Professional Documents
Culture Documents
Santos Main
Santos Main
configurations
The content of this course is based on the following sections of the book on
triangulations coauthored by Francisco Santos [1]: Chapter 2, Sections 2.1 to 2.5;
Chapter 4, Sections 4.1, 4.3 and 4.4; Chapter 5, Section 5.1. All throughout we
give pointers to that book so that the interested reader can find more details and
explanations.
Additional references that we recommend are:
• For polytope theory, including an introduction to oriented matroids similar in spirit
to the one in Section 2, we recommend Ziegler’s book “Lectures on polytopes”
[5].
Francisco Santos
Universidad de Cantabria, e-mail: francisco.santos@unican.es. Work of F. Santos was sup-
ported by grants MTM2017-83750-P and PID2019-106188GB-I00 of the Spanish State Research
Agency (AEI/10.13039/501100011033).
Mónica Blanco
e-mail: m.blanco.math@gmail.com
Federico Castillo
Max Planck Institute for Natural Sciences, e-mail: castillo@mis.mpi.de
1
2 Francisco Santos, Mónica Blanco and Federico Castillo
We start with a brief summary of the basic concepts of affine and convex combina-
tions. For more details see [1, Section 2.1] or the first two chapters of [5].
Definition 1
Let A = {𝑝 1 , . . . , 𝑝 𝑛 } ⊂ R𝑑 be a finite subset.
• An affine combination of the points of A is any point of the form 𝜆1 𝑝 1 + · · · +𝜆 𝑛 𝑝 𝑛
Í
with 𝜆𝑖 ∈ R and 𝜆𝑖 = 1.
• An affine dependence between the pointsÍ of A is a valid expression of the form
𝜆1 𝑝 1 + · · · + 𝜆 𝑛 𝑝 𝑛 = 0 with 𝜆𝑖 ∈ R and 𝜆𝑖 = 0. The points of A are affinely
independent if their only affine dependence is the trivial one 𝜆1 = · · · = 𝜆 𝑛 = 0,
and affinely dependent otherwise.
• A subset of R𝑑 is an affine (sub)-space, or a flat, if it is closed under affine
combinations.
• The affine span of A is the smallest affine space containing A. Equivalently, it
is the set of points that are affine combinations of points of A. It is denoted by
aff(A).
• The dimension of an affine space is the maximum number of affinely independent
points minus 1. The dimension of A is the dimension of aff (A).
a↵(p1 , p2 ) = { 1 p1 + 2 p2 : 1 + 2 = 1}
1 <0
p4
1 =0
2 >1 q
i <1 2 =1
1 >1
1 =1
2 =0
i >0 q p
2 <0 p2 p3
p1 p
p5
Notation
We will be constantly dealing with a set of points 𝑝 1 , . . . , 𝑝 𝑛 and its dif-
ferent subsets. In order to simplify notation, we will identify each point with
its label i ≡ 𝑝 𝑖 ≡ {𝑝 𝑖 }. Similarly, for 𝑖1 , . . . , 𝑖 𝑚 ∈ {1, . . . , 𝑛}, we will write
i1 . . . im ≡ {𝑝 𝑖1 , . . . , 𝑝 𝑖𝑚 } for the set of the points corresponding to those labels,
and conv(i1 . . . im ) ≡ conv{𝑝 𝑖1 , . . . , 𝑝 𝑖𝑚 } for their convex hull. This notation will
also appear in the figures.
The faces of the polygon P = conv{𝑝 1 , 𝑝 2 , 𝑝 3 }, with 𝑝 1 = (0, 0), 𝑝 2 = (1, 0) and
𝑝 3 = (0, 1) (see Figure 2), are:
• ∅ as the only face of dimension −1;
• P−𝑥−𝑦 = conv(1), P 𝑥 = conv(2) and P 𝑦 = conv(3) as its vertices;
• P−𝑦 = conv(12), P−𝑥 = conv(13) and P 𝑥+𝑦 = conv(23) as its edges.
• P = conv(123) as its only face of dimension 2.
The polytope that is of most interest to us is the one with the minimum number
of vertices:
−x − y x
3 3 3
y
1 2 1 2 1 2 3
x+y x
3 3 3
y 1 2
1 2 1 2 1 2
Fig. 2 Faces of a polygon and the normals of the corresponding functionals.
Remark
There is no (much) loss of generality in assuming that A is full-dimensional because
if A ⊂ R𝐷 and dim(A) = 𝑑 < 𝐷 then we can forget some coordinates and get a
𝑑-dimensional configuration A0 in R𝑑 that is affinely isomorphic to A; all we are
going to do is invariant under affine isomorphism.
Nonetheless, sometimes we need to consider a subset B of A as a configuration
in itself and, of course, it may be lower dimensional.
Running example
The point configuration depicted in Figure 3, will serve as our running example. The
faces of this point configuration are: ∅ of dimension −1; 1, 2, 3 and 4 of dimension
0; 12, 13, 24 and 34 of dimension 1; and the unique two-dimensional face 12345,
which includes the non-vertex 5.
6 Francisco Santos, Mónica Blanco and Federico Castillo
3 4
3 4 3 4 3 4
5 5
1 2 1 2 1 2
Remark
We may sometimes refer to subdivisions or triangulations of a polytope, in which
case we are refering to subdivisions on the point configuration consisting of the
vertices of the polytope.
8 Francisco Santos, Mónica Blanco and Federico Castillo
Problems
1.2 a. *Construct all the subdivisions of a triangular prism (the product of an equilateral triangle
and a segment perpendicular to it). Draw also the poset of subdivisions.
