Extension Complexity of Matroid Polytopes*

Ayush Agarwal Soham Joshi

[email protected] [email protected]
IIT Bombay IIT Bombay

May 14, 2024

Abstract

Given a matroid, or any polytope, its extension complexity was characterised Yannakakis
[Yan88]. Examining extension complexity of polytopes gives insight into time complexity of
solving optimisation problems using LP based methods. In this survey, we examine the extension
complexity of some matroid polytopes.

* This was a credited research project under the supervision of Prof. Rohit Gurjar at IIT Bombay.
Contents
1 Introduction 1

2 Preliminaries 1
2.1 Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1.1 Matroid Rank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1.2 Matroid Polytope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.3 Matroid Intersection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.4 Matroid Intersection Polytope . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.5 Chaining tight sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.6 Partition Matroid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.7 Partition Matroid Polytope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.8 Matroid Union . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Extension Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Non-negative rank characterization . . . . . . . . . . . . . . . . . . . . . . . . 6

3 Results 6
3.1 Dilworth truncation of matroids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.1.2 Lower truncated tranversal and 𝑘 − 𝑙 sparsity polymatroid . . . . . . . . . . . 7
3.1.3 Extension complexity of lower truncated transversal polymatroid . . . . . . . 8
3.2 Relating communication and extension complexity . . . . . . . . . . . . . . . . . . . . 9
3.2.1 Deterministic Communication Protocols . . . . . . . . . . . . . . . . . . . . . 9
3.2.2 Randomized Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2.3 Factorisation and protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.3 𝑘 − 𝑙 sparsity matroids (Alternate method) . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3.1 Protocol when 𝑘 ≥ 𝑙 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3.2 Protocol for 𝑘 ≤ 𝑙 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Hitting Set of Matroid Polytopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.4.1 Hitting Set and Extension Complexity . . . . . . . . . . . . . . . . . . . . . . . 14
3.4.2 Relation with extension of dual matroid . . . . . . . . . . . . . . . . . . . . . . 15
3.4.3 Relation with (co)-graphic matroid polytope . . . . . . . . . . . . . . . . . . . 16
3.5 2-sums of Matroid Polytopes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.5.1 Connection with hitting set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.5.2 Connection with regular matroids . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.6 Extension of Polygons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.6.1 𝑂(log 𝑛) extension √ for regular polygons . . . . . . . . . . . . . . . . . . . . . . 18
3.6.2 Lower bound of 2𝑛 for generic 𝑛-gons . . . . . . . . . . . . . . . . . . . . . . 19
3.7 Exact Spanning Tree, Matchings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.7.1 Exact weighted spanning tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.7.2 Exact or odd spanning tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2
4 Discussion 23

5 Acknowledgements 23

References 23
1 Introduction
The extension complexity of a polytope 𝑃 is the smallest number of facets among convex polytopes
𝑄 that can have 𝑃 as a projection. In this case, 𝑄 is called an extended formulation of 𝑃 and it may
have a higher dimension than 𝑃. The extension complexity of a polytope is a natural question to
ask since many problems can be encoded as a combination of linear constraints. If this admits
a small extended formulation, then standard techniques like linear program solvers can be used
to solve the problem. In 1980s, Swart [Swa87] attempted to prove that there is a polytope with
only polynomial many facets that projects down to the TSP polytope. Note that a polynomial size
extended formulation for the TSP polytope would imply 𝑃 = 𝑁P. The purported linear programs
were extremely complicated to analyze. However, in a breakthrough paper, Yannakakis [Yan88]
refuted all such attempts by showing that every symmetric linear program (LP) for the TSP poly-
tope has exponential size. Since the proposed LP by Swart was symmetric, that refuted Swart’s proof.

However, the question of extension complexity of non-symmetric LPs remained open. How-
ever, in 2010, Kaibel, Pashkovich, and Theis [KPT10] gave examples of polytopes that do not have
polynomial size symmetric extended formulations, but that do have polynomial size asymmetric
extended formulations. This rekindled interest in the lingering question. In another breakthrough
paper in 2012, Fiorini, Masser, Pokutta, Tiwary, and de Wolf [FMP+ 12] finally proved that the TSP
polytope does not admit any polynomial size extended formulation, symmetric or not.

In this survey we examine many matroid polytopes and give an overview of known exten-
sion complexities and open questions. We begin with k-l sparsity matroids, which can be viewed
as dilworth truncation of spanning trees which is known to have a small extension as shown in
[IKK+ 16]. Then, we examine extension complexity from the lens of hitting sets and communication
complexity ([Apr22], [FFGT15], [KLdW15]). Hitting set of polytopes is quite interesting and can be
possibly used as an alternative method for small extension of regular matroids. As a first step, we
see the effect of 2-sums on size of hitting set using [ACF18]. Moreover, we examine effect of this on
extension of regular matroids.

Moreover, we examine the extension complexity of polygons as shown in [FRT12]. A related

question was raised on Open Problem garden, which asks if there exist infinitely many 𝑛 such that a
convex polygon on 𝑛 vertices has extension complexity Ω(𝑛). This was answered by Shitov [Shi20],
who showed that every 𝑛-gon is a projection of polytope with atmost 147 · 𝑛 2/3 facets. Finally, we
examine the extension complexity of exact bipartite matching polytope which was shown to be
exponential by Svennson et. al. [JSY].

2 Preliminaries
Matroid theory originated in the middle of the 1930s. There is a huge literature on matroids by now.
For an introduction, see for example the excellent textbooks of Oxley [Oxl06] or Schrijver [S+ 03].

1
Moreover, our problem also deals with the extension complexity of matroids, for which refer the
excellent lecture notes by Li [Li18] as an introduction. Below we give some basic definitions and
facts about matroids and extension complexity of polytopes.

2.1 Matroid
A matroid 𝑀 is a pair 𝑀 = (𝐸, 𝐼), where 𝐸 is the finite ground set and 𝐼 ⊆ 𝑃(𝐸) is a nonempty
family of subsets of E that satisfies the following two axioms.

1. Closure under subsets. For every 𝐼 ∈ ℐ and 𝐽 ⊆ 𝐼 we have 𝐽 ∈ ℐ.

2. Augmentation property. For every 𝐼, 𝐽 ∈ ℐ where |𝐼 | < |𝐽 |, there is an 𝑗 ∈ 𝐽 such that 𝐼 ∪ 𝑗 ∈ ℐ.

We denote 𝑚 = |𝐸| throughout the paper. The sets in 𝐼 are called the independent sets of 𝑀. An
inclusion-wise maximal set 𝐵 ∈ ℐ is called a base. Note that by the augmentation property, all base
sets have the same size. Let ℬ ∈ ℐ denote the collection of base sets.

Definition 2.1 (Dual Matroid). The dual of a matroid 𝑀 is another matroid 𝑀 ∗ that has the same
elements as 𝑀, and in which a set is independent if and only if 𝑀 has a basis set disjoint from it.
Alternatively, basis sets of dual matroid are complements of basis sets of 𝑀.

Definition 2.2 (Circuits). For a matroid M, any inclusion-wise minimal dependent subset is called a
circuit.

Similarly, we can define the notion of co-circuits as circuits in the dual matroid.

2.1.1 Matroid Rank

. Motivated by Linear Algebra, there is a rank-function of a matroid that is defined for every subset
𝐴 ⊆ 𝐸 as the size of the largest independent set that is contained in 𝐴,

𝑟 𝑎𝑛 𝑘(𝐴) = 𝑚𝑎𝑥{|𝐼 | | 𝐼 ∈ ℐ and 𝐼 ⊆ 𝐴}

The size of every maximal independent set is rank(E). This number is called the rank of M. The
matroid problem is to compute a maximal independent set.
An important property of the rank-function is its submodularity. In general, a function
𝑓 : 𝑃(𝐸) → ℝ is called submodular, if for any sets 𝑆, 𝑇 ⊆ 𝐸, we have

𝑓 (𝑆) + 𝑓 (𝑇) ≥ 𝑓 (𝑆 ∪ 𝑇) + 𝑓 (𝑆 ∩ 𝑇)

2
2.1.2 Matroid Polytope

The polytopes we consider in this paper are convex polytopes defined as the convex hull of finitely
many points in ℝ𝑚 . Any convex polytope 𝑃 can be described as the intersection of halfspaces, i.e.,
as 𝑃 = {𝑥 ∈ ℝ𝑚 |𝐴𝑥 ≤ 𝑏}, for some matrix 𝐴 ∈ ℝ 𝑘×𝑚 and vector 𝑏 ∈ ℝ 𝑘 . A face of the polytope 𝑃 is
the set of points in 𝑃 minimizing or maximizing a linear function. If the polytope P is described by
′ ′ ′ ′
𝐴𝑥 ≤ 𝑏, then any face of 𝑃 can be described as {𝑥 ∈ 𝑃 | 𝐴 𝑥 ≤ 𝑏 }, where (𝐴 𝑏 ) is some subset of
the rows of (𝐴 𝑏). With every matroid, there is an associated matroid polytope. This polytope is
crucial for our arguments.
For a set 𝐼 ∈ 𝐸, its characteristic vector 𝑥 𝐼 ∈ 𝑅 𝐸 is defined as
(
1, if 𝑒 ∈ 𝐼
𝑥=
0, otherwise

For any collection of sets 𝐴 ⊆ 𝑃(𝐸), the polytope 𝑃(𝐴) ⊂ 𝑅 𝐸 is defined as the convex hull of the
characteristic vectors of the sets in 𝐴,

