,
e.g., simplicial polytopes, for which the Dehn–Sommerville relations give
necessary conditions.
Theorem 3.23 was part of Thomas Zaslavsky’s Ph.D. thesis, which started
the modern theory of hyperplane arrangements [86]. A nice survey article
on the combinatorics of hyperplane arrangements is [77]. One can study
hyperplane arrangements in C, which gives rise to numerous interesting
topological considerations [55].
The 3-dimensional case of the Brianchon–Gram relation (Theorem 3.24)
was discovered by Charles Julien Brianchon in 1837 [19] and—as far as
we know—independently reproved by Jørgen Gram in 1874 [32]. It is not
clear who first proved the general d-dimensional case of the Brianchon–Gram
relation; the oldest proofs we could find were from the 1960s [35, 46, 56, 67].
Exercises
69
Exercise 3.46 illustrates that the Brianchon–Gram relation can be understood
in terms of valuations; this point of view was pioneered by Jim Lawrence [49].
For (much) more about the wonderful combinatorial world of polyhedra,
including proofs of the Minkowski–Weyl Theorem, we recommend [35, 88].
Exercises
3.1 Prove (3.4), namely, that for a polyhedron given in the form (3.2),
\
aff(Q) =
{Hi : Q ⊆ Hi } .
3.2 Let Q be a polyhedron given in the form (3.2) and consider a face F
of Q. Renumber the hyperplanes H1 , H2 , . . . , Hk so that F ⊆ Hj for
1 ≤ j ≤ m and F 6⊆ Hj for j > m, for some index m. Prove that
F =
m
\
j=1
Hj ∩
k
\
Hj≤ .
j=m+1
3.3 Let F be a nonempty face of the polyhedron
n
o
Q = x ∈ Rd : hai , xi ≤ bi for i = 1, . . . , k
and let G be a face of Q that contains F . By Exercise 3.2, there is a
subset I ⊆ [k] of indices such that
F = {x ∈ Q : hai , xi = bi for i ∈ I} .
Show that there exists an inclusion-maximal subset J ⊆ I such that
G = {x ∈ Q : hai , xi = bi for i ∈ J} .
3.4 Let Q = {x ∈ Rd : hai , xi ≤ bi for i = 1, . . . , k} be a full-dimensional
polyhedron. For p ∈ Rd and > 0, let B(p, ) be the ball of radius centered at p. A point p ∈ Q is an interior point of Q if B(p, ) ⊆ Q
for some > 0. Show that (3.5) equals the set of interior points of Q.
3.5 Let Q = {x ∈ Rd : hai , xi ≤ bi for i = 1, . . . , k} be a polyhedron. For a
subset S ⊆ Q define I(S) = {i ∈ [k] : ai p = bi for all p ∈ S}.
(a) Show that aff(Q) = {x ∈ Rd : ai x = bi for all i ∈ I(Q)}.
(b) Let L = aff(Q). Show that a point p is in the relative interior of Q
if and only if
B(p, ) ∩ L ⊆ Q ∩ L
for some > 0 (see Exercise 3.4).
(c) Show that Q◦ = {x ∈ Q : hai , xi < bi for all i 6∈ I(Q)}.
3.6 Let Q ⊂ Rd be a nonempty full-dimensional polyhedron and H a
hyperplane such that Q◦ ∩ H 6= ∅. Show that dim(Q◦ ∩ H) = dim Q − 1.
70
3. Polyhedral Geometry
3.7 Prove that, if F and G are faces of a polyhedron that are unrelated in
the face lattice (i.e., F 6⊆ G and G 6⊆ F ), then
F ◦ ∩ G◦ = ∅ .
3.8 Let P be the convex hull of a finite set S. Prove that vert(P) is the
unique inclusion-minimal subset T ⊆ S with P = conv(T ).
3.9 Let P be the convex hull of {p1 , p2 , . . . , pm }. Prove that
(
)
m
X
P = λ1 p1 + λ2 p2 + · · · + λm pm :
λi = 1, λ1 , . . . , λm ≥ 0 .
i=1
3.10 Let P = conv{p1 , . . . , pm } be a polytope. Show that a point q is in the
relative interior of P if there are λ1 , . . . , λm > 0 such that
q = λ1 p1 + · · · + λm pm
and
λ1 + · · · + λm = 1.
For which polytopes is this an if and only if condition?
3.11 Let Q = {x ∈ Rd : A x ≤ b} be a nonempty polyhedron. Show that
lineal(Q) = ker(A). Infer that p + lineal(Q) ⊆ Q for all p ∈ Q.
3.12 The definition of lineality spaces makes sense for arbitrary convex sets
K in Rd . Show that in this more general situation, convexity implies
that p + lineal(K) ⊆ K for all p ∈ K.
3.13 Prove that a polyhedron Q ⊆ Rd is a cone if and only if
n
o
Q = x ∈ Rd : A x ≤ 0
for some irredundant matrix A.
3.14 Show that
(x, y, z) ∈ R3 : z ≥ x2 + y 2
is a cone but not polyhedral.
3.15 Show that a polyhedral cone C is pointed if and only if p, −p ∈ C
implies p = 0.
3.16 Prove that the vertices of a polyhedron Q are precisely the 0-dimensional
faces of Q.
3.17 Show that the i-faces of P are in bijection with the (i + 1)-faces of
hom(P), for all i = 0, 1, . . . , d.
3.18 Prove (3.7), namely, hom(P) = cone {(v, 1) : v ∈ vert(P)} . (Hint: you
will need Theorem 3.1.)
3.19 Let C ⊂ Rd+1 be a pointed polyhedral cone and let R ⊂ C \ {0} be a
generating set. Show that, if H ⊂ Rd+1 is a hyperplane such that for
every r ∈ R, there is a λr ≥ 0 with λr r ∈ H, then C ∩ H is a polytope.
Exercises
71
3.20 Let vert(Q1 ) = {u1 , u2 , . . . , um } and vert(Q2 ) = {v1 , v2 , . . . , vn }, and
consider a point s+t ∈ Q1 +Q2 ;P
then there
P are coefficients λ1 , . . . , λm ≥ 0
and µ1 , . . . , µn ≥ 0 such that i λi = j µj = 1 and
s+t =
m
X
λ i ui +
i=1
n
X
µj vj
j=1
Now set αij = λi · µj ≥ 0 for (i, j) ∈ [m] × [n]. Prove that
X
s+t =
αij (ui + vj ) .
(i,j)∈[m]×[n]
3.21 Show that if K1 , K2 ⊆ Rd are convex then so is the Minkowski sum
K1 + K2 .
3.22 Prove Corollary 3.2: Let Q ⊂ Rd be a polyhedron and φ(x) = A x + b
an affine projection Rd → Re . Then φ(Q) is a polyhedron. If P ⊂ Rd is
a polytope, then P ∩ Q is a polytope.
3.23 Recall from Theorem 3.1 that every polyhedron Q is of the form Q =
P + C where P is a polytope and C is a polyhedral cone. Prove that for
every nonempty face F Q, there are unique faces F 0 P and F 00 C
such that F = F 0 + F 00 .
3.24 Prove that a face of a (bounded) polyhedron is again a (bounded)
polyhedron.
3.25 Prove that every face of a polyhedron P is the intersection of some of
the facets of P.
3.26 Show that the unit disc B2 = {(x, y) ∈ R2 : x2 + y 2 ≤ 1} is not a
polyconvex set.
3.27 Prove Proposition 3.6, namely, that the function χ(H, ·) from PC(H)
to Z is a valuation.
3.28 Prove Proposition 3.11: Let P be a nonempty polytope with 0 ∈ P◦ .
For any proper face F ≺ P, the set C0 (F ) is a relatively open polyhedral
cone of dimension dim F + 1. Furthermore,
]
C0 (P) = aff(P) =
C0 (F ) .
∅≺F ≺P
3.29 Prove Corollary 3.12: If Q is a polyhedron with lineality space L =
lineal(Q), then
χ(Q) = (−1)dim L χ(Q/L) .
3.30 Let Q ⊂ Rd be a pointed polyhedron of dimension d. Let v a vertex
of Q and H = {x : ha, xi = δ} a hyperplane such that Q ⊆ H ≤ and
{v} = Q ∩ H. For > 0, define H = {x : ha, xi = δ − }.
72
3. Polyhedral Geometry
(a) For every sufficiently small > 0, show that the polytope P(H, v) =
Q ∩ H has the following property: There is a bijection between the
k-faces of P(H, v) and the (k + 1)-faces of Q containing v, for all
k = −1, . . . , d − 1.
This shows that the face lattice Φ(P(H, v)) is isomorphic to [v, Q]Φ(Q)
and, in particular, independent of the choice of H. We call any such
polytope P(H, v) the vertex figure of Q at v and denote it by Q/v.
(b) If you know about projective transformations (see, for example, [88,
Chapter 2.6]), prove that P(H, v) and P(H 0 , v) are projectively
isomorphic for any two supporting hyperplanes H, H 0 for v.(This
justifies calling Q/v the vertex figure of Q at v.)
(c) Show that Φ(Q/v) is isomorphic to [v, Q]Φ(Q) .
3.31 Continuing Exercise 3.30, we call a polytope P a face figure of Q at
the face G if Φ(P) is isomorphic to [G, Q]Φ(Q) . We denote any such
polytope with Q/G.
(a) Verify that the following inductive construction yields a face figure:
If dim G = 0 then Q/G is the vertex figure of Exercise 3.30. If
dim G > 0, pick a face F ⊂ G of dimension dim G − 1. Then Q/F
exists and, in particular, there is a vertex vG of Q/F corresponding
to G. We define the face figure Q/G as the vertex figure (Q/F )/vG .
(b) As an application of face figures, show that the M¨obius function of
a polyhedron is given by µΦ (F, G) = χ((G/F )◦ ).
3.32 Let Q ⊂ Rd be a polyhedron given as
Q = {x ∈ Rd : hai , xi ≤ bi for all i = 1, . . . , n} = P + C
where P is a polytope and C is a polyhedral cone.
(a) Show that hom(Q) is a polyhedral cone given by
hom(Q) = {(x, t) ∈ Rd × R : t ≥ 0, Ax ≤ tb}
and that hom(Q) is pointed whenever Q is.
For s > 0, consider the intersection of hom(Q) with the hyperplane
Hs := {(x, t) ∈ Rd × R : t = s}. It is clear from the definition that
hom(Q) ∩ Hs = sQ × {s}.
(b) Show that hom(Q) ∩ H0 is a face isomorphic to C.
(c) Show that the k-faces of Q are in bijection with the (k + 1)-faces of
hom(Q) not meeting H0 .
(d) Argue that χ(hom(Q)) = χ(C) − χ(Q) and infer Corollary 3.14 from
Proposition 3.13.
3.33 Let ∆ be a k-dimensional simplex. Show that the zeta polynomial of
the face lattice of ∆ is ZΦ(∆) (n) = nk+1 .
3.34 Show the following properties of Bernoulli polynomials and numbers.
Exercises
(a)
73
n−1
X
j k−1 =
1
k
(Bk (n) − Bk ) .
j=0
(b) Bk (n) =
k X
k
m=0
m
Bk−m nm .
(c) Bk (1 − n) = (−1)k Bk (n) .
(d) Bk = 0 for all odd k ≥ 3.
3.35 Prove that Theorem 3.19 is equivalent to (3.23):
d−1
X
k k+1
(−1)
fk = (−1)d−1 fm
m+1
k=m
for m = −1, 1, 3, . . . (Hint: Start with Theorem 2.11 for ZΦ(P) (n),
rewrite both sides of the identity in terms of Bernoulli polynomials as
in (3.22), and use Exercise 3.34(c).)
3.36 Show that χ(R) = 1.
3.37 Real-valued valuations φ : PCd → R on Rd form a vector space Val and
valuations that satisfy the properties of Theorem 3.5 constitute a vector
subspace U ⊆ Val. For the real line, what is the dimension of U ?
3.38 Let P ⊂ Rd be a full-dimensional polytope. Show that for every p ∈ Rd ,
there is at least one facet F such that p is beneath F .
3.39 Prove that any two elements in L(H) have a meet (we call such a poset
a meet semilattice). Furthermore, show that L(H) is a lattice if and
only if H is central.
3.40 Let H = {xj = 0 : 1 ≤ j ≤ d}, the d-dimensional Boolean arrangement consisting of the coordinate hyperplanes in Rd . (Figure 13 shows
an example.) Show that r(H) = 2d .
3.41 Let H = {xj = xk : 1 ≤ j < k ≤ d}, the d-dimensional real braid arrangement. Show that r(H) = d!.
3.42 Let H be an arrangement in Rd consisting of k hyperplanes in general
position, i.e., each j-dimensional flat of H is the intersection of exactly
d−j hyperplanes, and any d−j hyperplanes intersect in a j-dimensional
flat. (One example is pictured in Figure 16.) Then
k
k
k
k
r(H) =
+
+ ··· +
+
.
d
d−1
1
0
What can you say about b(H)?
3.43 Let H be an arrangement in Rd consisting of k hyperplanes. Show that
k
k
k
k
r(H) ≤
+
+ ··· +
+
,
d
d−1
1
0
74
3. Polyhedral Geometry
Figure 16. An arrangement of four lines in general position and its
intersection poset.
that is: the number of regions of H is bounded by the number of regions
created by k hyperplanes in Rd in general position.
3.44 Let H = {H1 , . . . , Hk } be a hyperplane arrangement in Rd and let L
be the inclusion-maximal linear subspace that is parallel to all hyperplanes in H. Let H0 = {H1 ∩ L⊥ , . . . , Hk ∩ L⊥ } be the arrangement of
hyperplanes in the orthogonal complement L⊥ of L.
(a) Show that L(H) ∼
= L(H0 ).
(b) We call a region R of H relatively bounded if R∩L⊥ is a bounded
region of H0 . Show that (−1)d χL(H) (−1) is the number of relatively
bounded regions.
3.45 Give an elementary proof for the Brianchon–Gram relation (Theorem 3.24) for simplices.
3.46 The dual of the polyhedron Q ⊆ Rd is
n
o
Q∨ := x ∈ Rd : hx, yi ≤ 1 for all y ∈ Q .
Suppose P ⊂ Rd is a d-polytope with 0 ∈ P◦ .
(a) Show that P∨ is a polytope.
(b) Suppose F is a nonempty face of P. Show that (TF (P))∨ is the
convex hull of 0 and
F dual := x ∈ P∨ : hx, yi = 1 for all y ∈ F .
(c) Prove that
[P∨ ] =
X
(−1)dim F [conv(0, F dual )] .
∅≺F ≺P
7→ [Q∨ ]
(d) Prove that [Q]
is a valuation.
(e) Conclude from (b)–(d) an alternative proof of the Brianchon–Gram
relation (Theorem 3.24).
Chapter 4
Rational Generating
Functions
Life is the twofold internal movement of composition and decomposition at once
general and continuous.
Henri de Blainville
We now return to a theme started in Chapter 1: counting functions that
are polynomials. Before we can ask about possible interpretations of these
counting functions at negative integers—the theme of this book—, we need
structural results such as Proposition 1.1, which says that the chromatic
polynomial is indeed a polynomial. We hope that we conveyed the message
in Chapter 1 that such a structural result can be quite nontrivial for a given
counting function. Another example is given by the zeta polynomials of
n (ˆ
Section 2.3—their definition ZΠ (n) := ζΠ
0, ˆ1) certainly does not hint at the
fact that ZΠ (n) is indeed a polynomial. One of our goals in this chapter is
to develop machinery that allows us to detect and study polynomials. We
will do so alongside introducing several families of counting functions and,
naturally, we will discover a number of combinatorial reciprocity theorems
along the way.
4.1. Matrix Powers and the Calculus of Polynomials
To warm up, we generalize in some sense the zeta polynomials from Section
2.3. A matrix A ∈ Cd×d is unipotent if A = I + B where I is the d × d
identity matrix and there exists a positive integer k such that Bk = 0 (that
is, B is nilpotent). The zeta functions from Chapter 2 are our motivating
examples of unipotent matrices. Let’s recall that, thinking of the zeta
function of a poset Π as a matrix, the entry ζΠ (ˆ0, ˆ1) was crucial in Chapter
75
76
4. Rational Generating Functions
2—the analogous entry in powers of ζΠ gave rise to the zeta polynomial
(emphasis on polynomial !)
n ˆ ˆ
ZΠ (n) = ζΠ
0, 1 .
We obtain a polynomial the same way from any unipotent matrix.
Proposition 4.1. Let A ∈ Cd×d be a unipotent matrix, fix indices 1 ≤ i, j ≤
d, and consider the sequence f (n) := (An )ij formed by the (i, j)-entries of
the nth powers of A. Then f (n) agrees with a polynomial in n.
Proof. We essentially repeat the argument behind (2.3) which gave rise to
Proposition 2.4. Suppose A = I + B where Bk = 0. Then
n k−1 X
X
n
n
f (n) = ((I + B)n )ij =
(Bm )ij =
(Bm )ij ,
m
m
m=0
m=0
which is a polynomial in n. Here, the second equality is the Binomial
Theorem; see Exercise 2.1.
n
At the heart of the above proof is the fact that m
is a polynomial in
n, and in fact, viewing this binomial coefficient as a polynomial in the “top
variable” lies at the core of much of the enumerative side of combinatorics,
starting with (0.1)—the very first sample counting function in this
book.
n
Our proof also hints at the fact that the binomial coefficients m
form a
basis for the space of polynomials. Much of what we will do in this chapter
can be viewed as switching bases in one way or another. Here are the key
players:
m
◦ {z
≤ d}
(M)-basis
m: 0≤m
d−m
◦ z (1 − z)
: 0≤m≤d
(F)-basis
z
◦ m
: 0≤m≤d
(H)-basis
◦ z+m
: 0≤m≤d .
(G)-basis
d
Proposition 4.2. The sets (M), (F), (H), and (G) are bases for the vector
space C[z]≤d := {f ∈ C[z] : deg(f ) ≤ d}.
Proof. The set (M) is the canonical basis of C[z]≤d . An explicit change of
basis from (H) to (M) is given in (1.2). For (F), we observe that
d X
d i
d
1 = (z + (1 − z)) =
z (1 − z)d−i
(4.1)
i
i=0
gives an explicit linear combination of z 0 . You will pursue this idea further in
Exercise 4.2. Finally, for (G), it is sufficient to show that its d + 1 polynomials
are linearly independent; see Exercise 4.2.
4.1. Matrix Powers and the Calculus of Polynomials
77
Looking back once more to Chapter 1, a geometric passage from (G) to
(H) is implicit in (1.6). This will be much clearer after Section 4.6.
Our proof of Proposition 4.1 is even closer to the principal structures
of polynomials than one might think. To this end, we consider three linear
operators on the vector space {(f (n))n≥0 } of all complex-valued sequences:
(If )(n) := f (n) ,
(∆f )(n) := f (n + 1) − f (n) ,
(identity operator)
(difference operator)
(Sf )(n) := f (n + 1) .
(shift operator)
They are naturally related through S = I + ∆. Aficionados of calculus will
anticipate the following result; we would like to emphasize the analogy of its
proof with that of Proposition 4.1.
Proposition 4.3. A sequence f (n) is given by a polynomial of degree ≤ d
if and only if (∆m f )(0) = 0 for m > d.
Proof. We start by noting that f (n) = (S n f )(0). If (∆m f )(0) = 0 for all
m > d, then
f (n) = ((I + ∆)n f ) (0)
n d X
X
n
n
m
=
(∆ f )(0) =
(∆m f )(0) ,
m
m
m=0
(4.2)
m=0
a polynomial of degree ≤ d. Conversely, if f (n) is a polynomial of degree ≤ d
P
n
then we can express it in terms of the (H)-basis, i.e., f (n) = dm=0 αm m
for some α0 , α1 , . . . , αd ∈ C. Now, observe that
n
n
∆
=
m
m−1
Pd
n
and hence ∆f = m=1 αm m−1
is a polynomial of degree ≤ d − 1. By
induction, this shows that ∆m f = 0 for m > d.
The numbers (∆m f )(0) play a central role in the above proof, namely,
they are the coefficients of the polynomial f (n) when expressed in terms of
the (H)-basis. They will appear time and again in this section, and thus we
define
d
X
(m)
m
(m) n
f
:= (∆ f )(0) ,
whence
f (n) =
f
.
(4.3)
m
m=0
We make one more observation, namely, that f (n) has degree d if and only
if f (d) 6= 0.
78
4. Rational Generating Functions
We now introduce the main tool of this chapter, the generating function of the sequence f (n): the formal power series
X
F (z) :=
f (n) z n .
n≥0
The set CJzK of all such series is a vector space with the natural addition
and scalar multiplication
X
X
X
f (n) z n + c
g(n) z n :=
(f (n) + c g(n)) z n .
n≥0
n≥0
n≥0
Moreover, formal power series constitute a commutative ring with 1 under
the multiplication
!
!
!
X
X
X X
f (n) z n ·
g(n) z n =
f (k)g(l) z n ,
n≥0
n≥0
n≥0
k+l=n
and Exercise 4.3 says which formal power series have multiplicative inverses.
Borrowing a leaf from calculus, we can also define the derivative
X
d X
f (n) z n :=
n f (n) z n−1 .
dz
n≥0
n≥1
Most of our use of generating functions is intuitively sensible but with the
above definitions also on solid algebraic foundation. For example, our favorite
counting function comes with a generating function that is (or rather, should
n
be) known to any calculus student: for f (n) = m
, we compute
X n
1 mX
F (z) =
zn =
z
n(n − 1) · · · (n − m + 1) z n−m
m
m!
n≥0
n≥m
m
1
zm
1 m d
z
=
.
(4.4)
=
m!
dz
1−z
(1 − z)m+1
A generating function, such as the above example, that can be expressed
as the quotient of two polynomials is called rational. Thanks to (4.3), this
sample generating function gives us a way to compute the generating function
for any polynomial f (n): with the notation of (4.3), we obtain
d
d
X
X
Xn
X
zm
n
(m)
n
F (z) =
f (n) z =
f
z =
f (m)
(4.5)
m
(1 − z)m+1
m=0
m=0
n≥0
n≥0
Pd
f (m) z m (1 − z)d−m
= m=0
.
(4.6)
(1 − z)d+1
The (equivalent) expressions (4.5) and (4.6) representing the rational generating function of the polynomial f (n) allow us to develop several simple but
powerful properties of polynomials and their generating functions.
4.1. Matrix Powers and the Calculus of Polynomials
79
Proposition 4.4. A sequence f (n) is given by a polynomial of degree ≤ d
if and only if
X
h(z)
f (n) z n =
(1 − z)d+1
n≥0
for some polynomial h(z) of degree ≤ d. Furthermore, f (n) has degree d if
and only if h(1) 6= 0.
Proof. The first part of the proposition follows from (4.6) and writing f (n)
in terms of the (H)-basis and h(z) in terms of the (F)-basis.
The second part follows from (4.3) which illustrates that f (n) has degree
d if and only if f (d) =
6 0. This, in turn, holds (by (4.6)) if and only if
h(1) 6= 0.
Proposition 4.4 can be rephrased in terms of recursions for the sequence
(f (n))n≥0 : given a polynomial f (n), we can rewrite the rational generating
function identity given in Proposition 4.4 as
d+1 X
X
X
d+1
d+1
n
h(z) = (1 − z)
f (n) z =
f (n)
(−1)j z n+j .
j
n≥0
n≥0
j=0
Since h(z) has degree ≤ d, the coefficients of z n for n > d on the right-hand
side must be zero, that is,
d+1 X
d+1
(−1)j f (n − j) = 0
(4.7)
j
j=0
for all n ≥ d + 1.
Reflecting some more on (4.7) naturally leads one to see the full strength
of rational generating functions: a sequence (f (n))n≥0 satisfies a linear
recurrence if for some c0 , . . . , cd ∈ C with c0 , cd 6= 0
c0 f (n + d) + c1 f (n + d − 1) + · · · + cd f (n) = 0
(4.8)
for all n ≥ 0. Thus, the sequence (f (n))n≥0 is fully described by the coefficients of the linear recurrence c0 , . . . , cd and the starting values f (0), . . . , f (d).
Proposition 4.5. Let (f (n))≥0 be a sequence of numbers. Then (f (n))n≥0
satisfies a linear recurrence of the form (4.8) if and only if
X
p(z)
F (z) =
f (n)z n =
d
cd z + cd−1 z d−1 + · · · + c0
n≥0
for some polynomial p(z) of degree deg(p) < d.
The reasoning that led to (4.7) also yields Proposition 4.5; see Exercises 4.5 and Exercises 4.6.
80
4. Rational Generating Functions
Let’s look at an example. The first recursion any student sees is likely
that for the Fibonacci numbers defined through
f (0) := 0 ,
f (1) = f (2) := 1 ,
f (n) := f (n − 1) + f (n − 2) for n ≥ 3 .
Proposition 4.5 says that the generating function for the Fibonacci numbers
is rational with denominator 1 − z − z 2 . The starting data give
X
z
f (n) z n =
.
1 − z − z2
n≥1
Exercise 4.7 shows how we can concretely compute formulas from this rational
generating function, but its use is not limited to that—in the next section
we will illustrate how one can derive interesting identities from generating
functions.
Coming back to our setup of f (n) being a polynomial of degree d, we
claim that (4.7) also has to hold for n < 0. Indeed, f (−n) is, of course,
also a polynomial in n of degree s and so comes with a rational generating
function. Towards computing it, we start by plugging in negative arguments
into (4.3):
X
d
d
X
(m) −n
(m)
m n+m−1
f (−n) =
f
=
f (−1)
,
(4.9)
m
m
m=0
m=0
by our very first combinatorial reciprocity instance (0.2). We define the
accompanying generating function as
X
F ◦ (z) :=
f (−n) z n .
n≥1
(There are several good reasons for starting this series only at n = 1; one
of them is (4.9).) Note that F ◦ (z) is rational, by Proposition 4.4. Here is a
blueprint for a formal reciprocity on the level of rational generating functions:
By (4.9),
d
XX
◦
(m)
m n+m−1
F (z) =
f (−1)
zn
m
n≥1 m=0
=
d
X
f
(m)
m
(−1)
m=0
=
d
X
m=0
X n + m − 1
n≥1
m
zn
d
f
(m)
X
( z1 )m
z
(m)
(−1)
=
−
f
= −F ( z1 ) ,
1 m+1
(1 − z)m+1
1
−
m=0
z
m
4.1. Matrix Powers and the Calculus of Polynomials
81
where the penultimate equation follows from Exercise 4.4. This is a formal
rather than a true combinatorial reciprocity, as the calculation does not tell
us if we can think of f (−n) as a genuine counting function.
If (f (n))n≥0 is a sequence satisfying a linear recurrence of the form (4.8),
then we can let the recurrence run backwards: for example, since we assume
that cd 6= 0, the value of f (−1) is determined by f (0), . . . , f (d − 1) by way
of (4.8). The sequence f ◦ (n) := f (−n) satisfies the linear recurrence
cd f ◦ (n + d) + cd−1 f ◦ (n + d − 1) + · · · + c0 f ◦ (n) = 0 .
(4.10)
For example, for the Fibonacci numbers, this gives
n
1
2 3
4 5
6 7
8 ...
1 −1 2 −3 5 −8 13 −21 . . .
f ◦ (n)
The corresponding rational generating function is
z
F ◦ (z) =
.
1 + z − z2
On the level of general rational generating functions, this formal reciprocity
reads like the following.
P
Theorem 4.6. Let F (z) = n≥0 f (n)z n be a rational generating function.
Then
X
F ◦ (z) =
f (−n)z n = −F ( z1 ) .
n≥1
Proof. Assume that
F (z) =
cd
zd
p(z)
+ cd−1 z d−1 + · · · + c0
where c0 , cd 6= 0 and p(z) is a polynomial of degree < d. It follows that
−F ( z1 ) = −
z d p( z1 )
,
c0 z d + c1 z d−1 + · · · + cd
also a rational generating function. The coefficients of −F ( z1 ) satisfy the
linear recurrence (4.10). To prove Theorem 4.6, we thus only have to verify
that the starting data of f ◦ (n) is encoded in the numerator of −F ( z1 ), that
is, we have to verify that
(c0 z d + c1 z d−1 + · · · + cd )F ◦ (z) = −z d p( z1 ) .
This is left to the reader (Exercise 4.8).
The assumption that F (z) is a rational generating function is essential—see
Exercise 4.10.
Theorem 4.7. to be...
82
4. Rational Generating Functions
4.2. Compositions
A nifty interpretation of nk goes as follows: since
n
o
n
= (a1 , a2 , . . . , ak ) ∈ Zk : 0 < a1 < a2 < · · · < ak < n + 1 ,
k
we can set a0 := 0, ak+1 := n + 1 and define bj := aj − aj−1 for 1 ≤ j ≤ k + 1,
which gives
n
o
n
= (b1 , b2 , . . . , bk+1 ) ∈ Zk+1
>0 : b1 + b2 + · · · + bk+1 = n + 1 .
k
We call a vector (b1 , b2 , . . . , bk+1 ) ∈ Zk+1
>0 such that b1 + b2 + · · · + bk+1 = n + 1
a composition of n + 1; the b1 , b2 , . . . , bk+1 are called the parts of the
composition. So nk equals the number of compositions of n + 1 with
k + 1 parts. This interpretation
also gives an alternative reasoning for the
n
generating function of k given in (4.4). Namely,1
X n
X
qn =
#(compositions of n + 1 with k + 1 parts) q n
k
n≥0
n≥0
=
1X
#(compositions of n + 1 with k + 1 parts) q n+1
q
n≥0
1
=
q
X
q b1 +b2 +···+bk+1
(4.11)
b1 ,b2 ,...,bk+1 ≥1