Clue: modulo symmetries there are only three subdivisions: triangulations (all equivalent
under symmetry), the trivial subdivision, and another class.
b. Describe all the triangulations of Δ 𝑘 × Δ1 , the product of a 𝑘-simplex and a segment. (See
[1, Sec. 6.2.1] for more details)
Clue: there is a natural bijection between triangulations of Δ 𝑘 × Δ1 and permutations of
{1, . . . , 𝑘 + 1}, where the 𝑘 + 1 symbols {1, . . . , 𝑘 + 1} correspond to the vertices of Δ 𝑘 .
1.3 *Let A consist of 4 points in R2 , not all in a line (that is, they affinely span R2 ). Show that
there are three subdivisions of A, no matter what the points are: the trivial subdivision and two
triangulations.
1.5 Do the same for 𝑑 + 2 points in R𝑑 , not all in a hyperplane. For this:
a. Observe that there is (modulo multiplication by a constant) a unique affine dependence
among the points. Let A+ and A− be the sets of points with positive and negative coefficient
in it.
b. Show that every full-dimensional simplex in A is of the form A \ { 𝑝𝑖 } where 𝑝𝑖 is in
A+ ∪ A− .
c. Show that two such simplices A \ { 𝑝𝑖 } and A \ { 𝑝 𝑗 } intersect properly (that is, they satisfy
(IP)) if and only if 𝑝𝑖 and 𝑝 𝑗 lie in the same subset A+ or A− .
d. Show that {A \ { 𝑝𝑖 } : 𝑝𝑖 ∈ A+ } and {A \ { 𝑝𝑖 } : 𝑝𝑖 ∈ A− } are two triangulations of A,
and conclude from the above that they are the only non-trivial subdivisions of it.
Clue: Lemma 20 below is essentially solving this exercise.
1.6 a. *Consider the vertex set of a 1 × 2 grid. That is, let A = { (𝑖, 𝑗) : 𝑖 ∈ {0, 1}, 𝑗 ∈ {0, 1, 2} }.
Construct all the triangulations of A. How many of them use all the points?
b. Consider the vertex set of a 1 × 𝑘 grid. That is, let A = { (𝑖, 𝑗) : 𝑖 ∈ {0, 1}, 𝑗 ∈
{0, 1, . . . , 𝑘 } }. Show that A has 2𝑘
𝑘 triangulations that use all points.
1.7 *Show that the 3-dimensional cube [0, 1] 3 has six symmetry classes of triangulations, one with
five tetrahedra and five with six tetrahedra. (See [1, Sec. 6.3.4] for more details)
Clue: first show that every triangulation uses either one of the big, regular, tetrahedra inscribed
in the cube, or one of the four long diagonals between opposite vertices. In the first case this
completely determines the triangulation; in the second case look at the possible configurations
of tetrahedra around that diagonal.
1.8 For each 𝑛 ≥ 3, let 𝑇𝑛 be the set of all triangulations of a convex 𝑛-gon. The goal is to show
1 2𝑛−4
that |𝑇𝑛 | equals 𝑛−1 𝑛−2 . For this:
a. Show that every triangulation of the 𝑛-gon has 2𝑛 − 3 edges (the 𝑛 edges of the 𝑛-gon plus
𝑛 − 3 internal diagonals). Deduce that the average degree of vertices in a triangulation is
4 − 𝑛6 .
Note: The degree here is in terms of graphs: the number of edges a particular vertex belongs
to.
b. Look at a particular vertex, say vertex 1. Explain why the average degree of *that vertex*
among *all triangulations* is also 4 − 𝑛6 .
c. Consider the map 𝑓 : 𝑇𝑛+1 → 𝑇𝑛 that “contracts” the edge {1, 𝑛 + 1}, converting each
triangulation of the (𝑛 + 1)-gon into a triangulation of the 𝑛-gon.2 Show that for each
T ∈ 𝑇𝑛 , the number of triangulations on 𝑓 −1 ( T) equals degT (1) (the degree of vertex 1 in
T).
d. Conclude from (b) and (c) that the average size of 𝑓 −1 ( T) is exactly 4𝑛−6
𝑛 .
e. Deduce that
4𝑛 − 6
|𝑇𝑛+1 | = |𝑇𝑛 |,
𝑛
and from this prove the formula for |𝑇𝑛 | by induction.
1 2𝑛
Remark: the numbers 𝐶𝑛 = 𝑛+1 𝑛 are called Catalan numbers. The exercise shows that |𝑇𝑛 |
equals the (𝑛 − 2)-nd Catalan number 𝐶𝑛−2 . 3
1.10 Show a bijection between the above two sets and Dyck paths of length 2𝑛: monotone paths
in the integer grid going from (0, 0) to (𝑛, 𝑛) and staying always above the diagonal 𝑥 =
𝑦. For example, for 𝑛 = 3 there are five paths: 𝑁 𝑁 𝑁 𝐸 𝐸 𝐸, 𝑁 𝑁 𝐸 𝐸 𝑁 𝐸, 𝑁 𝐸 𝑁 𝑁 𝐸 𝐸,
𝑁 𝐸 𝑁 𝐸 𝑁 𝐸, 𝑁 𝑁 𝐸 𝑁 𝐸 𝐸. Here 𝑁 and 𝐸 denote a “north” and “east” step, respectively.
