Polyhedral

ooovi

We work with a two mutually dual finite-dimensional vector spaces \( N \) and \( M \) over a linearly ordered field \( \Bbbk \). This means, we have a bilinear pairing \( \langle \cdot , \cdot \rangle \colon N \times M \to \Bbbk \) such that the induced maps \( M \to \operatorname{Hom}(N,\Bbbk ) \) and \( N \to \operatorname{Hom}(N,\Bbbk ) \) are both bijective (this is also known as a perfect pairing). Since \( \Bbbk \) is linearly ordered, we can define the non-negative scalars \(\Bbbk _{\geq 0} = \{ t \in \Bbbk \mid t \geq 0 \} \subseteq \Bbbk \). This is a semifield, i.e. a commutative semiring, in which every nonzero element has a multiplicative inverse.

Definition 1
#

A hyperplane is the perp space of a point.

A halfspace is the dual of a point. The dimension of the module the cone is a submodule of.

The dimension of a pointed cone.

A polyhedral cone is full-dimensional if its dimension is the same as the dimension of the ambient space.

The set of all \(k\)-dimensional faces of a polyheadral cone.

For each \(k\) gives the number of \(k\)-dimensional faces of a polyheadral cone.

0.1 Polyhedral cones

Definition 2
#

A pointed cone \( \sigma \subseteq N \) is a \( \Bbbk _{\geq 0} \)-submodule of \( N \).

Definition 3

A pointy cone \( \sigma \subseteq N \) is a \( \Bbbk _{\geq 0} \)-submodule of \( N \) that does not contain a nontrivial linear space.

Definition 4

For a set \( S \subseteq N \), the cone generated by \( S \) is the pointed cone

\[ \operatorname{Cone}(S) := \operatorname{Span}_{\Bbbk _{\geq 0}}(S) \subseteq N. \]
Definition 5

A pointed cone \( \sigma \subseteq N \) is called finitely generated, if \( \sigma = \operatorname{Cone}(S) \) for some finite set \( S \subseteq N \).

Definition 6

For a set \( S \subseteq M \), the dual cone of \( S \) is the pointed cone

\[ S^\vee := \{ x \in N \mid \forall y \in S, \langle x, y \rangle \geq 0 \} \subseteq N. \]
Definition 7

Another word for finitely generated cone with a finitely generated canonical dual.

0.1.1 Examples

Proposition 8

The zero-space \( \{ 0\} \subseteq N \) is a polyhedral cone.

Proof

If \( S \subseteq M \) is a finite \( \Bbbk \)-basis of \( M \), then \( \{ 0\} = (S \cup -S)^\vee \).

Proposition 9

The full space is a polyhedral cone.

Proposition 10

Any halfspace is a polyhedral cone.

Proposition 11

Any linear subspace is a polyhedral cone.

Proposition 12

If \( \sigma \subset N \) is finitely generated, then \( \sigma ^\vee \subset M \) is polyhedral.

Proof

Let \( \sigma = \operatorname{Cone}(S) \) for \( S \subseteq N \). We need to show \( \operatorname{Cone}(S)^\vee = S^\vee \). The inclusion “\( \subseteq \)” is obvious, as \( S \subseteq \operatorname{Cone}(S) \). The opposite inclusion can be shown by induction on the general element of \( \operatorname{Cone}(S) \).

Proposition 13 Double dual

If \( \sigma \subseteq N \) is polyhedral, then \( \sigma ^{\vee \vee } = \sigma \).

Proof

For any set \( S \subseteq N \) it is trivial to show \( S^{\vee \vee \vee } = S^{\vee } \). The claim follows since \( \sigma = S^{\vee } \) for a finite set \( S \subseteq N \).

Proposition 14 Fourier-Motkin elimination

If \( \sigma \subseteq N \) is a polyhedral cone and \( w \in N \), then \( \sigma + \operatorname{Cone}(\{ w\} ) \) is polyhedral too.

