Polyhedral computations in computational algebraic geometry and optimization

Polyhedral computations in computational
algebraic geometry and optimization
Vissarion Fisikopoulos
Computer Science Department, Algorithms Group
LSE, Lunchtime Seminar, 15 May 2015

Outline of the talk
Polytopes deﬁned by oracles: computation & combinatorics
Enumeration of 2-level polytopes

Motivation: resultant polytopes
Algebra:
complexity of resultant polynomial
Geometry:
generalize Birkhoﬀ polytopes
faces are Minkowski sums of resultant polytopes
vertices → triangulations ↔ subdivisions of Mink. sums
Applications:
support computation → interpolate implicit equation of
parametric hypersurface
compute resultants, discriminants
Examples of resultant polytopes

Newton polytope
Deﬁnition
Given polynomial f ∈ K[x1, . . . , xn] the Newton polytope N(f) of
f is the convex hull of the support, i.e. exponent vectors of
monomials with non-zero coeﬃcient.
3
2
1 2 3 5
5
f(x1, x2) = 8x2 + x1x2 − 24x2
2 −
16x2
1 + 220x2
1x2 − 34x1x2
2 −
84x3
1x2 +6x2
1x2
2 −8x1x3
2 +8x3
1x2
2 +
8x3
1 + 18x3
2
N(f)
1

Polytopes and Algebra
Definition
Given are polynomials f0, f1, . . . , fn ∈ K[x1, . . . , xn], s.t. the
supports define an essential family A0, A1, . . . , An ⊂ Zn, i.e. the
Ai generate Zn and any k-subset generates a sublattice of
dimension ≥ k.
The system’s (sparse) resultant R is the polynomial in the system’s
coefficients, defined up to sign, which vanishes iff the polynomials
have a common root in the corresponding toric variety X:
(K
∗
)n ⊂ X.
The resultant polytope N(R) is the Newton polytope of R.
A0
A1
N(R)R(a, b, c, d, e) = ad2
b + c2
b2
− 2caeb + a2
e2
f0(x) = ax2
+ b
f1(x) = cx2
+ dx + e

Birkhoﬀ polytope
Linear polynomials
A0
A1
N(R)
f0(x, y) = ax + by + c
f1(x, y) = dx + ey + f
f2(x, y) = gx + hy + iA2
a b c
d e f
g h i
4-dimensional Birkhoff polytope
R(a, b, c, d, e, f, g, h, i) =

Existing work
Resultants, secondary polytopes, Cayley trick [GKZ ’94]
TOPCOM [Rambau ’02] computes all vertices of secondary
polytope.
[Michiels & Verschelde DCG’99] coarse equivalence classes of
secondary polytope vertices.
[Michiels & Cools DCG’00] decomposition of Σ(A) in
Minkoski summands, including N(R).
Tropical geometry [Sturmfels-Yu ’08]: algorithms for resultant
polytope (GFan library) [Jensen-Yu ’11] and discriminant
polytope (TropLi software) [Rincn ’12].

Regularity
Regular subdivision of A ⊂ Rd are obtained by projecting the lower
(or upper) hull of A lifted to Rd+1 via a lifting function
w ∈ (R|A|)×.
w = (2, 1, 4)w = (2, 6, 4)
A

Resultant polytope vertices and mixed subdivisions
A subdivision S of A0 + A1 + · · · + An is
mixed when its cells have expressions as Minkowski sums of
convex hulls of point subsets in Ai’s,
ﬁne when each cell has dimension equal to the sum of its
summands dimensions.
Example
mixed subdivision S of A0 + A1 + A2
A0
A1
A2

Resultant polytope vertices and mixed subdivisions
A subdivision S of A0 + A1 + · · · + An is
mixed when its cells have expressions as Minkowski sums of
convex hulls of point subsets in Ai’s,
fine when each cell has dimension equal to the sum of its
summands dimensions.
Theorem [GKZ ’94, Sturmfels ’94]
many-to-one relation between regular fine mixed subdivisions
and N(R) vertices
one-to-one relation between regular fine mixed subdivisions
and secondary polytope Σ(A) vertices

The idea of the algorithm for N(R)
Input: A ∈ Z2n deﬁned by A0, A1, . . . , An ⊂ Zn
Simplistic method:
compute the secondary polytope Σ(A)
many-to-one relation between vertices of Σ(A) and N(R)
Cannot enum 1 representative/class by walking on secondary edges

The idea of the algorithm for N(R)
Input: A ∈ Z2n deﬁned by A0, A1, . . . , An ⊂ Zn
New Algorithm:
Vertex oracle: given direction vector compute a vertex of
N(R) by computing a subdivision using the direction as lifting
Output sensitive: computes only one subdivision of A per
N(R) vertex + one per N(R) facet
Computes projections of N(R) or Σ(A)

Incremental algorithm for N(R)
ﬁrst: compute conv.hull of d + 1 aﬀ. indep. vertices of N(R)
step: call the oracle with outer normal vector of a halfspace
→ either validate this halfspace
→ or add a new vertex to the convex hull
N(R)
Q

Theorem (Emiris,F,Konaxis,Penaranda)
Given P ⊆ Rd, H-, V-repr. & triang. T of N(R) can be computed
in
O(d5
ns2
) arithmetic operations + O(n + m) oracle calls
s is the number of cells of T.