2 Oriented matroids
𝜆 = (𝜆1 , . . . , 𝜆 𝑛 ) ∈ R𝑛 ,
Í
for 𝜆𝑖 ∈ R such that 𝜆𝑖 = 0 and 𝜆1 𝑝 1 + · · · + 𝜆 𝑛 𝑝 𝑛 = 0.
• For any affine functional 𝑓 : R𝑑 → R, the vector
𝑓 ( 𝑝 1 ), . . . , 𝑓 ( 𝑝 𝑛 ) ∈ R𝑛
That is, there is a bijection between the 3𝑛 possible signatures of vectors in R𝑛 and
pairs (𝐸 + , 𝐸 − ) of disjoint subsets of {1, . . . , 𝑛}.
Thus, 𝑝 is both a convex combination of the sets {𝑝 𝑖 : 𝜆𝑖 > 0} and {𝑝 𝑖 : 𝜆𝑖 < 0}.
In particular, it is a point in the intersection of their respective convex hulls.
That is, for each non-zero dependence 𝜆 = (𝜆1 , . . . , 𝜆 𝑛 ) ∈ R𝑛 of a point config-
uration A, its signature defines a pair (A+𝜆 , A−𝜆 ) where A+𝜆 (resp. A−𝜆 ) is the set of
points of A with positive (resp. negative) coefficients in 𝜆. These two sets have the
property that conv(A+𝜆 ) ∩ conv(A−𝜆 ) ≠ ∅.
In particular, we are interested in the affine dependences among the smallest
possible subsets of points, which are called circuits.
Definition 12 ( [1, Def. 2.4.1] ) A circuit is a set C of affinely dependent points such
that every proper subset is affinely independent.
In particular, for a circuit C = {𝑐 1 , . . . , 𝑐 𝑘 }, we have 𝑘 = dim(C) + 2 and there is a
unique (up to rescaling) affine dependence 𝜆 = (𝜆1 , . . . , 𝜆 𝑘 ) between the points of
C. Since every proper subset of C is affinely independent, 𝜆 has no zero entries. Its
signature determines a partition (C+ , C− ) of the point configuration C that is called
the Radon partition and is the unique partition of C into two disjoint subsets such
that conv(C+ ) ∩ conv(C− ) ≠ ∅.
A circuit {𝑐 1 , 𝑐 2 , 𝑐 3 , 𝑐 4 } with four elements has two options for its signature:
(12, 34), in which case conv(12) ∩ conv(34) ≠ ∅ (two segments that intersect); and
(1, 234), which implies 1 ∈ conv(234).
See Figure 6 for an example of each type of 2-dimensional circuit.
H− H0
H− H0
H+
3 4 3 4 H+
5 5
Fig. 7 Affine evaluation
signatures on A0 . The one 1 2 1 2
on the right has minimal
support, giving cocircuit 23
and oriented cocircuit (2, 3). (+, +, −, +, +) (0, +, −, 0, 0)
Notice that both the circuits and the cocircuits are merely subsets of A (those
with non-zero coefficient in a particular dependence or evaluation, respectively), and
that oriented circuits and cocircuits, as signatures, are completely determined by the
pairs (A+𝜆 , A−𝜆 ) or (A+𝑓 , A−𝑓 ), respectively. Also note that each circuit (resp. cocircuit)
produces two opposite oriented circuits (resp. cocircuits).
Notice that, for A of dimension 𝑑, any 𝑑 +2 points of A that span aff(A) determine
a 𝑑-dimensional circuit, and that any 𝑑 affinely independent points determine a
cocircuit.
(∅, 125), (∅, 135), (∅, 245), (∅, 345), (1, 24), (1, 34), (2, 3),
(125, ∅), (135, ∅), (245, ∅), (345, ∅), (24, 1), (34, 1), (3, 2).
Remark
Observe that from the oriented matroid of A we can recover several things:
• The facets of A (as a point configuration) are the complements of the positive
cocircuits. That is, F is a facet of A if and only if (A \ F, ∅) is a cocircuit.
14 Francisco Santos, Mónica Blanco and Federico Castillo
Proof (sketch) Let S = {S1 , . . . , S𝑚 } where the S𝑖 are subsets of A. Let us check
the properties of subdivisions as in Definition 8:
(DP) S𝑖 is 𝑑-dimensional if, and only if, it contains 𝑑 + 1 points that are affinely
independent. A subset of A is affinely independent if it does not contain a circuit.
(UP) It can be checked that, in the presence of (IP), property (UP) is equivalent to
the following: for every cell S𝑖 in S and every facet F of S𝑖 such that conv(F) is
not contained in the boundary of conv(A), there is another cell S 𝑗 containing the
same facet F and lying in the opposite side of the hyperplane containing conv(F).
Facets of every S𝑖 ∈ S and the existence of such an S 𝑗 for each of them can be
checked knowing the oriented matroid alone.
(IP) Claim: If 𝑖 ≠ 𝑗 then conv(S𝑖 ) ∩ conv(S 𝑗 ) = conv(S𝑖 ∩ S 𝑗 ) if, and only if,
there exists no (oriented) circuit (C+ , C− ) with C+ ⊆ S𝑖 and C− ⊆ S 𝑗 .