 
1  X b1   X b2   X bk+1 
=
q
q
···
q
q
b1 ≥1
=
1
q
q
1−q
b2 ≥1
k+1
=
bk+1 ≥1
qk
.
(1 − q)k+1
This computation might be an overkill for the binomial coefficient, but
it hints at the powerful machinery that generating functions provide. We
give another sample of this machinery next.
Proposition 4.8. Given a set A ⊆ Z>0 , let cA (n) denote the number of
compositions of n with parts in A. Then its generating function is
X
1
P
CA (q) := 1 +
cA (n) q n =
.
1 − m∈A q m
n≥1
1 The reader might wonder about our switching to the variable q starting with this generating
function. The reason will be given in Section 4.7.
4.3. Plane Partitions
83
Proof. Essentially by repeating the argument in (4.11), the generating func
P
m j.
tion for all compositions with exactly j parts (all in A) equals
q
m∈A
Thus
!j
X X
1
P
=
CA (q) = 1 +
qm
.
1 − m∈A q m
j≥1
m∈A
We show two pieces of art from the exhibition provided by this proposition.
First, suppose A consists of all odd positive integers. Then
1
1−
P
m≥1 odd q
m
=
1
1−
q
1−q 2
=
1 − q2
q
=1+
2
1−q−q
1 − q − q2
—smell familiar? Now let A consist of all integers ≥ 2. Then
1−
1
P
m≥2 q
m
=
1
1−
q2
1−q
=
1−q
q2
=
1
+
.
1 − q − q2
1 − q − q2
What these two generating functions prove, in two simple lines, is the
following:
Theorem 4.9. The number of compositions of n using only odd parts, and
the number of compositions of n − 1 using only parts ≥ 2 are both given by
the nth Fibonacci number.
4.3. Plane Partitions
Our next composition counting function concerns the simplest case of a plane
partition;2 namely, we will count all compositions n = a1 + a2 + a3 + a4
such that the integers a1 , a2 , a3 , a4 ≥ 0 satisfy the inequalities
a1
≥
≥
≥
a3
a2
≥
(4.12)
a4 .
(For a general plane partition, this array of inequalities can form a rectangle
of any size.) Let pl (n) denote the number of plane partitions of n of the
form (4.12). We will compute its generating function
X
X
Pl (q) :=
pl (n) q n =
q a1 +a2 +a3 +a4 ,
n≥0
2Despite their name, plane partitions are not partitions in the sense of Section 4.4 below, but
a special case of P -partitions which we will study in Section 6.4.
84
4. Rational Generating Functions
where the last sum is over all integers a1 , a2 , a3 , a4 ≥ 0 satisfying (4.12), from
first principles, using geometric series:
Pl (q) =
=
X
q a4
X
q a3
X
a4 ≥0
a3 ≥a4
a2 ≥a4
X
X
X
q a4
a4 ≥0
q a3
a3 ≥a4
X
q a2
a1 ≥max(a2 ,a3 )
q a2
a2 ≥a4
q max(a2 ,a3 )
1−q

=
=
1
1−q
1
1−q
1
=
1−q
=
X
q a4
X
a4 ≥0
a3 ≥a4
X
X
q a4
a4 ≥0
X
a4 ≥0
a3 ≥a4
q a1
q a3 q a3
aX
3 −1
a2 =a4

q a2 +
X
q 2a2 
a2 ≥a3
a4 − q a3
q 2a3
a3
a3 q
+
q
q
1−q
1 − q2
q 4a4
q 4a4
q 4a4
−
+
(1 − q)(1 − q 2 ) (1 − q)(1 − q 3 ) (1 − q 2 )(1 − q 3 )
X
1 + q2
1 + q2
4a4
q
=
(1 − q)(1 − q 2 )(1 − q 3 )
(1 − q)(1 − q 2 )(1 − q 3 )(1 − q 4 )
a4 ≥0
=
1
.
(1 − q)(1 − q 2 )2 (1 − q 3 )
(4.13)
(If this computation seems lengthy, or even messy, don’t worry: we will
shortly develop theory that will simplify computations tremendously.)
Next we’d like to compute an explicit formula for pl (n) from this generating function. One way—through a partial-fraction expansion of Pl (q)—is
outlined in Exercise 4.15. A second approach is through expanding both
numerator and denominator of Pl (q) by the same factor to transform the
denominator into a simple form:
Pl (q) =
q 16 + q 15 + 3 q 14 + 4 q 13 + 7 q 12 + 9 q 11 + 10 q 10 + 13 q 9
+12 q 8 + 13 q 7 + 10 q 6 + 9 q 5 + 7 q 4 + 4 q 3 + 3 q 2 + q + 1
(1 − q 6 )4
.
(4.14)
The reason for this transformation is that the new denominator yields an
easy power series expansion, by Exercise 4.4:
X n + 3
1
=
q 6n .
(1 − q 6 )4
3
n≥0
This suggests that the resulting plane-partition counting function pl (n) has
a certain periodic character, with period 6. Indeed, we can split up the
4.3. Plane Partitions
85
rational generating function Pl (q) into 6 parts as follows.
7 q 12 + 10 q 6 + 1 4 q 13 + 13 q 7 + q 3 q 14 + 12 q 8 + 3 q 2
+
+
(1 − q 6 )4
(1 − q 6 )4
(1 − q 6 )4
q 15 + 13 q 9 + 4 q 3 q 16 + 10 q 10 + 7 q 4 9 q 11 + 9 q 5
+
+
+
(1 − q 6 )4
(1 − q 6 )4
(1 − q 6 )4
X
7 n+1
+ 10 n+2
+ n+3
q 6n
=
3
3
3
Pl (q) =
n≥0
+
X
n+3
3
4
n+1
3
+ 13
n+2
3
+
3
n+1
3
+ 12
n+2
3
+3
q 6n+1
n≥0
+
X
n+3
3
q 6n+2
n≥0
+
X
n+1
3
+ 13
n+2
3
+4
n+3
3
q 6n+3
n+1
3
+ 10
n+2
3
+7
n+3
3
q 6n+4
n+3
3
n≥0
+
X
n≥0
+
X
9
n+2
3
+9
q 6n+5 .
n≥0
From this data, a quick calculation yields

1 2
2
1 3


72 n + 6 n + 3 n + 1



1 3
1 2
13
5


72 n + 6 n + 24 n + 18


 1 n3 + 1 n2 + 2 n + 8
6
3
9
pl (n) = 72
1 2
13
1
1 3

n
+
n
+
n
+

72
6
24
2



1 3
1 2
2
7

n
+
n
+
n
+

72
6
3
9


 1 n3 + 1 n2 + 13 n + 7
72
6
24
18
if
if
if
if
if
if
n ≡ 0 mod 6,
n ≡ 1 mod 6,
n ≡ 2 mod 6,
n ≡ 3 mod 6,
n ≡ 4 mod 6,
n ≡ 5 mod 6.
(4.15)
The counting function pl (n) is our first instance of a quasipolynomial, that
is, a function p : Z → C of the form
p(n) = cn (n) nd + · · · + c1 (n) n + c0 (n) ,
where c0 (n), c1 (n), . . . , cn (n) are periodic functions in n. In the case p(n) =
pl (n) of our plane-partition counting function, the coefficient c0 (n) has period
6, the coefficient c1 (n) has period 2, and c2 (n) and c3 (n) have period 1, i.e.,
they are constants. Of course, we can also think of the quasipolynomial pl (n)
as given by the list (4.15) of six polynomials, which we run through cyclically
as n increases. We will say more about basic properties of quasipolynomials
in Section 4.5.
86
4. Rational Generating Functions
We finish our study of plane partitions by observing a simple algebraic
relation for Pl (q), namely,
1
= q 8 Pl (q) .
(4.16)
Pl
q
This suggests some relation to Theorem 4.6. Indeed, this theorem says
X
1
pl (−n) q n .
(4.17)
Pl
= −
q
n≥1
But then equating the right-hand sides of (4.16) and (4.17) yields the following reciprocity relation for the plane-partition quasipolynomials:
pl (−n) = −pl (n − 8) .
(4.18)
This is a combinatorial reciprocity that we are after: up to a shift by 8, the
plane partitions are self reciprocal. We can think about this reciprocity
as follows: If we count the number of plane partitions of n for which all the
inequalities in (4.12) are strict and all parts are strictly positive, then this is
exactly the number of quadruples (a1 , . . . , a4 ) ∈ Z4≥0 such that
a1 + 3
≥
≥
≥
a3 + 2
a2 + 2
≥
(4.19)
a4 + 1 .
But these are just plane partitions with a1 + · · · + a4 = n − 8, which are
counted on the right-hand side of (4.18). There are many generalizations of
pl (n); one is given in Exercise 4.16.
4.4. Restricted Partitions
An (integer) partition of n is a sequence (a1 ≥ a2 ≥ · · · ≥ ak ≥ 1) of nonincreasing positive integers such that
n = a1 + a2 + · · · + ak .
(4.20)
The numbers a1 , a2 , . . . , ak are the parts of this partition. For example,
(4, 2, 1) and (3, 2, 1, 1) are partitions of 7.
Partitions have been around since at least Euler’s time. They provide a
fertile ground for famous theorems (see, e.g., the work of Hardy, Ramanujan,
and Rademacher) and open problems (e.g., nobody understands how the
parity of the number of partitions of n behaves), and they provide a justas-fertile ground for connections to other areas in mathematics and physics
(e.g., Young tableaux, which open a window to representation theory).
Our goal is to enumerate partitions with certain restrictions, which will
allow us to prove a combinatorial reciprocity theorem not unlike that for the
plane-partition counting function pl (n). (The enumeration of unrestricted
partitions is the subject of Exercise 4.23, but it will not yield a reciprocity
4.4. Restricted Partitions
87
theorem.) Namely, we restrict the parts of a partition to a finite set A :=
{a1 > a2 > · · · > ad } ⊂ Z>0 , that is, we only allow partitions of the form
(a1 , . . . , a1 , a2 , . . . , a2 , . . . , ad , . . . , ad ) .
(It is interesting—and related to topics to appear soon—to allow A to be a
multiset; see Exercise 4.20.) The restricted partition function for A is
n
o
pA (n) := (m1 , m2 , . . . , md ) ∈ Zd≥0 : m1 a1 + m2 a2 + · · · + md ad = n .
By now it will not come as a surprise that we will approach the counting
function pA (n) through its generating function
X
PA (q) :=
pA (n) q n .
n≥0
One advantage of this (and any other) generating function is that it allows
us, in a sense, to manipulate the sequence (pA (n))n≥0 algebraically:
X
PA (q) =
q m1 a1 +m2 a2 +···+md ad
m1 ,m2 ,...,md ≥0

= 

X
q m1 a1  
m1 ≥0
=

X

q m2 a2  · · · 
m2 ≥0

X
q md ad 
(4.21)
md ≥0
1
,
(1 − q a1 ) (1 − q a2 ) · · · (1 − q ad )
where the last identity comes from the geometric series. To see how the
generating function of this counting function helps us understand the latter,
let’s look at the simplest case when A contains only one positive integer a.
In this case
1
P{a} (q) =
= 1 + q a + q 2a + · · · ,
1 − qa
the generating function for
(
1 if a|n,
p{a} (n) =
0 otherwise
(as expected from the definition of p{a} (n)). Note that the counting function
p{a} (n) is a (very simple) instance of a quasipolynomial, namely, p{a} (n) =
c0 (n) where c0 (n) is the periodic function (with period a) that returns 1 if n
is a multiple of a and 0 otherwise.
Let’s look at another case, when A has two elements. The product
structure of the accompanying generating function
P{a,b} (q) =
1
(1 −
q a ) (1
− qb)
88
4. Rational Generating Functions
means that we can compute
p{a,b} (n) =
n
X
p{a} (s) p{b} (n − s) .
s=0
Note that we are summing a quasipolynomial here, and so p{a,b} (n) is
again a quasipolynomial by the next proposition, whose proof we leave as
Exercise 4.17.
P
Proposition 4.10. If p(n) is a quasipolynomial, so is r(n) := ns=0 p(s).
More generally, if f (n) and g(n) are quasipolynomials, so is their convolution
n
X
c(n) :=
f (s) g(n − s) .
s=0
We invite the reader to explicitly compute some examples of restricted
partition functions, such as p{1,2} (n) (Exercise 4.19). Naturally, we can
repeatedly apply Proposition 4.10 to deduce:
Corollary 4.11. The restricted partition function pA (n) is a quasipolynomial
in n.
Since pA (n) is a quasipolynomial, we are free to evaluate it at negative
integers. We define
n
o
p◦A (n) := (m1 , m2 , . . . , md ) ∈ Zd>0 : m1 a1 + m2 a2 + · · · + md ad = n ,
the number of restricted partitions of n such that every ai is used at least
once.
Theorem 4.12. If A = {a1 , a2 , . . . , ad } ⊂ Z>0 then
pA (−n) = (−1)d−1 pA (n − a1 − a2 − · · · − ad ) = (−1)d−1 p◦A (n) .
Proof. We first observe that the number of partitions of n in which ai is
used at least once equals pA (n − ai ). Thus, setting a := a1 + a2 + · · · + ad ,
we see that
p◦A (n) = pA (n − a) ,
and this gives the second equality. To prove the first, we use simple algebra
on (4.21) to obtain
1
q a1 +a2 +···+ad
PA
= a1
= (−1)d q a PA (q) .
(4.22)
q
(q − 1) (q a2 − 1) · · · (q ad − 1)
Now we use Theorem 4.6 for quasipolynomials, which gives
X
1
PA
=−
pA (−n) q n .
(4.23)
q
n≥1
Equating the right-hand sides of (4.22) and (4.23) proves the claim.
4.5. Quasipolynomials
89
4.5. Quasipolynomials
We have defined a quasipolynomial p(n) as a function Z → C of the form
p(n) = cn (n) nn + · · · + c1 (n) n + c0 (n) ,
(4.24)
where c0 , c1 , . . . , cd are periodic functions in n. Assuming that cd is not
the zero function, the degree of p(n) is d, and the least common period of c0 (n), c1 (n), . . . , cd (n) is the period of p(n).
Alternatively, for
a quasipolynomial p(n), there exist a positive integer k and polynomials
p0 (n), p1 (n), . . . , pk−1 (n) such that


p0 (n)
if n ≡ 0 mod k,



p1 (n)
if n ≡ 1 mod k,
p(n) =
.
..





pk−1 (n) if n ≡ k − 1 mod k.
The minimal such k is the period of p(n), and for this minimal k, the
polynomials p0 (n), p1 (n), . . . , pk−1 (n) are the constituents of p(n). Of
course, when k = 1, we need only one constituent and the coefficient functions
c0 (n), c1 (n), . . . , cd (n) are constants, and so p(n) is a polynomial. Yet another
perspective on quasipolynomials is explored in Exercise 4.27.
As we have seen in (4.14) and (4.21), the quasipolynomials arising from
plane partitions and restricted partitions can be encoded into rational generating functions with a particular denominator. The following proposition
asserts that the denominators of rational generating functions are the key to
detecting quasipolynomials.
Proposition 4.13. Let p : Z → C be a function with associated generating
function
X
P (z) :=
p(n) z n .
n≥0
Then p(n) is a quasipolynomial of degree ≤ d and period dividing k if and
only if
h(z)
P (z) =
(1 − z k )d+1
where h(z) is a polynomial of degree at most k(d + 1) − 1.
This proposition explains, in some sense, why we rewrote the generating
function Pl (q) for plane partitions in (4.14). Apparently the generating
function (4.21) for the restricted partition function is not of this form, but
multiplying numerator and denominator by appropriate terms yields the
denominator (1 − z k )d where k = lcm(a1 , a2 , . . . , ad ). For example,
P{2,3} (z) =
1
1 + z2 + z3 + z4 + z5 + z7
=
.
(1 − z 2 )(1 − z 3 )
(1 − z 6 )2
90
4. Rational Generating Functions
For the general recipe to convert (4.21) into a form fitting Proposition 4.13, we
refer to Exercise 4.28. In particular, the presentation of P (z) in Proposition
4.13 is typically not reduced, and by getting rid of common factors it can be
seen that h(z)
g(z) gives rise to a quasipolynomial if and only if each zero γ of g(z)
satisfies γ k = 1 for some k ∈ Z>0 . The benefit of having Proposition 4.13 is
that it gives a pretty effective way of showing that a function q : Z → C is a
quasipolynomial.
Proof of Proposition 4.13. Suppose p(n) is a quasipolynomial of degree ≤
d and period dividing k. So we can find polynomials p0 (n), p1 (n), . . . , pk−1 (n)
of degree ≤ d such that


p0 (n)
if n ≡ 0 mod k,



p1 (n)
if n ≡ 1 mod k,
p(n) =
.
..





pk−1 (n) if n ≡ k − 1 mod k.
Thus
P (z) =
k−1
XX
p(ak + b) z ak+b =
a≥0 b=0
k−1
X
b=0
zb
X
pb (ak + b) z ak ,
a≥0
and since pb (ak + b) is a polynomial in a of degree ≤ d, we can use Proposition 4.4 to conclude that
P (z) =
k−1
X
b=0
zb
hb (z k )
(1 − z k )d+1
P
b
k
for some polynomials hb (z) of degree ≤ d. Since k−1
b=0 z hb (z ) is a polynomial of degree ≤ k(d + 1) − 1, this proves the forward implication of
Proposition 4.13.
h(z)
For the converse implication, suppose P (z) = (1−z
k )d+1 , where h(z) is a
polynomial of degree ≤ k(d + 1) − 1, say
k(d+1)−1
h(z) =
X
m=0
cm z m =
d X
k−1
X
cak+b z ak+b .
a=0 b=0
Then
P (z) = h(z)
k−1 X
d
X d + j XX
d + j k(j+a)+b
z kj =
cak+b
z
d
d
j≥0 b=0 a=0
j≥0
=
k−1 X
d
XX
j≥0 b=0 a=0
cak+b
k−1
d + j − a kj+b X X
z
=
pb (kj + b) z kj+b ,
d
j≥0 b=0
4.6. Integer-point Transforms and Lattice Simplices
91
P
where pb (kj + b) = da=0 cak+b d+j−a
, a polynomial in j of degree ≤ d. In
d
other words, P (z) is the generating function of the quasipolynomial with
constituents p0 (n), p1 (n), . . . , pb−1 (n).
4.6. Integer-point Transforms and Lattice Simplices
It is time to return to geometry. The theme of this chapter has been
generating functions that we constructed from some (counting) function,
with the hope of obtaining more information about this function. Now we
will start with a generating function and try to interpret it geometrically as
an object in its own right.
Our setting is that of Section 1.4 but in general dimensions: we consider
a convex lattice polytope P ⊂ Rd and its Ehrhart function
ehrP (n) := nP ∩ Zd .
(4.25)
We call the accompanying generating function
X
EhrP (z) := 1 +
ehrP (n) z n
(4.26)
n≥1
the Ehrhart series of P. Our promised geometric interpretation of this
generating function uses a technique from Chapter 3: namely, we consider
the homogenization of P, defined in (3.7)
n
o
hom(P) = (p, t) ∈ Rd+1 : t ≥ 0, p ∈ tP = cone {(v, 1) : v ∈ vert(P)} .
One advantage of hom(P) is that we can see a copy of the dilate nP as
the intersection of hom(P) with the hyperplane xd+1 = n, as illustrated in
Figure 1; we will say that points on this hyperplane are at height n and
that the integer points in hom(P) are graded by the last coordinate. In
other words, the Ehrhart series of P can be computed through
X
EhrP (z) =
# (lattice points in hom(P) at height n) z n
(4.27)
n≥0
(and this also explains our convention of starting EhrP (z) with constant
term 1). This motivates the study of the arithmetic of a set S ⊆ Rd+1 by
computing a multivariate generating function, namely, its integer-point
transform
X
σS (z1 , . . . , zd+1 ) :=
zm ,
m∈S∩Zd+1
md+1
z1m1 z2m2 · · · zd+1
; we
where we abbreviated zm :=
will often write σS (z) to
shorten the notation involving the variables. In the language of integer-point
transforms, (4.27) can be rewritten as
EhrP (z) = σhom(P) (1, . . . , 1, z) .
(4.28)
92
4. Rational Generating Functions
Figure 1. Recovering dilates of P in hom(P).
An integer-point transform
is typically not a formal power series but
P
m ∈ CJz ±1 , . . . , z ±1 K. If we want to
a formal Laurent series
1
m∈Zd+1 z
d+1
multiply formal Laurent series (what we do want), we have to be a little
careful for which sets S we want to consider integer-point transforms; see
Exercise 4.30. In particular, we are safe if we work with convex, linefree sets
such as K = hom(P).
We illustrate the above train of thoughts with the down-to-earth example
from Section 1.4, the lattice triangle
∆ := (x1 , x2 ) ∈ R2 : 0 ≤ x1 ≤ x2 ≤ 1 .
Its vertices are (0, 0), (0, 1), and (1, 1), and so
0
0
1
hom(∆) = R≥0 0 + R≥0 1 + R≥0 1 .
1
1
1
(4.29)
The three generators of hom(∆)—call them v1 , v2 , v3 —form a lattice basis
of Z3 , i.e., every point in Z3 can be uniquely expressed as an integral linear
combination of v1 , v2 , v3 (see Exercise 4.29). In particular, every lattice
point in hom(∆) can be uniquely written as a nonnegative integral linear
combination of v1 , v2 , v3 . Hence
X
X
σhom(∆) (z) =
zm =
zk1 v1 +k2 v2 +k3 v3
m∈hom(∆)∩Z3
k1 ,k2 ,k3 ≥0
1
(1 − zv1 ) (1 − zv2 ) (1 − zv3 )
1
=
.
(1 − z3 ) (1 − z2 z3 ) (1 − z1 z2 z3 )
=
(4.30)
4.6. Integer-point Transforms and Lattice Simplices
93
With (4.28), we compute
X n + 2
1
Ehr∆ (z) = σhom(∆) (1, 1, z) =
zn
=
(1 − z)3
2
(4.31)
n≥0
and thus recover the Ehrhart polynomial of ∆ which we computed in Section 1.4.
Our computation of σhom(∆) (z) in (4.30) is somewhat misleading when
thinking about the integer-point transform of the homogenization of a general
lattice simplex, because the cone hom(∆) is unimodular: it is generated
by a lattice basis of Z3 . We change our example slightly and consider the
lattice triangle P := conv {(0, 0), (0, 1), (2, 1)}; note that
0
0
2
hom(P) = R≥0 0 + R≥0 1 + R≥0 1
(4.32)
1
1
1
is not unimodular. We can still implement the basic idea behind our computation of σhom(∆) (z), but we have to take into account some extra data. We
write w1 , w2 , w3 for the generators of hom(P) given in (4.32). We observe
that the lattice point q = (1, 1, 1) is contained in hom(P) but is not an
integer linear combination of the generators (mind you, the generators are
linearly independent). Hence, we will miss zq if we exercise (4.30) with P
instead of ∆. Stronger even, the linear independence of w1 , w2 , w3 yields
that none of the points
for a1 , a2 , a3 ∈ Z≥0
q + a1 w1 + a2 w2 + a3 w3
can be expressed as a nonnegative integer linear combination of w1 , w2 , w3 .
A little bit of thought (see Exercise 4.34 for more) brings the following insight:
We define the half-open parallelepiped
n 0
0
2
o
Π := λ1 0 + λ2 1 + λ3 1 : 0 ≤ λ1 , λ2 , λ3 < 1 ,
1
1
1
Then for every p ∈ hom(P) ∩ Z3 there are unique k1 , k2 , k3 ∈ Z≥0 and
r ∈ Π ∩ Z3 such that
p = k1 w1 + k2 w2 + k3 w3 + r .
Figure 2 illustrates a geometric reason behind our claim, namely, we can tile
the cone hom(P) with translates of Π. In our specific case it is easy to check
(Exercise 4.33) that
Π ∩ Z3 = {0, q}.
Thus, we compute
94
4. Rational Generating Functions
Figure 2. Tiling hom(P) with translates of its fundamental parallelepiped.
σhom(P) (z) =
X
zk1 w1 +k2 w2 +k3 w3 +
k1 ,k2 ,k3 ≥0
X
zq+k1 w1 +k2 w2 +k3 w3
k1 ,k2 ,k3 ≥0
z1 z2 z3
1
+
(1 − zw1 ) (1 − zw2 ) (1 − zw3 ) (1 − zw1 ) (1 − zw2 ) (1 − zw3 )
1 + z1 z2 z 3
.
=
(1 − z3 ) (1 − z2 z3 ) (1 − z1 z2 z3 )
=
This implies, again with (4.28),
EhrP (z) = σhom(P) (1, 1, z) =
1+z
.
(1 − z)3
Note that the form of this rational generating function, as that of (4.31),
ensures that ehrP (n) is a polynomial, by Proposition 4.4. Of course, this
merely confirms Theorem 1.15.
We can interpret the coefficients of the numerator polynomial of EhrP (z):
the constant is 1 (stemming from the origin of hom(P)), and the remaining
terms in the example above come from the fact that the fundamental parallelepiped of hom(P) contains precisely one integer point at height 1 and
none of larger height. Exercise 4.38 gives a few general interpretations of the
Ehrhart series coefficients.
The general case of the integer-point transform of a simplicial cone is
not much more complicated than our computations in this section so far;
we will treat it in Theorem 4.18 below, which will also contain a reciprocity
relation generalizing that of Theorem 1.15. One of its consequences will be
the following result, which gives a simplicial version of Theorem 1.15 in all
dimensions. Recall that ∆◦ denotes the relative interior of ∆.
4.7. Gradings of Cones and Rational Polytopes
95
Theorem 4.14. Suppose ∆ is a lattice simplex.
(a) For positive integers n, the counting function ehr∆ (n) agrees with a
polynomial in n of degree dim(∆) whose constant term is 1.
(b) When this polynomial is evaluated at negative integers, we obtain
ehr∆ (−n) = (−1)dim(∆) ehr∆◦ (n) .
As a first general example, we discuss the Ehrhart polynomial of a simplex
∆ = conv {v0 , v1 , . . . , vd } ⊂ Rd
that is unimodular, i.e., the vectors v1 − v0 , v2 − v0 , . . . , vd − v0 form
a Z-basis for Zd (which, by Exercise 4.29, is equivalent to the condition
det (v1 − v0 , v2 − v0 , . . . , vd − v0 ) = ±1). The following result, which can
be proved without assuming Theorem 4.14, is subject to Exercise 4.31:
Proposition 4.15. Let ∆ be the convex
hull of the origin and the d unit
n+d
d
vectors in R . Then ehr∆ (n) = d , and this polynomial satisfies the
conclusions of Theorem 4.14. More generally, ehr∆ (n) = n+d
for any
d
unimodular simplex ∆.
Before proving (a generalization of) Theorem 4.14, we will revisit the
partition and composition counting functions that appeared earlier in this
chapter. Again we can use a cone setting to understand these functions
geometrically, but instead of grading integer points by their last coordinate,
i.e., along (0, 0, 1), we will introduce a grading along a different direction.
4.7. Gradings of Cones and Rational Polytopes
In the last section when considering the cone K = hom(P), we said that K is
graded by height; this suggests that there are other ways to grade a cone.
Let K ⊂ Rd be a pointed, rational d-dimensional cone. A grading of K is a
vector a ∈ Zd such that ha, pi > 0 for all p ∈ K \ {0}. For a grading a ∈ Zd ,
we define
n
o
ha (K, n) := m ∈ K ∩ Zd : ha, mi = n .
Since K is pointed, it follows that ha (K, n) < ∞ for all n ∈ Z≥0 and thus we
can define the generating function
X
Ha (K; z) := 1 +
ha (K, n) z n .
n≥1
One benefit of integer transforms is the following.
Proposition 4.16. Let K ⊂ Rd be a pointed, rational cone and a ∈ Zd a
grading of K. Then
Ha (K; z) = σK (z1a1 , . . . , zdad ) .
96
4. Rational Generating Functions
Proof. We compute
σK (z1a1 , . . . , zdad ) =
X
X
z ha,mi
m∈K∩Zd \{0}
m∈K∩Zd
= 1+
X
z a1 m1 +···+ad md = 1 +
ha (K, n)z n .
n≥1
Consider again the cone defined by (4.29),
 