Proof

Write \( \sigma = S^{\vee } \) for a finite \( S \subseteq M \). Define

\[ T\ :=\ \{ x \in S \mid 0 \leq \langle x, w \rangle \} \ \cup \ \{ y \langle x, w \rangle - x \langle y, w \rangle \mid x, y \in S, 0 \leq \langle x, w \rangle , \langle y, w \rangle {\lt} 0 \} \ \subseteq \ M. \]

Note that \( T \) is finite. We claim \( S^{\vee } + \operatorname{Cone}(\{ w\} ) = T^{\vee } \). Clearly \( T \subseteq \operatorname{Cone}(S) \), hence \( S^{\vee } = \operatorname{Cone}(S)^{\vee } \subseteq T^{\vee } \). Also clear is \( w \in T^{\vee } \), hence \( S^{\vee } + \operatorname{Cone}(\{ w\} ) \subset T^{\vee } \). For the reverse inclusion, let \( v \in T^\vee \). We need to show \( v - uw \in S^\vee \) for some \( u \in \Bbbk _{\geq 0} \). If there is no \( y \in S \) with both \( \langle y, w \rangle {\lt} 0 \) and \( \langle y, v \rangle {\lt} 0 \), we can easily see that \( u = 0 \) works, i.e. \( T^{\vee } \subseteq S^{\vee } \). Otherwise, we set

\[ u\ :=\ \max _{\substack {y \in S \\ \langle y, w \rangle {\lt} 0}} \frac{\langle y, v \rangle }{\langle y, w \rangle }, \]

which is then well-defined and non-negative. We need to show \( u \langle z, w \rangle \leq \langle z, v \rangle \) for all \( z \in S \). If \( \langle z, w \rangle {\lt} 0 \), this follows from maximality of \( u \). If \( \langle z, w \rangle = 0 \), the claim follows from \( z \in T \) and \( v \in T^{\vee } \). Lastly, assume \( \langle z, w \rangle {\gt} 0 \) and fix \( y \in S \) such that \( u = \frac{\langle y, v \rangle }{\langle y, w \rangle } \). In particular, \( \langle y, w \rangle {\lt} 0 \). Then the claim \( u \langle z, w \rangle \leq \langle z, v \rangle \) is equivalent to

\[ \langle y, v \rangle \langle z, w \rangle \geq \langle z, v \rangle \langle y, w \rangle . \]

But \( y \langle z, w \rangle - z \langle y, w \rangle \in T \), hence the claim follows from \( v \in T^{\vee } \).

Proposition 15

If \( \sigma \in N \) is finitely generated, then it is polyhedral.

Proof

Write \( \sigma = \operatorname{Cone}(S) \) and use induction on the size of \( S \): For \( S = \emptyset \), we use 8 and the induction step is 14.

Proposition 16

For a finite set \( S \subseteq N \), we have \( S^{\vee \vee } = \operatorname{Cone}(S) \).

Proof

The cone \( \sigma := \operatorname{Cone}(S) \) is finitely generated, hence polyhedral by 15. Furthermore, we have \( \sigma ^{\vee \vee } = S^{\vee \vee } \). Now apply 13.

Proposition 17 Polyhedral = Finitely generated

A pointed cone \( \sigma \subseteq N \) is finitely generated iff it is polyhedral.

Proof

The implication “\( \Rightarrow \)” is Proposition 15. For the converse, assume \( \sigma = S^{\vee } \) for some finite \( S \subseteq M \). Then \( \operatorname{Cone}(S) \) is finitely generated, hence polyhedral by 15. Thus we find a finite \( T \subseteq N \) with \( \operatorname{Cone}(S) = T^\vee \). We obtain

\[ \sigma = S^{\vee } = \operatorname{Cone}(S)^\vee = T^{\vee \vee } = \operatorname{Cone}(T), \]

where the last equality is due to 16. Thus \( \sigma \) is finitely generated.

If \( \sigma \subseteq N \) is polyhedral, then \( \sigma ^\vee \subseteq M \) is polyhedral too.

Proof

12 \( + \) 17.

Proposition 19

Every polyhedral cone can be written as the minkowski sum of a pointy cone and a linear subspace.

0.1.2 Operations

Proposition 20

If \( \sigma \subseteq N \) is polyhedral and \(f : N -{\gt} M\) is a linear map, then \( f.map \sigma \subseteq M \) is polyhedral too.