See Figure 8 for an example when the (IP) property does not hold: the circuit
(C+ , C− ), with C+ := i2 i3 ⊂ 𝑆𝑖 and C− := j1 j2 ⊂ 𝑆 𝑗 is such that the point 𝑝 :=
conv(i2 i3 ) ∩ conv(j1 j2 ) lies in conv(S𝑖 ) ∩ conv(S 𝑗 ) but not in conv(S𝑖 ∩ S 𝑗 ) = ∅:
j2
i2
Si = i1 i2 i3 p Sj = j1 j2 j3
j1 j3
Fig. 8 Setting of proof of
Theorem 16. i1 i3
Triangulations of polytopes and point configurations 15
Oriented matroids can be defined (and are usually defined) for vector configurations
rather than point configurations. If V = {𝑣 1 , . . . , 𝑣 𝑛 } ⊂ R𝐷 is a vector configuration
(a finite subset of vectors of R𝐷 ) then we define its dependences, evaluations, circuits
and cocircuits as above, simply changing the word affine to linear everywhere. Of
course, one can say that a point configuration and a vector configuration are the same
thing, a finite subset of R𝐷 . What makes a difference is whether we are doing affine
geometry or linear algebra in R𝐷 . In particular, 𝑘 linearly independent vectors span
a space of dimension 𝑘, whereas 𝑘 + 1 affinely independent points are needed for an
affine space of dimension 𝑘.
We typically represent a point configuration A = {𝑝 1 , . . . , 𝑝 𝑛 } ⊂ R𝑑 as the
(𝑑 + 1) × 𝑛 matrix
𝑝 . . . 𝑝𝑛
𝑀A = 1 .
1 ... 1
and a vector configuration V = {𝑣 1 , . . . , 𝑣 𝑛 } ⊂ R𝐷 as the 𝐷 × 𝑛 matrix
𝑀V = 𝑣 1 . . . 𝑣 𝑛 .
With this matrix representation, we have the following, which explains and re-
proves Lemma 10:
Lemma 17 ( [1, Lem. 4.1.31] ) Affine dependences of A are the kernel of 𝑀A and
affine evaluations of A are the rowspace of 𝑀A (the linear subspace generated by
rows, which is the same as the image of 𝑀A𝑡 ).
Linear dependences of V are the kernel of 𝑀V and linear evaluations are the
rowspace of 𝑀V .
Remark
Definition 18 ( [1, Def. 4.1.35] ) We say that V and W are oriented matroid duals if
the oriented circuits of V are the oriented cocircuits of W, and viceversa.
We say that W is the Gale transform of V if the row spaces of 𝑀V and 𝑀W are
orthogonal complements to one another.
0303 1
𝑀V0 = 0 0 3 3 1® .
© ª
«1 1 1 1 1¬
Hence its columns are the elements 𝑤 𝑖 of a vector configuration W which is the Gale
transform of V0 . This vector configuration is drawn in Figure 9.
Observe that vectors 𝑤 2 and 𝑤 3 coincide. This is in agreement with the fact
that (2, 3) was an oriented cocircuit in V0 (corresponding to the linear evaluation
(0, 3, −3, 0, 0) of the functional 𝑥 − 𝑦), so (2, 3) must now be an oriented circuit of
W and (0, 3, −3, 0, 0) its linear dependence in W. These two “equal vectors” must be
considered different elements in the oriented matroid and in the vector configuration,
distinguished by their labels.
Triangulations of polytopes and point configurations 17
w1 w4
w5
Problems
2.1 *Show that there are only three possible oriented matroids for five points in
general position (=no 3 collinear) in the plane.
2.2 *Write down the whole list of circuits and cocircuits of the first two point config-
urations of Problem 1.1.
6
Clue: since the configurations are in general position there should be exactly 4 pairs of opposite
circuits and 62 pairs of opposite cocircuits in each of them.
2.3 Write down the circuits and cocircuits of the following vector configuration of
rank 3:
{(0, 2, 1), (0, −1, −2), (1, 0, 2), (−2, 0, −1), (2, 1, 0), (−1, −2, 0)}.
2.4 Show that the above list of vectors is actually a Gale transform of the set of vertices
of the hexagon (the second configuration in Problem 1.1) with coordinates given
as follows:
𝑝 1 = (0, 2, 1), 𝑝 2 = (0, 1, 2), 𝑝 3 = (1, 0, 2),
𝑝 4 = (2, 0, 1), 𝑝 5 = (2, 1, 0), 𝑝 6 = (1, 2, 0)
Remark: the extra coordinate that you have (the configuration is two dimensional but it is
embedded in an affine plane in R3 ) implies that you can already think of the vertices of the
hexagon as vectors, without the need of an additional coordinate constantly equal to 1 (see
Remark 2.2 in the notes).
2.5 *Construct a Gale transform of the m.o.a.e. (the first configuration in Problem
1.1), with coordinates given as follows:
2.6 *Consider versions of the m.o.a.e. where the inner triangle is rotated with respect
to the outer one. (That is, you have the six vertices of two equilateral triangles
with the same barycenter and one inside the other, but the two triangles are no
18 Francisco Santos, Mónica Blanco and Federico Castillo
longer parallel to one another). Convince yourself of the following behavior: if the
rotation is small, neither the oriented matroid nor the set of triangulations of your
point set changes. As you keep rotating, there is a precise point when both the
oriented matroid and the set of triangulations change. Try to describe the change.
(How many and which circuits, cocircuits, and triangulations are affected?).
2.7 Oriented matroid of a directed graph. Let 𝐺 be a directed graph: a pair (𝑉, 𝐸)
where 𝑉 = {1, . . . , 𝑘 } is a set of 𝑘 vertices and 𝐸 a list of 𝑛 directed edges (each
→−
directed edge is an ordered pair 𝑖 𝑗 := (𝑖, 𝑗) with 𝑖, 𝑗 ∈ 𝑉).