 
 
0
0
1





0
1
1 .
K := R≥0
+ R≥0
+ R≥0
1
1
1
In the last section we computed its integer-point transform
1
σK (z) =
,
(1 − z3 ) (1 − z2 z3 ) (1 − z1 z2 z3 )
1
and from the grading (0, 0, 1) we obtained the Ehrhart series (1−z)
3 of the
triangle ∆ := {(x1 , x2 ) ∈ R2 : 0 ≤ x1 ≤ x2 ≤ 1}. If we instead grade the
points in K along (1, 1, 1), we meet another friendly face:3
H(1,1,1) (q) = σK (q, q, q) =
1
,
(1 − q)(1 − q 2 )(1 − q 3 )
the generating function for the restricted partition function with parts 1, 2,
and 3, as well as (by Exercise 4.25) for the number of partitions into at most
three parts. We remark the obvious, namely, that this generating function is
quite different from the Ehrhart series (4.31) of ∆; yet they both come from
the same cone and the same integer-point transform, specialized along two
different gradings.
Let’s try some of the other composition/partition counting functions that
appeared earlier in this chapter. The compositions into k + 1 parts, which—
up to a factor of 1q —we enumerated in (4.11), live in the open (unimodular)
cone Rk+1
>0 which comes with the integer-point transform
z1
z2
zk+1
σRk+1 (z) =
···
,
>0
1 − z1 1 − z2
1 − zk+1
k+1
≥
≥
q
and indeed, σRk+1 (q, q, . . . , q) = (1−q)
k+1 confirms (4.11).
>0
The plane partitions of Section 4.3 are the integer points in the cone


x1 ≥ x2 

C := x ∈ R4≥0 :
.


x3 ≥ x4
3We follow the convention of using the variable q when grading points along (1, 1, . . . , 1). This
also explains our mysterious shift from z to q in Section 4.2.
4.7. Gradings of Cones and Rational Polytopes
97
The computation of its integer-point transform is less trivial, as C is not
simplicial. However, thinking about possible orderings of x2 and x3 in the
definition of C gives the inclusion–exclusion formula
σC (z) = σC2≥3 (z) + σC3≥2 (z) − σC2=3 (z)
(4.33)
where
x ∈ R4≥0 : x1 ≥ x2 ≥ x3 ≥ x4 ,
:= x ∈ R4≥0 : x1 ≥ x3 ≥ x2 ≥ x4 , and
:= x ∈ R4≥0 : x1 ≥ x2 = x3 ≥ x4 .
C2≥3 :=
C3≥2
C2=3
Each of these cones is unimodular (and, in fact, can be interpreted as
containing partitions featured in Exercise 4.25), and so
1
σC2≥3 (z) =
(4.34)
(1 − z1 z2 z3 z4 )(1 − z1 z2 z3 )(1 − z1 z2 )(1 − z1 )
1
σC3≥2 (z) =
(1 − z1 z2 z3 z4 )(1 − z1 z2 z3 )(1 − z1 z3 )(1 − z1 )
1
σC2=3 (z) =
.
(1 − z1 z2 z3 z4 )(1 − z1 z2 z3 )(1 − z1 )
Combining these integer-point transforms according to (4.33) gives
σC (z) =
1 − z12 z2 z3
(1 − z1 z2 z3 z4 )(1 − z1 z2 z3 )(1 − z1 z2 )(1 − z1 z3 )(1 − z1 )
(4.35)
and σC (q, q, q, q) = (1−q)(1−q12 )2 (1−q3 ) confirms the generating function (4.13)
for the plane partitions.
Finally, we interpret the restricted partition function from Section 4.4
geometrically. There are two ways to go about that: first, we can interpret
n
o
pA (n) = (m1 , m2 , . . . , md ) ∈ Zd≥0 : m1 a1 + m2 a2 + · · · + md ad = n as counting lattice points in the cone Rd≥0 along the grading (a1 , a2 , . . . , ad );
this comes with the integer-point transform
1
σRd (z) =
≥0
(1 − z1 )(1 − z2 ) · · · (1 − zd )
which specializes to
σRd (q a1 , q a2 , . . . , q ad ) =
≥0
1
,
(1 − q a1 ) (1 − q a2 ) · · · (1 − q ad )
confirming (4.21). Possibly the more insightful geometric interpretation of
pA (n) goes back to the viewpoint through Ehrhart series. Namely, restricted
partitions are integer points in the cone
n
o
K := (x1 , x2 , . . . , xd , xd+1 ) ∈ Rd+1
≥0 : x1 a1 + x2 a2 + · · · + xd ad = xd+1
98
4. Rational Generating Functions
and now we’re interested in the “Ehrhart grading” along (0, 0, . . . , 0, 1). In
fact, K is the homogenization of a polytope, but this polytope is not a lattice
polytope anymore: the generators of K can be chosen as
1  
 
0
0
a1
0
0 1
   a2 
 
0 0
 .. 
   
 
 ..  ,  ..  , . . . ,  .  .
. .
0
   
1
0 0
 
ad
1
1
1
Indeed, from the look of the generating function of pA (n) we should not
expect the Ehrhart series of a lattice polytope, as the denominator of PA (q)
is not simply a power of (1 − q).
These considerations suggest a variant of Theorem 4.14 for rational
polytopes—those with vertices in Qd . We leave its proof as Exercise 4.39
(with which you should wait until understanding our proof of Theorem 4.18
below).
Theorem 4.17. If ∆ is a rational simplex, then for positive integers n, the
counting function ehr∆ (n) is a quasipolynomial in n whose period divides
the least common multiple of the denominators of the vertex coordinates of
∆. When this quasipolynomial is evaluated at negative integers, we obtain
ehr∆ (−n) = (−1)dim(∆) ehr∆◦ (n) .
The different types of grading motivate the study of integer-point transforms as multivariate generating functions. Our next goal is to prove structural results for them, in the case that the underlying polyhedron is a
simplicial cone.
4.8. Stanley Reciprocity for Simplicial Cones
≥
≥
We return once more to the plane partitions of Section 4.3, and again we
think of them as integer points in the 4-dimensional cone


x1 ≥ x2 

C := x ∈ R4≥0 :
.


x3 ≥ x4
In (4.33) we computed the integer-point transform of C via inclusion–exclusion,
through the integer-point transforms of three unimodular cones (two 4dimensional and one 3-dimensional cone). Now we will play a variation of
the same theme, however, we will only use full-dimensional cones. Namely,
4.8. Stanley Reciprocity for Simplicial Cones
99
if we let
C2≥3 :=
C3>2 :=
x ∈ R4≥0 : x1 ≥ x2 ≥ x3 ≥ x4
x∈
R4≥0
and
: x1 ≥ x3 > x2 ≥ x4 ,
then C2≥3 ] C3>2 = C (recall that this symbol denotes a disjoint union), and
thus
σC (z) = σC2≥3 (z) + σC3>2 (z) ,
(4.36)
from which we can compute Pl (q) once more by specializing z = (q, q, q, q).
Thus, compared to (4.33) we have fewer terms to compute, but we are paying
a price by having to deal with half-open cones. There is a general philosophy
surfacing here, which we will see in action time and again: it is quite useful
to decompose a polyhedron in different ways (in this case, giving rise to
(4.33) and (4.36)), each having its own advantages.
We have computed the integer-point transform of C2≥3 already in (4.34):
σC2≥3 (z) =
1
.
(1 − z1 z2 z3 z4 ) (1 − z1 z2 z3 ) (1 − z1 z2 ) (1 − z1 )
The analogous computation for C3>2 has a little twist stemming from the
fact that C3>2 is half open: in terms of generators,
 
 
 
 
1
1
1
1
1
1
0
0

 
 
 
C3>2 = R≥0 
1 + R≥0 1 + R>0 1 + R≥0 0 .
1
0
0
0
The integer-point transform of its closure C3≥2 we also computed in (4.34),
from which we deduce
z1 z3
σC3>2 (z) = z1 z3 σC3≥2 (z) =
.
(1 − z1 z2 z3 z4 ) (1 − z1 z2 z3 ) (1 − z1 z3 ) (1 − z1 )
Putting it all together, we recover with (4.36) the integer-point transform
for C which we computed in (4.35):
σC (z) =
1 − z12 z2 z3
.
(1 − z1 z2 z3 z4 )(1 − z1 z2 z3 )(1 − z1 z2 )(1 − z1 z3 )(1 − z1 )
There are two features that made the “geometric computations” of σC (z)
through (4.33) and (4.36) possible: the fact that we could decompose C into
simplicial cones (in the first instance C2≥3 , C3≥2 , and C2=3 , in the second
C2≥3 and C3>2 ), and the fact that these cones were unimodular. We will
see in Chapter 5 that any pointed cone can be similarly decomposed into
simplicial cones (this is called a triangulation of the cone), and the following
important theorem shows that the integer-point transform of any simplicial
cone (not just a unimodular one) is nice. Even better, it gives our first
reciprocity theorem for a multivariate generating function.
100
4. Rational Generating Functions
Theorem 4.18. Fix linearly independent vectors v1 , v2 , . . . , vk ∈ Zd and
1 ≤ m ≤ k and define the two half-open cones
b := R≥0 v1 + · · · + R≥0 vm−1 + R>0 vm + · · · + R>0 vk
K
and
q := R>0 v1 + · · · + R>0 vm−1 + R≥0 vm + · · · + R≥0 vk .
K
Then σKb (z) and σKq (z) are rational generating functions in z1 , z2 , . . . , zd which
are related via
σKq z1 = (−1)k σKb (z)
where
1
z
:= ( z11 , z12 , . . . , z1d ).
b and use a tiling argument, which we hinted at
Proof. We start with K
already in Figure 2, to compute its generating function. Namely, let
b := [0, 1) v1 + · · · + [0, 1) vm−1 + (0, 1] vm + · · · + (0, 1] vk ,
Π
b see Figure 3 for an example. Then
the fundamental parallelepiped of K;
Figure 3. The fundamental parallelogram of a 2-cone.
b by translates of Π,
b as we invite the reader to prove in Exerwe can tile K
cise 4.34:
] b=
b
K
j 1 v1 + · · · + j k vk + Π
(4.37)
j1 ,...,jk ≥0
(see Figure 4). This can be translated into the language of generating
functions as

 

X
X
σ b (z) = 
zj1 v1  · · · 
zjk vk  σ b (z)
Π
K
j1 ≥0
=
(1 −
σΠ
b (z)
v
1
z ) · · · (1
jk ≥0
− zvk )
.
(4.38)
4.8. Stanley Reciprocity for Simplicial Cones
101
Figure 4. Translates of the fundamental parallelogram tile the cone.
Completely analogously, we compute
σKq (z) =
σ (z)
(1 −
q
Π
v
1
z ) · · · (1
− zvk )
where
q := (0, 1] v1 + · · · + (0, 1] vm−1 + [0, 1) vm + · · · + [0, 1) vk .
Π
q then
The two fundamental parallelepipeds are intimately related: if x ∈ Π
we can write x = λ1 v1 + · · · + λk vk for some 0 < λ1 , . . . , λm−1 ≤ 1 and
0 ≤ λm , . . . , λk < 1. But then
v1 + v2 + · · · + vk − x = (1 − λ1 )v1 + · · · + (1 − λk )vk
b Since v1 , v2 , . . . , vk are linearly independent, this proves
is a point in Π.
q
b = v 1 + v2 + · · · + v k − Π
Π
(4.39)
(illustrated in Figure 5) or, in terms of generating functions,
1
v1 +v2 +···+vk
σΠ
σΠq
.
b (z) = z
z
This yields
z−v1 −v2 −···−vk σΠ
σΠq ( z1 )
1
b (z)
=
σKq
=
−v
−v
−v
z
(1 − z 1 ) · · · (1 − z k )
(1 − z 1 ) · · · (1 − z−vk )
σΠ
b (z)
= (−1)k
= (−1)k σKb (z) .
v
1
(1 − z ) · · · (1 − zvk )
Theorem 4.18 is at the heart of this book. Once we have developed the
machinery of decomposing polyhedra in Chapter 5, we will be able to extend
102
4. Rational Generating Functions
Figure 5. The geometry of (4.39) in dimension 2.
it to any pointed rational cone (Theorem 5.13). For now, we limit ourselves
to consequences for simplicial cones. For k = 0, observe that in Theorem 4.18
b=K
K
and
q = K◦ .
K
This extreme case yields a reciprocity for simplicial cones equipped with an
arbitrary grading.
Corollary 4.19. Let K ⊂ Rd+1 be a rational simplicial cone with generators
v1 , . . . , vk and fundamental parallelepiped Π. For a grading a ∈ Zd+1 ,
X
Ha (Π; z)
(4.40)
Ha (K; z) =
ha (K, n) z n =
ha,v
1 − z 1 i · · · 1 − z ha,vk i
n≥0
and
1
Ha K;
= (−1)k+1 Ha (K◦ ; z) .
z
One interesting fact to note is that the fundamental parallelepiped
depends on a choice of generators and hence is only unique up to scaling of
the generators. Technically, this means that the right-hand side of (4.40)
depends on a choice of generators, whereas the left-hand side does not; see
Exercise 4.35.
Corollary 4.19 also furnishes a proof of Theorem 4.14.
Proof of Theorem 4.14. Recall our setup for this result, given at the
beginning of Section 4.6: To a given r-dimensional lattice simplex ∆ ⊂ Rd
with vertices v1 , . . . , vr+1 , we associate its homogenization K = hom(∆).
This is a simplicial cone with generators wi = (vi , 1) for i = 1, . . . , r and
associated parallelepiped
Π := [0, 1)w1 + [0, 1)w2 + · · · + [0, 1)wk+1 .
Notes
103
For the grading a = (0, . . . , 0, 1), we can write the Ehrhart series of ∆ as
X
Ha (Π; z)
ehr∆ (n) z n = Ha (K; z) =
Ehr∆ (z) = 1 +
(1 − z)r+1
n≥1
where the last equality is Corollary 4.19. The numerator polynomial
h∗ (z) := Ha (Π; z) = h∗0 + h∗1 z + · · · + h∗r z r
enumerates the number of lattice points in Π according to their height. We
observe that there are no lattice points of height > r and that the origin 0 is
the unique lattice point in Π of height zero. Hence h∗i ≥ 0 for i = 0, . . . , r
and h∗0 = 1. Using Proposition 4.4, this shows that ehr∆ (n) is a polynomial
of degree at most r and, in fact, of degree equal to r as h∗ (1) > 0. By the
same token, we observe that ehr∆ (0) = Ha (K; 0) = h∗ (0) = 1 which proves
part (a).
For part (b) we combine Theorem 4.6 with Corollary 4.19 to infer
X
Ehr◦∆ (z) =
ehrP (−n)z n = − Ehr∆ ( z1 ) = (−1)r Ehr∆ (z)
n≥1
which is the combinatorial reciprocity that we were after.
Notes
We have barely started to touch on the useful and wonderful world of
generating functions. We heartily recommend [50] and [85] if you’d like to
explore more.
Partition and composition counting functions form a major and longrunning theme in number theory; again we barely scratched the surface in
this chapter. See, e.g., [2, 3, 38] for further study.
The earliest reference for Proposition 4.8 we are aware of is William
Feller’s book [30, p. 311] on probability theory, which was first published
in 1950. The earliest combinatorics paper that includes Proposition 4.8
seems to be [54]. A two-variable generalization of Proposition 4.8, which
seems to have first appeared in [41], is described in Exercise 4.14. It is not
clear who first proved Theorem 4.9. The earliest reference we are aware of
is [41] but we suspect that the theorem has been known earlier. Arthur
Cayley’s collected works [22, p. 16] contains the result that c{j∈Z: j≥2} (n)
equals the nth Fibonacci number, but establishing a bijective proof of the
fact that c{j∈Z: j≥2} (n) = c{2j+1: j≥0} (n) (which is part of Theorem 4.9) is
nontrivial [68]. Other applications of Proposition 4.8 include recent theorems,
e.g., [61, 62, 87]; see Exercises 4.12 and 4.13.
Plane partitions were introduced by Percy MacMahon about a century
ago, who proved a famous generating-function formula for the general case
of an m × n plane partition [52]. There are various generalizations of plane
104
4. Rational Generating Functions
partitions, for example, the plane partition diamonds given in Exercise 4.16,
due to George Andrews, Peter Paule, and Axel Riese [4].
Theorem 4.12 is due to Eug`ene Ehrhart [27], who also proved Theorem 4.17, which gives a vast generalization of the reciprocity featured in
Theorem 4.12, from one to an arbitrary (finite) number of linear constraints.
Restricted partition functions are closely related to a famous problem
in combinatorial number theory: namely, what is the largest integer root
of pA (n) (the Frobenius number associated with the set A)?4 This problem,
first raised by Georg Frobenius in the 19th century, is often called the coinexchange problem—it can be phrased in lay terms as looking for the largest
amount of money that we cannot change given coin denominations in the set
A. Exercise 4.21 (which gives a formula for the restricted partition function
in the case that A contains two elements; it goes back to an 1811 book on
elementary number theory by Peter Barlow [8, p. 323–325]) suggests that the
Frobenius problem is easy for |A| = 2 (and you may use this exercise to find
a formula for the Frobenius number in this case), but this is deceiving: the
Frobenius problem is much harder for |A| = 3 (though there exist formulas of
sorts [25]) and certainly wide open for |A| ≥ 4. The Frobenius problem is also
interesting from a computational perspective: while the Frobenius number is
known to be polynomial-nime computable for fixed |A| [45], implementable
algorithms are harder to come by with (see [14] for the current state of the
art). For much more on the Frobenius problem, we refer to [59].
Eug`ene Ehrhart laid the foundation for lattice-point enumeration in
rational polyhedra, starting with Theorems 4.14 and 4.17 in 1962 [26] as a
teacher at a lyc´ee in Strasbourg, France. (Ehrhart received his doctorate
later, at age 60 on the urging of some colleagues.) The proof we give here
follows Ehrhart’s original lines of thought; an alternative proof from first
combinatorial principles can be found in [64].
Richard Stanley developed much of the theory of Ehrhart (quasi-)polynomials, initially from a commutative-algebra point of view. Theorem 4.18
appeared in [73], the paper that coined the term combinatorial recirocity
theorem. Another one of his famous theorems says that the numerator
polynomial of an Ehrhart series has nonnegative integral coefficients [74]; we
indicated the simplicial case of this theorem in our proof of Theorem 4.14.
These inequalities serve as the starting point when trying to classify Ehrhart
polynomials, though a complete classification is known only in dimension
two [11]. The current state of the arts regarding inequalities among Ehrhart
coefficients is [79, 80].
For (much) more about Ehrhart (quasi-)polynomials and solid-angle
enumeration, see [9, 12, 39, 78].
4For this question to make sense, we need to assume that the elements of A are relatively
prime.
Exercises
105
Exercises
4.1 Prove the following extension of Proposition 4.1: Let A ∈ Cd×d , fix
indices 1 ≤ i, j ≤ d, and consider the sequence a(n) := (An )ij formed
by the (i, j)-entries of the nth powers of A. Then a(n) agrees with a
polynomial in n if and only if A is unipotent. (Hint: Consider the
Jordan normal form of A.)
4.2 Complete the proof of Proposition 4.2 that (M), (F), (H), and (G) are
bases for the vector space C[z]≤d = {f ∈ C[z] : deg(f ) ≤ d}:
(a) Give an explicit change of bases from (F) to (M) in the spirit of (4.1).
d
For example, consider z j dz
(z + w)d and set w = 1 − z.
(b) Assume that there are numbers α0 , . . . , αd such that
z
z+1
z+d
α0
+ α0
+ · · · + αd
= 0.
d
d
d
Argue, by specializing z, that αj = 0 for all j.
(c) Can you find explicit changes of bases for the sets (M), (F), (H),
and (G) and give them combinatorial meaning?
P
4.3 Show that for F (z) = n≥0 f (n)z n there exists a power series G(z)
such that F (z) G(z) = 1 if and only if f (0) 6= 0. (This also explains
why we do not allow A to contain the number 0 in Proposition 4.8.)
4.4 Show that for integers m ≥ k ≥ 0,
X n + m − k zk
.
zn =
(1 − z)m+1
m
n≥k
4.5 Prove Proposition 4.5: Let (f (n))≥0 be a sequence of numbers. Then
(f (n))n≥0 satisfies a linear recurrence of the form (4.8) if and only if
X
p(z)
F (z) =
f (n)z n =
cd z d + cd−1 z d−1 + · · · + c0
n≥0
for some polynomial p(z) of degree < d.
4.6 Let (f (n))n≥0 be a sequence of numbers. Show that f (n) satisfies a linear
P
recursion for sufficiently large n > 0 if and only if n≥0 f (n)z n = p(z)
q(z)
for some polynomials p(z) and q(z), with no restriction on the degree
of p(z).
4.7 Show that, if f (n) is the sequence of Fibonacci numbers, then
X
z
f (n) z n =
.
1 − z − z2
n≥1
106
4. Rational Generating Functions
Expand this rational function into partial fractions to give a closed
formula for f (n).
4.8 Let f ◦ (n) := f (−n) be the sequence satisfying the linear recurrence (4.10)
with starting
P values f (0), f (1), . . . , f (d−1). Compute the numerator for
F ◦ (z) = n≥1 f ◦ (n)z n and thereby complete the proof of Theorem 4.6.
4.9 Let A ∈ Cd×d , and for 1 ≤ i, j ≤ d define f (n) := (An )ij .
(a) Show that f (n) satisfies a linear recurrence.
(b) If A is invertible, show that f ◦ (n) = (A−n )ij (in the language of
Exercise 4.8). What is f ◦ (n) if A is not invertible?
(c) If A is diagonalizable with eigenvalues λ1 , . . . , λd , show that there
are a1 , . . . , ad ∈ C such that
f (n) = a1 λn1 + · · · + ad λnd .
P
n
4.10 Consider the formal power series F (z) = n≥0 zn! .
(a) Prove that F (z) is not rational.
(b) Show that F ( z1 ) is not a generating function. (Hint: This time you
might want to think about F ( z1 ) as a function.)
4.11 Give an P
alternative proof of Proposition 4.8 by utilizing the fact that
cA (n) = m∈A cA (n − m).
4.12 Let A := {n ∈ Z>0 : 3 - n}. Compute the generating function for cA (n)
and derive from it both a recursion and closed form for cA (n). Generalize.
4.13 Fix positive integers a and b. Prove that the number of compositions of
n with parts a and b equals the number of compositions of n + a with
parts in {a + bj : j ≥ 0} (and thus, by symmetry, also the number of
compositions of n + b with parts in {aj + b : j ≥ 0}).
4.14 Towards a two-variable generalization of Proposition 4.8, let cA (n, m)
denote the number of compositions of n with precisely m parts in the
set A, and let
X
CA (x, y) :=
cA (n, m) xn y m ,
n,m≥0
where we set cA (0, 0) := 1. Prove that
CA (x, y) =
1−y
1
P
m∈A x
m
.
Compute a formula for cA (n, m) when A is the set of all positive odd
integers.
4.15 Compute the quasipolynomial pl (n) through a partial-fraction expansion
a(q)
b(q)
c(q)
of Pl (q) of the form (1−q)
4 + (1−q 2 )2 + 1−q 3 . Compare your formula for
pl (n) with (4.15).
Exercises
107
4.16 Show that the generating function for plane partition diamonds
a1 ≥ a2
≥
≥
a3 ≥ a4 ≥ a5
≥
≥
a6 ≥ a7
..
.
a3n−2 ≥ a3n−1
≥
≥
a3n
≥ a3n+1
is
(1 + q 2 )(1 + q 5 )(1 + q 8 ) · · · (1 + q 3n−1 )
.
(1 − q)(1 − q 2 ) · · · (1 − q 3n+1 )
Derive the reciprocity theorem for the associated plane-partition-diamond
counting function.
4.17 Prove
Pn Proposition 4.10: If p(n) is a quasipolynomial, so is r(n) :=
s=0 p(s). More generally, if f (n) and g(n) are quasipolynomials then
so is their convolution
c(n) =
n
X
f (s) g(t − s) .
s=0
4.18 Continuing Exercise 4.17, let c(n) be the convolution of the quasipolynomials f (n) and g(n). What can you say about the degree and the
period of c(n), given the degrees and periods of f (n) and g(n)?
4.19 Compute the quasipolynomial pA (n) for the case A = {1, 2}.
4.20 How does your computation of both the generating function and the
quasipolynomial pA (n) change when we switch from Exercise 4.19 to
the case of the multiset A = {1, 2, 2}?
4.21 Suppose a and b are relatively prime positive integers. Define the
integers α and β through
bβ ≡ 1 mod a
and
aα ≡ 1 mod b ,
and denote by {x} the fractional part of x, defined through
x = bxc − {x} ,
where bxc is the largest integer ≤ x. Prove that
n o
βn
n
αn
p{a,b} (n) =
−
−
+ 1.
ab
a
b
108
4. Rational Generating Functions
4.22 The (unrestricted) partition function p(n) counts all partitions of
n. Show that its generating function is
X
Y 1
1+
p(n) q n =
.
1 − qk
n≥1
k≥1
4.23 Let d(n) denote the number of partitions of n into distinct parts (i.e.,
no part is used more than once), and let o(n) denote the number of
partitions of n into odd parts (i.e., each part is an odd integer). Compute
the generating functions of d(n) and o(n), and prove that they are equal
(and thus d(n) = o(n) for all positive integers n).
4.24 In this exercise we consider the problem of counting partitions of n with
an arbitrary but finite number of parts, restricting the maximal size of
each part. That is, let
(m ≥ a1 ≥ a2 ≥ · · · ≥ ak ≥ 1) :
.
p≤m (n) := k ∈ Z>0 and a1 + a2 + · · · + ak = n Prove that
p≤m (n) = p{1,2,...,m} (n) .
4.25 Let pk (n) denote the number of partitions of n into at most k parts.
(a) Show that
X
1
1+
pk (n) q n =
(1 − q)(1 − q 2 ) · · · (1 − q k )
n≥1
and conclude that pk (n) is a quasipolynomial in n.
(b) Prove that (−1)k−1 pk (−n) equals the number of partitions of n into
exactly k distinct parts.
4.26 Compute the constituents and the rational generating function of the
quasipolynomial q(n) = n + (−1)n .
4.27 Recall that ζ ∈ C is a root of unity if ζ m = 1 for some m ∈ Z>0 .
(a) Prove that if c : Z → C is a periodic function with period k,
then there are roots of unity ζ0 , ζ1 , . . . , ζk−1 ∈ C and coefficients
c0 , c1 , . . . , ck−1 ∈ C such that
n
c(n) = c0 ζ0n + c1 ζ1n + · · · + ck−1 ζk−1
.
(b) Show that q : Z≥0 → C is a quasipolynomial if only if
p(n) =
m
X
ci ζin nki
i=1
where ci ∈ C, ki ∈ Z≥0 , and ζi are roots of unity.
Exercises
109
4.28 For A = {a1 , a2 , . . . , ad } ⊂ Z>0 let k = lcm(a1 , a2 , . . . , ad ) be the least
common multiple of the elements of A. Provide an explicit polynomial
hA (q) such that generating function PA (q) for the restricted partitions
with respect to A is
X
hA (q)
PA (q) =
pA (n) q n =
.
(1 − q k )d
n≥0
4.29 Recall that v1 , v2 , . . . , vd ∈ Zd form a lattice basis of Zd if every point
in Zd can be uniquely expressed as an integral linear combination of
v1 , v2 , . . . , vd . Let A be the matrix with columns v1 , v2 , . . . , vd . Show
that v1 , v2 , . . . , vd is a lattice basis if and only if det(A) = ±1. (Hint:
use Cramer’s rule.)
±1
±1
4.30 (a) For w ∈ Zd+1 , define Tw : CJz1±1 , . . . , zd+1
K → CJz1±1 , . . . , zd+1
K by
w
T (f ) := (1 − z )f . Show thatP
Tw is an invertible linear transform.
(b) For w ∈ Zd+1 show that f := t∈Z ztw equals zero.
(c) Let S ⊂ Rd+1 and a ∈ Zd+1 such that for any δ ∈ Z the sets
Sδ := {m ∈ S ∩ Zd+1 : ha, mi = δ} are all finite. Show that the
specialization σS (t a) ∈ CJt±1 K is well defined.
(d) Let S ⊂ Rd+1 be a convex, line-free set. Prove that σS (z) is
nonzero. Moreover, if S1 , S2 ⊂ Rd+1 are line-free convex sets, show
that σS1 (z) σS2 (z) is well defined.
4.31 Prove Proposition 4.15 (without assuming Theorem 4.14): Let ∆ be
the convex hull of the origin and the d unit vectors in Rd . Then
ehr∆ (n) = t+d
polynomial satisfies Theorem 4.14. More
d , and this
t+d
generally, ehr∆ (n) = d for any unimodular simplex ∆.
4.32 Show that the cone C2≥3 = x ∈ R4≥0 : x1 ≥ x2 ≥ x3 ≥ x4 has generators
       