Theorem (Emiris,F,Konaxis,Penaranda)
Given P ⊆ Rd, H-, V-repr. & triang. T of N(R) can be computed
in
O(d5
ns2
) arithmetic operations + O(n + m) oracle calls
s is the number of cells of T.
BUT: s can be O n d/2

ResPol package
Towards high-dimensional CGAL (Computational Geometry
Algorithms Library)
Hashing of determinantal predicates scheme: optimizing
sequences of similar determinants (x100 speed-up)
Computes 5-, 6- and 7-dimensional polytopes with 35K, 23K
and 500 vertices, respectively, within 2hrs
Computes polytopes of many important surface equations
encountered in geometric modeling in < 1sec, whereas the
corresponding secondary polytopes are intractable
http://sourceforge.net/projects/respol

Combinatorics of resultant polytopes
[GKZ’90] Univariate case, general-dimensional N(R):
The Ai are multisets from Z: |A0| = k0 + 1, |A1| = k1 + 1 ⇒
⇒ dim N(R) = k0 + k1 − 1, k0+k1
k0
vertices, k0k1 + 3 facets.
[Sturmfels’94] Multivariate case / up to 3 dimensions
The only resultant polytopes up to dimension 3

One step beyond: 4-dimensional N(R)
f-vector of face cardinalities: of vertices, edges, ridges, facets.
Some f-vectors (generic input):
(5, 10, 10, 5): 4-simplex
(6, 15, 18, 9): Birkhoﬀ
(8, 20, 21, 9)
(9, 22, 21, 8)
. . .
(10, 26, 25, 9): Sylvester, ki ∈ {2, 3}
. . .
(17, 50, 50, 17)
(18, 51, 48, 15)
(18, 51, 49, 16)
(18, 52, 50, 16)
(18, 52, 51, 17)
(18, 53, 51, 16)
(18, 53, 53, 18)
(18, 54, 54, 18)
(19, 54, 52, 17)
(19, 55, 51, 15)
(19, 55, 52, 16)
(19, 55, 54, 18)
(19, 56, 54, 17)
(19, 56, 56, 19)
(19, 57, 57, 19)
(20, 58, 54, 16)
(20, 59, 57, 18)
(20, 60, 60, 20)
(21, 62, 60, 19)
(21, 63, 63, 21)
(22, 66, 66, 22)

Combinatorics of 4-dim resultant polytopes
Theorem (Dickenstein,Emiris,F)
Given essential family A0, A1, . . . , An ⊂ Zn, with N(R) of
dimension 4, N(R) is (a degeneration of) any of the following
polytopes:
(i) |Ai| : 2 . . . 2, 5, N(R) is the 4-simplex, f-vector (5, 10, 10, 5).
(ii) |Ai| : 2 . . . 2, 3, 4, N(R) f-vector (10, 26, 25, 9).
(iii) |Ai| : 2 . . . 2, 3, 3, 3, N(R) has maximal face numbers
˜f3 = 22, ˜f2 = 66, ˜f1 = ˜f0 + 44, and 22 ≤ ˜f0 ≤ 28.
Degenarations can only decrease the number of faces.
Previous upper bound for vertices yields 6608 [Sturmfels’94].
Focus on new case (iii): reduces to n = 2 and
|A0| = |A1| = |A2| = 3

Mixed subdivisions and N(R) faces (I)
Proposition (GKZ,Sturmfels)
Consider the regular mixed subdivision S of A0 + A1 + · · · + An,
obtained by a lifting deﬁned by w ∈ Rm. Then, S deﬁnes a face of
N(R) which has w as outer normal, equal to the Newton polytope
of
σ∈S
R(f0|σ, . . . , fn|σ)dσ
,
i.e. the Minkowski sum of the resultant polytopes of subsystems
{f0|σ, . . . , fn|σ} correspods to cells σ ∈ S, where dσ is the
normalized volume of σ.
Mink. sum of N(R) triangle and N(R) segmentsubd. S of A0 + A1 + A2

Genericity maximizes complexity (II)
Theorem
The number of N(R) faces for 3 triangles is maximized for generic
triangles, namely 2-d, without parallel edges.
N(R∗
) f-vector: (18, 52, 50, 16)
N(R) f-vector: (14, 38, 36, 12)
p
p∗
A0 A1 A2
A0 A1 A2