Consider the directed incidence matrix 𝐴 of 𝐺: the 𝑘 × 𝑛 matrix whose 𝑙-th
column is the vector 𝑒 𝑖 − 𝑒 𝑗 , where 𝑒 𝑠 ∈ {0, 1} 𝑘 is the vector with all zeroes and
→
−
coordinate 𝑠 equal to 1, and the 𝑙-th edge of the graph is 𝑖 𝑗 .
Let X := {𝑎 1 , . . . , 𝑎 𝑛 } ⊂ R 𝑘 be the set of columns of 𝐴 (so that the elements of
X are in bijection with edges of 𝐺, and the number of coordinates is the number
of vertices of 𝐺). Show that:
a. A subset of X is linearly dependent if, and only if, the corresponding set of
edges of 𝐺 contains a (not necessarily well-directed) cycle.
Clue: for the if part, show how to get a dependence from the cycle; for the only if part, argue
that if the set of edges contains no cycle then some vertex is used in a single edge, and use
that to induct on the number of edges: removing that edge from the list reduces the rank of
your set of vectors by one.
b. Circuits of X are in bijection to (not necessarily well-directed) cycles of 𝐺.
c. For each circuit, the two sides of the corresponding oriented circuit are the
edges pointing to one or the other direction along the cycle.
2.8 a. *Write down all the oriented circuits and cocircuits of the point configuration
A = {(6, 0), (3, 0), (2, 2), (0, 0), (0, 3), (0, 6)}.
Lemma 20 ( [1, Lem. 2.4.2] ) Let C = (C+ , C− ) be an oriented circuit. Then C has
exactly two triangulations:
which we will refer to as the positive and negative triangulations of the circuit.
T ∈ C \ {𝑐 𝑖 } | 𝑐 𝑖 ∈ C+ , ∀T ∈ T
or
T ∈ C \ {𝑐 𝑖 } | 𝑐 𝑖 ∈ C− ,
∀T ∈ T
Since both those sets are triangulations of C, that completes the proof.
TC+ = {123, 234} TC− = {124, 134} TD+ = {124, 134, 234} TD− = {123}
3 4 3 4 3 3
1 2 1 2 1 2 1 2
Fig. 10 Triangulations of 2-dimensional circuits.
20 Francisco Santos, Mónica Blanco and Federico Castillo
Definition 22 ( [1, Defs. 2.4.7 and 4.4.2, and Thm. 4.4.1] ) Let T be a triangulation
of A. Suppose there is an oriented circuit C = (C+ , C− ) in A such that T contains
the positive triangulation TC+ of C, and suppose further that the links in T of all the
cells of TC+ are identical. Let L be that link.
Then it is possible to construct a new triangulation T 0 of A as follows:
T 0 = T \ T ∪ L | T ∈ TC+ , L ∈ L ∪ T ∪ L | T ∈ TC− , L ∈ L .
That is, we remove TC+ (together with its link) from T and insert TC− (with the
same link). We say that T 0 is obtained from T by a (geometric bistellar) flip on the
circuit (C+ , C− ), and that T and T 0 are adjacent or flip-neighbors.
Remark
When the circuit C is full-dimensional (consists of exactly 𝑑 + 2 points), then the
cells of TC+ are full-dimensional and the condition about the links is vacuous. Hence
T 0 = T \ T ∈ TC+ ∪ T ∈ TC− .
However, when the circuit is not full-dimensional (the circuit involves less points)
the link condition has to be verified. Notice that in this case, when we say that “T
contains’ TC+ ”, we mean that every cell (or face) of TC+ is a face (but not necessarily
a cell) of T .
Definition 23 ( [1, Def. 2.4.8] ) The set of all triangulations of A, under adjacency
by flips, forms the graph of triangulations, or flip-graph, of A.
Triangulations of polytopes and point configurations 21
T1 T2 T3 T4
3 4 3 4 3 4 3 4
5 5
1 2 1 2 1 2 1 2
Fig. 11 Triangulations of A0 .
• The circuit (14, 23) is full-dimensional, and its two triangulations are contained,
respectively, in triangulations T1 and T2 of A0 .
• The circuit (123, 5) is full-dimensional, and its two triangulations are contained,
respectively, in triangulations T2 and T3 of A0 .
• The circuit (23, 45) is full-dimensional, and its two triangulations are contained,
respectively, in triangulations T3 and T4 of A0 .
• The circuit (14, 5) is lower dimensional, and it is a bit more complicated to
study. The two triangulations of 145 are {14} and {15, 45}. Notice that these are
contained in T1 and T4 , respectively, and that the link of 14, 15 and 45 in both
triangulations is the same, namely {2, 3}. Hence the conditions are met and these
two triangulations are adjacent via a flip on the circuit.
3 4 3 4
(14, 23)
1 2 1 2
(14, 5) (123, 5)
3 4 3 4
T1 = 1236, 1348, 1368, 1568, 3678
8 7
5 6
8 7
5 6
4 3
4 1 2
3
1 2
T2 = 1238, 1268, 1348, 1568, 2368, 3678}
Notice that both T1 and T2 contain the (2-dimensional) triangulation {123, 134}
of the circuit (24, 13). However, on T1 these two triangles have different links: the
link of 123 is {6} whereas that of 134 is {8}. Thus, in T1 we cannot perform a flip
on that circuit. On the other hand, in T2 the link of both triangles is {8} so we can
perform a flip to obtain a third triangulation T3 : remove the two tetrahedra 1238 and
1348 and add 1248, 2348 (see Figure 14).