1
1
1
1
 1   1   1   0 
 ,  ,  ,  .
 1   1   0   0 
1
0
0
0
4.33 Let P := conv {(0, 0), (0, 1), (2, 1)}. Show that every lattice point in
hom(P) can be uniquely written as either
 
 
 
0
0
2
k1  0  + k2  1  + k3  1 
1
1
1
or
 
 
 
 
1
0
0
2
 0  + k1  0  + k2  1  + k3  1  ,
0
1
1
1
for some nonnegative integers k1 , k2 , k3 .
110
4. Rational Generating Functions
4.34 Prove (4.37), i.e., fix linearly independent vectors v1 , v2 , . . . , vk ∈ Zd
and define
b := R≥0 v1 + · · · + R≥0 vm−1 + R>0 vm + · · · + R>0 vk
K
and
b := [0, 1) v1 + · · · + [0, 1) vm−1 + (0, 1] vm + · · · + (0, 1] vk .
Π
Then
b=
K
]
b .
j 1 v1 + · · · + j k vk + Π
j1 ,...,jk ≥0
4.35 Let K be a rational, simplicial cone with generators v1 , . . . , vk ∈ Zd
and corresponding parallelepiped Π. Let Π0 be the parallelepiped for
vi0 = λi vi , i = 1, . . . , k. Write σΠ0 (z) in terms of σΠ (z) and verify that
the right-hand side of (4.40) is independent of a particular choice of
generators.
b and Π
b is
b be as in (the proof of) Theorem 4.18. Prove that if K
4.36 Let K
b contains precisely one lattice point.
unimodular then Π
4.37 Pick four concrete points in Z3 and compute the Ehrhart polynomial of
their convex hull.
4.38 Let ∆ be a lattice d-simplex and write
Ehr∆ (z) =
h∗d z d + h∗d−1 z d−1 + · · · + h∗0
.
(1 − z)d+1
Prove that
(a) h∗d = ∆◦ ∩ Zd ;
(b) h∗1 = ∆ ∩ Zd − d − 1 ;
(c) h∗0 + h∗1 + · · · + h∗d = d! vol(∆) .
4.39 Prove Theorem 4.17: If ∆ is a rational simplex, then for positive integers
n, the counting function ehr∆ (n) is a quasipolynomial in n whose period
divides the least common multiple of the denominators of the vertex
coordinates of ∆. When this quasipolynomial is evaluated at negative
integers, we obtain
ehr∆ (−n) = (−1)dim(∆) ehr∆◦ (n) .
4.40 Let S be an m-dimensional subset of Rd (i.e., the affine span of S has
dimension m). Then we define the relative volume of S to be
1 vol S := lim m nS ∩ Zd .
n→∞ n
(a) Convince yourself that vol S is the usual volume if m = d.
(b) Show that, if ∆ ⊂ Rd is an m-dimensional rational simplex, then
the leading coefficient of ehr∆ (n) (i.e., the coefficient of nm ) equals
vol ∆.
Chapter 5
Subdivisions
Let no-one ignorant of geometry enter.
Inscription at the door of Plato’s academy
[under construction]
An idea that prevails in many parts of mathematics is to decompose
a complex object into simpler ones, do computations on the simple pieces,
and put together the local information to get the ‘whole picture’. This was
our strategy, e.g., in Section 1.4 where we proved Ehrhart’s theorem and
Ehrhart–Macdonald reciprocity for lattice polygons. In the plane we could
appeal to the reader’s intuition that every lattice polygon can be triangulated
into lattice triangles. This technique of handwaiving fails in dimensions ≥ 3.
It is no longer obvious that very (lattice) polytope can be decomposed into
(lattice!) simplices. However, it can be done, even elegantly so.
5.1. Decomposing a Polytope
The first challenge in asserting that every polytope can be decomposed into
simpler polytopes is to make precise what we actually mean by that. It is easy
to see that counting lattice points through E(S) := |S ∩ Zd | is a valuation
in the sense of (3.8): For any two bounded (polyconvex) sets S, T ⊂ Rd , we
have
E(S ∪ T ) = E(S) + E(T ) − E(S ∩ T ).
111
112
5. Subdivisions
Recall that for a lattice polytope P ⊂ Rd , the Ehrhart function of P is given
by
ehrP (n) := E(nP)
for all n ≥ 1. For lattice simplices P, Theorem 4.14 asserts that ehrP (n)
agrees with a polynomial of degree dim P. Thus, for general lattice polytopes
we could wish for the following: If for a lattice polytope P ⊂ Rd , we can
find lattice simplices P1 , . . . , Pm such that P = P1 ∪ · · · ∪ Pm , then by the
inclusion-exclusion principle (see Exercise 2.2.12 ),
X
ehrP (n) =
(−1)|I|−1 ehrPI (n)
(5.1)
∅6=I⊆[m]
T
where PI = i∈I Pi . This attempt is bound to fail as the polytopes PI are
in general not simplices and, even worse, not lattice polytopes anymore. We
need more structure.
A dissection of a polytope P ⊂ Rd is a collection of polytopes P1 , . . . , Pm
of the same dimension such that
P = P1 ∪ · · · ∪ Pm
and
P◦i ∩ P◦j = ∅ for all i 6= j.
(5.2)
This, at least, opens the door to induction on the dimension although the
problem with PI not being a lattice polytope prevails. As we will see in
Section 5.3, a dissection into lattice simplices is all we need but there are
alternatives.
We will pursue a more combinatorial (and conservative) approach in
this section. A polyhedral complex S is a nonempty finite collection of
polyhedra in Rd such that the following two conditions are satisfied:
◦ if F is a face of P ∈ S, then F ∈ S;
(containment property)
◦ if P, Q ∈ S, then P ∩ Q is a face of P and Q.
(intersection property)
S
The support of a polyhedral complex S is |S| := P∈S P ⊆ Rd , the pointset
underlying the polyhedral complex. We already know two seemingly trivial
instances of polyhedral complexes.
Proposition 5.1. If Q is a nonempty polyhedron, then the collection of faces
Φ(Q) = {F ⊆ Q : F face} is a polyhedral complex. Moreover, the collection
of bounded faces
Φbnd (Q) := {F ∈ Φ(Q) : F bounded}
is a polyhedral complex.
Proof. By Lemma ??, a face of a face of Q is a face of Q, and the intersection
of faces of Q is again a face. This is exactly the containment and the
intersection property. As every face of a bounded face is necessarily bounded,
we get the second claim.
5.1. Decomposing a Polytope
113
A subdivision of a polytope P ⊂ Rd is a polyhedral complex S such
that P = |S|. Moreover, we call S a triangulation of P if all F ∈ S are
simplices. The first two dissections in Figure ?? are subdivisions, the second
one is even a triangulation.
The benefit of the intersection property is evident: if P1 , . . . , Pm are
the inclusion-maximal
T polytopes in a subdivision S, then for every I ⊆ [m],
the polytope PI = i∈I Pi is a lattice polytope whenever the polytopes
P1 , . . . , Pm are. For lattice triangulations, that is, triangulations into
lattice simplices, Theorem 4.14 gives the following.
Corollary 5.2. Let P ⊂ Rd be a lattice polytope that admits a lattice triangulation. Then ehrP (n) agrees with a polynomial of degree dim P.
Proof. Let P1 , . . . , Pm be the inclusion-maximal elements of a lattice triangulation T . By the intersection property, every element in T is a lattice
simplex and Theorem 4.14 guarantees that (5.1) is an alternating sum of
polynomials. Exercise 5.15 states that dim Pi = dim P for all i, and since
Pi ∩ Pj ⊆ Pi is a proper face whenever i 6= j we conclude from Theorem 4.14,
that deg(ehrPI (n)) < dim P for all |I| > 1. Together with the fact that all
ehrPi (n) are polynomials of degree dim P with positive leading coefficient,
this proves our claim about the degree.
The reason for requiring the containment property is that we can appeal
to Chapter 2 and express an Ehrhart polynomial as follows.
Corollary 5.3. Let S be a subdivision of a lattice polytope P. Then Sb :=
S ∪ {P} is a poset with respect to inclusion and with maximal element ˆ1 = P
and
X
ehrP (n) =
−µSb(F, P) ehrF (n).
(5.3)
F ∈S
Proof. We define a function f on Sb by f (F ) := ehrF ◦ (n) for F ∈ S and
f (ˆ
1) := 0. Then for F ∈ S
X
f (F ) =
f (G) = ehrF (n)
GSbF
ˆ = ehrP (n). By
by Lemma 3.3 and, by the definition of a subdivision, f (1)
M¨
obius inversion (Theorem 2.12),
X
0 = f (ˆ
1) = f (ˆ1) +
µSb(F, P) ehrF (n) .
F ∈S
In the next section, we will determine the M¨obius function of a subdivision.
This is only a good plan for proving Ehrhart’s theorem and Ehrhart–
Macdonald reciprocity provided we can show that every lattice polytope has
114
5. Subdivisions
a lattice triangulation. Our first answer to that is that we actually already
know how to do this. Here is the setup: We fix a finite set V ⊂ Rd such
that P = conv(V ) is full-dimensional. Of course, it would suffice to assume
that V is the vertex set of P but as we will see, it will pay off to have some
flexibility. For a function ω : V → R, we write
V ω := {(v, ω(v)) ∈ Rd+1 : v ∈ V }
for the graph of the function ω. For now, we write ↑R := {(0, t) ∈ Rd+1 : t ≥ 0}
and π for the coordinate projection π(x1 , . . . , xd , xd+1 ) = (x1 , . . . , xd ), so
that π(↑R ) = {0}. Finally, we define
E ω (V ) := conv(V ω )+ ↑R
which we call the convex epigraph of ω for reasons that will become clear
in a second. By the Minkowski–Weyl Theorem 3.1, this a genuine polyhedron
in Rd+1 with bounded and unbounded faces.
Proposition 5.4. Let H = {(x, xd+1 ) : ha, xi + ad+1 xd+1 = b} be a hyperplane in Rd+1 . If E ω (V ) ⊆ H ≤ , then ad+1 ≤ 0. Moreover, the face E ω (V )∩H
is bounded if and only if ad+1 < 0.
Proof. For every p ∈ E ω (V ), we have p+ ↑R ⊂ E ω (V ). Hence, if E ω (V ) ⊆
H ≤ , then also p+ ↑R ⊂ H ≤ which implies ad+1 ≤ 0. The nonempty face
F = E ω (V ) ∩ H is unbounded if and only if F + ↑R ⊆ F . This happens if and
only if for every point (q, qd+1 ) ∈ H, we have (q, qd+1 + t) ∈ H for all t ≥ 0.
This last condition is satisfied if and only if ad+1 = 0.
It is clear that π(E ω (V )) = P. The crucial insight now is that P is already
the image of the collection of bounded faces of E ω (V ).
Theorem 5.5. Let P = conv(V ) and for ω : V → R, let E ω (V ) be its convex
epigraph. Then
S ω := {π(F ) : F ∈ Φbnd (E ω (V ))}
is a subdivision of P with vertices in V .
Proof. For a point p ∈ P, let us denote by ω(p) the smallest real number
ˆ = (p, h) ∈ E ω (V ). Then p
ˆ is contained in the boundary of
h such that p
ω
ˆ ∈ F ◦.
E (V ) (see Figure 1) and there is a unique face F ⊂ E ω (V ) with p
We claim that F is bounded. Indeed, F is unbounded if and only if all
ˆ are unbounded. However, Proposition 5.4 then implies
faces that contain p
ω
(p, ω(p) − ) ∈ E (V ) for some > 0. We now claim that π restricted
to the union of bounded faces |Φbnd (E ω (V ))| is injective. Since for every
p ∈ P there is exactly one h such that (p, h) is contained in a bounded
face, if suffices to show that π is injective when restricted to the relative
interior of every bounded face F ⊂ E ω (V ). Again by Proposition 5.4, we
know that F = E ω (V ) ∩ H for some supporting hyperplane H = {(x, xd+1 ) :
5.1. Decomposing a Polytope
115
Figure 1. The epigraph of a polytope.
ha, xi − xd+1 = δ}. Now observe that the map s : Rd → H given by
s(x) = (x, ha, xi − δ) is a linear inverse to π|H : H → Rd . Thus, π|H is an
isomorphism and, in particular, injective on F ◦ . This shows that S ω is a
polyhedral complex with |S ω | = P.
For a polytope P = conv(V ), we call a subdivision S of P regular
or coherent if there is a function ω : V → R such that S = S ω . In
Exercise 5.16 you will check that not all subdivisions of a polytope are
regular. Nevertheless, this construction proves the main result of this section.
Corollary 5.6. Every (lattice) polytope has a (lattice) triangulation.
Proof. Let V ⊂ Zd be a subset containing the vertices of P. For any
ω : V → R, Theorem 5.5 yields a subdivision S ω whose cells are all lattice
polytopes. To get a triangulation, we appeal to Exercise 5.17 for a suitable
choice of ω.
We now live up to our promise and explain why we call E ω (V ) the
convex epigraph. In the proof of Theorem 5.5, we defined for every p ∈ P
the number ω(p) as the smallest h such that (p, h) ∈ Pω . This yields a
continuous function ω : P → R.
116
5. Subdivisions
Proposition 5.7. Let P = conv(V ) be a polytope. For any ω : V → R, the
function ω : P → R is convex such that ω(v) ≤ ω(v) for all v ∈ V .
Proof. Let p, q ∈ P and 0 ≤ λ ≤ 1. Since E ω (V ) ⊂ Rd+1 is a convex
polyhedron, we know that (1 − λ)(p, ω(p)) + λ(q, ω(q)) ∈ E ω (V ) and hence
(1 − λ)ω(p) + λω(q) ≤ ω((1 − λ)p + λq).
Moreover, by definition (v, ω(v)) ∈ Pω and therefore ω(v) ≤ ω(v).
Hence E ω (V ) is the epigraph {(p, t) : p ∈ P, t ≥ ω(p)} of ω. Now there
are many convex functions f : P → R such that f (v) ≤ ω(v). However, ω is
special. It can be shown that ω is the unique convex function that minimizes
the volume of the convex body Kf := {(p, t) : p ∈ P, f (p) ≤ t ≤ M } where
M := max{ω(v) : v ∈ V }; see the at the end of the chapter.
simply cut along hyperplanes (idea from polyconvex sets)...difficult to
maintain the property of, say, lattice polytope (exercise, example?)
second answer: beneath–beyond appeal to points beneath a face
very algorithmic (in fact, yields first half of Minkowski–Weyl)
5.2. M¨
obius Functions of Subdivisions
Let P ⊂ Rd be a full-dimensional lattice polytope and T a triangulation
of P into lattice simplices, whose existence is vouched for by Corollary 5.6.
For a M¨obius-inversion-type representation (5.3) of ehrP (n) with respect to
Tb := T ∪ {P}, we need to know more about the M¨obius function µTb if we
want to make a more refined statement about ehrP (n).
If T = T ω is a regular triangulation, then T is isomorphic to the subposet Φbnd (E ω (V )) ⊂ Φ(E ω (V )) and our knowledge about M¨
obius functions
of polyhedra (Theorem 3.15) gives almost everything; see Exercise 5.1. However, as we know from Exercise 5.16, not all subdivisions are regular. It
turns out that exactly the methods developed in Section 3.4 will help us to
determine µTb .
Theorem 5.8. Let S be a subdivision of a polytope P and set Sb = S ∪ {P}.
Then for F, G ∈ Sb with F ⊆ G