𝑃(𝐴) = 𝑐𝑜𝑛𝑣{𝑥 𝐼 | 𝐼 ∈ 𝐴}

For a matroid 𝑀 = (𝐸, 𝐼), its matroid polytope is defined as 𝑃(𝐼) ⊂ ℝ𝐸 , i.e., the convex hull of the
characteristic vectors of the independent sets. The points {𝑥 𝐼 | 𝐼 ∈ ℐ} are the corners of the matroid
polytope 𝑃(𝐼). Edmonds [Edm70] gave a simple description of this polytope which uses the rank
function of the matroid (see also [Sch03]). For convenience, we define for any 𝑥 ∈ ℝ𝐸 and 𝑆 ⊆ 𝐸,
Õ
𝑥(𝑆) = 𝑥𝑒
𝑒∈𝑆

Lemma 2.3 ([Edm03]). For a matroid (𝐸, 𝐼) with rank function 𝑟, a point 𝑥 ∈ ℝ𝐸 is in 𝑃(𝐼) iff

𝑥 𝑒 ≥ 0 ∀ 𝑒 ∈ 𝐸(1) (1)
𝑥(𝑆) ≤ 𝑟(𝑆) ∀ 𝑆 ⊆ 𝐸 (2)

It is easy to see that any 0-1 corner of the polytope given by (1) and (2) corresponds to an
independent set in ℐ. The nontrivial part is to show that the described polytope does not have a
non-integral corner. Let ℬ be the family of base sets of the matroid (𝐸, ℐ). Let n be the rank of the
matroid, i.e., the size of any base set. The matroid base polytope, defined as 𝑃(𝐵), is clearly a face
of the matroid polytope 𝑃(𝐼). Putting the following equation together with (1) and (2) will give a
description of 𝑃(𝐵),

𝑥(𝐸) = 𝑛 (3)

2.1.3 Matroid Intersection

The matroid intersection problem is, given two matroids 𝑀1 = (𝐸, 𝐼1 ) and 𝑀2 = (𝐸, 𝐼2 ) over the
same ground set, compute a maximum size set in 𝐼1 ∩ 𝐼2 , the common independent sets. Note that
in general (𝐸, 𝐼1 ∩ 𝐼2 ) is not a matroid anymore.

3
2.1.4 Matroid Intersection Polytope

The intersection of two matroids also has an easy polytope description: Edmonds [Edm70] showed
a surprising result that one can describe the matroid intersection polytope 𝑃(𝐼1 ∩ 𝐼2 ) just by putting
together the constraints of the two matroid polytopes 𝑃(𝐼1 ) and 𝑃(𝐼2 ) (see also [S+ 03])

Theorem 2.4 ([Edm03]). For two matroids (𝐸, 𝐼1 ) and (𝐸, 𝐼2 ),

𝒫(𝑀1 ∩ 𝑀2 ) = 𝒫(𝑀1 ) ∩ 𝒫(𝑀2 )

That is, a point 𝑥 ∈ ℝ𝐸 is in the polytope 𝒫(𝑀1 ∩ 𝑀2 ) iff

𝑥𝑒 ≥ 0 ∀ 𝑒 ∈ 𝐸 (4)
𝑥(𝑆) ≤ 𝑟1 (𝑆) ∀ 𝑆 ⊆ 𝐸 (5)
𝑥(𝑆) ≤ 𝑟2 (𝑆) ∀ 𝑆 ⊆ 𝐸, (6)

where 𝑟1 and 𝑟2 are the rank functions of the two matroids, respectively.

Let ℬ1 and ℬ2 be the families of base sets of the matroids (𝐸, 𝐼1) and (𝐸, 𝐼2), respectively. Note
that there can be a common base set only if the two matroids have same rank, say n. To obtain the
common base polytope 𝑃(𝐵1 ∩ 𝐵2 ) one just needs to put the constraint 3 together with inequalities
4, 5 and 6.

2.1.5 Chaining tight sets

In this section, we state a lemma (see [S+ 03]) which expresses any tight constraint as a linear
combination of tight constraints obtained by a chain of sets.

Definition 2.5 (Tight sets). Let ℳ be a matroid with the rank function 𝑟. Let 𝒙 ∈ 𝒫(ℳ) (matroid
polytope). We call a set 𝑆 a tight set of 𝒙 with respect to 𝑟 if 𝑥(𝑆) = 𝑟(𝑆)

Lemma 2.6 (Uncrossing operation). Let ℳ be a matroid with rank function 𝑟, and let 𝒙 ∈ 𝒫(ℳ). If 𝑆
and 𝑇 are tight sets of 𝒙 with respect to 𝑟, then so are 𝑆 ∪ 𝑇 and 𝑆 ∩ 𝑇.

Lemma 2.7 (Maximal Chain of Tight Sets). Let ℳ be a matroid with rank function 𝑟, and let 𝒙 ∈ 𝒫(ℳ).
Let 𝒞 = {𝐶1 , · · · , 𝐶 𝑘 } with 𝜙 ⊂ 𝐶1 ⊂ · · · ⊂ 𝐶 𝑘 be an inclusion-wise maximal chain of tight sets of 𝒙 with
respect to 𝑟. Then every tight set 𝑇 of 𝒙 with respect to 𝑟 must satisfy 𝜒𝑇 ∈ span{𝜒𝐶 : 𝐶 ∈ 𝒞}

2.1.6 Partition Matroid

Partition matroid is a matroid in which 𝐸 is partitioned into (disjoint) sets 𝐸1 , 𝐸2 , .., 𝐸 𝑙 and

𝐼 = {𝑋 ⊆ 𝐸 : |𝑋 ∩ 𝐸 𝑖 | ≤ 𝑘 𝑖 ∀ 𝑖 = 1, .., 𝑙},

for some given parameters 𝑘 1 , .., 𝑘 𝑙

4
2.1.7 Partition Matroid Polytope

For a given partition matroid 𝑃 = (𝐸, ℐ) and partition 𝐸1 , 𝐸2 , .., 𝐸 𝑙 , a point 𝑥 ∈ 𝑅 𝐸 is in the polytope
𝑃(𝐼1 ∩ 𝐼2 ) iff

𝑥𝑒 ≥ 0 ∀ 𝑒 ∈ 𝐸 (7)
𝑥(𝐸 𝑖 ) ≤ 𝑟1 (𝐸 𝑖 ) ∀ 𝑖 (8)

2.1.8 Matroid Union

Matroid union of 𝑀1 , 𝑀2 , 𝑀3 , .., 𝑀 𝑘 is defined as

𝑘 𝑘
𝑀 = 𝑀1 ∨ 𝑀2 ∨ .. ∨ 𝑀 𝑘 = (∪𝑖=1 𝑆 𝑖 , ℐ = {∪𝑖=1 𝐼 𝑖 |𝐼 𝑖 ∈ ℐ})

We can show that M is a matroid, and derive the rank function for it which is given by,

𝑘
" #
Õ
𝑟𝑚 (𝑈) = min 𝑈\𝑇 + 𝑟 𝑀𝑖 (𝑇 ∩ 𝑆 𝑖 )
𝑇⊆𝑈
𝑖=1

2.2 Extension Complexity

Given a polytope 𝑃 ⊆ ℝ𝑛 as a convex hull of finitely many points (called vertices) in ℝ𝑛 , by
Minkowski- Weyl theorem, it is equivalent to a bounded polyhedron 𝑄 that is described by a system
of linear equalities. Let |𝑄| denote the number of linear inequalities in 𝑄’s description. There
might be different LP formulations for the same polytope. Now, LP solvers can give a solution to an
optimization problem on this polytope if the LP formulation is polynomial in |𝑄|. Hence, there are
two directions to investigate, as given in [Li18].
′
1. Find another description 𝑄 with polynomially many inequalities;

2. Find a higher dimension polytope (bounded polyhedron) 𝐻 ⊆ ℝ 𝑘 (where 𝑘 > 𝑛) such that it
projects to 𝑃, and 𝐻 has a 𝑝𝑜𝑙 𝑦(𝑛) description.

In some cases, it can be shown that 1 is not possible by showing that any description in ℝ𝑛 has
exponential size, hence we can only turn to 2. Note that any optmization problem over 𝑃 can be
done by optimizing the same objective function over 𝐻.

Definition 2.8. Let 𝑃 ⊆ ℝ𝑛 be a polytope, a polytope 𝐻 ⊆ ℝ 𝑘 in a higher dimensional space is

called an extended formulation of 𝑃 if 𝜋(𝐻) = 𝑃, where 𝜋 : ℝ 𝑘 → ℝ𝑛 is the projection map. The
extension complexity of 𝑃 denoted by 𝑥𝑐(𝑃), is defined to be :

𝑥𝑐(𝑃) := min |𝑄|

𝑄: 𝑄 is an extended formulation of 𝑃

We say 𝑃 has a compact formulation if 𝑥𝑐(𝑃) = 𝑝𝑜𝑙 𝑦(𝑛).

5
 Now, linear programs with compact formulations as defined above can be solved in PTIME by
using an LP solver on the extended formulation of the LP. Note that if a polytope does not have a
compact formulation, that does not imply that the problem cannot be solved in PTIME, one of the
most famous examples being the matching polytope. Finding a maximum matching for a general
graph has a polynomial time algorithm, but it has been shown recently that the matching polytope
has no compact formulation [Rot17].

2.2.1 Non-negative rank characterization

A characterization of the extension complexity was given by [Yan88] in terms of the non-negative
rank of the “slack matrix" associated with the polytope.

Definition 2.9 (Slack matrix). Given a polytope 𝑃 ⊆ ℝ𝑛 suppose it is described by 𝑃 := {𝑥 ∈ ℝ𝑛 :

𝐴𝑥 ≤ 𝑏} where 𝐴𝑚×𝑛 is a matrix and 𝑏 ∈ ℝ𝑚 is a vector. Let the set of vertices of 𝑃 be {𝑧 1 , 𝑧2 , · · · , 𝑧 𝑘 }.
The slack matrix of 𝑃 under the above description is defined to be the matrix

𝑆𝑚×𝑘 := (𝑏 − 𝐴𝑧 1 , 𝑏 − 𝐴𝑧 2 , · · · , 𝑏 − 𝐴𝑧 𝑘 ) ∈ ℝ𝑚×𝑘
≥0

Definition 2.10 (Non-negative rank). Given a matrix 𝑆𝑚×𝑘 , its non-negative rank, denoted by
rank+ (𝑆), is defined to be the least integer 𝑟 ∈ ℕ, such that 𝑆 = 𝐿𝑚×𝑟 𝑅 𝑟×𝑘 where 𝐿 and 𝑅 are two
non-negative matrices.