T2 = 1238, 1268, 1348, 1568, 2368, 3678
8 7
5 6
8 7
5 6
4 3
4 1 2
3
1 2
T3 = 1248, 1268, 1568, 2348, 2368, 3678
Flip-graph of moae
The moae (the “mother of all examples”) is the point configuration of Figure 15.
Figure 16 shows its graph of triangulations.
4 6
Fig. 15 The mother of all
examples. 1 3
Fig. 16 The graph of flips of the moae. The two non-regular triangulations are in blue and red, and
the flips joining them to the regular ones are dashed.
24 Francisco Santos, Mónica Blanco and Federico Castillo
It is a fact that the flips among triangulations are in bijection with subdivisions
whose proper refinements are triangulations. See Section 2.4.2 of [1] for more on
this.
Definition 24 Let A be a point configuration, and let C = (C+ , C− ) be an oriented
circuit of A. A flip on (C+ , C− ) is said to be of type (𝑖, 𝑗), where |C+ | = 𝑖 and
|C− | = 𝑗.
In Figure 12, we have one lower dimensional circuit of type (1, 2), a full-
dimensional circuit of type (1, 3), and two full-dimensional circuits of type (2, 2).
2 3
T1
6 4
1 T3 T4
type (2, 1)
2 3 2 3
5
type (3, 1)
4 4
2 3 1 1
7
T2
6 4 5 5
1
2 3 2 3 3
6 6 6 4
1 1
3 2 3 3
4 4 4
1 1 1
7
Fig. 18 Flippable circuits of
type (2, 2). 5
Problems
3.1 *Show that all the triangulations of a convex 𝑛-gon have exactly 𝑛 − 𝑑 − 1 = 𝑛 − 3
flips, where 𝑛 is the number of vertices and 𝑑 = 2 is the dimension.
3.2 *Show that for the triangular prism Δ2 × Δ1 all triangulations have exactly 𝑛 −
𝑑 − 1 = 2 flips, where 𝑛 = 6 is the number of vertices and 𝑑 = 3 is the dimension.
3.3 If you did Problem 1.2b, show that flips among triangulations of Δ 𝑘 × Δ1 corre-
spond to permutations that differ in a transposition of consecutive elements. In
particular, all triangulations have exactly 𝑛 − 𝑑 − 1 = 𝑘 flips, where 𝑛 = 2𝑘 + 2 is
the number of vertices and 𝑑 = 𝑘 + 1 is the dimension.
3.4 *Show that m.o.a.e. has triangulations with three and with four flips. (Here 𝑛−𝑑−1
is three).
3.5 *Show that the 3-cube [0, 1] 3 has triangulations with four, and with six flips.
(Here 𝑛 − 𝑑 − 1 is four).
26 Francisco Santos, Mónica Blanco and Federico Castillo
3.6 *Draw the graphs of flips among triangulations of the configuration you chose in
Problem 1.1.
3.7 *Draw the graph of flips among triangulations of the following six points: the five
vertices of a regular pentagon, together with its center.
3.8 Let T1 and T2 be two triangulations of the convex 𝑛-gon. Show that:
a. *If one of them is the triangulation that joins one vertex to all the others, then
we can go from T1 to T2 in 𝑛 − 3 or less flips.
b. *No matter what they are, we can go from T1 to T2 in 2𝑛 − 6 or less flips.
c. In fact, we can go from T1 to T2 in 2𝑛 − 10 or less flips. (Clue: [1, Prop. 1.1.5]).
d. Let 𝑛 be even and suppose that T1 uses all the even ears4 and T2 uses all the
odd ears5. Show that it is impossible to go from T1 to T2 in less than 32 𝑛 − 5
flips.
Clue: classify the possible diagonals of an 𝑛-gon in three types: even-even, even-odd, and
odd-odd. This gives you, for each triangulation, a vector of length three that counts the
number of diagonals of each type. Study how that vector changes by a flip (there are several
cases), and what change you need in order to go from T1 to T2 .
3.9 Let A = {±𝑒 𝑖 ± 𝑒 𝑗 : 𝑖, 𝑗 ∈ {1, 2, 3}} ∪ {(0, 0, 0)} (the vertices and barycenter
of a regular cube-octahedron; see [1, Exm. 3.6.16 and Prop. 3.6.17]). Let T be
a triangulation in which all tetrahedra use (0, 0, 0) as a vertex (put differently, T
is one of the 64 triangulations obtained triangulating the 6 quadrilateral facets of
conv(A) and coning the boundary to (0, 0, 0)). Show that:
a. The number of flips in T equals 6 plus twice the number of vertices of degree
four that you see in the triangulation of the boundary.
b. There are two triangulations with only 6 flips (observe that 𝑛 − 𝑑 − 1 = 9 in
this example).
c. Are there any triangulations with 8 flips?
(−1,0,1)
(0,−1,1)
(0,1,1)
(1,0,1) (−1,−1,0)
(0,0,0) (−1,1,0)
(1,−1,0)
(1,1,0) (−1,0,−1)
(0,−1,−1)
(0,1,−1)
(1,0,−1)
Figure 19 shows two liftings of the point configuration 12345 which have same
lifted convex hull but yield two different regular subdivisions. Observe that the
subdivision on the right is a refinement of the subdivision on the left.