dim G−dim F

if G 6= P,
(−1)
dim
P−dim
F
−1
µSb(F, G) = (−1)
if G = P and F 6⊆ ∂P,


0
if G = P and F ⊆ ∂P.
If G 6= P, then the interval [F, G]Sb is contained in Φ(G) and hence the
first case is subsumed by Theorem 3.15. For G = P, we need to ponder
the structure of the collection of cells K ∈ S that surround F , that is,
5.2. M¨
obius Functions of Subdivisions
117
{K ∈ S : F ⊆ K}. Recall from Section 3.4 that for a polyhedron Q and a
point q ∈ Q the tangent cone of Q at q is given by
Tq (Q) = {q + u : q + u ∈ Q for all > 0 sufficiently small} .
According to Proposition 3.16, Tq (Q) is the translate of a polyhedral cone
that depends only on the unique face F ⊆ Q that contains q in its relative
interior. In Chapter 3, tangent cones helped us to understand the invervals
[F, G]Φ(Q) . The following is a strengthening of Lemma 3.17 from polyhedra
to polyhedral subdivisions.
Lemma 5.9. Let P be a nonempty polytope with polyhedral subdivision S.
For a point q ∈ P, the set
Tq (S) := {Tq (G) : G ∈ S}
is a polyhedral complex with support |Tq (S)| = Tq (P). Moreover, if F ∈ S is
the unique face with q in its relative interior, then
[F, P]Sb ∼
= Tq (S)
given by K 7→ Tq (K).
Proof. Let u ∈ Rd . If q + u ∈ P for all suffiently small, then there is
a unique cell G ∈ S such that q + u ∈ G◦ for all < 0 . In particular,
this implies that F ⊆ G is a face and that Tq (G)◦ ∩ Tq (G0 )◦ = ∅ whenever
G 6= G0 . The containment property is a statement about polytopes and
hence a consequence of Lemma 3.17. This shows that
Tq (S) = {Tq (G) : F ⊂ G ∈ S} ∼
= [F, P]Sb
and |Tq (S)| = Tq (P).
Observe that Tq (S) does not depend on the choice of a point q ∈ F ◦
and hence we write TF (S). We are ready to prove Theorem 5.8.
Proof of Theorem 5.8. We only need to consider the case that G = P.
Hence, for a cell F ∈ S we compute
X
X
(−1)dim K−dim F = (−1)dim F
(−1)dim C = χ(TF (P))
F SbK≺SbP
C∈TF (S)
and therefore µSb(F, G) = −(−1)dim F χ(Tq (G)).
Let q ∈ F ◦ be a point in the relative interior of F . Now, if F ∈ S is
contained in the boundary of P, then q is contained in a proper face of
P and, by Lemma 3.17, TF (P) is a proper polyhedron. Hence, our Euler
characteristic computations (Corollaries 3.12 and 3.14) gives χ(Tq (P)) = 0.
On the other hand, if F 6⊆ ∂P, then q ∈ P◦ . In this case Tq (P) = aff(P) and
χ(Tq (P)) = (−1)dim P finishes the proof.
118
5. Subdivisions
With this machinery we can extend Ehrhart–Macdonald reciprocity from
lattice simplices to all lattice polytopes.
Theorem 5.10. Let P ⊂ Rd be a lattice polytope and ehrP (n) its Ehrhart
polynomial. Then
(−1)dim P ehrP (−n) = |nP◦ ∩ Zd |
for all n > 0. Moreover, ehrP (0) = χ(P) = 1.
Proof. Let S be a triangulation of P into lattice polytopes. This exists by
Theorem 5.6. M¨
obius inversion on Sb = S ∪ {P} yields
X
ehrP (n) =
−µSb(F, P) ehrF (n).
F ∈S
By Theorem 5.8, this simplifies to
X
ehrP (n) =
(−1)dim P−dim F ehrF (n).
F ∈S,F 6⊆∂P
Ehrhart–Macdonald reciprocity for simplices (Theorem 4.14) asserts that
ehrF (−n) = (−1)dim F |nF ◦ ∩ Zd | for every lattice simplex F . Hence
X
(−1)dim P ehrP (n) =
ehrF ◦ (n)
F ∈S,F 6⊆∂P
where the right-hand side counts the number of lattice points in the relative
interior of nP, by Lemma 3.3 and the definition of a subdivision. For every
lattice simplex F , the constant term of ehrF ◦ (n) is (−1)dim F and setting
n = 0, this is just the computation of the Euler characteristic of P.
5.3. Beneath, beyond, and half-open decompositions
We promised a second technique to obtain nontrivial subdivisions of a
given (lattice) polytope. The technique presented in this section is not only
algorithmic but it also furnishes a general methodology that dissections
actually do suffice for most of our purposes.
Let P ⊂ Rd be a full-dimensional polytope with vertex set V = vert(P ).
Fix a vertex v ∈ V and let P0 = conv(V \ v). Suppose that we have a
subdivision S 0 of P0 . How can we extend S 0 to a subdivision of P? Recall
from Section 3.6 that the point p is beyond a face F ⊆ P0 if p 6∈ TF (P0 ).
From the definition of tangent cones, it follows that p is beyond F if and
only if for any q ∈ F
[p, q] ∩ P0 = {q}.
In lay terms, this means that every point of F is visible from p. Let us define
the collection of faces for which p is beyond.
Bp (P0 ) := {F P0 : p is beyond F }.
5.3. Beneath, beyond, and half-open decompositions
119
Proposition 5.11. The collection of polytopes Bp (P0 ) is a polyhedral complex.
Proof. The collection Bp (P0 ) is a subset of the boundary complex of P0 .
Hence, the intersection property is automatically satisfied and we only need
to check the containment property. This, follows from TG (P0 ) ⊆ TF (P0 ) for
faces G ⊆ F .
We can now give a way of extending a subdivision of P0 to P. For that
we have to fill in the part P \ P0 .
Theorem 5.12. Let P0 be a polytope with subdivision S 0 . For a point v 6∈ P0 ,
a subdivision of P = conv(P0 ∪ v) is given by
S := S 0 ∪ {conv(C ∪ v) : C ∈ S 0 , C ⊆ F for some F ∈ Bv (P0 )}.
Proof. We write Cv for conv(C ∪ v) where C is a cell satisfying the stated
conditions. For any K ∈ S 0 \ Bv (P0 ) it is clear that K ∩ Cv = ∅ as for every
q ∈ Cv the half-open segment [v, q] does not meet P0 unless q ∈ |Bv (P0 )|.
.
This gives a practical algorithm for producing a (lattice) triangulation
of a (lattice) polytope P: Assume that P ⊂ Rd+1 is a full-dimensional
polytope with vertices v1 , v2 , . . . , vn such that P0 := conv(v1 , . . . , vd+1 ) is a
d-simplex. In particular, a triangulation of P0 is given by T0 := Φ(P0 ). For
every i = 1, . . . , n − d − 1, use Theorem 5.12 to obtain a triangulation Ti of
Pi = conv(Pi−1 ∪ vd+1+i ) from Ti−1 . Exercise 5.2 deals with the necessary
linear algebra behind and Exercise 5.3 utilizes Theorem 5.12 to give a proof
of the first half of the Minkowski–Weyl Theorem 3.1.
every point in P0 \ P expressible by v and a point in visible complex
restriction yields subdiv/triang of visible complex
“cone” over this subcomplex extends the subdivison
turn this into algorithm:
given v1 , . . . , vn , assume that v1 , . . . , vd+1 are affinely indep add point,
keep track of triangulation
Exercise: this also yields first half of minkowski–weyl
in fact, turn every ’new’ simplex into a half-open simplex. that only
records the ‘new’ points in P0 .
rather special subdivison
visibility: night-and-day
and half-open decompositions
There is another, more algorithmic way to obtain a (lattice) triangulation
of a (lattice) polytope.
120
5. Subdivisions
day and night sides of polytopes and connection to beneath beyond
night-sides are bright-sides
Proof of Stanley-reciprocity for general polytopes (and hence yet another
proof of Ehrhart–Macdonald)
5.4. h∗ -vectors and unimodular triangulations
What are Ehrhart-polynomials
h∗ -vectors
Stanley’s nonnegativity (and monotonicity??) via half-open decompositions
unimodular triangulations and combinatorics; connections to h-vectors
from earlier
5.5. Solid angles?
Exercises
5.1 Let S = S ω be a regular subdivision of a polytope P. Use the fact
that Sb ∼
= Φbnd (E ω (V )) ∪ {E ω (V )} ⊂ Φ(E ω (V )) and Theorem ?? to infer
Theorem 5.8 for regular subdivisions.
5.2 Let P be a full-dimensional polytope and v 6∈ P. Let F1 , . . . , Fm be the
facets of P.
(a) Show that v is beyond a face F ⊆ P if and only if v is beyond some
facet Fi containing F .
(b) Assume that Fi = P ∩ {x : hai , xi = bi } and such that hai , pi ≤ bi
for all p ∈ P. Show that v is beyond Fi if and only if hai , vi > bi .
(c) For a given triangulation T of P describe how to algorithmically
obtain a triangulation for conv(P ∪ v).
5.3 (a) Let Q ⊂ Rd be a simplex with vertices v1 , . . . , vd . Write Q as an
intersection of halfspaces.
(b) Let P be a polytope with a triangulation T given in terms of
vertices of P. Give a description of P in terms of halfspaces.
5.6. Ehrhart–Macdonald Reciprocity
121
5.6. Ehrhart–Macdonald Reciprocity
Theorem 5.13. Given a pointed rational cone K, fix a generic vector s ∈
aff(K). We call a facet F of K illuminated if there exist x ∈ F and > 0
such that x + s ∈ K, let Day(K) denote the collection of all illuminated
facets, and Night(K) the collection of the remaining facets of K. Then as
rational functions,
σK\S Day(K) z1 = (−1)dim(K) σK\S Night(K) (z) .
Figure 2 illustrates how we partition the facets of a cone into “day” and
“night” collections.
Figure 2. Illuminating certain facets of a cone.
Proof. Fix a triangulation T of K without introducing new generators, and
let S1 , S2 , . . . , Sk denote the simplicial cones in T of maximal dimension
(i.e., dim(K)). Given our generic vector s ∈ aff(K), let Day(Sj ) denote the
collection of all illuminated facets of Sj and Night(Sj ) the collection of the
remaining facets of Sj . Then Exercise 5.7 (see also Figure 3) says that
Figure 3. Illuminating a triangulation.
K\
[
Day(K) =
k [
j=1
Sj \
[
Day(Sj )
(5.4)
122
5. Subdivisions
and
K\
[
k [
Night(K) =
Sj \
[
Night(Sj )
(5.5)
j=1
where the unions with running index j are disjoint. Thus, with Theorem 4.18
and Exercise 5.6,
σK\S Day(K)
1
z
=
k
X
σSj \S Day(Sj )
1
z
k
X
=
(−1)dim(Sj ) σSj \S Night(Sj ) (z)
j=1
j=1
dim(K)
= (−1)
σK\S Night(K) (z) ,
as claimed.
A simpler version of Theorem 5.13 follows immediately
S from the special
case that s points into the cone, which means that Day(K) = ∂K and
Night(K) = ∅:
Corollary 5.14. Let K be a pointed rational cone. Then as rational functions,
σK◦ z1 = (−1)dim(K) σK (z) .
Completely analogous to our argument in Section 4.8 that derived (??)
from (4.28), Corollary 5.14 yields:
Corollary 5.15. Let P be a lattice polytope. Then as rational functions,
EhrP z1 = (−1)dim(P)+1 EhrP◦ (z) .
Theorem 5.16. Suppose P is a lattice polytope.
(a) For positive integers n, the counting function ehrP (n) is a polynomial in
n.
(b) When this polynomial is evaluated at negative integers, we obtain
ehrP (−n) = (−1)dim(P) ehrP◦ (n) .
Proof. Part (a) follows from triangulating P and Theorem 4.14(a).
Part (b) follows, similarly to our proof of Theorem 4.14(b), with Theorem 4.6 and Corollary 5.15:
X
ehrP (−n) z n = − EhrP z1 = (−1)dim(P) EhrP◦ (z)
n≥1
= (−1)dim(P)
X
ehrP◦ (n) z n .
n≥1
Now compare coefficients.
With Theorem 5.16, we now know what the polynomial ehrP (n) evaluates
to for any n ∈ Z \ {0}; our next result completes this picture.
5.7. Solid Angles
123
Theorem 5.17. If P is a lattice polytope then ehrP (0) = χ(P) = 1.
Proof. Fix a triangulation of P, and denote the open faces of the simplices of
this triangulation by ∆◦1 , ∆◦2 , . . . , ∆◦m ; note that these are all lattice simplices
(of varying dimension) with vertices coming from the vertices of P. Since P
is the disjoint union of ∆◦1 , ∆◦2 , . . . , ∆◦m ,
ehrP (n) =
m
X
ehr∆◦j (n) .
j=1
To compute its constant term ehrP (0), we think of P = ∆◦1 ∪ ∆◦2 ∪ · · · ∪ ∆◦m
as a polyconvex set. By Theorem 4.14,
ehrP (0) =
m
X
ehr∆◦j (0) =
j=1
m
X
dim ∆j
(−1)
ehr∆j (0) =
j=1
m
X
(−1)dim ∆j .
j=1
But the summands on the right-hand side are the Euler characteristics of
∆◦1 , ∆◦2 , . . . , ∆◦m , by Corollary 3.10. Thus the left-hand side ehrP (0) is the
Euler characteristic of P, which is 1.
Our proof of Theorem 5.16 can be generalized for a rational polytope,
that is, the convex hull of points in Qd . We invite the reader to prove the
following generalization of Theorem 5.16 in Exercise 4.39:
Theorem 5.18. If P is a rational d-polytope, then for positive integers n,
the counting function ehrP (n) is a quasipolynomial in n whose period divides
the least common multiple of the denominators of the vertex coordinates of
P. When this quasipolynomial is evaluated at negative integers, we obtain
ehrP (−n) = (−1)d ehrP◦ (n) .
Theorems 5.16 and 5.18 were proved in 1962 by Eug`ene Ehrhart, in
whose honor we call ehrP (n) the Ehrhart (quasi-)polynomial of P.
We can also compute Ehrhart quasipolynomials of rational polytopal
complexes C: If ehrC (n) denotes the number of lattice points in the n-dilate
of the union of polytopes in C, then ehrC (n) is a quasipolynomial whose
constant term is ehrC (0) = χ(C), by repeating the argument given in the
proof of Theorem 5.17.
5.7. Solid Angles
Suppose P ⊆ Rd is a polyhedron. The solid angle ωP (x) of a point x (with
respect to P) is a real number equal to the proportion of a small ball centered
at x that is contained in P. That is, we let B (x) denote the ball of radius centered at x and define
vol (B (x) ∩ P)
ωP (x) :=
vol B (x)
124
5. Subdivisions
for all sufficiently small . We note that when x ∈
/ P, ωP (x) = 0; when
x ∈ P◦ , ωP (x) = 1; and when x ∈ ∂P, 0 < ωP (x) < 1. Computing solid
angles is somewhat nontrivial, as the reader can see, e.g., in Exercise 5.10.
Now let P be a polytope. We define
X
AP (n) :=
ωnP (m) ,
m∈nP∩Zd
the sum of the solid angles at all integer points in nP; recalling that ωP (x) = 0
if x ∈
/ P, we may also write
X
AP (n) =
ωnP (m) .
m∈Zd
The following theorem can be proved along the exact same lines as our proof
of Theorem ??. We invite the reader to do so in Exercise 5.11.
Theorem 5.19. Suppose P is a lattice d-simplex. Then AP (n) is a polynomial in n of degree d whose constant term is 0. Furthermore, AP (n) is either
even or odd:
AP (−n) = (−1)d AP (n) .
Because solid angles behave additively when we glue two polytopes
together (and because we do not have to take into account lower-dimensional
intersections), this result effortlessly1 extends to general lattice polytopes,
i.e., convex hulls of finitely many points in Zd :
Corollary 5.20. Suppose P is a lattice d-polytope. Then AP (n) is a polynomial in n of degree d whose constant term is 0. Furthermore, AP (n) is
either even or odd:
AP (−n) = (−1)d AP (n) .
Note that the analogous step from Theorem 5.19 to Corollary 5.20 is not
as easy for Ehrhart polynomials: polynomiality does extend (as we have
shown in Section 5.6), but for reciprocity, we need to introduce additional
machinery (which we will do in the next chapter). Assuming the general
case of Ehrhart–Macdonald Reciprocity (Theorem 5.22) for the moment, we
can derive the following classical result, a higher-dimensional analogue of the
fact that the angles in a triangle add up to 180◦ , as the reader should show
in Exercise 5.12.
Theorem 5.21 (Brianchon, Gram). Suppose P is a rational d-polytope.
Then
X
(−1)dim F ωP (F) = 0 ,
F
1 As in Section 5.6, here we need the fact that every polytope can be triangulated into
simplices, which is the statement of Corollary 5.6.
5.8. Ehrhart–Macdonald Reciprocity Revisited [old]
125
where the sum is taken over all faces F of P, and ωP (F) := ωP (x) for any x
in the relative interior of F.
5.8. Ehrhart–Macdonald Reciprocity Revisited [old]
The M¨obius function we computed in Corollary ?? allows us to deduce
the general Ehrhart–Macdonald reciprocity theorem from the simplex case
(Proposition ??).
Theorem 5.22 (Ehrhart, Macdonald). If P is a rational d-polytope, then
for positive integers t, the counting function ehrP (n) is a quasipolynomial in
t whose period divides the least common multiple of the denominators of the
vertex coordinates of P. When this quasipolynomial is evaluated at negative
integers, we obtain
ehrP (−n) = (−1)d ehrP◦ (n) .
Proof. Compute a triangulation lattice Φ of P. Proposition ?? gives the
quasipolynomial character of ehrP (n) and the statement about its period.
We now show how the general reciprocity relation follows from the simplex
case (Proposition ??). We will use M¨obius inversion (Theorem 2.12) on the
triangulation lattice Φ for the functions
(
ehrF (n) if F 6= 1,
f (F) :=
ehrP (n) if F = 1,
and
(
ehrF◦ (n)
f (F) :=
0
if F 6= 1,
if F = 1.
Because every point in P is in the interior of a unique face,2
X
X
f (1) = ehrP (n) =
ehrF◦ (n) =
f (F) .
F∈Φ
F∈Φ\{1}
By M¨
obius inversion,
X
X
f (1) = 0 =
µ(F, 1) f (F) = ehrP (n) +
(−1)d+1−dim F ehrF (n) ,
F∈Φ
F∈Φ\{1}
F6⊂∂P
that is,
ehrP (n) = (−1)d
X
(−1)dim F ehrF (n) .
F∈Φ\{1}
F6⊂∂P
2 Here we mean relative interior; in particular, F◦ = F if F is a vertex.
(5.6)
126
5. Subdivisions
Now we evaluate these quasipolynomials at negative integers and use Proposition ?? for the simplices F ∈ Φ \ {1}:
X
X
ehrP (−n) = (−1)d
(−1)dim F ehrF (−n) = (−1)d
ehrF◦ (n)
F∈Φ\{1}
F6⊂∂P
F∈Φ\{1}
F6⊂∂P
= (−1)d ehrP◦ (n) .
5.9. Unimodular Triangulations
Now we define the face numbers of a triangulation (lattice) Φ of a polytope
P, in sync with Chapter 3: let fk denote the number of k-dimensional
simplices in Φ that are not contained in the boundary of P. In Chapter 3, we
collected these numbers in a vector (the f -vector); for a change of scenery,
this time, we introduce the polynomial
fΦ (n) =
d
X
X
fk nk =
k=0
ndim F .
F∈Φ\{∅,1}
F6⊂∂P
However, this f-polynomial is not nearly as handy as the h-polynomial,
defined through
hΦ (n) =
d
X
hk nk =
k=0
d
X
fk (n − 1)d−k .
k=0
The h-polynomial makes a miraculous appearance in the Ehrhart series
X
EhrP (z) =
ehrP (n) z n
n≥0
of a lattice polytope that admits a unimodular triangulation, that is, a
triangulation into unimodular simplices.
Theorem 5.23 (Stanley). Suppose P is a d-dimensional lattice polytope that
has a unimodular triangulation Φ. Then
hΦ (z)
EhrP (z) =
.
(1 − z)d+1
Proof. We use (5.6):
ehrP (n) = (−1)d
X
(−1)dim F ehrF (n)
F∈Φ\{1}
F6⊂∂P
F
and Proposition 4.15, which says that ehrF (n) = n+dim
for a unimodular
dim F
simplex F. Thus
X
d
X
d−dim F n + dim F
d−k n + k
ehrP (n) =
(−1)
=
(−1)
fk ,
dim F
k
F∈Φ\{∅,1}
F6⊂∂P
k=0
Notes
127
and so
X
n
ehrP (n) z =
n≥0
d
X
d−k
(−1)
n≥0
k=0
Pd
=
d
X n + k X
(−1)d−k fk
n
fk
z =
k
(1 − z)k+1
k=0
(z − 1)d−k
(1 − z)d+1
k=0 fk
and the definition of hΦ (z) finishes our proof.
(By far) not every lattice polytope has a unimodular triangulation—
consider, e.g., the tetrahedron with vertices (1, 0, 0), (0, 1, 0), (0, 0, 1), and
(1, 1, 1). Nevertheless, if P is a lattice polytope, we can always write EhrP (z)
as a rational function with denominator (1 − z)d+1 , and in Exercise 5.2 we
explored a few properties of the numerator coefficients.
Notes
Triangulations of polyhedra and manifolds form an active area of research
with numerous interesting open problems, many of which are described
in [23].
Theorem 5.22 was conjectured (and proved in several special cases) by
Eug`ene Ehrhart and proved by I. G. Macdonald [51]. Its multivariate
generalization (Exercise 5.1) is due to Richard Stanley [73].
The reciprocity theorems for Ehrhart polynomials and their solid-angle
counterparts (Theorem 5.22 and Corollary 5.20) are particular cases of
a reciprocity relation for simple lattice-invariant valuations due to Peter
McMullen [53], who also proved a parallel extension of Ehrhart–Macdonald
reciprocity to general lattice-invariant valuations.
Ehrhart–Macdonald reciprocity takes on a special form for reflexive
polytopes which we define and study in Exercise 5.4. The term reflexive
polytope was coined by Victor Batyrev, who motivated these polytopes by
applications of mirror symmetry in string theory [10]. That the Ehrhart
series of a reflexive polytope exhibits an unexpected symmetry (Exercise 5.4)
was discovered by Takayuki Hibi [40]. The number of reflexive polytopes in
dimension d is known only for d ≤ 4 [47, 48]; see also [1, Sequence A090045].
Theorem 5.23 is due to Richart Stanley [74]. As we mentioned already,
lattice polytopes that admit unimodular triangulations are quite special.
There are other, less restrictive, classes of lattice polytopes which come with
numerous applications (e.g., to semigroup algebras) and open problems [21].
128
5. Subdivisions
Exercises
5.1 Extend Exercise ?? to general rational cones, i.e., cones whose generators
are in Qd : prove that if C is a rational cone in Rd , then σC (z1 , z2 , . . . , zd )
is a rational function which satisfies
σC z11 , z12 , . . . , z1d = (−1)dim C σC ◦ (z1 , z2 , . . . , zd ) .
5.2 For a lattice d-polytope P ⊂ Rd , let
X
EhrP (z) =
ehrP (n) z n
and
EhrP◦ (z) =
X
ehrP◦ (n) z n
t>0
n≥0
(as in Corollary 5.15). By Proposition 4.13 and Theorem 5.22, we can
write
hd z d + hd−1 z d−1 + · · · + h0
EhrP (z) =
(1 − z)d+1
for some h0 , h1 , . . . , hd . Prove:
(a) h0 = 1
(b) h1 = P ∩ Zd − d − 1
(c) hd = P◦ ∩ Zd (d) h0 + h1 + · · · + hd = d! vol(P) .
5.3 Compute the Ehrhart polynomial of the pyramid with vertices (0, 0, 0),
(2, 0, 0), (0, 2, 0), (2, 2, 0), and (1, 1, 1).
5.4 A reflexive polytope is a lattice polytope P such that the origin is
an interior point of P and3
ehrP◦ (n) = ehrP (n − 1)
for all n .
(5.7)
Prove that if P is an lattice d-polytope that contains the origin in its
interior and that has the Ehrhart series
EhrP (z) =
hd z d + hd−1 z d−1 + · · · + h1 z + h0
,
(1 − z)d+1
then P is reflexive if and only if hk = hd−k for all 0 ≤ k ≤ d2 .
5.5 Compute the Ehrhart polynomial of the octahedron with vertices
(±1, 0, 0), (0, ±1, 0), and (0, 0, ±1). Generalize to the d-dimensional
cross polytope, the convex hull of the unit vectors in Rd and their
negatives.
3More generally, if the 1 on the right-hand side of (5.7) is replaced by an arbitrary fixed
positive integer, we call P Gorenstein. You may think about how Exercise 5.4 can be extended
to Gorenstein polytopes.
Exercises
129
5.6 Fix linearly independent vectors v1 , v2 , . . . , vk ∈ Zd and let
ˆ := R≥0 v1 + · · · + R≥0 vm−1 + R>0 vm + · · · + R>0 vk .
K
Show that there exist a cone K and a vector s ∈ aff(K) with whose help
we can define illuminated facets of K as in Theorem 5.13, such that
ˆ = K \ Day(K).
K
5.7 As in the setup and proof of Theorem 5.13, let K be a pointed rational
cone and fix a generic vector s ∈ aff(K) determining Day(K) and
Night(K). Let S1 , S2 , . . . , Sk denote the simplicial cones of maximal
dimension in a triangulation of K. Prove (5.4) and (5.5), i.e.,
K\
[
Day(K) =
k [
Sj \
[
Day(Sj )
Sj \
[
Night(Sj )
j=1
and
K\
[
Night(K) =
k [
j=1
and that the unions with running index j are disjoint.
5.8 Show that if P is a d-dimensional lattice polytope in Rd , then the degree
of its Ehrhart polynomial ehrP (n) is d and the leading coefficient is the
volume of P. What can you say if P is not full dimensional?
5.9 Find and prove an interpretation of the second leading coeffient of
ehrP (n) for a lattice polytope P. (Hint: start by computing the Ehrhart
polynomial of the boundary of P.)
5.10 Compute the solid angles of all points in the tetrahedron with vertices
(0, 0, 0), (1, 0, 0), (0, 1, 0), and (0, 0, 1). (You may use Theorem 5.21.)
5.11 Prove Theorem 5.19: Suppose P is a lattice d-simplex. Then AP (n) is
a polynomial in n of degree d whose constant term is 0. Furthermore,
AP (n) is either even or odd:
AP (−n) = (−1)d AP (n) .
5.12 Prove Theorem 5.21: Suppose P is a rational d-polytope. Then
X
(−1)dim F ωP (F) = 0 ,
F
where the sum is taken over all faces F of P, and ωP (F) := ωP (x) for
any x in the relative interior of F.
5.13 State and prove a generating-function analogue of Corollary 5.20 along
the lines of Corollary 5.15.
5.14 Show that the interpretations for h∗d , h∗1 , and h∗0 + h∗1 + · · · + h∗d given
in Exercise 4.38 hold for any lattice d-polytope (not just a simplex).
130
5. Subdivisions
5.15 Let S be a subdivision of P. Show that if F ∈ S such that dim F <
dim P, then there is G ∈ S with F ⊂ G.
5.16 Consider the following subdivision with ...
5.17 A stronger condition that all bounded faces of E ω (V ) are simplices is
that no hyperplane in Rd+1 contains more than d + 1 points of V ω .
Indeed, any supporting hyperplane of E ω (V ) will then contain k ≤ d + 1
points which are then the vertices of a (k − 1)-simplex.
Let us assume that V = {v1 , v2 , . . . , vn } ⊂ Rd is a configuration of
n ≥ d + 2 such that P = conv(V ) is full-dimensional. We claim that
there is a M 0 such that ω(vi ) := M i does the job.
(a) Assume that n = d + 2. Then V ω is contained in a hyperplane
if and only if the points V ω are affinely dependent, that is, if the
(d + 2) × (d + 2)-matrix


v1 v2 · · · vd+2
M M 2 · · · M d+2 
1
1 ···
1
has determinant = 0. Use Laplace expansion to show that this can
only happen for finitely many values of M .
(b) Argue that this holds true for n > d + 2 by considering all (d + 2)subsets of V .
Chapter 6
Partially Ordered Sets,
Geometrically
Order is not sufficient. What is required, is something much more complex. It is
order entering upon novelty; so that the massiveness of order does not degenerate
into mere repetition; and so that the novelty is always reflected upon a background
of system.
Alfred North Whitehead
[under construction]
Now we go back to the beginnings, the posets of Chapter 2. The central
object of this chapter is the order cone for a given a finite poset Π,
KΠ :=
x ∈ RΠ
≥0 : j k ⇒ xj ≤ xk .
The philosophy we have in mind is that a point x ∈ KΠ is, by definition, a
map Π → R≥0 that is order preserving. (Viewing R≥0 as another poset with
the relation ≤, we can think of such a map as a poset homomorphism.) On
the other hand, RΠ is a vector space of dimension |Π|, and so we can view
KΠ as a geometric object (indeed, a pointed cone) in this space, an object
that realizes the poset Π geometrically. This point of view will shed new
light on several results from Chapters 2 and 4, with a healthy dose of input
from Chapters 3 and 5.
131
132
6. Partially Ordered Sets, Geometrically
6.1. The Geometry of Order Cones
Our first goal is to study the geometry of KΠ , alongside a family of natural
subdivisions. We warm up with three example posets:
• the chain Γd consisting of d totally ordered elements;
• the antichain Ad consisting of d elements with no relations;
• the diamond poset 3 pictured in Figure 1.
Figure 1. The diamond poset 3.
The order cones of these posets are all old friends: the one for a chain,
n
o
KΓd = x ∈ Rd : x1 ≥ x2 ≥ · · · ≥ xd ≥ 0
made several appearances in Chapter 4, starting with our study of partition
functions; its generators are
   