Proposition 21

If \( \sigma , \tau \subseteq N \) are polyhedral, then \( \sigma + \tau \) is polyhedral too.

Proposition 22

If \( \sigma , \tau \subseteq N \) are polyhedral, the direct sum \( \sigma \oplus _p \tau \) is polyhedral too.

Proposition 23

If \( \sigma , \tau \subseteq N \) are polyhedral, their join is polyhedral too.

Proposition 24

If \( \sigma , \tau \subseteq N \) are polyhedral, then \( \sigma \cap \tau \) is polyhedral too.

Proposition 25

If \( \sigma , \tau \subseteq N \) are polyhedral, then the cartesian product \( \sigma \prod \tau \) is polyhedral too.

Definition 26 Convex hull
#

For a set \(S \subseteq N\), the convex hull of \(S\) is

\[ \operatorname{Conv}(S) := \left\{ \sum _{u \in S} \lambda _u | \lambda _u \ge 0, \sum _u \lambda _u = 1\right\} \]
Proposition 27

If \( \sigma , \tau \subseteq N \) are polyhedral, then the convex hull of the two is polyhedral too.

Polyhedral cones for a lattice under intersections and convex hulls (its the same lattice as on pointed cones).

0.1.3 Special cones

Proposition 29

A simplicial cone \(\sigma \) is a polyhedral cone that can be generated by \(dim (\sigma )\) many generators.

0.1.4 Faces and the Face Lattice

We define faces of polyhedral cones. The main goal is the partial order relation on the set of faces and showing that there is a bijection between faces of \( \sigma \) and faces of \( \sigma ^{\vee } \).

Definition 30
#

For any subset \( A \subseteq M \), we define the perp space

\[ A^{\perp }\ :=\ \{ v \in N\ \mid \ \forall u \in A, \langle v, u \rangle = 0\} , \]

which is a pointed cone in \( N \). We also write \( u^{\perp } := \{ u\} ^{\perp } \).

Lemma 31

If \( A \subseteq M \) is closed under negation (for instance, \( A \subseteq M \) a \( \Bbbk \)-submodule), then \( A^{\perp } = A^{\vee } \).

Proof

\( A^{\perp } \subseteq A^{\vee } \) is clear. Let \( x \in A^{\vee } \) and \( a \in A \). then \( \langle x, a \rangle \geq 0 \) and \( -\langle x, a \rangle = \langle x, -a \rangle \geq 0 \). Hence \( \langle x, a \rangle = 0 \) and \( x \in A^{\perp } \).

Definition 32 Face of a cone

A face of a pointed cone \( \sigma \subseteq N \) is the intersection of \( \sigma \) with some hyperplane \( u^{\perp } \) for \( u \in \sigma ^{\vee } \). We write \( \tau \preceq \sigma \) if \( \tau \) is a face of \( \sigma \) and \( \tau \prec \sigma \) if \( \tau \) is a proper face, i.e. \( \tau \preceq \sigma \) and \( \tau \neq \sigma \).

Proposition 33

For any subset \( S \subseteq N \) and \( u \in S^{\vee } \), we have

\[ \operatorname{Cone}(S) \cap u^{\perp } = \operatorname{Cone}(S \cap u^{\perp }). \]
Proof

Both inclusions should be a simple Span-induction.

A face of a polyhedral cone is polyhedral.

Proof

Use 33 and the fact that if \( S \) is finite, so is \( S \cap u^{\perp } \).

Lemma 35
#

A polyhedral cone has only finitely many faces.

Proof

By 34, a face of a polyhedral cone \( \sigma = \operatorname{Cone}(S) \) is of the form \( \operatorname{Cone}(S') \) for some \( S' \subseteq S \). Being finite, \( S \) has only finitely many subsets, which shows the claim.

Lemma 36 Intersection of faces

If \( \sigma \) is a pointed cone, then the intersection of two faces of \( \sigma \) is a again face of \( \sigma \).