1̃ 5̃ 1̃ 3̃ 5̃
2̃ 3̃ 4̃ 2̃ 4̃
1 2 3 4 5 1 2 3 4 5
3 4
Fig. 20 A regular triangula-
tion of A0 , with lifting vector 5
(4, 3, 3, 14, 2). 1 2
we need 𝛼1 < 𝛼3 . The same argument for the edges 35 and 24 leads us to the
contradiction 𝛼1 < 𝛼3 < 𝛼2 < 𝛼1 . Hence, no lifting vector yields the subdivision
of Figure 21. Section 7.1.2 of [1] contains twelve proofs of non-regularity of this
triangulation.
4
6
Fig. 21 A non-regular subdi-
vision. 1 3
For example, if 𝛼 = 0 then we get the trivial subdivision {A}. If 𝛼 has a single
non-zero coordinate 𝛼𝑖 , corresponding to a certain point 𝑝 𝑖 ∈ A, then we get the
following two subdivisions, depending on the sign of 𝛼𝑖 .
6 6
pulling 1 pushing 1
5 5
7
4 1 4 1
8
trivial subdivision
3 6 3
2 5 2
{123, 1345, 156} 7 {128, 167, 178, 2345678}
4 1
8
6 3 6
pulling 4 2 pushing 4
5 5
{12345678} 7
4 1 1
8
3 3
2 2
{124, 146, 234, 456} {1235678}
𝛼 𝑠𝑢𝑏𝑑𝑖𝑣𝑖𝑠𝑖𝑜𝑛
(−1, 0, 0, 0, 0) {124, 134}
(0, −1, 0, 0, 0) {123, 234}
(0, 0, −1, 0, 0) {123, 234}
(0, 0, 0, −1, 0) {124, 134}
(0, 0, 0, 0, −1) {125, 135, 245, 345}
(1, 0, 0, 0, 0) {125, 135, 2345}
(0, 1, 0, 0, 0) {1245, 1345}
(0, 0, 1, 0, 0) {1245, 1345}
(0, 0, 0, 1, 0) {1235, 234}
(0, 0, 0, 0, 1) {1234}
10 2 14 4 7
1 12 1 15 13 12 1 15 12
Lemma 27 ([1, Lemma 4.3.12]) The lexicographic subdivision coincides with the
regular subdivision obtained choosing positive numbers 𝜆1 · · · 𝜆 𝑘 0 and
taking 𝛼𝑖 = 𝜖𝑖 𝜆𝑖 for each point in the list and 𝛼𝑖 = 0 for each point not in the list.
The actual numbers 𝜆𝑖 in the lemma are not important, as long as each one is
sufficiently larger than the next one.
Problems
4.2 *Deduce from the previous problem that to construct all the regular subdivisions
of a set A, or to check whether a particular subdivision is regular or not, there is
no loss of generality in choosing a priori 𝑑 + 1 affinely independent points of A
(a 𝑑-simplex) and prescribing those 𝑑 + 1 coordinates in the lifting vector to be
zero.
2 2
5 5
4 6
6 4
32 Francisco
1 Santos, Mónica
3 Blanco
1 and Federico Castillo
3
4.4 *Show that, except for the non-regular triangulation in Fig. 21 and the mirror
image of it, all other triangulations of the sets in Problem 1.1 are lexicographic
(hence regular).
Prove that
4.7 Let B1 and B2 be two subsets ofthe
A following regular1triagulation
with conv(B ) ∩ conv(B is 2not
) =lexicographic:
∅. Show that
there is a subdivision of A (in fact, a regular one) in which both B1 and B2 are
faces.
4.8 Show that the same is not necessarily true for *three* subsets: taking B the vertex
set of a triangular prism, find three subsets B1 , B2 , B3 of it with conv(B𝑖 ) ∩
conv(B 𝑗 ) = ∅ for all 𝑖, 𝑗 ∈ {1, 2, 3}, but such that no subdivision of A has B1 , B2
and B3 as cells.
Remark: the m.o.a.e. is an example where no regular subdivision exists using B1 , B2 and B3 .
Triangulations of polytopes and point configurations 33
The importance of regular triangulations is that the graph of flips among regular
triangulations of a 𝑑-dimensional point configuration A with 𝑛 points is the graph
of a polytope of dimension 𝑛 − 𝑑 − 1, where the graph of a polytope is the collection
of its vertices and edges. For example, in Figure 16 of the graph of triangulations
of the moae, if we remove the two non-regular subdivisions (i.e. the red and blue
colored ones), the remaining graph is planar. That is, it is the graph of a polytope of
dimension 3.
More specifically, it turns out that the whole poset of regular subdivisions of A
under refinement has the following very nice structure.
The two conclusions (≥ 𝑛 − 𝑑 − 1 flips and connected graph) hold also for non-
regular triangulations in dimension two, but not in higher dimensions: triangulations
with less than 𝑛 − 𝑑 − 1 flips exist starting in dimension 3 (there is one in Problem
3.9), and triangulations with no flips at all exist in dimension 6. Point sets with
disconnected graphs of triangulations exist in dimension 5 and higher. Whether they
exist in dimensions three and four is open.
Definition 29 The polytope 𝚺(A) is called the secondary polytope of the point
configuration A.
We finish by discussing briefly the main ingredients for the proof of Theorem 28.
We keep studying our running example. The set of all lift vectors 𝛼 equals R5 but
we can pick any three non-collinear points in A and assume they have height zero
(see Problem 5.1). Then depending on 𝑥 and 𝑦 we get a different subdivision. See
Figure 25.
x y
In Table 1 we write down all possible subdivisions and the corresponding sec-
ondary cone (depending only on 𝑥, 𝑦). The first four are triangulations, the next four
are coarsenings, and the last one is the trivial subdivision.