Possible facets
Lemma
resultant facet: 3-d N(R): octagon in S,
prism: 2-d N(R) (triangle) + 1-d N(R): heptagon and
hexagon,
zonotope: 1-d N(R) + 1-d N(R) + 1-d N(R): 3 hexagons.
3D
2D

Counting facets and duality of mixed subdivisions (III)
Lemma
Maximal face numbers are as follows:
9 resultant facets: 9 octagons with |Ai| = 3, 3, 2.
9 prisms: 9 hexagon-heptagon pairs:
unique subdivision if common edge
picked, i.e., common dual ray ﬁxed.
4 zonotopes: 4 triplets of tri-chromatic points.

Extensions - Open problems
Algorithmic
Total polynomial algorithms for CH (edge-directions
[Emiris-F-Gartner])
Volume computation (randomized implementation [Emiris-F])
Lattice points enumeration
Combinatorial
The maximum f-vector of a 4d N(R) is (22, 66, 66, 22)
Explain symmetry of maximal f-vectors

Enumeration of 2-level polytopes
Joint work with:
Adam Bohn (now in Tailand)
Yuri Faenza (now at EPFL)
Samuel Fiorini (ULB)
Marco Macchia (ULB)
Kanstantsin Pashkovich (now at Waterloo)

H0 H1
Deﬁnition (#1)
A polytope P is 2-level if ∀ facet-deﬁning hyperplane H0 ∃ a
parallel hyperplane H1 such that: vert(P) ⊆ H0 ∪ H1
4, 6, 4 5, 8, 5 6, 9, 5 6, 12, 8 8, 12, 6
7, 12, 7 5, 9, 6 6, 11, 7

Deﬁnition (#2)
A polytope P is 2-level iﬀ its slack matrix is 0/1 (perhaps after
scaling some facets)
Not invariant under polarity:
P = P∆
=
S(P) =






0 0 0 2 2 2
2 2 2 0 0 0
0 0 3 0 0 3
0 3 0 0 3 0
3 0 0 3 0 0






S(P ) =








0 2 0 0 3
0 2 0 3 0
0 2 3 0 0
2 0 0 0 3
2 0 0 3 0
2 0 3 0 0









Motivations for studying 2-level polytopes
Algebraic combinatorics / Erhart polynomials (Stanley ’80)
Statistical disclosure elimination (Sullivant ’06)
Centrally symmetric polytopes (Sanyal, Werner, Ziegler ’09)
Theta bodies (Gouveia, Parrilo & Thomas ’10)
Communication complexity (log-rank conjecture)
Combinatorial optimization (what do 2-level polytopes
capture?)

Examples of 2-level polytopes
Birkhoﬀ polytopes := convhull of permutation matrices
Hanner polytopes := iterated products / free sums of
segments
Stable set polytope STAB(G) with G perfect
Hansen polytopes := twisted prisms over STAB(G), G perfect
{x ∈ [0, 1]d | Ax = b} where A is totally unimodular and b
integer

General properties of 2-level polytopes (2LPs)
A d-dim 2LP has at most 2d vertices and facets (GPT ’10)
A polytope is 2-level iﬀ its Theta rank is 1
(it is a projection of a spectahedron) (GPT ’10)
Every face of a 2LP is a 2LP

General properties of 2-level polytopes (2LPs)
A d-dim 2LP has at most 2d vertices and facets (GPT ’10)
A polytope is 2-level iff its Theta rank is 1
(it is a projection of a spectahedron) (GPT ’10)
Every face of a 2LP is a 2LP
The combinatorial type of a 2LP determines its affine type
A 2LP has a simple vertex iff it is STAB(G), G perfect

Simplicial cores
In 2-level polytope P, pick
d + 1 vertices v1, . . . , vd+1
d + 1 facets F1, . . . , Fd+1
such that ∀i : vi /∈ Fi and vi+1, . . . , vd+1 ∈ Fi

Simplicial cores
d + 1 vertices v1, . . . , vd+1
d + 1 facets F1, . . . , Fd+1
v1
F1

Simplicial cores
d + 1 vertices v1, . . . , vd+1
d + 1 facets F1, . . . , Fd+1
v1
F1
Equivalently, ﬁnd following submatrix in S(P):









1 0 0 0 · · · 0 0
∗ 1 0 0 · · · 0 0
∗ ∗ 1 0 · · · 0 0
...
...
...
∗ ∗ ∗ ∗ · · · 1 0
∗ ∗ ∗ ∗ · · · ∗ 1










Embeddings









1 0 0 0 · · · 0 0
∗ 1 0 0 · · · 0 0
∗ ∗ 1 0 · · · 0 0
.
.
.
.
.
.
.
.
.
∗ ∗ ∗ ∗ · · · 1 0
∗ ∗ ∗ ∗ · · · ∗ 1