Now, with these definitions, we state Yannakakis’s characterization of the extension complexity
of a polytope.

Theorem 2.11 ([Yan88]). For any polytope 𝑃 = {𝑥 ∈ ℝ𝑛 : 𝐴𝑥 ≤ 𝑏} whose dimension ≥ 1, let 𝑆 be its slack
matrix, one has

𝑥𝑐(𝑃) = rank+ (𝑆)

3 Results
In this section, we examine the existing literature on the extension complexity of k-l sparsity
matroids, regular matroids, polygons and methods of evaluating extension complexity such as
hitting set, communication complexity.

3.1 Dilworth truncation of matroids

It is interesting to examine the extension complexity of matroids obtained upon some operation on
a matroid with small extension. One of such operations is truncation of a matroid. In this section,
we see the definition of truncation of a matroid and show that for transversal matroids, truncation
has a small extension complexity via flows, as shown in [HIH21]. As a corollary

6
3.1.1 Definitions

Definition 3.1 (Set function). A set function 𝜌 : 2𝐸 → ℝ is a 𝛽0 -function if it satisfies the following:
1. 𝜌(𝑋) ≥ 0 for 𝜙 ≠ 𝑋 ⊆ 𝐸

2. 𝜌(𝑌) ≤ 𝜌(𝑋) for 𝑌 ⊆ 𝑋 ⊆ 𝐸 (monotonicity)

3. 𝜌(𝑋) + 𝜌(𝑌) ≥ 𝜌(𝑋 ∪ 𝑌) + 𝜌(𝑋 ∩ 𝑌) for 𝑋 , 𝑌 ⊆ 𝐸 (submodularity)

Definition 3.2 (Polymatroid). Given a set function 𝜌 : 2𝐸 → ℝ, a polytope
𝑃(𝜌) = {𝑥 ∈ ℝ𝐸 : 𝑥 ≥ 0, 𝑥(𝑋) ≤ 𝜌(𝑋)(𝜙 ≠ 𝑋 ⊆ 𝐸)}
is a polymatroid where 𝑥(𝑋) = 𝑥𝑒 .
Í
𝑒∈𝑋

Definition 3.3 (𝑙-lower truncated polymatroid). For this polymatroid 𝑃(𝜌), consider 𝑙 satisfying
0 ≤ 𝑙 ≤ min{𝜌({𝑒})|𝑒 ∈ 𝐸}. Then, 𝜌′ : 2𝐸 → ℝ defined by 𝜌′(𝑋) = 𝜌(𝑋) − 𝑙 (𝑋 ⊆ 𝐸), (𝐸, 𝜌′) is a
𝛽 0 -function, which defines a polymatroid 𝑃(𝜌′). This polymatroid is called an 𝑙-lower-truncated
polymatroid obtained from (𝐸, 𝜌), simply a lower-truncated polymatroid.

3.1.2 Lower truncated tranversal and 𝑘 − 𝑙 sparsity polymatroid

In this section, we shall see the connection of extension complexity of lower truncated transversal
matroid with network flow.
However, first recall the notion of transversal matroid.
Definition 3.4 (Transversal Matroid). Let 𝐺 = (𝑈 , 𝑊; 𝐴) be a bipartite graph with left vertex set
𝑈,right vertex set 𝑊 and directed edge set 𝐴 ⊆ 𝑈 × 𝑊 where edges are directed from 𝑈 to 𝑊.
Define Γ(𝑋) to be {𝑤|(𝑢, 𝑤) ∈ 𝐴, 𝑢 ∈ 𝑋}. Define (𝑈 , ℐ) as the transversal matroid where ℐ contains
𝑋 ⊆ 𝑈 that can be covered by a matching.
By Hall’s theorem, 𝑋 ⊆ 𝑈 can be covered by a matching iff |𝑌| ≤ Γ(𝑌) for all 𝑌 ⊆ 𝑋.
Definition 3.5 (Lower truncated transversal polymatroid). For 𝐺 = (𝑈 , 𝑊; 𝐴),let 𝑘, 𝑙 be positive
integers with 𝑑′ 𝑘 − 𝑙 > 0, where 𝑑′ is the minimum degree of a vertex in 𝑈. Define 𝛾𝑘,𝑙 : 2𝑈 → ℝ by
𝛾𝑘,𝑙 (𝑋) = 𝑘𝛾𝑁 (𝑋) − 𝑙 (𝑋 ⊆ 𝑈). (𝑈 , 𝛾𝑘,𝑙 ) is a polymatroid, called a (𝑘, 𝑙)-lower-truncated transversal
polymatroid.
First, we establish relation between the 𝑘 − 𝑙 transversal polymatroid and 𝑘 − 𝑙 sparsity matroid.
Consider a case of |Γ({𝑢})| = 2(𝑢 ∈ 𝑈) in the bipartite graph 𝐺. Let 𝐺′ = (𝑉 , 𝐸) be an undirected
graph with vertex set 𝑉 = 𝑊 and edge set 𝐸 = 𝑈. 𝐺′ can be regarded as a graph obtained from 𝐺
by subdividing each edge 𝑒 of 𝐺′ by a vertex 𝑒 of 𝐺. (see fig. 1)
Then, we define :
Definition 3.6 (𝑘 − 𝑙 sparsity matroid). For 𝐺′ and positive integers 𝑘, 𝑙 with 2𝑘 − 𝑙 > 0, {𝑋 : |𝑌| ≤
𝑘|𝑉(𝑋)| − 𝑙(𝜙 ≠ 𝑌 ⊆ 𝑋)} is the set of independent sets of a matroid, called (𝑘, 𝑙)-sparsity matroid or
count matroid, where 𝑉(𝑋) denotes the set of vertices on which edges in 𝑋 are incident.
Observe that 𝑀(𝐺; 𝑘, 𝑙) is isomorphic to 𝑘 − 𝑙 sparsity matroid, since 𝑉(𝑋) in 𝐺′ is the same as
Γ(𝑋) in 𝐺 for 𝑋 ⊆ 𝐸. So, we only need to examine 𝑀(𝐺; 𝑘, 𝑙) now.

7
 𝐸 𝑉

1 1
1
1 2 2
2
3 3
2 3
4 4 4
3 4

𝐺′ 𝐺

Figure 1: Relating 𝐺 and 𝐺′ for an example graph

3.1.3 Extension complexity of lower truncated transversal polymatroid

In this section, we examine the extension complexity of 𝑀(𝐺; 𝑘, 𝑙). Recall that given the polymatroid
𝑃(𝜌), consider 𝑙 satisfying 0 ≤ 𝑙 ≤ min{𝜌({𝑒})|𝑒 ∈ 𝐸}. Let 𝜌′ : 2𝐸 → ℝ be defined by 𝜌′(𝑋) = 𝜌(𝑋)− 𝑙
(𝑋 ⊆ 𝐸). Let us define 𝜒𝑒 (𝑒 ∈ 𝐸) to be the unit vector on the underlying set 𝐸 with 𝜒𝑒 (𝑒) = 1 and
others 0.
Lemma 3.7. 𝑥 ∈ 𝑃(𝜌′) iff 𝑥 + 𝑙𝜒𝑒 ∈ 𝑃(𝜌) for each 𝑒 ∈ 𝐸.
Proof. Suppose 𝑥 ∈ 𝑃(𝜌′). For each non empty 𝑋 ⊆ 𝐸,we have 𝑥(𝑋) ≤ 𝜌′(𝑋) = 𝜌(𝑋) − 𝑙,and hence
(𝑥 + 𝑙𝜒𝑒 )(𝑋) ≤ 𝑥(𝑋) + 𝑙 ≤ 𝜌(𝑋), which implies 𝑥 + 𝑙𝜒𝑒 ∈ 𝑃(𝜌).
Conversely, if 𝑥 + 𝑙𝜒𝑒 ∈ 𝑃(𝜌) for 𝑒 ∈ 𝐸 , then, for 𝑋 with 𝑒 ∈ 𝑋 ⊆ 𝐸, we have 𝑥(𝑋) + 𝑙 ≤ 𝜌(𝑋). This
holds for each 𝑒 ∈ 𝐸, and the lemma follows. □

Theorem 3.8. 𝑥𝑐(𝑃(𝜌′)) ≤ |𝐸| · 𝑥𝑐(𝑃(𝜌)).