In Figure 26 we show all possible lifts for each subdivision. We omit the trivial
subdivision at the origin.
y
y = 2x > 0
∑︁
𝑧(T ) = (𝑧1 , . . . , 𝑧 𝑛 ) ∈ R𝑛 , 𝑧𝑖 = vol(F),
𝑝𝑖 ∈F
Lemma 30 ([1, Thm. 5.2.16]) The secondary fan of A equals the normal fan of the
polytope 𝚺(A).
Some examples
In dimension 1, since all subdivisions are regular, it is not hard to see that for
the one dimensional point configuration A = {0, 1, 2, · · · , 𝑛 + 1} ⊂ R1 we have
that 𝚺( 𝐴) is combinatorially equivalent to an 𝑛-cube: there are 2𝑛 triangulations,
all regular, one for each possible subset of {1, . . . , 𝑛}. (The subset indicates which
of the intermediate points are used as vertices). Do explicitly the computation of
GKZ-vectors for 𝑛 = 2 and perhaps 𝑛 = 3.
Another especial example is given by the vertices of Δ 𝑘 ×Δ1 , with Δ 𝑘 a 𝑘-simplex.
In this case all triangulations are unimodular (all simplices have the same volume),
36 Francisco Santos, Mónica Blanco and Federico Castillo
and you can compute that the secondary polytope is the permutohedron. (See [1, Sec
6.2.1] for details).
For a less trivial example, if we take A to be the vertices of an 𝑛-gon, the secondary
polytope is the so-called (𝑛 − 2)-associahedron.
The secondary polytope of the vertices of the standard cube [0, 1] 𝑑 has only been
computed for 𝑑 ≤ 4.
Remark
For any triangulation T , its GKZ vector will be a point in 𝚺(A) but not a
vertex unless T is regular. Moreover, the GKZ vector is unique for each regular
triangulation but may not be unique for non-regular ones (for example, the two non-
regular subdivisions of the moae have the same GKZ-vector; this cannot happen for
regular ones).
Problems
5.3 Draw the graph of flips among *regular* triangulations of the first two configura-
tions in Problem 1.1 and understand how they are the graphs of two 3-dimensional
polytopes (the corresponding secondary polytopes).
b. To show that dim 𝚺( 𝐴) ≤ 𝑛−𝑑−1 show that the following 𝑑+1 affine equations
are independent and are satisfied by the GKZ-vector of every triangulation:
𝑛
∑︁
𝑧 𝑖 = (𝑑 + 1) vol(P),
𝑖=1
𝑛
∑︁
𝑗
𝑧𝑖 𝑝 𝑖 = = (𝑑 + 1) vol(P)𝑐 𝑗 , ∀ 𝑗 ∈ {1, . . . , 𝑑}.
𝑖=1
5.5 In the first configuration of Problem 1.1, consider the GKZ-vectors of the eight
triangulations that use all points.
Remark: Since the secondary polytope is three dimensional, you can look at only three coordi-
nates; put differently, you are projecting R6 → R3 forgetting three coordinates. For the sake of
symmetry, use the coordinates of the three outer points).
With or without doing any computations show that:
a. These eight points lie in an affine plane.
b. The two non-regular triangulations produce the same point.
c. The mid-point of any two “opposite” triangulations is the same.
By symmetry, this implies the six regular triangulations produce the vertices of a
regular hexagon and the two non-regular ones both produce the barycenter of it.
5.6 Now consider the same configuration but with one of the two concentric triangles
in it slightly rotated with respect to the other one.
a. Are the eight GKZ-vectors still in an affine plane?
b. Do the two (formerly) non-regular triangulations still produce the same point?
c. Are the mid-points of any two “opposite” triangulations still the same?
By answering these (or other) questions argue that the eight points now are the
vertices of a 3-dimensional parallelepiped. What effect has this on the secondary
polytope, and on the set of regular triangulations?
Acknowledgements We thank Viviana Márquez for kindly sharing with us her board pictures of
Santos’ lectures, which have been of great help for producing these notes.
38 Francisco Santos, Mónica Blanco and Federico Castillo
References
1. Jesús A. De Loera, Jörg Rambau and Francisco Santos. Triangulations: Structures for Al-
gorithms and Applications. Algorithms and Computation in Mathematics, vol. 25, Springer,
2010. DOI 10.1007/978-3-642-12971-1
2. Branko Grünbaum, Convex polytopes. Second edition revised and updated by V. Kaibel,
V. Klee and G. M. Ziegler. Graduate Text in Mathematics, vol. 221. Springer, 2003. DOI
10.1007/978-1-4613-0019-9
3. Carl Lee and Francisco Santos. Subdivisions and triangulations of polytopes.
In “Handbook of Discrete and Computational Geometry”, Third Edition, edited
by Csaba D. Toth, Joseph O’Rourke, Jacob E. Goodman, CRC Press, Novem-
ber 2017, pp 415–447. ISBN 9781498711395 A preprint version is available at
http://www.csun.edu/ ctoth/Handbook/HDCG3.html.
4. Rekha R. Thomas, Lectures in Geometric Combinatorics. Student Mathematical Library, vol.
33. IAS/Park City Mathematical Subseries. American Mathematical Society, Providence, RI;
Institute for Advanced Study (IAS), Princeton, NJ, 2006. viii+143 pp. ISBN: 0-8218-4140-8
5. Günter M. Ziegler, Lectures on polytopes. Graduate Texts in Mathematics, vol. 152. Springer,
1995. DOI 10.1007/978-1-4613-8431-1