Proof

Let \( \tau = \sigma \cap u^{\perp } \) and \( \tau ' = \sigma \cap \tau '^{\perp } \) be faces. We claim that \( \sigma \cap u^{\perp } \cap u'^{\perp } = \sigma \cap (u + u')^{\perp } \). The inclusion “\( \subseteq \)” is obvious. For the converse, suppose \( v \in \sigma \) with \( \langle v, u \rangle + \langle v, u' \rangle = 0 \). Since \( u, u' \in \sigma ^{\vee } \), both summands are nonnegative, hence \( \langle v, u \rangle = \langle v, u' \rangle = 0 \).

Lemma 37

Let \( \tau \preceq \sigma \) be a face of a pointed cone. Then \( \tau ^{\vee } = \tau ^{\perp } + \sigma ^{\vee } \).

Proof

It’s easy to show that \( \tau = \operatorname{Span}_{\Bbbk }(\tau ) \cap \sigma ^{\vee } \). Dualising gives \( \tau ^{\vee } = \operatorname{Span}_{\Bbbk }(\tau )^{\vee } + \sigma ^{\vee } \). Apply 31 to get \( \operatorname{Span}_{\Bbbk }(\tau )^{\vee } = \operatorname{Span}_{\Bbbk }(\tau )^{\perp } = \tau ^{\perp } \).

Lemma 38 Face of a face

A face of a face of a polyhedral cone \( \sigma \) is again a face of \( \sigma \).

Proof

Let \( \tau = \sigma \cap u^{\perp } \) with \( u \in \sigma ^{\vee } \) and \( \rho = \tau \cap v^{\perp } \) with \( v \in \tau ^{\vee } \). By 37, we can write \( v = v_1 + v_2 \) with \( v_1 \in \tau ^{\perp } \) and \( v_2 \in \sigma ^{\vee } \). Then \( u + v_2 \in \sigma ^{\vee } \) and we claim \( \rho = \sigma \cap (u + v_2)^{\perp } \). For “\( \subseteq \)”, let \( x \in \rho = \tau \cap v^{\perp } \). Since \( x \in \tau \), we get \( \langle x, u \rangle = 0 \) and \( \langle x, v_1 \rangle = 0 \). Since \( x \in v^{\perp } \), also \( \langle x, v \rangle = 0 \), which by \( v = v_1 + v_2 \) implies \( \langle x, v_2 \rangle = 0 \). In total, \( \langle x, u + v_2 \rangle = 0 \). For the converse, assume \( x \in \sigma \cap (u + v_2)^{\perp } \), i.e. \( \langle x, u \rangle + \langle x, v_2 \rangle = 0 \). Since both \( u, v_2 \in \sigma ^{\vee } \) and \( x \in \sigma \), both summands are nonnegative, hence \( \langle x, u \rangle = \langle x, v_2 \rangle = 0 \). This implies \( x \in \tau \), thus also \( \langle x, v_1 \rangle = 0 \). In total, \( \langle x, v \rangle = \langle x, v_1 \rangle + \langle x, v_2 \rangle = 0 \), thus \( x \in \rho \).

In particular, Lemma 38 establishes a partial order on the set of faces of a polyhedral cone \( \sigma \) (written \( \operatorname{faces}(\sigma ) \)). It is finite by  35.

Lemma 39 Face order is a lattice

Given a polyhedral cone, the partial order on its faces is a lattice. We call it the face lattice of the cone.

Proof

Top is the entire cone, bot is the empty cone.

Lemma 40 Face lattice is graded
#

Given a polyhedral cone, the partial order on its faces is graded.

Proof

The grading is given by the number of faces of this face.

Definition 41 Combinatorial Equivalence
#

Two polyhedral cones are combinatorially equivalent if there is an order isomorphism between their face lattices.

Proposition 42

Face lattices of dual cones are dual to each other.

Proposition 43

The face lattice of a face is isomorphic to a lower interval of the face lattice of a polyhedral cone.

Proposition 44

Linear transformaions preserve combinatorial equivalence. There is an iso that should be provided

0.1.5 Polarity

Definition 45 Polar Cone

For a pointed cone \( \sigma \), the polar cone of \( \sigma \) is the pointed cone

\[ S^\circ := \{ x \in \sigma \mid \forall y \in \sigma , \langle x, y \rangle \leq 0 \} \]

0.2 Polytopes

Definition 46 Polytope

A polytope is a set that can be written as \(\operatorname{Conv}(S)\) for some finite set \(S\).