 
1
1
1
 1 
 0   1 
   
 
 0   0 
 
 , ,..., 1 .
 ..   .. 
 .. 
 . 
 .   . 
0
0
1
The order cone of an antichain is the nonnegative orthant:
KAd = Rd≥0 ,
≥
≥
generated by the unit vectors e1 , e2 , . . . , ed in Rd . Both KΓd and KAd are
simplicial (in fact, even unimodular). This is not true for the order cone of
the diamond poset,


x1 ≥ x2 

K3 = x ∈ R4≥0 :
,


x3 ≥ x4
which made an appearance in Section 4.8, where we painted a geometric
picture of plane partitions. There we implicitly computed the generators of
K3 , namely,
         
1
1
1
1
1
 0   1   0   1   1 
 , , , , .
 0   0   1   1   1 
0
0
0
0
1
6.1. The Geometry of Order Cones
133
Our examples hint at a general geometric structure of order cones. Our
next goal is to describe the extreme rays of KΠ , but we first show that we
can restrict ourselves to a connected poset, that is, a poset whose Hasse
diagram is a connected graph. The following result follows essentially from
the definition of an order cone.
Lemma 6.1. If a finite poset can be written as Π1 ∪ Π2 with no relations
between elements of Π1 and Π2 , then
KΠ1 ∪Π2 = KΠ1 × KΠ2 .
We recall from Chapter 2 that a subset F ⊆ Π is a filter if k ∈ F and
j k implies j ∈ F . For example, 3 has five nonempty filters, as many as
there are generators of K3 . This is not a coincidence:
P
Theorem 6.2. If Π is finite and connected, then j∈F ej , for all nonempty
connected filters F of Π, form a set of generators of KΠ .
Proof. Exercise 6.3 says that the facets of KΠ are precisely the sets KΠ ∩
{xj = xk } for some j ≺· k and KΠ ∩ {xj = 0} for some minimal j ∈ Π. Thus
an extreme ray of KΠ is a 1-dimensional set of points whose coordinates
come in exactly two flavors: those that are zero and the nonzero coordinates,
which are all equal. The latter are necessarily
P indexed by a connected filter
F , and so the extreme ray is generated by j∈F ej .
P
Conversely, given a connected filter F of Π, the ray generated by j∈F ej
(which is certainly in KΠ ) equals
all xj for j ∈ F equal
Π
R := x ∈ R≥0 :
.
xj = 0 for j ∈
/F
Now choose real numbers aj , for j ∈ F , that sum to zero and satisfy aj > ak
whenever j k. Then all points on R are on the hyperplane