The proof of 3.8 follows easily, since we add constraints corresponding to 𝑥 + 𝑙𝜒𝑒 for all 𝑒 ∈ 𝐸
using 3.7, hence we have |𝐸| · 𝑥𝑐(𝑃(𝜌)) constraints. In order to examine the transversal polymatroid,
we examine the network flow polytope.
Definition 3.9 (Network flow). Let 𝑁˜ = (𝑉˜ , 𝐴;˜ 𝑆, 𝑡; 𝑐˜) be a network with a vertex set 𝑉, ˜ a directed
˜ ˜ ˜ 𝐴˜
edge set 𝐴,a set of sources 𝑆 ⊂ 𝑉, a unique sink 𝑡 ∈ 𝑉 − 𝑆, and a capacity 𝑐˜ ∈ ℝ where 𝑐˜(𝑎)(≥ 0) is
a capacity of 𝑎 ∈ 𝐴.˜ For 𝑓 ∈ ℝ𝐴˜ , we define 𝛿+ 𝑓 ∈ ℝ by 𝛿+ 𝑓 (𝑣) = − Í
𝑒=(𝑢,𝑣)∈ 𝐴˜ 𝑓 (𝑒) + 𝑒=(𝑣,𝑤)∈ 𝐴˜ 𝑓 (𝑒)
Í
˜ The restriction of 𝛿+ 𝑓 to a subdomain 𝑋(⊆ 𝑉)
for 𝑣 ∈ 𝑉. ˜ is denoted by 𝛿+ 𝑓 | 𝑋 . 𝑓 ∈ ℝ𝐴˜ is a flow if
˜ 𝛿+ 𝑓 |𝑉−(𝑆∪{𝑡}) = 0 and 𝛿+ 𝑓 | 𝑆 ≥ 0.
0 ≤ 𝑓 (𝑎) ≤ 𝑐˜(𝑎)(𝑎 ∈ 𝐴),
˜
Definition 3.10 (Cut function). A cut function 𝑐˜ : 2𝑉−{𝑡} → ℝ is defined by 𝑐˜(𝑌) = 𝑐˜(𝑒)
Í
˜
𝑒=(𝑢,𝑤)∈ 𝐴,𝑢∈𝑌,𝑤∈ ˜
𝑉−𝑌
for 𝑌 ⊆ 𝑉˜ − {𝑡}.
Definition 3.11 (Network flow polymatroid). Define the function 𝛾𝑁˜ : 2𝑆 → ℝ as 𝛾𝑁˜ (𝑋) =
min{ 𝑐˜(𝑌)|𝑌 ⊆ 𝑉˜ − {𝑡}, 𝑌 ∩ 𝑆 = 𝑋} for 𝑋 ⊆ 𝑆. Then, (𝑆, 𝛾𝑁˜ ) is a polymatroid.

8
 With this definition, we have the following well-known result.

Lemma 3.12. For 𝑥 ∈ ℝ𝑆 , 𝑥 ∈ 𝑃(𝛾𝑁˜ ) iff there is a flow 𝑓 with 𝑥 = 𝛿+ 𝑓 | 𝑆 . Hence 𝑥𝑐(𝑃(𝛾𝑁˜ )) ≤ 2| 𝐴|
˜ + |𝑆|.

Now, we create a flow network to get the extension complexity of transversal polymatroid.
Consider the bipartite graph 𝐺 = (𝑈 , 𝑊; 𝐴) with left vertex set 𝑈,right vertex set 𝑊 and directed
edge set 𝐴 ⊆ 𝑈 × 𝑊 where edges are directed from 𝑈 to 𝑊, with an extra vertex 𝑡 and directed
edges from 𝑊 to 𝑡. Let 𝑐(𝑒) = ∞ for 𝑒 ∈ 𝐴, 𝑐(𝑒) = 1 for (𝑤, 𝑡). Let this network be 𝑁, then we have
𝛾𝑁 (𝑋) = Γ(𝑋). So, as a consequence we get

Lemma 3.13. 𝑥𝑐(𝑃(𝛾𝑁 )) = |𝐴| + |𝑊 |. For the transversal matroid over 𝑈 of bipartite graph 𝐺, its
independence polytope has extension complexity of |𝐴| + |𝑈 | + |𝑊 |.

Moreover, using 3.13 and 3.8 we get

Theorem 3.14. 𝑥𝑐(𝑃(𝛾𝑘,𝑙 )) ≤ |𝑈 |(|𝐴| + |𝑊 |).

However, we can also get an extension directly using the network flow idea. For the network
𝑁 = (𝑈 ∪ 𝑊 ∪ {𝑡}, 𝐴′; 𝑈 , 𝑡; 𝑐), we replace 𝑐 with a new capacity 𝑐 𝑤 for 𝑤 ∈ 𝑊 defined by 𝑐 𝑤 (𝑒) = +∞
(𝑒 ∈ 𝐴), 𝑐 𝑤 (𝑒) = 𝑘 (𝑒 = (𝑤 ′ , 𝑡), 𝑤 ′ ∈ 𝑊 − {𝑤}), and 𝑐 𝑤 (𝑒) = 𝑘 − 𝑙 (𝑒 = (𝑤, 𝑡)).

Lemma 3.15. For 𝑥 ∈ ℝ𝑈 , 𝑥 ∈ 𝑃(𝛾𝑘 , 𝑙) iff there is a flow 𝑓 with 𝑥 = 𝛿+ 𝑓 | 𝑆 in network 𝑁𝑤 =

(𝑈 ∪ 𝑊 ∪ {𝑡}, 𝐴′; 𝑈 , 𝑡; 𝑐 𝑤 ) for each 𝑤 ∈ 𝑊.

Using this lemma when 𝑘 ≥ 𝑙, we get a better bound,

Theorem 3.16. When 𝑘 ≥ 𝑙, 𝑥𝑐(𝑃(𝛾𝑘,𝑙 )) ≤ |𝑊 |(|𝐴| + |𝑊 |).

Applying these results for 𝑘 − 𝑙 sparsity matroids we get

Theorem 3.17. For the independence polytope of a (𝑘, 𝑙)-sparsity matroid of a graph 𝐺′ = (𝑉 , 𝐸), there is an
extended formulation of size |𝐸|(2|𝐸| + |𝑉 | + 1) in general and that of size |𝐸| + |𝑉 |(2|𝐸| + |𝑉 |) when 𝑘 ≥ 𝑙.

3.2 Relating communication and extension complexity

In the previous section, we had an approach to get extended formulation for 𝑘 − 𝑙 sparsity matroids
using an isomorphism with truncated transversal matroid, which in-turn was solved using network
flows and a characterisation of membership of truncated transversal matroids in general. However,
a similar result can be obtained as shown in [IKK+ 16] a more general hammer. This hammer is the
relation between communication complexity and extended formulations as shown in [FFGT15].
But first, we set up basics of communication complexity, as shown in [FFGT15]

3.2.1 Deterministic Communication Protocols

Let 𝑋, 𝑌 , and 𝑍 be arbitrary finite sets with 𝑍 ⊆ ℝ+ , and let 𝑓 : 𝑋 × 𝑌 → 𝑍 be a function. Suppose
that there are two players Alice and Bob who wish to compute 𝑓 (𝑥, 𝑦) for some inputs 𝑥 ∈ 𝑋 and
𝑦 ∈ 𝑌 . Alice knows only 𝑥 and Bob knows only 𝑦. They must therefore exchange information to be

9
able to compute 𝑓 (𝑥, 𝑦). (We assume that each player possesses unlimited computational power.)
The communication is carried out as a protocol that is agreed upon beforehand by Alice and Bob,
on the sole basis of the function 𝑓 . At each step of the protocol, one of the players has the token.
Whoever has the token sends a bit to the other player, that depends only on their input and on
previously exchanged bits. This is repeated until the value of 𝑓 on (𝑥, 𝑦) is known to both players.
The minimum number of bits exchanged between the players in the worst case to be able to evaluate
𝑓 by any protocol is called the communication complexity of 𝑓 .

3.2.2 Randomized Protocols

We define a randomized protocol (with private random bits and nonnegative outputs) as a rooted
binary tree with some extra information attached to its nodes. Let 𝑋 and 𝑌 be finite sets, as
above. Each node of the tree has a type, which is either 𝑋 or 𝑌 . To each node 𝑣 of type 𝑋 are
attached two function 𝑝 0,𝑣 , 𝑝1,𝑣 : 𝑋 → [0, 1]; to each node 𝑣 of type 𝑌 are attached two functions
𝑞 0,𝑣 , 𝑞1,𝑣 : 𝑌 → [0, 1]; and to each leaf 𝑣 is attached a nonnegative vector Λ𝑣 that is a column vector of
size |𝑋 | for leaves of type 𝑋 and a row vector of size |𝑌| for leaves of type 𝑌 . The functions 𝑝 𝑖,𝑣 and
𝑞 𝑗,𝑣 define transition probabilities, and we assume that 𝑝 0,𝑣 (𝑥) + 𝑝 1,𝑣 (𝑥) ≤ 1 and 𝑞0,𝑣 (𝑦) + 𝑞1,𝑣 (𝑦) ≤ 1.
An example is shown in fig. 2. In the example, |𝑋 | = |𝑌| = 4.

Figure 2: Example of randomized communication protocol

An execution of the protocol on input (𝑥, 𝑦) ∈ 𝑋 × 𝑌 is a random path that starts at the root
and descends to the left child of an internal node 𝑣 with probability 𝑝 0,𝑣 (𝑥) if 𝑣 is of type 𝑋
and 𝑞0,𝑣 (𝑦) if 𝑣 is of type 𝑌 , and to the right child of 𝑣 with probability 𝑝1,𝑣 (𝑥) if 𝑣 is of type
𝑋 and 𝑞1,𝑣 (𝑦) if 𝑣 is of type 𝑌 . With probability 1 − 𝑝 0,𝑣 (𝑥) − 𝑝 1,𝑣 (𝑥) and 1 − 𝑞0,𝑣 (𝑦) − 𝑞1,𝑣 (𝑦)
respectively, the execution stops at 𝑣. For an execution stopping at leaf 𝑣 with vector Λ𝑣 , the value
of the execution is defined as the entry of Λ𝑣 that corresponds to input 𝑥 ∈ 𝑋 if 𝑣 is of type 𝑋,
and 𝑦 ∈ 𝑌 if 𝑣 is of type 𝑌 . For an execution stopping at an internal node, the value is defined to be 0.