=






0
M
.
.
.
0
∗ · · · ∗ 1






Lemma
For 2L P an x-embedding has 0/1 facets:
P = {x ∈ Rd
| ∀E ∈ E : 0
i∈E
xi 1}
for some E with subsets of [d] and vert(P) ⊆ M−1{0, 1}d ⊆ Zd
(E contains all the subsets of [d] if P is the simplex)
Lemma
For 2L P a y-embedding is P = conv(X) for X ⊆ {0, 1}d
Remark: The two embeddings linked: y = Mx ⇐⇒ x = M−1y

A proxy for 2LP: closed sets
Deﬁnition
I := M−1 · {0, 1}d then A ⊆ I is closed if clI(A) = A.
E(A) :=
x∈A
{E ⊆ [d] | 0 ≤ x(E) ≤ 1} .
clI(A) := {x ∈ I | 0 ≤ x(E) ≤ 1 for every E ∈ E(A)} .
Lemma
If 2L P in x-embedding then the vertex set of P is a closed set wrt
M−1 · {0, 1}d.

The enumeration algorithm
Input: List Ld−1 of 2L polytopes & simplicial cores
1. Foreach P1 ∈ Ld−1 & simplicial core Γ1: Md−1 := M(Γ1)
1.1 Complete Md−1 to a (d × d)-matrix in the following way:
Md :=







1 0 · · · 0
0
b1
...
bd−2
Md−1







where b = (b1, . . . , bd−2) ∈ {0, 1}d−2
.
1.2 Foreach b1, . . . , bd−2 ∈ {0, 1}:
1.2.1 Using the Ganter-Reuter algorithm, compute the list A of
closed sets wrt M−1
d · ({1} × {0, 1}d−1
)
1.2.2 Foreach A ∈ A, let P := conv ({0} × P1) ∪ A .
1.2.3 If P is 2-level & not isomorphic to any P ∈ Ld, add P to Ld.

Experimental results
d 2L ∆-f STAB polar CS Birk 0/1 closed sets
3 5 4 4 4 2 4 8 19
4 19 12 11 12 4 11 192 350
5 106 41 33 42 13 33 1,048,576 21239
6 1150 248 148 276 45 129 ∼ 1.8 · 1019
1.05 · 108
7 - - 906 - 238 661 - -
8 - - 8887 - - 4530 - -
Combinatorially equivalent 0/1 polytopes and 2L polytopes
∆-f: with on simplicial facet
STAB: stable sets of perfect graphs [Hougardy06]
polar: 2-level polytopes whose polar is 2-level
CS: centrally symmetric
Birk: Birkhoﬀ polytope faces [Paﬀenholz13]
’-’: exact numbers are unknown.

Statistics: facets vs vertices
0
10
20
30
40
50
60
70
0 10 20 30 40 50 60 70
facets
vertices
2-level
centrally-symmetric
f(x) = 64*12/x
The relation between the number of facets and the number of
vertices of 6-dim 2-level polytopes

Statistics: number of 2L
0
20
40
60
80
100
120
0 10 20 30 40 50 60 70 80
2Lpolytopes
vertices
no simplicial facet
one simplicial facet
The number of 6-dim 2-level polytopes and the class with the ones
with a simplicial facet as a function of the number of vertices.

Open questions
Output-sensitive enumeration algorithm for 2L polytopes
(Hint: better proxy)
g(d) := #(d-dimensional 2L polytopes, up to isomorphism)
Is g(d) = 2poly(d)?
Known: g(d) 2Ω(d2) (e.g., STAB(G) with G bipartite)

Open questions
Output-sensitive enumeration algorithm for 2L polytopes
(Hint: better proxy)
g(d) := #(d-dimensional 2L polytopes, up to isomorphism)
Is g(d) = 2poly(d)?
Known: g(d) 2Ω(d2) (e.g., STAB(G) with G bipartite)
xc(P) = extension complexity = min # facets of a lift of P
f(d) := max{xc(P) | P is d-dimensional 2L polytope}
Is f(d) = 2polylog(d)? (“log-rank conj. for slack matrices”)
xc(STAB(G)) 2O(log2
n)
for n-vertex perfect G (Yannakakis’91)
f(d) 2
˜O(
√
d) for d-dimensional 2LP (Lovett ’14)
g(d) 2O[poly(d)f2(d)] (Rothvoss ’11)

Polyhedral computations in computational algebraic geometry and optimization

More Related Content

What's hot (20)

Similar to Polyhedral computations in computational algebraic geometry and optimization (20)

More from Vissarion Fisikopoulos (19)

Recently uploaded (20)

Polyhedral computations in computational algebraic geometry and optimization