X
X
x ∈ RΠ :
aj xj +
xj = 0 ,


j∈F
whereas any point x ∈ KΠ \ R satisfies
means that R is an extreme ray of KΠ .
j ∈F
/
P
j∈F
aj xj +
P
j ∈F
/
xj > 0. This
Given two posets Π and Γ on the same ground set, we say that Γ refines
Π if
j Π k =⇒ j Γ k .
If, in addition, Γ is a chain then we call Γ a linear extension of Π.
Our next goal is to study subdivisions of an order cone. We have already
seen such a subdivision (in fact, a triangulation) of K3 in (4.33):
K3 = x ∈ R4≥0 : x1 ≥ x2 ≥ x3 ≥ x4 ∪ x ∈ R4≥0 : x1 ≥ x3 ≥ x2 ≥ x4 .
134
6. Partially Ordered Sets, Geometrically
The two cones on the right-hand side are order cones in their own right,
namely, those of the chains {1 2 3 4} and {1 3 2 4}. Figure 2
illustrates that these two chains are the two linear extensions of 3.
Figure 2. The two linear extensions of 3.
For a second example, we consider KAd = Rd≥0 . A point in Rd≥0 has
coordinates that follow some order, and since every possible permutation of
orders is possible, we can write
o
[ n
Rd≥0 =
x ∈ Rd : xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0 ,
(6.1)
σ∈Sd
where Sd stands, as usual, for the set of permutations on d letters. The cones
on the right should look suspiciously familiar from Chapter 4: they are all
unimodular and they, in fact, triangulate Rd≥0 (Exercise 6.5). Again, we can
interpret these cones as order cones, namely, those for all possible chains on
d elements. Since Ad is an antichain, these are precisely the linear extensions
of Ad . So (6.1) represents the unimodular triangulation
[
K Ad =
KΓ
(6.2)
Γ
where the union is over all chains Γ on d elements. We can do something
more general. Namely, if Π1 , Π2 , . . . , Πm are posets on [d] with the property
that for every j there is a unique d-chain that refines Πj , then
K A d = Π 1 ∪ Π2 ∪ · · · ∪ Πm
is a regular subdivision of KAd = Rd≥0 (Exercise 6.7). This is a special case
of the following theorem.
Theorem 6.3. Let Π be a finite poset and Π1 , . . . , Πm a collection of distinct
posets refining Π. Then we have the regular subdivision
KΠ = KΠ1 ∪ · · · ∪ KΠm
if and only if for every linear extension Γ of Π there is a unique j such that
Γ refines Πj .
6.2. Order Polytopes
135
Proof.
Corollary 6.4. Let Π be a finite poset. Then we have the regular unimodular
triangulation
[
KΠ =
KΓ
Γ
where the union is over all linear extensions Γ of Π.
You can check that ‘refinement’ is an order relation on
Nn := {Π = ([n], Π poset on [n]}
observe that Π1 Π2 implies OΠ1 ⊂ OΠ2 . So we could hope that for every
poset Π there is a collection of posets Π1 , . . . , ΠM so that
OΠ = OΠ1 ∪ · · · ∪ OΠM
is a dissection.
6.2. Order Polytopes
Given a finite poset Π, a bounded variant of the order cone KΠ is the order
polytope
OΠ := x ∈ [0, 1]Π : j k ⇒ xj ≤ xk .
Our two polyhedral constructions stemming from Π are intimately connected,
ˆ through adding
as can be most prominently seen by constructing the poset Π
an element ˆ
1 to Π that is than any member of Π and keeping the remaining
relations. Its order cone is
j k ⇒ xj ≤ xk and
ˆ
Π
KΠˆ = x ∈ R :
= hom (OΠ ) ,
(6.3)
0 ≤ xj ≤ xˆ1 for all j ∈ Π
the homogenization of the order polytope of Π. This relation yields, for
example, an almost immediate description of the vertices of OΠ from Theorem 6.2.
TheoremP6.5. If Π is finite, then the vertices of the order polytope OΠ are
precisely j∈F ej , for all filters F of Π.
Proof. Equation (6.3) implies that
OΠ = KΠˆ ∩ xˆ1 = 1 ,
and so the vertices of OΠ are exactly the generators of KΠˆ with the xˆ1 ˆ is connected, as are all its nonempty
coordinate removed. We note that Π
filters (since they all contain ˆ1). The result now follows with Theorem 6.2. 136
6. Partially Ordered Sets, Geometrically
Theorem 6.5 says, in particular, that an order polytope is a lattice polytope. (Even more, its vertices have coordinates only among 0 and 1, and so
an order polytope is a 0/1-polytope.) Thus, by Ehrhart’s Theorem (Corollary 5.2), ehrOΠ (n) is a polynomial in n. In fact, we know this polynomial
quite well:
Proposition 6.6. Let Π be a finite poset. Then
ΩΠ (n) = ehrOΠ (n − 1) .
Proof. Let f ∈ (n − 1)OΠ ∩ ZΠ . Then, by definition, f is an order preserving
map with range f (Π) ⊆ {0, 1, . . . , n − 1}. So, if we define the map φ by
φ(p) := f (p) + 1 for p ∈ Π, then φ is order preserving with φ(Π) ⊆ [n]. This
argument works also in the other direction and proves that the lattice points
in (n − 1)OΠ are in bijection with order preserving maps into the n-chain. Proposition 6.6, with assistance from Theorem 6.5 and Corollary 5.2,
gives an alternative, geometric proof that ΩΠ (n) agrees with a polynomial
of degree n (Proposition 1.11). This geometric point of view also yields an
alternative proof for the reciprocity theorem for order polynomials.
Second proof of Theorem 1.12. Essentially by the same argument as in
our proof of Proposition 6.6, we can realize the order polynomial Ω◦Π (n) as
counting interior lattice points in a dilate of the order polytope OΠ . More
precisely,
Ω◦Π (n) = ehrO◦Π (n + 1)
(6.4)
(Exercise 6.4). Thus, by Ehrhart–Macdonald reciprocity (Theorem ??),
Ω◦Π (−n) = ehrO◦Π (−n + 1) = (−1)|Π| ehrOΠ (n − 1) = ΩΠ (n) .
Our next goal is to study subdivisions of order polytopes. Thanks once
more to (6.3), Theorem 6.3 has an immediate analogue for order polytopes.
Corollary 6.7. Let Π be a finite poset and Π1 , . . . , Πm a collection of distinct
posets refining Π. Then we have the regular subdivision
OΠ = OΠ1 ∪ · · · ∪ OΠm
if and only if for every linear extension Γ of Π there is a unique j such that
Γ refines Πj .
Corollary 6.8. Let Π be a finite poset. Then we have the regular unimodular
triangulation
[
OΠ =
OΓ
Γ
where the union is over all linear extensions Γ of Π.
Corollary 6.9. Let Π be a finite poset. Then the volume of OΠ equals
times the number of linear extensions of Π.
1
|Π|!
6.3. Generating Functions and Permutation Statistics
137
Proposition 6.10. The unimodular triangulation of OΠ given in Corollary 6.8, viewed as a poset, is isomorphic to the Birkhoff lattice J (Π).
Proof. Vertices of OΠ are in bijection with order ideals.
Suffices to show that maximal cells of T are in bijection with maixmal
chains of order ideals.
The number of k-dimensional faces of T is exactly sk+1 (Π), the number
of surjective order preserving maps φ : Π [k + 1]. Argue via partition of
OΠ into relatively open simplices and sum up the open order polynomials.
sinces this is a unimodular triangulation, we see the f-vector.
6.3. Generating Functions and Permutation Statistics
Our next goal is to compute the integer-point transform of an order cone. A
natural starting point is Corollary 6.4, since the triangulation given in this
corollary is unimodular. As in Chapter 5, we’d like to make this triangulation
half open, so that the union in Corollary 6.4 becomes disjoint.
Again we first discuss a special case, our warm-up situation that Π = Ad
is an antichain on d elements, for which Corollary 6.4 becomes the unimodular
triangulation (6.1) of the nonnegative orthant, equivalently expressed in (6.2).
With the definition
n
o
Kσ := x ∈ Rd : xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0
this triangulation can be written in the shorthand
[
Rd≥0 =
Kσ .
σ∈Sd
Towards a half-open version of this triangulation, we use a concept familiar
from Chapter 5: we employ s := (d, d − 1, . . . , 1) to obtain visible facets of
each cone Kσ , followed by a night-and-day triangulation of the form (5.5).
For starters, Night(Rd≥0 ) = ∅. The facet Kσ ∩{xσ(j) = xσ(j+1) } is illuminated
if and only if (see Figure 3)
d + 1 − σ(j) = sσ(j) > sσ(j+1) = d + 1 − σ(j + 1) ,
i.e., σ(j + 1) > σ(j); in this case we call j an ascent of the permutation σ;
analogously, j is a descent of σ if σ(j + 1) < σ(j). So the night facets of
Kσ correspond precisely to the set Des(σ) of descents of σ, in the sense that
[
e σ := Kσ \ Night(Kσ )
K
(6.5)
xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0
d
= x∈R :
.
xσ(j) > xσ(j+1) if j ∈ Des(σ)
138
6. Partially Ordered Sets, Geometrically
Figure 3. Determining whether the facet Kσ ∩ {xσ(j) = xσ(j+1) } is illuminated.
Thus in this situation (5.5) becomes
]
Rd≥0 =
eσ
K
(6.6)
σ∈Sd
where this union is now disjoint and so, in particular, we can translate (6.6)
into an identity of rational generating functions. The integer-point transform
of the left-hand side is easy:
1
σRd (z) =
.
≥0
(1 − z1 ) (1 − z2 ) · · · (1 − zd )
And from our experience gained in Chapter 4, we effortlessly see that
X
X
eσ =
K
R>0 eσ(1) + · · · + eσ(j) +
R≥0 eσ(1) · · · + eσ(j) ,
j∈Des(σ)
j∈[d]\Des(σ)
and thus
Q
σKe σ (z) = Q
j∈Des(σ) zσ(1) zσ(2) · · · zσ(j)
j∈Des(σ)
1 − zσ(1) zσ(2) · · · zσ(j)
(6.7)
(Exercise 6.5). So the integer-transform version of (6.6) is
Q
X
1
j∈Des(σ) zσ(1) zσ(2) · · · zσ(j)
=
. (6.8)
Qd
(1 − z1 ) (1 − z2 ) · · · (1 − zd )
1
−
z
z
·
·
·
z
σ(1)
σ(2)
σ(j)
j=1
σ∈Sd
Thinking back to Section 4.7, at this point we cannot help specializing
this identity for z1 = z2 = · · · = zd = q, which gives
P
Q
j
1
σ∈Sd
j∈Des(σ) q
=
.
(6.9)
(1 − q)d
(1 − q)(1 − q 2 ) · · · (1 − q d )
This motivates the definition
maj(σ) :=
X
j∈Des(σ)
j,
6.3. Generating Functions and Permutation Statistics
139
the major index of σ. Now the terms in the numerator on the right-hand
side of (6.9) are simply q maj(σ) , and in fact, (6.9) gives a distribution of this
statistic:
X
q maj(σ) = (1 + q)(1 + q + q 2 ) · · · (1 + q + · · · + q d−1 ) .
(6.10)
σ∈Sd
With a small modification, we can see another permutation statistic
appearing. Namely, the order cone of an antichain appended by a maximal
element is
n
o
KAˆ d = x ∈ Rd+1 : x0 ≥ x1 , x2 , . . . , xd ≥ 0 .
The contemplations of the previous two pages apply almost verbatim to
this modified situation. The analogue of (6.6) is the disjoint unimodular
triangulation
] x0 ≥ xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0
d+1
KAˆ d =
x∈R
:
, (6.11)
xσ(j) > xσ(j+1) if j ∈ Des(σ)
σ∈Sd
and the analogue of (6.7), i.e., the integer-point transforms of one of the
cones on the right-hand side, is
Q
j∈Des(σ) z0 zσ(1) zσ(2) · · · zσ(j)
Qd
j=0 1 − z0 zσ(1) zσ(2) · · · zσ(j)
On the other hand, the integer-point transform of KAˆ d is, practically by
definition,
X
σAˆ d (z) =
(1 + z1 + · · · + z1n ) · · · (1 + zd + · · · + zdn ) z0n
n≥0
and so (6.11) implies
X
(1 + z1 + · · · + z1n ) · · · (1 + zd + · · · + zdn ) z0n
n≥0
(6.12)
Q
=
j∈Des(σ) z0 zσ(1) zσ(2) · · · zσ(j)
Qd
1
−
z
z
z
·
·
·
z
0
σ(1)
σ(2)
σ(j)
j=0
σ∈Sd
X
.
ˆ d , that the role of z0 is different
It is clear, already from the definition of A
from that of the other variables, and so we now specialize (6.12) to z0 = z
and z1 = z2 = · · · = zd = q:
P
Q
j
X
σ∈Sd
j∈Des(σ) zq
n d n
.
(1 + q + · · · + q ) z =
(1 − z)(1 − zq)(1 − zq 2 ) · · · (1 − zq d )
n≥0
(6.13)
140
6. Partially Ordered Sets, Geometrically
In the numerator of the right-hand side we again see the major index of the
permutation σ appearing, alongside its descent number
des(σ) := |Des(σ)| .
With the abbreviation [n + 1]q := 1 + q + · · · + q n , we can thus write (6.13)
more compactly in the following form.
P
des(σ) q maj(σ)
X
σ∈Sd z
d n
Theorem 6.11.
[n + 1]q z =
.
(1 − z)(1 − zq)(1 − zq 2 ) · · · (1 − zq d )
n≥0
6.4. Order Polynomials and P -Partitions
The computations in the previous section are surprisingly close to those for a
general finite poset Π, even though an antichain seems to be a rather special
poset. The reason for this similarity is geometric: in the general case, the
unimodular triangulation (6.2) gets replaced by the very-similarly-looking
Corollary 6.4. The next step is to make this triangulation disjoint, analogous
to (6.6). We would like to keep the handy description of the half-open cones
e σ by way of the descent set of σ, which came from our choice of the vector
K
s := (d, d − 1, . . . , 1) yielding the night-and-day triangulation (6.6). It turns
out we can use that same vector if we are careful how we name the elements
of Π. (In the case that Π is an antichain, this did not matter, since every
element looks the same.) Namely, we need to make sure that Night(KΠ ) = ∅,
that is, every facet KΠ ∩ {xj = xk }, for some j ≺· k, is illuminated by s. But
this means
d + 1 − j = sj < sk = d + 1 − k ,
that is, j > k. Thus, from now on we will assume that the elements 1, 2, . . . , d
of Π satisfy
j Π k
⇐⇒
j ≥R k .
(6.14)
In plain English: we label the elements of Π by the integers 1, 2, . . . , d in
such a way that larger elements (in terms of Π) get smaller labels (viewed as
real numbers). With this convention, we can repeat the steps leading to (6.6)
and (6.8) to obtain the following decomposition theorem for order cones and
their integer-point transforms.
Theorem 6.12. Let Π be a poset on [d] satisfying (6.14). Then
]
x
≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0
KΠ =
x ∈ Rd : σ(1)
xσ(j) > xσ(j+1) if j ∈ Des(σ)
σ
Q
X
j∈Des(σ) zσ(1) zσ(2) · · · zσ(j)
σKΠ (z) =
Qd
σ
j=1 1 − zσ(1) zσ(2) · · · zσ(j)
and
where the (disjoint) union and sum are over all σ ∈ Sd that satisfy σ(j) Π
σ(k) ⇒ j ≥R k.
6.4. Order Polynomials and P -Partitions
141
Proof. Corollary 6.4 gives the (regular unimodular) triangulation
KΠ =
[
KΓ
Γ
where the union is over all linear extensions Γ of Π. Our convention (6.14)
ensures that, using the light source s := (d, d−1, . . . , 1), we have Night(KΠ ) =
∅. We can rewrite the order cone KΓ of a linear extension Γ of Π in terms
of a permutation σ ∈ Sd as in (6.1), and our light source yields a half-open
variant of this cone as in (6.5). This gives the disjoint triangulation
KΠ =
]
x
≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0
x ∈ Rd : σ(1)
xσ(j) > xσ(j+1) if j ∈ Des(σ)
σ
where the union is over all permutations σ yielding a linear extension of
Π, i.e., if σ(j) σ(k) in Π then the same relation must hold in the linear
extension; but the latter is equivalent to j ≥ k (as real numbers).
Equations (6.11) and (6.12) also have natural analogues when Ad is
replaced by a general finite poset Π. Essentially by repeating the proofs of
(6.11) and (6.12), we obtain the following variant of Theorem 6.12.
Theorem 6.13. Let Π be a poset on [d] satisfying (6.14). Then
]
x0 ≥ xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0
d+1
KΠˆ =
x∈R
:
xσ(j) > xσ(j+1) if j ∈ Des(σ)
σ
Q
X
j∈Des(σ) z0 zσ(1) zσ(2) · · · zσ(j)
σKΠˆ (z) =
Qd
σ
j=0 1 − z0 zσ(1) zσ(2) · · · zσ(j)
and
where the (disjoint) union and sum are over all σ ∈ Sd that satisfy σ(j) Π
σ(k) ⇒ j ≥R k.
A natural next step, again parallel to those in Section 6.3, is to specialize
the variables in the integer-point transforms in Theorems 6.12 and 6.13. This
gives two important applications of these theorems; the first concerns order
polynomials. Given a poset Π on [d], consider
KΠˆ =
x∈R
d+1
0 ≤ xj ≤ x0 for all 1 ≤ j ≤ d
:
j Π k ⇒ xj ≤ xk
.
Integer points in this order cone with a fixed x0 -coordinate are precisely
order-preserving maps Π → [0, x0 ] ∩ Z. Thus
142
6. Partially Ordered Sets, Geometrically
σKΠˆ (z, 1, 1, . . . , 1) =
X
#(order-preserving maps Π → [0, n] ∩ Z) z n
n≥0
=
X
#(order-preserving maps Π → [0, n − 1] ∩ Z) z n−1
n≥1
=
1X
#(order-preserving maps Π → [1, n] ∩ Z) z n
z
n≥1
=
1X
ΩΠ (n) z n ,
z
(6.15)
n≥1
and Theorem 6.13 gives the following formula for the rational generating
function of an order polynomial.
Corollary 6.14. Let Π be a poset on [d] satisfying (6.14). Then
P des(σ)
1X
σz
ΩΠ (n) z n =
z
(1 − z)d+1
n≥1
where the sum in the numerator is over all σ ∈ Sd that satisfy σ(j) Π
σ(k) ⇒ j ≥R k.
This corollary also allows us to give yet another proof of the reciprocity
theorem of order polynomials.
Third proof of Theorem 1.12. One object of Exercise 6.9 is the integerpoint transform of K◦Πˆ , whose specialization involves the number asc(σ) of
ascents of a permutation σ:
P 2+asc(σ)
σz
◦
σK ˆ (z, 1, 1, . . . , 1) =
.
Π
(1 − z)d+1
On the other hand, a computation exactly parallel to (6.15) gives (Exercise 6.10)
X
σK◦ˆ (z, 1, 1, . . . , 1) = z
Ω◦Π (n) z n ,
(6.16)
Π
n≥0
and so we obtain
X
n≥0
Ω◦Π (n) z n
z 1+asc(σ)
.
(1 − z)d+1
P
=
σ
Now Asc(σ) ] Des(σ) = [d − 1] for any σ ∈ Sd , and so by Theorem 4.6,
P −1−asc(σ)
P des(σ)−d
d+1
X
◦
n
d z
σz
σz
Ω (−n) z = −
= (−1)
d+1
(1 − z)d+1
1 − z1
n≥1
P des(σ)+1
X
d
d
σz
= (−1)
=
(−1)
ΩΠ (n) z n .
(1 − z)d+1
n≥1
6.4. Order Polynomials and P -Partitions
143
Our second application brings back memories of Chapter 4. The theory
of P -partitions1 allows us to interpolate between compositions and partitions;
that is, here we study compositions of n that satisfy some of the inequalities
imposed on the parts of a partition. A natural way to introduce such a
subset of inequalities is through a poset.
Given a poset Π on [d], we call (a1 , a2 , . . . , ad ) ∈ Zd≥0 a Π-partition of
n if
n = a1 + a2 + · · · + ad
and
j Π k
=⇒
aj ≥ ak ,
that is, the map Π → Z≥0 given by (a1 , a2 , . . . , ad ) is order reversing.
Here are three examples of Π-partitions we are already familiar with:
• If Π is a chain with d elements, a Π-partition of n is a partition of
n in the sense of Section 4.4.
• If Π is an antichain with d elements, a Π-partition of n is a composition of n (Section 4.2).
• The plane partitions of Section 4.3 are situated between these two
extremes: in their simplest instance (4.12), they are Π-partitions
where Π = 3, depicted in Figure 1.
Let pΠ (n) be the number of Π-partition of n, with associated generating
function PΠ (q). In sync with our geometric interpretations of partitions and
compositions in Chapter 4, this generating function is an evaluation of the
integer-point transform of the order cone of q, the poset obtained from Π
by reversing each relation:
PΠ (q) = σKq (q, q, . . . , q)
This means Theorem 6.12 will have consequences for Π-partitions: its statement
Q
X
j∈Des(σ) zσ(1) zσ(2) · · · zσ(j)
σKq (z) =
,
Qd
σ
j=1 1 − zσ(1) zσ(2) · · · zσ(j)
where the sum is over all σ ∈ Sd that satisfy σ(j) Π σ(k) ⇒ j ≥R k,
translates with the specialization z = (q, q, . . . , q) to the following result.
Corollary 6.15. Let Π be a poset on [d] satisfying (6.14). Then
P maj(σ)
σq
PΠ (q) =
,
(1 − q)(1 − q 2 ) · · · (1 − q d )
where the sum is over all σ ∈ Sd that satisfy σ(j) Π σ(k) ⇒ j ≥R k.
1Here P stands for a specific poset—for which we tend to use greek letters such as Π to avoid
confusions with polyhedra.
144
6. Partially Ordered Sets, Geometrically
Since our first specialization of σK (z) already yielded a combinatorial
reciprocity theorem (that for order polynomials), it should come as no surprise
that the specialization to PΠ (q) does likewise. The reciprocal concept is a
strict Π-partition of n: a d-tuple (a1 , a2 , . . . , ad ) ∈ Zd≥0 such that
n = a1 + a2 + · · · + ad
and
j ≺Π k
=⇒
aj > ak .
We count all strict Π-partitions with the function p◦Π (n), with accompanying
generating function PΠ◦ (q).
Theorem 6.16. The rational generating functions PΠ (q) and PΠ◦ (q) are
related by
PΠ ( 1q ) = (−q)|Π| PΠ◦ (q) .
Proof. Let Π be a poset on d elements. Corollary 6.15 has an analogue for
strict Π-partitions, namely (Exercise 6.11)
P amaj σ
σq
PΠ◦ (q) =
,
(6.17)
(1 − q)(1 − q 2 ) · · · (1 − q d )
where the sum is over all σ ∈ Sd that satisfy σ(j) Π σ(k) ⇒ j ≥R k. Here
X
amaj σ :=
j,
j∈Asc σ
the ascent version of the major index of σ. Now by Corollary 6.15,
P − maj σ
1
σq
PΠ ( q ) =
(1 − q −1 )(1 − q −2 ) · · · (1 − q −d )
P
q 1+2+···+d σ q − maj σ
= (−1)d
(1 − q)(1 − q 2 ) · · · (1 − q d )
d
P
q d σ q (2)−maj σ
d
= (−1)
(1 − q)(1 − q 2 ) · · · (1 − q d )
P amaj σ
d
σq
,
= (−q)
(1 − q)(1 − q 2 ) · · · (1 − q d )
where each sum is taken over all σ ∈ Sd that satisfy σ(j) Π σ(k) ⇒ j ≥R k.
The last equation holds because
d−1
X
d
maj σ + amaj σ =
j =
.
2
j=1
The result now follows with (6.17).
6.5. Chromatic and Order Polynomials
145
6.5. Chromatic and Order Polynomials
At the end of Section 1.3, we already hinted at a relation between the
chromatic polynomial of a graph and certain order polynomials. Let’s recall
the underlying construction. An acyclic orientation ρ of a graph G can be
viewed as a poset Π(ρ G) by starting with the binary relations given by ρ G
and adding the necessary transitive and reflexive relations. Corollary 1.14
stated that the chromatic polynomial can be expressed as
X
χG (n) =
Ω◦Π(ρ G) (n)
(6.18)
ρ
where the sum is over all acyclic orientations ρ of G. With the insight gained
in this chapter, we will now prove Theorem 1.5, the combinatorial reciprocity
theorem for chromatic polynomials. We recall its statement: (−1)|V | χG (−n)
equals the number of compatible pairs (ρ, c) where c is a n-coloring and ρ is
an acyclic orientation.
We start by drawing a picture of (6.18) (the case G = K2 is depicted in
Figure 4): The order polynomials Ω◦Π(ρ G) (n) on the right-hand side of (6.18)
are, by (6.4), the Ehrhart polynomials of the interiors of the respective order
polytopes, evaluated at n + 1. Since every proper coloring gives rise to a
compatible acyclic orientation, these dilated order polytopes combined must
contain all proper n-colorings of G as lattice points. In other words,
[
(n + 1)O◦Π(ρ G)
(6.19)
ρ acyclic
equals (0, n + 1)V minus any point x with xj = xk for jk ∈ E.
Figure 4. The geometry behind (6.18) for G = K2 .
146
6. Partially Ordered Sets, Geometrically
Proof of Theorem 1.5. By (6.18) and Theorem 1.12,
X
X
ΩΠ(ρ G) (n)
(−1)|V | Ω◦Π(ρ G) (−n) =
(−1)|V | χG (−n) =
ρ
ρ
where the sums are over all acyclic orientations ρ of G. We can interpret
each order polynomial ΩΠ(ρ G) (n) on the right-hand side as counting precisely
the n-colorings of G that are compatible with ρ. (Note that, since we do not
demand that our colorings are proper, a given n-coloring might be compatible
with more than one orientation.) Summing over all acyclic ρ hence gives the
theorem.
There is a picture reciprocal to Figure 4 that underlies our proof of
Theorem 1.5. Namely, with Proposition 6.6, we may think of
X
ΩΠ(ρ G) (n)
(6.20)
ρ acyclic
as a sum of the Ehrhart polynomials of the order polytopes OΠ(ρ G) evaluated
at n − 1. As in our geometric interpretation of Corollary 1.14, we can think
of a lattice point in (n − 1)OΠ(ρ G) as an n-coloring of G (where we shifted
the color set to 0, 1, . . . , n − 1) that is compatible with ρ. Contrary to the
picture behind Corollary 1.14, here the dilated order polytopes overlap, and
so every n-coloring gets counted with multiplicity equal to the number of its
compatible acyclic orientations (see Figure 5).
Figure 5. The geometry behind our proof of Theorem 1.5 for G = K2 .
6.6. Applications: Extension Problems and Marked Order
Polytopes
Notes
P -partitions were introduced in Richard Stanley’s PhD thesis [71]...
Exercises
147
Descent statistics of permutations go further back to at least Percy
MacMahon, who proved Theorem 6.11 and numerous variations of it [52].
Back then maj σ was called the greater index of σ; in the 1970’s Dominique
Foata suggested to rename it major index in honor of MacMahon, who was
a major in the British army (and usually published his papers including this
title) [31].
Exercises
Throughout the exercises, Π is a finite poset.
6.1 Show the order cone KΠ is pointed.
6.2 Show that dim(KΠ ) = dim(OΠ ) = |Π|.
6.3 Show that the facets of the order cone KΠ are precisely the sets KΠ ∩
{xj = xk } for some j ≺· k and KΠ ∩ {xj = 0} for some minimal j ∈ Π.
6.4 Prove (6.4):
Ω◦Π (n) = ehrO◦Π (n + 1) .
6.5 Recall that, for σ ∈ Sd , we defined
n
o
Kσ = x ∈ Rd : xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0 and
x
≥
x
≥
·
·
·
≥
x
≥
0
σ(1)
σ(2)
σ(d)
d
eσ = x ∈ R :
K
.
xσ(j) > xσ(j+1) if j ∈ Des(σ)
(a) Show that {Kσ : σ ∈ Sd } is a triangulation of Rd≥0 .
(b) Show that
X
eσ =
K
R>0 eσ(1) + · · · + eσ(j)
j∈Des(σ)
+
X
R≥0 eσ(1) · · · + eσ(j)
j∈[d]\Des(σ)
e σ is unimodular and
and conclude that K
Q
j∈Des(σ) zσ(1) zσ(2) · · · zσ(j)
.
σKe σ (z) = Q
j∈Des(σ) 1 − zσ(1) zσ(2) · · · zσ(j)
6.6 Show that
Rd>0 =
] σ∈Sd
x ∈ Rd :
xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) ≥ 0
xσ(j) > xσ(j+1) if j ∈ Asc(σ)
Here Asc(σ) denotes the set of ascents of σ.
.
148
6. Partially Ordered Sets, Geometrically
6.7 Prove, without referring to Theorem 6.3, that if Π1 , Π2 , . . . , Πm are
posets on [d] with the property that for every j there is a unique d-chain
that refines Πj , then
K A d = Π 1 ∪ Π2 ∪ · · · ∪ Πm
is a regular subdivision of KAd = Rd≥0 .
6.8 Let inv(σ) := |{(j, k) : j < k and σ(j) > σ(k)}|, the number of inversions of σ ∈ Sd . Show that
X
q inv(σ) = (1 + q)(1 + q + q 2 ) · · · (1 + q + · · · + q d−1 ) .
σ∈Sd
P
P
(Looking back at (6.10), this implies σ∈Sd q maj(σ) = σ∈Sd q inv(σ) ,
i.e., the statistics maj and inv are equidistributed.)
6.9 Prove the following “open analogues” of Theorems 6.12 and 6.13: Let
Π be a poset on [d] satisfying (6.14). Then
]
xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) > 0
◦
d
KΠ =
x∈R :
xσ(j) > xσ(j+1) if j ∈ Asc(σ)
σ
Q
X z1 z2 · · · zd j∈Asc(σ) zσ(1) zσ(2) · · · zσ(j)
σK◦Π (z) =
Qd
σ
j=1 1 − zσ(1) zσ(2) · · · zσ(j)
]
x0 > xσ(1) ≥ xσ(2) ≥ · · · ≥ xσ(d) > 0
d+1
◦
x∈R
:
KΠˆ =
xσ(j) > xσ(j+1) if j ∈ Asc(σ)
σ
Q
X z02 z1 z2 · · · zd j∈Asc(σ) z0 zσ(1) zσ(2) · · · zσ(j)
σK◦ˆ (z) =
Qd
Π
σ
j=0 1 − z0 zσ(1) zσ(2) · · · zσ(j)
where the (disjoint) unions and sums are over all σ ∈ Sd that satisfy
σ(j) Π σ(j) ⇒ σ(j) Γσ σ(j).
6.10 Show (6.16):
σK◦ˆ (z, 1, 1, . . . , 1) = z
X
Π
Ω◦Π (n) z n .
n≥0
6.11 Prove (6.17): Let Π be a poset on d elements. Then
P amaj σ
◦
σz
PΠ (z) =
,
(1 − z)(1 − z 2 ) · · · (1 − z d )
where the sum is over all σ ∈ Sd that satisfy σ(j) Π σ(k) ⇒ j ≥R k.
6.12 Recompute the generating function for plane partition diamonds from
Exercise 4.16 in Chapter 4 by way of Π-partitions.
6.13 Derive from Theorem 6.16 a reciprocity theorem that relates the
quasipolynomials pΠ (t) and p◦Π (t).
Exercises
149
6.14 Show that the generating function
X
ΩΠ (n)z n−1 =
n≥1
hΠ (z)
(1 − z)d+1
has a palindromic numerator polynomial hΠ if and only if Π is graded,
i.e., there is a function rank : Π → Z≥0 so that for every cover relation
j ≺ · k we have rank(j) + 1 = rank(k).
Chapter 7
Hyperplane
Arrangements
The purely formal language of geometry describes adequately the reality of space.
We might say, in this sense, that geometry is successful magic. I should like to state
a converse: is not all magic, to the extent that it is successful, geometry?
Rene Thom
[under construction]
We start this chapter by picking up the thread from Section 6.5, in which we
proved the reciprocity theorem for chromatic polynomials (Theorem 1.5) via
order polynomials stemming from the acyclic orientations of a given graph G.
Our starting point was the geometric realization that
[
O◦Π(ρ G) = (0, 1)V \ x ∈ RV : xj = xk for jk ∈ E .
(7.1)
ρ acyclic
(This is the un-dilated polyhedral version of (6.19). Dilating either side
of this equation by n + 1 and then intersecting with ZV yields precisely
the proper n-colorings of G.) While we read this equation from left to
right in Section 6.5, we now start on the right-hand side of (7.1): what we
subtract here from the open unit cube in RV is the union of the hyperplane
arrangement
HG := {xj = xk : jk ∈ E} ,
the graphical arrangement corresponding to G (see Figure 1 for an example where G = K2 , the graph with exactly two adjacent nodes). Thus each
151
152
7. Hyperplane Arrangements
x2
n+1
x1 = x2
6
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
s
- x1
n+1
Figure 1. The integer points n-color the graph K2 , with n = 6.
proper n-coloring of G corresponds to a lattice point in
[
(n + 1)P◦ \ HG ∩ ZV ,
(7.2)
where P := [0, 1]V , and we can compute the chromatic polynomial of G
geometrically as
[
χG (n) = (n + 1)P◦ \ HG ∩ ZV (7.3)
◦ [
1
V
Z ,
= P \
HG ∩
n+1
reminiscent of Ehrhart theory, where we count lattice points as a polytope P
gets dilated or, alternatively, the integer lattice gets shrunk.
7.1. Inside-out Polytopes
Our next goal is to study (7.3) for a general rational polytope P ⊂ Rd and a
general rational1 hyperplane arrangement H in Rd . Let
[ 1 IP,H (n) := P \ H ∩ Zd ,
n
in words: IP,H (n) counts those points in n1 Zd that are in the polytope P
but off the hyperplanes in H. In the absence of H, this is just the Ehrhart
quasipolynomial of P counting lattice points as we dilate P or, equivalently,
shrink the integer lattice Zd . We refer to (P, H) as an inside-out polytope
because we think of the hyperplanes in H as additional boundary inside P.
1 A hyperplane arrangement is rational if all of its hyperplanes can be described by linear
equalities with rational (or, equivalently, integer) coefficients.
7.1. Inside-out Polytopes
153
Towards computing IP,H (n), we recall from Chapter 3 that a flat of
the hyperplane arrangement H is a nonempty intersection of some of the
hyperplanes
S in H, and a region of H is the closure of a connected component
of Rd \ H. A face of any of the regions is called a face of H . (So the
closed regions of H are the maximal faces). Given a flat F of a hyperplane
arrangement H, we can create the hyperplane arrangement induced by H
on F , namely,
HF := {H ∩ F : H ∈ H, H ∩ F 6= ∅} .
In Chapter 3 we defined the characteristic polynomial of H as
X
χH (n) =
µ(Rd , F ) ndim F .
F ∈L(H)
(Actually, we defined, more generally, characteristic polynomials of graded
posets, which specializes to the above for the lattice of flats of H.) It turns
out the counting function IP,H (n) can also be computed through the M¨obius
function µ(Rd , F ) if we know the Ehrhart quasipolynomial of P ∩ F for each
flat F ∈ L(H):
Theorem 7.1. Suppose P ⊂ Rd is a d-polytope and H is a hyperplane
arrangement in Rd . Then
X
IP,H (n) =
µ(Rd , F ) ehrP∩F (n) .
F ∈L(H)
Proof. Given a flat F ∈ L(H), we can compute ehrP∩F (n) by counting the
lattice points in the faces of HF in the following way: each lattice point is in
the interior of a unique minimal face f of HF , and f is a region of HG for
some flat G ⊆ F (more precisely, G is the affine span of f ). Thus our lattice
point is one of the points counted by IP∩G,HG (n), and grouping the lattice
points in P ∩ F according to this scheme yields
X
ehrP∩F (n) =
IP∩G,HG (n) .
G⊆F
By M¨
obius inversion (Theorem 2.12),
X
IP∩F,HF (n) =
µ(F, G) ehrP∩G (n) ,
G⊆F
and so in particular for F =
IP,H (n) =
Rd ,
X
µ(Rd , G) ehrP∩G (n) .
G∈L(H)
If P and H are rational, P ∩ F is a rational polytope for each flat F , and
so by Theorem ?? we conclude effortlessly:
154
7. Hyperplane Arrangements
Corollary 7.2. If P ⊂ Rd is a rational polytope and H is a rational hyperplane arrangement in Rd , then IP,H (n) is a quasipolynomial in n.
You have undoubtedly noticed the similarities shared by the definition
of the characteristic polynomial χH (n) and the formula for IP,H (n) given
in Theorem 7.1. In fact, the two functions are intimately connected, as the
following first application shows.
Corollary 7.3. Let (P, H) be an inside-out polytope for which there exists a
function φ(n) such that ehrP∩F (n) = φ(n)dim F for any flat F ∈ L(H). Then
IP,H (n) = χH (φ(n)) .
Proof. Let (P, H) be an inside-out polytope satisfying the conditions of
Corollary 7.3. Then, by Theorem 7.1,
X
X
IP,H (n) =
µ(Rd , F ) ehrP∩F (n) =
µ(Rd , F ) φ(n)dim F
F ∈L(H)
F ∈L(H)
= χH (φ(n)) .
Corollary 7.3 allows us to compute the characteristic polynomial of H by