10
 For each fixed input (𝑥, 𝑦) ∈ 𝑋 × 𝑌 , the value of an execution on input (𝑥, 𝑦) is a random
variable. If we let 𝑍 ⊆ ℝ+ as before, we say that the protocol computes a function 𝑓 : 𝑋 × 𝑌 → 𝑍 in
expectation if the expectation of this random variable on each (𝑥, 𝑦) ∈ 𝑋 × 𝑌 is precisely 𝑓 (𝑥, 𝑦). The
complexity of a protocol is the height of the corresponding tree. Given an ordering 𝑥1 , · · · , 𝑥 𝑚 of the
elements of 𝑋, and 𝑦1 , · · · , 𝑦𝑛 of the elements of 𝑌 , we can visualize the function 𝑓 : 𝑋 × 𝑌 → 𝑍 as
a 𝑚 × 𝑛 nonnegative matrix 𝑆 = 𝑆( 𝑓 ) such that 𝑆 𝑖,𝑗 = 𝑓 (𝑥 𝑖 , 𝑦 𝑗 ) for all (𝑖, 𝑗) ∈ [𝑚] × [𝑛]. The matrix 𝑆
is called the communication matrix of 𝑓 . Below, as is natural, we will not always make a distinction
between a function and its communication matrix.

3.2.3 Factorisation and protocols

In [FFGT15], the main theorem about non-negative rank and communication complexity is given
below.

Theorem 3.18. If there exists a randomized protocol of complexity 𝑐 computing a matrix 𝑆 ∈ ℝ+𝑋×𝑌 in
expectation, then lg rank+ (𝑆) ≤ 𝑐. Conversely, if the nonnegative rank of matrix 𝑆 ∈ ℝ+𝑚×𝑛 is 𝑟, then there
exists a randomized protocol computing 𝑆 in expectation, whose complexity is at most ⌈lg r⌉. In other words,
if 𝑐 𝑚𝑖𝑛 (𝑆) denotes the minimum complexity of a randomized protocol computing 𝑆 in expectation, we have

𝑐 𝑚𝑖𝑛 (𝑆) = lg rank+ (𝑆)

 

Since Yannakakis showed that nonnegative rank of slack matrix is equal to extension complexity,
theorem 3.18 implies the following (see corollary 3.19).

Corollary 3.19. Let 𝑃 be a polytope with associated slack matrix 𝑆 = 𝑆(𝑃), such that 𝑃 is neither empty or a
point. If there exists a randomized protocol of complexity 𝑐 computing 𝑆 in expectation, then 𝑥𝑐(𝑃) ≤ 2𝑐 .
Conversely, if 𝑥𝑐(𝑃) = 𝑟, then there exists a randomized protocol computing S in expectation, whose
complexity is at most ⌈lg r⌉. In other words, if 𝑐 𝑚𝑖𝑛 (𝑆) denotes the minimum complexity of a randomized
protocol computing 𝑆 in expectation, we have

𝑐 𝑚𝑖𝑛 (𝑆(𝑃)) = lg rank+ (𝑆)

 

We do not cover the proof of theorem 3.18 in detail (see [FFGT15]), however here is a proof
sketch,

1. Given input (𝑥, 𝑦), probability of reaching a leaf in the tree of randomized protocol is given
by probability of taking “correct" branches that make the path.

2. Since protocol computes matrix 𝑆, compute 𝑆(𝑥, 𝑦) as sum of probability of getting a leaf
times value at that leaf. 𝑆(𝑥, 𝑦) = 𝑣∈𝐿𝑥 Λ𝑣 (𝑥)𝑃𝑣 (𝑥, 𝑦) + 𝑤∈𝐿 𝑦 Λ𝑤 (𝑦)𝑃𝑤 (𝑥, 𝑦), where 𝑃𝑣 (𝑥, 𝑦)
Í Í
denotes probability of reaching leaf 𝑣 given input (𝑥, 𝑦).

3. 𝑃𝑣 is a rank one matrix. This is because,

𝑘
Ö 𝑙
Ö
𝑃𝑣 (𝑥, 𝑦) = 𝑝 𝛼 𝑖 ,𝑣 𝑖 (𝑥) · 𝑞 𝛽 𝑗 ,𝑤 𝑗 (𝑦)
𝑖=1 𝑗=1

11
 𝑦

𝑙
Ö
𝑝 𝛽 𝑗 ,𝑤 𝑗 (𝑦)
𝑖𝑗= 1
𝑥 𝑘
Ö
𝑝 𝛼 𝑖 ,𝑣 𝑖 (𝑥)
𝑖=1

Figure 3: 𝑃𝑣 as a rank 1 matrix

where 𝛼 𝑖 , 𝛽 𝑗 ∈ {0, 1} are fixed given 𝑣. Then, 𝑃𝑣 is a product of column and row matrix shown
in fig. 3. Hence, we get 𝑆 as a sum of 2𝑐 rank one matrices.

4. Given nonnegative rank of 𝑆 = 𝐴𝐵 given by 𝑟. Let 𝛿 𝑖 := 𝑘 𝐴 𝑖,𝑘 ≤ 1. Alice selects a column

Í
index 𝑘 ∈ [𝑟] according to the probabilities found in row 𝑖 of matrix A, sends this index to
Bob, and Bob outputs the entry of B in row 𝑘 and column 𝑗. With probability 1 − 𝛿 𝑖 Alice
does not send any index to Bob and the computation stops with implicit output zero. Since,
𝑘 𝐴 𝑖,𝑘 𝐵 𝑘,𝑗 = 𝑆 𝑖,𝑗 , we get that this protocol computes 𝑆 in expectation. Moreover, 𝛿 𝑖 ≤ 1
Í
without loss of generality, since we can scale 𝐴, 𝐵 accordingly.

Now that we have this relation set up, we examine the extension complexity of 𝑘 − 𝑙 sparsity
matroids using this technique.

3.3 𝑘 − 𝑙 sparsity matroids (Alternate method)

The bound given by [HIH21] was predated by one with randomized communication protocol,
shown by [IKK+ 16]. Given a polytope 𝑃, in order to obtain 𝑥𝑐(𝑃) we need a communication protocol
to compute the slack matrix of 𝑃. That is, let Alice have a facet defining inequality 𝑎 𝑇𝑖 𝑥 ≤ 𝑏 𝑖 and
Bob have a vertex 𝑣 𝑗 we need a protocol to return 𝑏 𝑖 − 𝑎 𝑇𝑖 𝑣 𝑗 .

3.3.1 Protocol when 𝑘 ≥ 𝑙

In 𝑘 − 𝑙 sparsity matroids, inequalities are given by 𝑥(𝐸(𝑋)) ≤ 𝑘|𝐸(𝑋)| − 𝑙 where 𝑋 ⊆ 𝑉. Let us

denote the in-degree of vertex 𝑣 by 𝜌(𝑣). So, given a subset of vertices with Alice (facet defining
inequality) and subset of edges (a vertex of polytope), the corresponding slack is given by,

𝑆𝑋 ,𝐹 = 𝑘|𝑋 | − 𝑙 − |𝐹 ∩ 𝐸(𝑋)|

12
 Then, protocol in [IKK+ 16] is given by,

1. Alice sends a vertex 𝑥 ∈ 𝑋 to Bob (Note that by vertex we mean vertex of the graph rather
than that of polytope)

2. Bob orients the edges of 𝐹 in such a way that 𝜌(𝑥) = 𝑘 − 𝑙 and 𝜌(𝑧) = 𝑘 for all 𝑧 ∈ 𝑉\{𝑥}. Then,
Bob picks an oriented edge (𝑢, 𝑣) of 𝐹 uniformly at random and sends it to Alice.

3. Alice outputs 𝑘|𝑉 | − 𝑙 if 𝑢 ∈ 𝑋 and 𝑣 ∈ 𝑋, and outputs 0 otherwise.

The following key lemma will enable us to show correctness of the protocol.

Lemma 3.20. Let 𝑘, 𝑙 be integers such that 0 ≤ 𝑙 ≤ 𝑘, 𝐻 = (𝑉 , 𝐹) be a (𝑘, 𝑙) tight graph (vertex of polytope)
with |𝑉 | ≥ 2, and 𝑥 ∈ 𝑉 be an arbitrary vertex of 𝐻. Then, 𝐻 has an orientation 𝐻 = (𝑉 , 𝐹) such that
𝜌(𝑥) = 𝑘 − 𝑙 and 𝜌(𝑧) = 𝑘 for all 𝑧 ∈ 𝑉\{𝑥}.

In order to show the lemma, here is a well known result by Hakimi [Hak65],

Lemma 3.21 (Hakimi). Let 𝐻 = (𝑉 , 𝐹) be an undirected graph and 𝑚(𝑣) be a non-negative integer for
every 𝑣 ∈ 𝑉 . Then, 𝐻 has an orientation such that 𝜌(𝑣) = 𝑚(𝑣) for all 𝑣 if and only if |𝐹| = 𝑣∈𝑉 𝑚(𝑣)
Í
and |𝐹(𝑋)| ≤ 𝑣∈𝑋 𝑚(𝑣) for all 𝑋 ⊆ 𝑉.
Í

Now, to prove lemma 3.20, it suffices to show condition in Hakimi’s lemma. We have,
𝑣𝑖𝑛𝑉 𝑚(𝑣) = 𝑘(|𝑉 | − 1) + 𝑘 − 𝑙 = 𝑘|𝑉 | − 𝑙 = |𝐹|, and 𝑣∈𝑋 𝑚(𝑣) ≥ 𝑘(|𝑋 | − 1) + 𝑘 − 𝑙 = 𝑘|𝑋 | − 𝑙 ≥ 𝐹(𝑋).
Í Í
Hence, we get that the protocol is feasible. Now, to show correctness, we compute expected value
of function computed by protocol.
First, we have
Õ
𝜌(𝑣) = 𝑘|𝑋 | − 𝑙
𝑣∈𝑋

𝜌(𝑣) counts the total in-degree of vertices in 𝑋 that is, number of edges with head in 𝑋.
Í
But, 𝑣∈𝑋
Õ
𝜌(𝑣) = |{(𝑢, 𝑣) : 𝑣 ∈ 𝑋 , 𝑢 ∈ 𝑋}| + |{(𝑢, 𝑣) : 𝑢 ∉ 𝑋 , 𝑣 ∈ 𝑋}|
𝑣∈𝑋
= |𝐹 ∩ 𝐸(𝑋)| + |{(𝑢, 𝑣) : 𝑢 ∉ 𝑋 , 𝑣 ∈ 𝑋}|

Hence, we get |{(𝑢, 𝑣) : 𝑢 ∉ 𝑋 , 𝑣 ∈ 𝑋}| = 𝑆𝑋 ,𝐹 . Finally, the protocol gives expected value
𝑆𝑋 ,𝐹
(𝑘|𝑉 | − 𝑙) · 𝑃𝑟(𝑢 ∉ 𝑋 , 𝑣 ∈ 𝑋) = (𝑘|𝑉 | − 𝑙) ·
|𝐹|
= 𝑆𝑋 ,𝐹

Hence, the protocol is correct. Finally, amount of bits exchanged are,

1. 𝑙𝑜𝑔|𝑉 | bits to send a vertex 𝑥 ∈ 𝑋

2. 𝑙𝑜𝑔|𝐸| bits to send a random edge 𝑒 ∈ 𝐸

Hence, total bits used is 𝑙𝑜𝑔(|𝑉 ||𝐸|), hence extension complexity is given by 𝑂(|𝑉 ||𝐸|).