counting lattice points. Let’s work through some examples.
Corollary 7.4. Let H = {xj = 0 : 1 ≤ j ≤ d}, the Boolean arrangement
in Rd . Then χH (n) = (n − 1)d .
Proof. Let P be the d-dimensional unit cube [0, 1]d . Then IP,H (n) = nd ,
S
since (nP \ H) ∩ Zd contains precisely those lattice points in nP that have
nonzero coordinates. On the other hand, a k-dimensional flat F ∈ L(H) is
defined by d − k equations of the form xj = 0, and so
ehrP∩F (n) = (n + 1)k .
Thus φ(n) = n + 1 in Corollary 7.3 gives χH (n + 1) = IP,H (n) = nd .
Note that, with Zaslavksy’s Theorem 3.23, Corollary 7.4 recovers Exercise 3.40, namely that the Boolean arrangement in Rd has 2d regions.
Corollary 7.5. Let H = {xj = xk : 1 ≤ j < k ≤ d}, the real braid arrangement in Rd . Then
χH (n) = n(n − 1)(n − 2) · · · (n − d + 1) .
Proof. Again let P be the d-dimensional
unit cube [0, 1]d . We can pick a
S
point (x1 , x2 , . . . , xd ) ∈ (nP \ H) ∩ Zd by first choosing x1 (for which there
are n + 1 choices), then choosing x2 6= x1 (for which there are n choices),
etc., down to choosing xd =
6 x1 , x2 , . . . , xd−1 (for which there are n − d + 2
choices), and so
IP,H (n) = (n + 1) n (n − 1) · · · (n − d + 2) .
7.2. Graph Colorings and Acyclic Orientations
155
On the other hand, a k-dimensional flat F ∈ L(H) is defined by d − k
equations of the form xi = xj , and so again
ehrP∩F (n) = (n + 1)k .
Thus we take φ(n) = n + 1 in Corollary 7.3, whence
χH (n + 1) = IP,H (n) = (n + 1) n (n − 1) · · · (n − d + 2) .
Again we can use Zaslavksy’s Theorem 3.23, with which Corollary 7.5
recovers Exercise 3.41.
Corollary 7.6. Let H = {xj = ±xk , xj = 0 : 1 ≤ j < k ≤ d}. Then
χH (n) = (n − 1)(n − 3) · · · (n − 2d + 1) .
the cube [−1, 1]d . We can pick a point (x1 , x2 , . . . , xd ) ∈
Proof.
S Let P be
6 0 (for which there are 2n choices),
(nP \ H) ∩ Zd by first choosing x1 =
then choosing x2 6= ±x1 , 0 (for which there are 2n − 2 choices), etc., down
to choosing xd 6= ±x1 , ±x2 , . . . , ±xd−1 , 0 (for which there are 2n − 2d + 2
choices), and so
IP,H (n) = 2n(2n − 2) · · · (2n − 2d + 2) .
On the other hand, a k-dimensional flat F ∈ L(H) is defined by d − k
equations of the form xi = ±xj or xj = 0, and so ehrP∩F (n) = (2n + 1)k .
Thus we take φ(n) = 2n + 1 in Corollary 7.3, whence
χH (2n + 1) = IP,H (n) = 2n(2n − 2) · · · (2n − 2d + 2) .
The hyperplane arrangements in Corollaries 7.5 and 7.6 are examples
of Coxeter arrangements: their hyperplanes correspond to root systems
or, more generally, finite reflection groups. The real braid arrangements in
Corollary 7.5 correspond to root systems of type A, the arrangements in
Corollary 7.6 to root systems of type B and C, and the arrangements in
Exercise 7.3 correspond to root systems of type D.
7.2. Graph Colorings and Acyclic Orientations
It’s time to return to colorings of a graph G = (V, E). As in the beginning of
this chapter, let P = [0, 1]V , the unit cube in RV . Just like in the case of real
braid arrangements (see the proof of Corollary 7.5), any k-dimensional flat
F of the graphical arrangement HG will give rise to the Ehrhart polynomial
ehrP∩F (n) = (n + 1)k , and with Corollary 7.3 we obtain
IP,HG (n) = χHG (n + 1) .
We can play the same game with the open unit cube P◦ = (0, 1)V : now if F
is a k-dimensional flat of HG then ehrP◦ ∩F (n) = (n − 1)k , and so
IP◦ ,HG (n) = χHG (n − 1) .
(7.4)
156
7. Hyperplane Arrangements
Now we recall the left-hand side of (7.1), a union of order polytopes
corresponding to the acyclic orientations of G. We can see these orientations
just as well in our hyperplane picture: Each region of the graphical arrangement HG is determined by inequalities of the form xi > xj or xi < xj , one
for each edge ij ∈ E. Each of these inequalities, in turn, can be thought of
as orienting the edge: if xi > xj then orient ij from j to i, and if xi < xj
then orient ij from i to j. In this way, each region of HG gives rise to
an orientation of G; see Figure 2 for an illustration in the case G = K3 .
The orientations that we can associate with the regions of HG are precisely
Figure 2. The regions of HK3 (projected to the plane x1 + x2 + x3 = 0)
and their corresponding acyclic orientations.
the acyclic ones—a directed circle would correspond to a sequence of the
nonsensical inequalities xi1 > xi2 > · · · > xik > xi1 . Let’s summarize:
Lemma 7.7. The regions of HG are in one-to-one correspondence with the
acyclic orientations of G. In particular, (−1)|V | χHG (−1) equals the number
of acyclic orientations of G.
Here the second statement follows from Zaslavksy’s Theorem 3.23.
We started this chapter by realizing colorings as integer points in an
inside-out polytope; in this language, (7.2) says that
χG (n) = IP◦ ,HG (n + 1)
[0, 1]V
RV .
(7.5)
where P =
is the unit cube in
This observation has some
immediate consequences. For example, we can combine it with (7.4) to
conclude that the coloring counting function χG (n) equals the characteristic
polynomial of the graphical hyperplane arrangement HG :
Corollary 7.8. For any graph G we have χG (n) = χHG (n) .
7.2. Graph Colorings and Acyclic Orientations
157
In particular, this reconfirms that χG (n) is a polynomial, as we have known
since Section 1.1.2 There are a few more immediate consequences:
Corollary 7.9. The chromatic polynomial χG (n) of G is monic, has degree
|V | and constant term 0. Its evaluation (−1)|V | χG (−1) equals the number
of acyclic orientations of G.
Proof. Since the unit cube P = [0, 1]V has volume one, the polynomial
χG (n) = IP◦ ,HG (n + 1) is monic, of degree |V |, and has constant term
χG (0) = IP◦ ,HG (1) = 0. The final statement in Corollary 7.9 follows from
Lemma 7.7 whose last part can now be restated as: (−1)|V | χG (−1) equals
the number of acyclic orientations of G.
We emphasize that Lemma 7.7 seemingly effortlessly gave us an instance
of the reciprocity theorem for chromatic polynomials (Theorem 1.5). In fact,
we can give a second proof of this reciprocity theorem with the machinery
developed in this chapter. We’ll do so for the more general class of insideout Ehrhart quasipolynomials and specialize to chromatic polynomials soon
thereafter.
Let P ⊂ Rd be a rational S
polytope, and H be a rational hyperplane
arrangement in Rd . Then P◦ \ H is the union of the (relative) interiors of
some rational polytopes, each of dimension dim(P):
[
P◦ \ H = Q◦1 ∪ Q◦2 ∪ · · · ∪ Q◦k .
(For the graphic case, this is (7.1).) Thus IP◦ ,H (n) =
Ehrhart–Macdonald reciprocity (Theorem 5.22) gives
IP◦ ,H (−n) = (−1)dim(P)
k
X
Pk
ehrQj (n) .
j=1 ehrQj (n),
◦
and
(7.6)
j=1
The sum on the right can be interpreted purely in terms of the inside-out
P
polytope (P, H). Namely, each point in n1 Zd that is counted by kj=1 ehrQj (n)
lies in P, and it gets counted with multiplicity equal to the number of closed
regions of (P, H) that contain it. Here, by analogy with the hyperplane
arrangement terminology, a S
(closed) region of (P, H) is (the closure of) a
connected component of P \ H. (In the graphic case, (7.6) is (6.20), up to
a minus sign.)
2 There are much easier ways to prove that χ (n) is a polynomial; one was given in Section
G
1.1. On the other hand, Corollary 7.8 is less trivial; all proofs we are aware of are at least as
complex as ours.
158
7. Hyperplane Arrangements
These observations motivate the following definitions. The multiplicity
of x ∈ Rd with respect to (P, H) is3
multP,H (x) := # closed regions of (P, H) that contain x .
Note that this definition implies multP,H (x) = 0 if x ∈
/ P. Thus
X
OP,H (n) :=
multP,H (x)
1 d
x∈ n
Z
equals the sum
theorem.
Pk
j=1 ehrQj (n)
in (7.6), which gives the following reciprocity
Theorem 7.10. Suppose P ⊂ Rd is a rational polytope, and H is a rational
hyperplane arrangement in Rd . Then OP,H (n) and IP◦ ,H (n) are quasipolynomials that satisfy
IP◦ ,H (−n) = (−1)dim P OP,H (n) .
We can say more: First, the periods of the quasipolynomials OP,H (n) and
IP◦ ,H (n) divide the least common multiple of the denominators of the vertex
coordinates of all of the regions of (P, H). Second, the leading coefficient of
both OP,H (n) and IP◦ ,H (n) equals the volume of P.4 Finally, since ehrQj (0) =
1 (by Theorem ??), we conclude:
Corollary 7.11. The constant term OP,H (0) equals the number of regions
of (P, H).
Now we apply the inside-out machinery to graph coloring.
Second proof of Theorem 1.5. Recall that (7.5) stated that χG (n) =
IP◦ ,HG (n + 1), where P = [0, 1]V . With Theorem 7.10 we thus obtain
(−1)|V | χG (−n) = OP,HG (n − 1) .
In the absence of the hyperplane arrangement HG , the function on the right
counts the integer lattice points in the cube [0, n − 1]V ; each such lattice
point can naturally be interpreted as an n-coloring. Now taking HG into
account, OP,HG (n − 1) counts each lattice point with multiplicity equal to
the number of closed regions the point lies in. But Lemma 7.7 asserts that
these regions correspond exactly to the acyclic orientations of G. So if we
think of a lattice point in [0, n − 1]V as an n-coloring, the region multiplicity
gives the number of compatible acyclic orientations. Because OP,HG (n − 1)
takes these multiplicities into account, the theorem follows.
3 In the definition of mult
P,H (x), we can replace the phrase “closed regions of (P, H)” by
“closed regions of H” if P and H are transversal, i.e., if every flat of L(H) that intersects P also
intersects P◦ .
4 If P is not full dimensional, volume hear means relative volume, as in Exercise 4.40.
Exercises
159
7.3. Graph Flows and Totally Cyclic Orientations
Notes
Inside-out polytopes were introduced in [13], initially motivated by Figure 1,
(an isomorphic copy of) which appeared in [17]. Corollary 7.3 is the Euclidean
analogue of the finite-field method [7] (see Exercise 7.4 for details); in fact,
the proof of Corollary 7.3 is essentially that of [7].
Lemma 7.7 is due to Curtis Greene [33, 34]. The proof of Theorem 1.5
via inside-out polytopes appeared (together with Theorem 7.10) in [13].
Exercises
7.1 Prove that all the coefficients of OP,HG (n−1) are nonnegative. Conclude
that, for any graph G, the coefficients of χG (n) alternate in sign.
7.2 Let H be an arrangement in Rd consisting of k hyperplanes in general
position. Prove that
k d
k
k d−2
d−1
d k
χH (n) =
n −
n
+
n
− · · · + (−1)
.
0
1
2
d
(Note that this recovers Exercise 3.42, by Zaslavksy’s Theorem 3.23.)
7.3 Let H = {xj = ±xk : 1 ≤ j < k ≤ d}. Then5
χH (n) = (n − 1)(n − 3) · · · (n − 2d + 5)(n − 2d + 3)(n − d + 1) .
7.4 This exercise gives an alternative to Corollary 7.3 for computing characteristic polynomials. Given a rational hyperplane arrangement H in
Rd , let q be a prime power such that H, when viewed in Fdq , yields the
same semilattice of flats as H viewed in Rd . Prove that
[ χH (q) = Fdq \ H .
Use this to re-compute some of the characteristic polynomials above.
5The last factor of h (n) is not a typographical error.
H
Bibliography
1. The On-Line Encyclopedia of Integer Sequences,
http://oeis.org, 2014.
published electronically at
2. George E. Andrews, The Theory of Partitions, Cambridge Mathematical Library,
Cambridge University Press, Cambridge, 1998, reprint of the 1976 original.
3. George E. Andrews and Kimmo Eriksson, Integer Partitions, Cambridge University
Press, Cambridge, 2004.
4. George E. Andrews, Peter Paule, and Axel Riese, MacMahon’s partition analysis VIII.
Plane partition diamonds, Adv. in Appl. Math. 27 (2001), no. 2-3, 231–242.
5. K. Appel and W. Haken, Every planar map is four colorable. I. Discharging, Illinois
J. Math. 21 (1977), no. 3, 429–490. MR 0543792 (58 #27598a)
6. K. Appel, W. Haken, and J. Koch, Every planar map is four colorable. II. Reducibility,
Illinois J. Math. 21 (1977), no. 3, 491–567. MR 0543793 (58 #27598b)
7. Christos A. Athanasiadis, Characteristic polynomials of subspace arrangements and
finite fields, Adv. Math. 122 (1996), no. 2, 193–233.
8. Peter Barlow, An Elementary Investigation of the Theory of Numbers, J. Johnson &
Co., London, 1811.
9. Alexander Barvinok, Integer points in polyhedra, Zurich Lectures in Advanced Mathematics, European Mathematical Society (EMS), Z¨
urich, 2008.
10. Victor V. Batyrev, Dual polyhedra and mirror symmetry for Calabi-Yau hypersurfaces in toric varieties, J. Algebraic Geom. 3 (1994), no. 3, 493–535,
arXiv:alg-geom/9310003.
11. Matthias Beck, Jes´
us A. De Loera, Mike Develin, Julian Pfeifle, and Richard P.
Stanley, Coefficients and roots of Ehrhart polynomials, Integer Points in Polyhedra—
Geometry, Number Theory, Algebra, Optimization, Contemp. Math., vol. 374, Amer.
Math. Soc., Providence, RI, pp. 15–36, arXiv:math.CO/0402148.
12. Matthias Beck and Sinai Robins, Computing the Continuous Discretely: Integer-point
Enumeration in Polyhedra, Undergraduate Texts in Mathematics, Springer, New York,
2007, electronically available at http://math.sfsu.edu/beck/ccd.html.
13. Matthias Beck and Thomas Zaslavsky, Inside-out polytopes, Adv. Math. 205 (2006),
no. 1, 134–162, arXiv:math.CO/0309330.
161
162
Bibliography
14. Dale Beihoffer, Jemimah Hendry, Albert Nijenhuis, and Stan Wagon, Faster algorithms for Frobenius numbers, Electron. J. Combin. 12 (2005), no. 1, Research Paper
27, 38 pp.
15. Garrett Birkhoff, Lattice theory, third ed., American Mathematical Society Colloquium Publications, vol. 25, American Mathematical Society, Providence, R.I., 1979.
16. George D. Birkhoff, A determinant formula for the number of ways of coloring a map,
Ann. of Math. (2) 14 (1912/13), no. 1-4, 42–46.
17. Andreas Blass and Bruce E. Sagan, Characteristic and Ehrhart polynomials, J. Algebraic Combin. 7 (1998), no. 2, 115–126.
18. Felix Breuer and Raman Sanyal, Ehrhart theory, modular flow reciprocity, and the
Tutte polynomial, Math. Z. 270 (2012), no. 1-2, 1–18, arXiv:0907.0845.
19. Charles J. Brianchon, Th´eor`eme nouveau sur les poly`edres, J. Ecole (Royale) Polytechnique 15 (1837), 317–319.
20. Heinz Bruggesser and Peter Mani, Shellable decompositions of cells and spheres, Math.
Scand. 29 (1971), 197–205 (1972).
21. Winfried Bruns and Joseph Gubeladze, Polytopes, rings, and K-theory, Springer
Monographs in Mathematics, Springer, Dordrecht, 2009.
22. Arthur Cayley, The Collected Mathematical Papers. Volume 10, Cambridge Library
Collection, Cambridge University Press, Cambridge, 2009, reprint of the 1896 original.
23. Jes´
us A. De Loera, J¨
org Rambau, and Francisco Santos, Triangulations, Algorithms
and Computation in Mathematics, vol. 25, Springer-Verlag, Berlin, 2010.
24. Max Dehn, Die Eulersche Formel im Zusammenhang mit dem Inhalt in der nichteuklidischen Geometrie, Math. Ann. 61 (1905), 279–298.
25. Graham Denham, Short generating functions for some semigroup algebras, Electron.
J. Combin. 10 (2003), Research Paper 36, 7 pp.
26. Eug`ene Ehrhart, Sur les poly`edres rationnels homoth´etiques a
` n dimensions, C. R.
Acad. Sci. Paris 254 (1962), 616–618.
27.
, Sur la partition des nombres, C. R. Acad. Sci. Paris 259 (1964), 3151–3153.
28. Leonhard Euler, Demonstatio nonnullarum insignium proprietatum, quibus solida
hedris planis inclusa sunt praedita, Novi Comm. Acad. Sci. Imp. Petropol. 4 (1752/53),
140–160.
29.
, Elementa doctrinae solidorum, Novi Comm. Acad. Sci. Imp. Petropol. 4
(1752/53), 109–140.
30. William Feller, An Introduction to Probability Theory and Its Applications. Vol. I,
Third Edition, John Wiley & Sons Inc., New York, N.Y., 1968.
31. Dominique Foata, Distributions eul´eriennes et mahoniennes sur le groupe des permutations, Higher Combinatorics (M. Aigner, ed.), NATO Adv. Study Inst. Ser., Ser. C:
Math. Phys. Sci., vol. 31, Reidel, Dordrecht, 1977, With a comment by Richard P.
Stanley, pp. 27–49.
32. Jørgen P. Gram, Om rumvinklerne i et polyeder, Tidsskrift for Math. (Copenhagen)
4 (1874), no. 3, 161–163.
33. Curtis Greene, Acyclic orientations, Higher Combinatorics (M. Aigner, ed.), NATO
Adv. Study Inst. Ser., Ser. C: Math. Phys. Sci., vol. 31, Reidel, Dordrecht, 1977,
pp. 65–68.
34. Curtis Greene and Thomas Zaslavsky, On the interpretation of Whitney numbers
through arrangements of hyperplanes, zonotopes, non-Radon partitions, and orientations of graphs, Trans. Amer. Math. Soc. 280 (1983), no. 1, 97–126.
Bibliography
163
35. Branko Gr¨
unbaum, Convex Polytopes, second ed., Graduate Texts in Mathematics,
vol. 221, Springer-Verlag, New York, 2003, Prepared and with a preface by Volker
Kaibel, Victor Klee, and G¨
unter M. Ziegler.
36. Philip Hall, The Eulerian functions of a group, Quart. J. Math. 7 (1936), 134–151.
37. Allen Hatcher, Algebraic topology, Cambridge University Press, Cambridge, 2002.
38. Silvia Heubach and Toufik Mansour, Combinatorics of Compositions and Words, Discrete Mathematics and its Applications (Boca Raton), CRC Press, Boca Raton, FL,
2010.
39. Takayuki Hibi, Algebraic Combinatorics on Convex Polytopes, Carslaw, 1992.
40.
, Dual polytopes of rational convex polytopes, Combinatorica 12 (1992), no. 2,
237–240.
41. Verner E. Hoggatt, Jr. and D. A. Lind, Fibonacci and binomial properties of weighted
compositions, J. Combinatorial Theory 4 (1968), 121–124.
42. June Huh, Milnor numbers of projective hypersurfaces and the chromatic polynomial
of graphs, J. Amer. Math. Soc. 25 (2012), no. 3, 907–927.
43. June Huh and Eric Katz, Log-concavity of characteristic polynomials and the Bergman
fan of matroids, Math. Ann. 354 (2012), no. 3, 1103–1116.
44. Fran¸cois Jaeger, Flows and generalized coloring theorems in graphs, J. Combin. Theory
Ser. B 26 (1979), no. 2, 205–216.
45. Ravi Kannan, Lattice translates of a polytope and the Frobenius problem, Combinatorica 12 (1992), no. 2, 161–177.
46. Victor Klee, A combinatorial analogue of Poincar´e’s duality theorem, Canad. J. Math.
16 (1964), 517–531.
47. Maximilian Kreuzer and Harald Skarke, Classification of reflexive polyhedra in three dimensions, Adv. Theor. Math. Phys. 2 (1998), no. 4, 853–871, arXiv:hep-th/9805190.
48.
, Complete classification of reflexive polyhedra in four dimensions, Adv. Theor.
Math. Phys. 4 (2000), no. 6, 1209–1230, arXiv:hep-th/0002240.
49. Jim Lawrence, Valuations and polarity, Discrete Comput. Geom. 3 (1988), no. 4, 307–
324.
50. L´
aszl´
o Lov´
asz, Combinatorial Problems and Exercises, second ed., North-Holland Publishing Co., Amsterdam, 1993.
51. Ian G. Macdonald, Polynomials associated with finite cell-complexes, J. London Math.
Soc. (2) 4 (1971), 181–192.
52. Percy A. MacMahon, Combinatory Analysis, Chelsea Publishing Co., New York, 1960,
reprint of the 1915 original.
53. Peter McMullen, Lattice invariant valuations on rational polytopes, Arch. Math.
(Basel) 31 (1978/79), no. 5, 509–516.
54. Leo Moser and E. L. Whitney, Weighted compositions, Canad. Math. Bull. 4 (1961),
39–43.
55. Peter Orlik and Hiroaki Terao, Arrangements of hyperplanes, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol.
300, Springer-Verlag, Berlin, 1992.
56. Micha A. Perles and Geoffrey C. Shephard, Angle sums of convex polytopes, Math.
Scand. 21 (1967), 199–218.
57. Georg Alexander Pick, Geometrisches zur Zahlenlehre, Sitzenber. Lotos (Prague) 19
(1899), 311–319.
164
Bibliography
58. Henri Poincar´e, Sur la g´en´eralisation d’un theorem d’Euler relatif aux poly`edres, C. R.
Acad. Sci. Paris (1893), 144–145.
59. Jorge L. Ram´ırez Alfons´ın, The Diophantine Frobenius Problem, Oxford Lecture Series
in Mathematics and its Applications, vol. 30, Oxford University Press, Oxford, 2005.
60. Ronald C. Read, An introduction to chromatic polynomials, J. Combinatorial Theory
4 (1968), 52–71.
61. Neville Robbins, On Tribonacci numbers and 3-regular compositions, Fibonacci
Quaterly 52 (2014), 16–19.
62.
, On r-regular compositions, Journal for Combinatorial Mathematics and Combinatorial Computing, to appear.
63. Gian-Carlo Rota, On the foundations of combinatorial theory. I. Theory of M¨
obius
functions, Z. Wahrscheinlichkeitstheorie und Verw. Gebiete 2 (1964), 340–368.
64. Steven V Sam, A bijective proof for a theorem of Ehrhart, Amer. Math. Monthly 116
(2009), no. 8, 688–701, arXiv:0801.4432v5.
65. Ludwig Schl¨
afli, Theorie der vielfachen Kontinuit¨
at, Ludwig Schl¨
afli, 1814–1895,
Gesammelte Mathematische Abhandlungen, Vol. I, Birkh¨
auser, Basel, 1950, pp. 167–
387.
66. Paul D. Seymour, Nowhere-zero 6-flows, J. Combin. Theory Ser. B 30 (1981), no. 2,
130–135.
67. Geoffrey C. Shephard, An elementary proof of Gram’s theorem for convex polytopes,
Canad. J. Math. 19 (1967), 1214–1217.
68. Andrew V. Sills, Compositions, partitions, and Fibonacci numbers, Fibonacci Quart.
49 (2011), no. 4, 348–354.
69. Duncan M. Y. Sommerville, The relation connecting the angle-sums and volume of
a polytope in space of n dimensions, Proc. Roy. Soc. London, Ser. A 115 (1927),
103–119.
70. Eugene Spiegel and Christopher J. O’Donnell, Incidence algebras, Monographs and
Textbooks in Pure and Applied Mathematics, vol. 206, Marcel Dekker Inc., New York,
1997.
71. Richard P. Stanley, Ordered structures and partitions, American Mathematical Society,
Providence, R.I., 1972, Memoirs of the American Mathematical Society, No. 119.
72.
, Acyclic orientations of graphs, Discrete Math. 5 (1973), 171–178.
73.
, Combinatorial reciprocity theorems, Advances in Math. 14 (1974), 194–253.
, Decompositions of rational convex polytopes, Ann. Discrete Math. 6 (1980),
74.
333–342.
75.
, Some aspects of groups acting on finite posets, J. Combin. Theory Ser. A 32
(1982), no. 2, 132–161.
76.
, Two poset polytopes, Discrete Comput. Geom. 1 (1986), no. 1, 9–23.
77.
, An introduction to hyperplane arrangements, Geometric combinatorics,
IAS/Park City Math. Ser., vol. 13, Amer. Math. Soc., Providence, RI, 2007, pp. 389–
496.
78.
, Enumerative Combinatorics. Volume 1, second ed., Cambridge Studies in
Advanced Mathematics, vol. 49, Cambridge University Press, Cambridge, 2012.
79. Alan Stapledon, Additive number theory and inequalities in Ehrhart theory, Preprint
(arXiv:0904.3035v2).
80.
, Inequalities and Ehrhart δ-vectors, Trans. Amer. Math. Soc. 361 (2009),
no. 10, 5615–5626, arXiv:math/0801.0873.
Bibliography
165
81. Ernst Steinitz, Polyeder und Raumeinteilungen, Encyclop¨
adie der mathematischen
Wissenschaften, Band 3 (Geometrie), Teil 3AB12 (1922), 1–139.
82. William T. Tutte, A ring in graph theory, Proc. Cambridge Philos. Soc. 43 (1947),
26–40.
83. Hassler Whitney, A logical expansion in mathematics, Bull. Amer. Math. Soc. 38
(1932), no. 8, 572–579.
84. Herbert S. Wilf, Which polynomials are chromatic?, Colloquio Internazionale sulle
Teorie Combinatorie (Roma, 1973), Tomo I, Accad. Naz. Lincei, Rome, 1976, pp. 247–
256.
85.
, generatingfunctionology, second ed., Academic Press Inc., Boston, MA, 1994,
electronically available at http://www.math.upenn.edu/∼wilf/DownldGF.html.
86. Thomas Zaslavsky, Facing up to arrangements: face-count formulas for partitions of
space by hyperplanes, Mem. Amer. Math. Soc. 154 (1975).
87. Doron
Zeilberger,
The
composition
enumeration
reciprocity
theorem,
Personal Journal of Shalosh B. Ekhad and Doron Zeilberger (2012),
http://www.math.rutgers.edu/∼zeilberg/mamarim/mamarimhtml/comp.html.
88. G¨
unter M. Ziegler, Lectures on Polytopes, Springer-Verlag, New York, 1995.
Index
P -partition, 143
acyclic, 4
affine hull, 41
affine hyperplane, 40
Andrews, George, 104
anti-symmetric, 9
antichain, 11, 26, 31, 143
Appel, Kenneth, 2, 18
arrangement of hyperplanes, 48, 63
Barlow, Peter, 104
Batyrev, Victor, 127
Bell, Eric Temple, 35
beneath, 66
Bernoulli number, 60
Bernoulli polynomial, 60
beyond, 66
Binomial Theorem, 76
binomial theorem, 36
Birkhoff lattice, 26
Birkhoff’s theorem, 33
Birkhoff, Garrett, 35
Birkhoff, George, 2, 18
Boolean arrangement, 73
characteristic polynomial of, 154
Boolean lattice, 30
braid arrangement
characteristic polynomial of, 154
Breuer, Felix, 18
Brianchon, Charles Julien, 68
Brianchon–Gram relation, 66
Brianchon–Gram theorem, 124
bridge, 6
Bruggesser, Heinz, 68
c, 9
Cayley, Arthur, 103
chain, 10, 143
characteristic polynomial
of a graded poset, 63
of a hyperplane arrangement, 64
chromatic polynomial, 3, 152
coin-exchange problem, 104
color gradient, 4
coloring, 1
combinatorial reciprocity theorem, xii
for binomial coefficients, xii
for chromatic polynomials, 5
for flow polynomials, 9
for lattice polygons, 14
for order polynomials, 12, 31
for zeta polynomials of Eulerian
posets, 34
for zeta polynomials of finite
distributive lattices, 33
complete bipartite graph, 21
complete graph, 20
composition, 82
part of, 82
cone, 43
generator of, 44
pointed, 43
rational, 44
reciprocity theorem for integer-point
transform of, 128
167
168
unimodular, 93
conical hull, 44
connected component, 6
constituent, 89
contraction, 3
convex, 40
convex cone, 43
convex hull, 43
convex lattice polygon, 13
convex polytope, 13
convolution, 88, 107
cover relation, 10
Coxeter arrangements, 155
cross polytope, 128
cycle, 20
d, 4
Dedekind, Richard, 35
degree, 89
Dehn, Max, 68
Dehn–Sommerville relations, 61
deletion, 3
difference operator, 77
dilate, 13
dimension, 42
of a polyhedron, 42
distributive lattice, 33
dual, 74
dual graph, 7
dual order ideal, 26
edge, 45
boundary, 15
contraction of, 3
deletion of, 3
interior, 15
of a graph, 1
of a polygon, 13
Ehrhart function, 14, 91
Ehrhart polynomial, 17, 93, 110
reciprocity theorem for, 125
Ehrhart quasipolynomial, 123
Ehrhart series, 91
Ehrhart’s theorem, 94, 122
Ehrhart, Eug`ene, 104, 127
Ehrhart–Macdonald reciprocity, 125,
157
Euler characteristic, 51
Euler, Leonhard, 68, 86
Euler–Poincar´e formula, 52
Eulerian poset, 34, 59
Index
f-polynomial
of a triangulation, 126
f -vector, 46
face, 14, 45
boundary, 15
interior, 15
of a hyperplane arrangement, 153
proper, 45
face figure, 72
face lattice, 46
face number, 46
of a triangulation, 126
facet, 45
Feller, William, 103
Fibonacci number, 80
filter, 26
finite reflection group, 155
finite-field method, 159
flat, 64, 153
flow, 6
flow polynomial, 9
fractional part, 107
Frobenius number, 104
Frobenius problem, 104
Frobenius, Georg, 104
fundamental parallelepiped, 100
general position, 73
generating function, 78
rational, 78
generator, 44
geometric series, 87
Gorenstein polytope, 128
graded poset, 34
grading, 91, 95
Gram, Jørgen, 68
graph, 1
chromatic polynomial of, 3
complete, 20
complete bipartite, 21
connected component of, 6
dual, 7
isomorphic, 19
orientation on, 4
planar, 2
graphical arrangement, 151
characteristic polynomial of, 156
greater index, 147
greatest lower bound, 33
Greene, Curtis, 159
Guthrie, Francis, 2
Index
169
h-polynomial, 126
Haken, Wolfgang, 2, 18
halfspace, 40
irredundant, 41
Hall, Philip, 35
Hardy, Godfrey Harold, 86
Hasse diagram, 10
height, 91
Hibi, Takayuki, 127
homogenization, 43, 91
Huh, June, 19
hyperplane, 40
supporting, 45
hyperplane arrangement, 48, 63
central, 64
Coxeter, 155
essential, 64
flat of, 63, 153
general position, 73
graphical, 151
induced, 153
rational, 152
real braid, 154
region, 65
face, 46
of order ideals, 26
lattice (poset)
Birkhoff, 26
Boolean, 30
of order ideals, 26
lattice basis, 92, 109
lattice length, 22
lattice of flats, 63
lattice polytope, 44
lattice segment, 22
Laurent series, 92
Lawrence, Jim, 68
least upper bound, 33
length of a poset interval, 34
line free, 42
line segment, 40
lineality space, 42
linear extension, 28
linear hyperplane, 40
linear recurrence, 79
linearly ordered, 10
log-concave, 19
loop, 1
identity operator, 77
incidence algebra, 26
inclusion–exclusion formula, 48
inclusion–exclusion principle, 38
inclusion-exclusion, 15
indicator function, 66
induced hyperplane arrangement, 153
inside-out polytope, 152
reciprocity theorem for, 158
integer-point transform, 91
interior, 13, 42
relative, 125
interior point, 69
intersection poset, 38
interval, 31
irredundant halfspace, 41
isomorphic posets, 31
M¨
obius function, 16
M¨
obius inversion, 153
Macdonald, I. G., 127
MacMahon, Percy, 103, 147
Mani, Peter, 68
map coloring, 18
maximum, 28
McMullen, Peter, 127
meet, 33
meet semilattice, 73
minimum, 28
Minkowski sum, 44
M¨
obius function, 29
M¨
obius inversion, 35
multichain, 26
multiplicity, 158
Jaeger, Fran¸cois, 18
join, 33
join irreducible, 37
Jordan normal form, 105
Katz, Eric, 19
lattice
(poset), 33
distributive, 33, 48
nilpotent, 36, 75
node, 1
nowhere-zero flow, 6
octahedron, 128
order ideal, 26
principal, 26
order polynomial, 12, 28
order preserving, 10
order preserving map, 25
170
orientation, 4
acyclic, 4, 156
induced by a coloring, 4
totally cyclic, 9
part, 86
partially ordered set, 9
partition, 36, 86
function, 108
part of, 86
partition function, 108
parts, 82
path, 20
Paule, Peter, 104
period, 89
Petersen graph, 21
Pick, Georg, 19
planar, 2
plane partition, 83, 143
diamond, 107
plane partition diamonds, 107
Poincar´e, Henri, 68
pointed, 42
polyconvex, 47
polygon, 13
polyhedral cone, 43
polyhedron, 40
line-free, 42
pointed, 42
polynomial, 28, 75, 89
polytope, 13, 44
Gorenstein, 128
lattice, 44
rational, 44, 98, 123
reflexive, 128
poset, 9, 25
direct product, 37
Eulerian, 34, 59
graded, 34
intersection, 38
interval, 31
isomorphism, 10
of partitions, 36
rank of, 34
P -partition, 143
reciprocity theorem for, 144
product, 27
proper coloring, 2
proper face, 45
pyramid, 46, 128
quasipolynomial, 85
Index
constituents of, 89
convolution of, 88, 107
degree of, 89
period of, 89
Rademacher, Hans, 86
Ramanujan, Srinivasa, 86
rank, 38
of a poset, 34
of an element, 38
rational cone, 44
rational generating function, 78
rational hyperplane arrangement, 152
rational polytope, 44, 98, 123
Read, Ronald, 19
real braid arrangement, 73
reciprocity theorem
for P -partitions, 144
for Ehrhart polynomials, 125
for inside-out polytopes, 158
for integer-point transforms of cones,
128
for restricted partition functions, 88
for solid-angle polynomials, 124
refinement, 36
reflexive, 9
reflexive polytope, 128
region, 65
bounded, 65
of a graphical arrangement, 156
of an inside-out polytope, 157
relatively bounded, 74
relative interior, 42
relative volume, 110
relatively bounded, 74
restricted partition function, 87
reciprocity theorem for, 88
Riese, Axel, 104
right action, 38
Riordan, John, xii
root system, 155
Rota’s crosscut theorem, 31
Rota, Gian–Carlo, 35
Sanyal, Raman, 18
Schl¨
afli, Ludwig, 68
self reciprocal, 86
Seymour, Paul, 9
shift operator, 77
simplex, 44
unimodular, 95, 126
simplicial, 59
Index
simplicial polytope, 68
solid angle, 123
solid-angle polynomial, 124
reciprocity theorem for, 124
Sommerville, D. M. Y., 68
Stanley, Richard, xii, 4, 18, 19, 36, 68,
104, 127, 146
Steinitz’s theorem, 68
Stirling number of the second kind, 11,
37
Stirling numbers of the second kind, 11
strict order polynomial, 10
strictly order preserving, 10
strictly order preserving map, 26
supporting hyperplane, 45
tangent cone, 57
tetrahedron, 127
total order, 10
totally cyclic, 9
totally ordered, 10
transitive, 9
transversal, 158
triangle, 13
triangulation, 14
unimodular, 126
Tutte, William, 8
unimodal, 19
unimodular, 95
unimodular cone, 93
unimodular triangulation, 126
unipotent, 75
valuation, 15, 48, 127
vector space
of formal power series, 78
of polynomials, 76
vertex, 13, 44, 45
vertex figure, 72
wheel, 20
Whitney, Hassler, 2, 18
Wilf, Herbert, 19
Zaslavsky, Thomas, 68
zeta function, 27
zeta polynomial, 32, 75
Zn -flow, 6
171