13
3.3.2 Protocol for 𝑘 ≤ 𝑙

Similar to the previous set-up, Alice has a subset of vertices 𝑋 ⊆ 𝑉, Bob has a set of edges.

1. Alice sends two arbitrary vertices 𝑥, 𝑦 ∈ 𝑋 to Bob.

2. Bob orients the edges of 𝐹 in such a way that 𝜌(𝑥) = 0, 𝜌(𝑦) = 2𝑘 − 𝑙 and 𝜌(𝑧) = 𝑘 for all
𝑧 ∈ 𝑉\{𝑥, 𝑦}. Then, Bob picks an oriented edge (𝑢, 𝑣) of 𝐹 uniformly at random and sends it
to Alice.

3. Alice outputs 𝑘|𝑉 | − 𝑙 if 𝑢 ∈ 𝑋 and 𝑣 ∈ 𝑋, and outputs 0 otherwise.

The feasibility and correctness of this protocol also follows similarly using Hakimi’s theorem
and a similar analysis to the protocol of 𝑘 ≥ 𝑙. Note that here 2 vertices are sent by Alice, hence
extension complexity is 𝑂(|𝑉 | 2 |𝐸|).

3.4 Hitting Set of Matroid Polytopes

In the previous section, we have obtained a general hammer in the form of communication protocols
to obtain extension complexity of polytopes ([FFGT15]). In this section we see a special case of this
protocol where we can obtain a communication protocol in terms of hitting set of a polytope as
described in [Apr22]. Hence, we get another way of getting upper bounds for extension complexity
by describing communication protocols. In this section, we shall look at hitting set of polytopes,
its relation with extension complexity, and getting small extended formulations for graphic and
co-graphic matroids using this.

3.4.1 Hitting Set and Extension Complexity

First, let us define the hitting set of a base polytope.

Definition 3.22 (Base Polytope). The base polytope of 𝑀 is 𝐵(𝑀) = 𝑐𝑜𝑛𝑣𝜒 𝐵 : 𝐵 ∈ ℬ, where 𝜒 𝐵 ∈ ℝ𝐸
is the incidence vector of basis 𝐵. The following is a description of 𝐵(𝑀):

𝐵(𝑀) = {𝑥 ∈ ℝ+𝐸 : 𝑥(𝐸) = 𝑟 𝑘(𝐸), 𝑥(𝐹) ≤ 𝑟 𝑘(𝐹)∀flat 𝐹 ⊆ 𝐸}

where we write 𝑥(𝐹) = 𝑥𝑒 .

Í
𝑒∈𝐹

A base 𝐵 has slack 0 with respect to flacet 𝐹 iff |𝐵 ∩ 𝐹| = 𝑟 𝑘(𝐹).

Definition 3.23 (Hitting family / Hitting set). Let 𝒱 be a family of bases of 𝑀. We say that 𝑉
is a hitting family for 𝑀 if, for any flacet 𝐹 of 𝑀, there is at least one basis 𝐵 ∈ 𝑉 that has full
intersection with 𝐹.

Slightly abusing notation, we denote by ℎ(𝑀) the size of the smallest hitting family for 𝑀.
Notice that this does not correspond exactly to ℎ(𝐵(𝑀)), as we are ignoring non-negativity facets,
but the two numbers are the same up to an additive 𝑛 factor, hence the difference is irrelevant to the

14
purpose of this section.
Note that independence set polytope has base polytope as a facet (adding equality constraint),
hence 𝑥𝑐(𝐵(𝑀)) ≤ 𝑥𝑐(𝑃(𝑀)). It also holds that 𝑥𝑐(𝑃(𝑀)) ≤ 𝑥𝑐(𝐵(𝑀)) + 2|𝐸|. This fact can be ob-
tained by adding 0 ≤ 𝑦 𝑒 ≤ 𝑥 𝑒 constraints and taking polytope formed by 𝑦 as independence polytope.

Now that we have these definitions, we state the main result of [Apr22].

Theorem 3.24. For a matroid M of n elements and rank r,

𝑥𝑐(𝐵(𝑀)) = 𝑂(ℎ(𝑀) · 𝑛 · 𝑟).

In this report, we omit the proof of this result, but rather focus on the consequences on extension
complexity, some of which are described in [Apr22].

3.4.2 Relation with extension of dual matroid

First, let us recall the definition the dual of a matroid 𝑀 in terms of the base sets.

Definition (Dual Matroid). The dual of a matroid 𝑀 is another matroid 𝑀 ∗ that has the same
elements as 𝑀, and in which a set is independent if and only if 𝑀 has a basis set disjoint from it.
Alternatively, basis sets of dual matroid are complements of basis sets of 𝑀.

It is well known that given a matroid 𝑀 = (𝐸, ℐ),

𝑟 𝑀 ∗ (𝑆) = |𝑆| + 𝑟 𝑀 (𝐸\𝑆) − 𝑟 𝑀 (𝐸)

In order to analyse extension complexity of dual matroid, we need to analyse the hitting set of the
dual matroid. Let the set of bases of 𝑀 be 𝔹.

Lemma 3.25. 𝐵 ∈ 𝔹 makes constraint corresponding to 𝑆 tight in 𝑀 iff 𝐸\𝐵 makes constraint corresponding
to 𝐸\𝑆 tight in 𝑀 ∗ .

Proof. Let 𝑥 𝑒 = 1 iff 𝑒 ∈ 𝐵, 0 otherwise. Also, since 𝑆 is tight, 𝑥(𝑆) = 𝑟(𝑆). Hence, 𝑆 ⊆ 𝐵. Hence, we
get 𝐸\𝐵 ⊆ 𝐸\𝑆. Now, consider the indicator vector for 𝐸\𝑆. Let 𝑦 𝑒 = 1 iff 𝑒 ∉ 𝐵, 0 otherwise. So,
𝑦(𝐸\𝑆) = |𝐸\𝐵|. Hence,

𝑟 𝑀 ∗ (𝐸\𝑆) = |𝐸\𝑆| + 𝑟 𝑀 (𝑆) − 𝑟 𝑀 (𝐸)

= |𝐸\𝑆| + 𝑟 𝑀 (𝑆) − 𝑟 𝑀 (𝐵)
= |𝐸\𝑆| + |𝑆| − |𝐵| since 𝑆 ⊆ 𝐵
= |𝐸\𝐵| = 𝑦(𝐸\𝑆)

So, we get the desired result. The converse is obtained by considering original matroid as 𝑀 ∗
instead of 𝑀. □

Lemma 3.26. If 𝒱 is a hitting set for 𝑀, 𝒱 ′ = {𝐸\𝐵|𝐵 ∈ 𝒱} is a hitting set for 𝑀 ∗ .

15
Proof. For any constraint in 𝑀 ∗ , say corresponding to set 𝑆. Some 𝐵 ∈ 𝒱 makes 𝐸\𝑆 tight. Hence,
𝐸\𝐵 makes 𝑆 tight (using lemma 3.25), and 𝐸\𝐵 ∈ 𝒱 ′. Hence, completing the proof. □

So, given a hitting set for 𝑀, we can get an equal sized hitting set for the dual matroid, and
vice-versa. So, we get

Lemma 3.27. ℎ(𝑀) = ℎ(𝑀 ∗ )

3.4.3 Relation with (co)-graphic matroid polytope

In this section, we see that as shown in [Apr22], graphic matroid polytopes admit a small hitting
set and the upper bound obtained on extension complexity matches the best known bounds. As a
corollary using lemma 3.27, we get a small extended formulation for the co-graphic matroid as well.
For the graphic matroid base polytope (denote as 𝑃(𝐺)), we have constraints given by

𝑥(𝐸(𝑈)) ≤ |𝑈 | − 1 𝜙≠𝑈 ⊂𝑉
𝑥(𝐸(𝑉)) = |𝑉 | − 1

where 𝐸(𝑈) is set of edges with both endpoints in 𝑈. Let the graph 𝐺 = 𝐾 𝑛 . Now, for any other
𝐺′, we get that base polytope for 𝐺′ is a facet of 𝐺, hence adding 𝑥 𝑒 = 0 constraints enables the
extension complexity bound to carry over. So, considering 𝐺 = 𝐾 𝑛 is enough. Let the edge set of
star graph with edges (𝑣, 𝑢) for all 𝑢 ∈ 𝑉 − {𝑣} be denoted as 𝛿 𝑣 . Now, we have the following claim,

Lemma 3.28. 𝒱 = {𝛿 𝑣 |𝑣 ∈ 𝑉 } is a hitting set for 𝑃(𝐺).

Proof. For any 𝑈 ⊆ 𝐸, with 𝑢 ∈ 𝑈 the facet inducing inequality is given by,

𝑥(𝐸(𝑈)) ≤ |𝑈 | − 1

Putting the vector 𝑥 corresponding to 𝛿 𝑢 , we get 𝑥(𝐸(𝑈)) = 𝑒∈𝐸(𝑈) 𝑒 ∈ 𝛿(𝑢) = 𝑑𝑢 = |𝑈 | − 1.

Í
Moreover, all star graphs are indeed spanning trees, i.e. base sets, hence showing the result. □

So, using lemmas 3.28 and 3.24, we get that for a graph 𝐺 = (𝑉 , 𝐸) with |𝑉 | = 𝑛, ℎ(𝑀) = 𝑛,
number of elements = 𝑛2 , rank 𝑟 = 𝑛 − 1. Hence, we get an extended formulation of size 𝑂(𝑛 4 ).

However, this approach can be made more efficient as shown in [Apr22] by bypassing a step in the
derivation of theorem 3.24 to get an 𝑂(𝑛 3 ) extension. However, we omit this modification in this
report.

Another approach using hitting sets examined in [Apr22] is extension for regular matroids.
However we come back to this in the next section after setting up some machinery for handling
2-sums of matroids.

16
3.5 2-sums of Matroid Polytopes
Before diving into why this operation matters, let us define direct sum, 2-sum of matroids.

Definition 3.29 (Direct Sum). Consider matroids 𝑀1 = (𝐸1 , 𝔹1 ) and 𝑀2 = (𝐸2 , 𝔹2 ), with non-empty
base sets. If 𝐸1 ∩ 𝐸2 = 𝜙, we can define the direct sum 𝑀1 ⊕ 𝑀2 as the matroid with ground set
𝐸1 ∪ 𝐸2 and base set 𝐵1 × 𝐵2 .

Definition 3.30 (2-sums). Consider matroids 𝑀1 = (𝐸1 , 𝔹1 ) and 𝑀2 = (𝐸2 , 𝔹2 ), with non-empty
base sets. If 𝐸1 ∩ 𝐸2 = {𝑝}, we can define the 2-sum 𝑀1 ⊕2 𝑀2 as the matroid with ground set
𝐸1 ∪ 𝐸2 − 𝑝 and base set {𝐵1 ∪ 𝐵2 − 𝑝 : 𝐵1 ∈ 𝔹1 , 𝐵2 ∈ 𝔹2 , 𝑝 ∈ 𝐵1 Δ𝐵2 }

where Δ denotes symmetric difference. In [ACF18] the following lemma is shown,

Lemma 3.31. Let 𝑀1 (𝐸1 , 𝔹1 ), 𝑀2 (𝐸2 , 𝔹2 ) be matroids with 𝐸1 ∩ 𝐸2 = {𝑝} and let 𝑀 = 𝑀1 ⊕2 𝑀2 .
Then 𝐵(𝑀) is linearly isomorphic to 𝐵(𝑀1 ) × 𝐵(𝑀2 ) ∩ {𝑥 ∈ ℝ𝐸1⊎𝐸2 : 𝑥 𝑝1 + 𝑥 𝑝2 = 1}, where 𝐸1 ⊎ 𝐸2 =
𝐸1 ∪ 𝐸2 ∪ {𝑝 1 , 𝑝2 } − 𝑝 is the disjoint union of 𝐸1 and 𝐸2 , with 𝑝 1 and 𝑝 2 corresponding to 𝑝 ∈ 𝐸1 and
𝑝 ∈ 𝐸2 respectively.

3.5.1 Connection with hitting set

In this section we show the following lemma,

Lemma 3.32. Consider matroids 𝑀1 = (𝐸1 , 𝔹1 ), 𝑀2 = (𝐸2 , 𝔹2 ) with 𝐸1 ∩ 𝐸2 = {𝑝}, then, ℎ(𝑀1 ⊕2 𝑀2 ) ≤
ℎ(𝑀1 ) + ℎ(𝑀2 ).

Proof. Consider the smallest hitting sets for 𝑃(𝑀1 ) and 𝑃(𝑀2 ) (say 𝒱1 , 𝒱2 ). Then, consider the
polytope 𝑃 given by 𝐵(𝑀1 )× 𝐵(𝑀2 )∩ {𝑥 ∈ ℝ𝐸1 ⊎𝐸2 : 𝑥 𝑝1 + 𝑥 𝑝2 = 1}. By lemma 3.31, 𝑃 and 𝑃(𝑀1 ⊕2 𝑀2 )
are affinely isomorphic. WLOG, 𝑀1 ⊕2 𝑀2 is non-empty, hence there exist bases 𝐵1 ∈ 𝔹1 , 𝐵2 ∈ 𝔹2
such that 𝑝 ∈ 𝐵1 , 𝑝 ∉ 𝐵2 . Let there exist bases 𝑝 ∉ 𝐵′1 ∈ 𝔹1 , 𝑝 ∈ 𝐵2 ∈ 𝔹2 . Then, consider the set
{[𝑉1 , 𝐵′2 ] : 𝑝 ∈ 𝑉1 ∈ 𝒱1 } ∪ {[𝑉1′ , 𝐵2 ] : 𝑝 ∉ 𝑉1′ ∈ 𝒱1 } ∪ {[𝐵1 , 𝑉2 ] : 𝑝 ∉ 𝑉2 ∈ 𝒱2 } ∪ {[𝐵′1 , 𝑉2′] : 𝑝 ∈ 𝑉2′ ∈
𝒱2 }. Then for any facet defining constraint in 𝐵(𝑀1 ), we get some base 𝐵1 ∈ 𝒱1 such that 𝐵1 is tight
with respect to the constraint. Then, an appropriate base in 𝒱2 can be appended to 𝐵1 to make the
constraint tight in the new polytope. Hence, we get a hitting set of size ℎ(𝑀1 ) + ℎ(𝑀2 ).
If all bases in 𝑀1 contain 𝑝 (or do not contain 𝑝), then there exists a base in 𝑀2 not containing
(containing) 𝑝. A similar idea still works since for all constraints that can become tight by a base set
in 𝑀2 containing 𝑝 and not by base sets not containing 𝑝, the constraint will not be facet defining in
the new polytope. □

3.5.2 Connection with regular matroids

First, recall that regular matroids are defined as linear matroids, when the matrix is totally
unimodular. It is well known that 𝑂(𝑛 6 ) extension complexity holds for regular matroids as shown
in [AF19]. This result relies on using seymour’s decomposition,

17
Theorem 3.33 (Seymour’s decomposition). A matroid is regular if and only if it is obtained by means of
1−, 2− and 3-sums, starting from graphic and cographic matroids and copies of a certain 10-elements matroid
𝑅 10 .

Hence, an approach to getting small extension complexity for regular matroids is showing that,
𝑥𝑐(𝑃(𝑀1 ⊕𝑡 𝑀2 )) ≤ 𝑥𝑐(𝑃(𝑀1 )) + 𝑥𝑐(𝑃(𝑀2 )) for 𝑡 = 1, 2, 3. However, for 𝑡 = 1 it holds trivially since
its a cartesian product. Similarly for 𝑡 = 2, its the intersection of cartesian product with a single
hyperplane. 3-sum is handled by the rest of the paper.
But, we can have two other approaches as shown in [Apr22],

1. Show that ℎ(𝑀1 ⊕𝑡 𝑀2 ) is small for 𝑡 = 1, 2, 3. For 𝑡 = 1, 2 we know that ℎ(𝑀1 ⊕𝑡 𝑀2 ) ≤

ℎ(𝑀1 ) + ℎ(𝑀2 ) as shown in lemma 3.32. But, for 3-sum it is not clear how to approach this.

2. Analyse size of hitting sets for regular matroid bypassing seymour’s decomposition. Similar to
proof shown for graphic matroids, we only need to consider the “maximal" regular matroids.
So, we only need to consider complete regular matroids which is known to have atmost
𝑟(𝑟 + 1)/2 elements, if rank is 𝑟.

3.6 Extension of Polygons

In this section we examine the extension complexity of polygons as shown in [FRT12]. This paper
shows 3 main results of which, we will discuss the first two.

1. Extension complexity of regular 𝑛-gons is 𝑂(log 𝑛).

√
2. Extension complexity of generic 𝑛-gons is lower bounded by 2𝑛.

q exist 𝑛-gons whose vertices lie on an 𝑂(𝑛)×𝑂(𝑛 ) integer grid with extension complexity
3. There 2

𝑛
Ω( log 𝑛 )

3.6.1 𝑂(log 𝑛) extension for regular polygons

In this section we show an 𝑂(log 𝑛) extension for regular 𝑛-gons. The construction of the extended
formulation is very similar to binary search.

Theorem 3.34. Let 𝑃 be a regular 𝑛-gon in ℝ2 . Then 𝑥𝑐(𝑃) = 𝑂(log 𝑛).

We describe the sketch of the proof here (see [FRT12] for details).
Without loss of generality, we may assume that the origin is the barycenter of 𝑃. After
numbering the vertices of 𝑃 counter-clockwise as 𝑣1 , · · · , 𝑣 𝑛 , we define a sequence 𝑙0 , · · · , 𝑙 𝑞−1 of
axes of symmetry of 𝑃, as

1. Initialize 𝑖 to 0, and 𝑘 to 𝑛. While 𝑘 > 1, repeat the following steps:

(a) define 𝑙 𝑖 as the line through the origin and the midpoint of vertices 𝑣 ⌈ 𝑘 ⌉ and 𝑣 ⌈ 𝑘+1 ⌉
2 2
 𝑘+1 
(b) replace 𝑘 by 2

18
 Figure 4: Construction of folding sequence given an 𝑛-gon

(c) increment 𝑖

For each 𝑖 = 0, · · · , 𝑞 − 1, one of the two closed halfplanes bounded by 𝑙 𝑖 contains 𝑣1 . We denote it
𝑙 𝑖+ . We denote the other by 𝑙 𝑖− (𝑞 = 𝑂(log 𝑛)).
Now, consider a vertex 𝑣 of 𝑃. We define the folding sequence 𝑣 (0) , 𝑣 (1) , · · · , 𝑣 (𝑞) of 𝑣 as follows.
We let 𝑣 (0) := 𝑣, and for 𝑖 = 0, · · · , 𝑞 − 1, we let 𝑣 (𝑖+1) denote the image of 𝑣 (𝑖) by the reflection with
respect to 𝑙 𝑖 if 𝑣 (𝑖) is not in the halfspace 𝑙 + 𝑖 , otherwise we let 𝑣 (𝑖+1) := 𝑣 (𝑖) . Similarly define the
reflection of a facet 𝐹 given by 𝑎 𝑇 𝑥 ≤ 𝑏 as conditional reflection of 𝑎 (𝑖) along 𝑙 𝑖 . Initially, start with
any facet say 𝑎 𝑇 𝑥 ≤ 𝛽 (facet 𝑖 is (𝑎 (𝑖) )𝑇 𝑥 ≤ 𝛽). Since final vertex of any folding sequence is 𝑣 1 , get
that final facet corresponds to edge [𝑣 1 , 𝑣2 ] or [𝑣 𝑛 , 𝑣1 ]. Upon analysis of slack, we get,
𝛽 − (𝑎 (𝑖) )𝑇 𝑣 (𝑖) = 𝛽 − (𝑎 (𝑖+1) )𝑇 𝑣 (𝑖+1) if 𝑣 (𝑖) , 𝑎 (𝑖) lie in same side of 𝑙 𝑖
𝛽 − (𝑎 (𝑖) )𝑇 𝑣 (𝑖) = 𝛽 − (𝑎 (𝑖+1) )𝑇 𝑣 (𝑖+1) + 2𝑑(𝑎 (𝑖) , 𝑙 𝑖 )𝑑(𝑣 (𝑖) , 𝑙 𝑖 ) otherwise
Now, we construct factorisation of slack as,
𝑆(𝑃) = 𝑇𝑈

√
In the left factor, row corresponding √ to facet 𝐹 is of form, (𝑡 0 , · · · , 𝑡 𝑞−1 ), where 𝑡 𝑖 := ( 2𝑑(𝑎 (𝑖) , 𝑙 𝑖 ), 0)
if 𝑎 (𝑖) is not in 𝑙 𝑖+ and 𝑡 𝑖 := (0, 2𝑑(𝑎 (𝑖) , 𝑙 𝑖 )) otherwise. Similarly define column corresponding to
√ √
corner 𝑣 as (𝑢0 , · · · , 𝑢𝑞−1 )𝑇 , where 𝑢𝑖 := (0, 2𝑑(𝑣 (𝑖) , 𝑙 𝑖 ))𝑇 if 𝑣 (𝑖) is not in 𝑙 𝑖+ and 𝑢𝑖 := ( 2𝑑(𝑣 (𝑖) , 𝑙 𝑖 ), 0)𝑇
otherwise. Now, slack of 𝑣 (𝑞) with respect to 𝐹 (𝑞) is always 0 (since it is the log 𝑛 step of this process).
Hence, applying telescopic sum on both sides of the slack equations, we get the desired result.
√
3.6.2 Lower bound of 2𝑛 for generic 𝑛-gons

This bound uses algebraic methods for bounding extension complexity.

19
Definition 3.35. A subset 𝑋 of 𝐿 is said to be algebraically independent over 𝐾 if no non-trivial
polynomial relation with coefficients in 𝐾 holds among the elements of 𝑋. The transcendence
degree of the field extension 𝐿/𝐾 is defined as the largest cardinality of an algebraically independent
subset of 𝐿 over 𝐾. It is also the minimum cardinality of a subset 𝑌 of 𝐿 such that 𝐿/𝐾(𝑌) is algebraic.
√
Theorem 3.36. If 𝑃 is a generic convex 𝑛-gon in ℝ2 then 𝑥𝑐(𝑃) ≥ 2𝑛

Recall that a generic convex 𝑛-gon is defined as the coordinates of its vertices being algebraically
independent over ℚ. Algebraically independent means there is no non-zero polynomial over ℚ
with variables 𝑥1 , 𝑥2 , · · · , 𝑥2𝑛 such that putting in coordinates of vertices makes it 0.
Let 𝛼1 , · · · , 𝛼 2𝑛 denote 𝑛 vertices of 𝑃, listed in any order. So, 𝑋 = {𝛼1 , · · · , 𝛼 2𝑛 } is alg. independent
over ℚ.
Let 𝑃 be the projection of a 𝑑-dimensional polytope 𝑄 with 𝑘 facets. Without loss of generality,
assume that 𝑃 is projection of 𝑄 along first two coordinates (if not, a change of basis gives a new
polytope 𝑄 ′ whose first 2-coordinates project to 𝑃).
Consider a linear description of 𝑄. We need 𝑘 inequalities for facets and each inequality has
𝑑 + 1 coefficients. Let us denote the coefficients as 𝛽 𝑖,𝑗 for 𝑖 ∈ {1, · · · , 𝑘}, 𝑗 ∈ {1, · · · , 𝑑 + 1}. Then by
cramer’s rule, the corners have coordinates as rational function of 𝛽 𝑖,𝑗 coordinates. Hence, 𝛼 𝑗 are
rational function of 𝛽 𝑖,𝑗 . Hence each 𝛼 𝑗 belongs to the extension field 𝐿(ℚ, 𝛽). Since 𝑋 is algebraically
independent over ℚ, 𝑋 ⊆ 𝐿, atmost transcedence degree of 𝐿/ℚ is atleast 2𝑛. But, atmost 𝑘(𝑑 + 1)
algebraically independent elements have been added to field ℚ. Hence transcedence degree is
upper bounded by 𝑘(𝑑 + 1). (If not, there exists an element of ℚ in the chosen set, implying that
setting all other coefficients to 0 in the polynomial gives us a non-trivial polynomial attaining 0
on the chosen √ points). Hence, 𝑘(𝑑 + 1) ≥ 2𝑛. But, since polytopes are bounded, we get 𝑑 + 1 ≤ 𝑘.
Hence, 𝑘 ≥ 2𝑛.

3.7 Exact Spanning Tree, Matchings

In this section, we discuss the progress in exact matchings made by [JSY]. However, the primary
objective of this inquiry is to ask the following questions.

1. What is the extension complexity of the exact spanning tree polytope i.e. polytope of spanning
trees having exactly 𝑘-red edges (Alternatively the polytope of spanning trees having odd red
edges)?

2. What is the extension complexity of polytope of spanning trees having weight exactly 𝑊?

Both of the above questions remain open, however we discuss some small observations here.

3.7.1 Exact weighted spanning tree

The first observation to be noted is that the problem is NP-complete. There is a simple reduction
from subset-sum. Let there be a set of objects with weights {𝑤 1 , · · · , 𝑤 𝑛 } and target weight 𝑊.
Then, consider a graph given by fig. 5, and target weight 𝑇 = 𝑊 + (𝑛 + 1)𝜀.

20
 𝑤1 + 𝜀 𝑤𝑛 + 𝜀

𝜀 𝜀
𝜀
···

Figure 5: Reduction from subset sum

Lemma 3.37. The graph shown in fig. 5 has a spanning tree of weight 𝑊 + (𝑛 + 1)𝜀 iff there is a subset of
{𝑤 1 , · · · , 𝑤 𝑛 } with target sum 𝑊.

Proof. Let there exist a subset satisfying target sum. Then choose red edges for all elements in that
subset, and choose blue edges for all other elements. Choose an additional blue edge for exactly
one element of chosen subset to complete the spanning tree. It is clear that the weight of the tree is
𝑊 + (𝑛 + 1)𝜀.
Let there exist a spanning tree with target weight. Since the spanning tree must have 𝑛 + 1 edges,
total degree must be 2𝑛 + 2. Now, every edge chosen increases the degree of upper vertices in the
figure by 1, hence degree of upper vertices is 𝑛 + 1. Now, every vertex in bottom row must have
degree atleast one and combined degree of 𝑛 vertices is 𝑛 + 1. Hence 𝑛 − 1 vertices have degree
1, and exactly one vertex has degree 2 in the bottom row. Let there be 𝑘 red edges incident on
the set 𝑆 in bottom row and 𝑛 − 𝑘 + 1 blue edges. So, 𝑖∈𝑆 𝑤 𝑖 + (𝑛 + 1)𝜀 = 𝑊 + (𝑛 + 1)𝜀, hence
Í
𝑊 = 𝑖∈𝑆 𝑤 𝑖 .
Í
□

Hence, for general weights we cannot expect a small extended formulation since that would
give us an LP based algorithm for subset sum running in PTIME. Another contributing factor to
why such√ a result cannot be expected is that the 0 − 1 knapsack polytope has extension complexity
atleast 2 𝑛 as shown in [PV13] (knapsack polytope is defined as convex hull of points in {0, 1} 𝑛
such that 𝑎 𝑇 𝑥 ≤ 𝑏). However, we can expect a small extended formulation for small weights.
The obvious first attempt is adding a target weight constraint to the spannning tree polytope. That
is,

𝑥(𝐸(𝑈)) ≤ |𝑈 | − 1 𝜙≠𝑈 ⊂𝑉
𝑥(𝐸(𝑉)) = |𝑉 | − 1
Õ
𝑥𝑒 𝑤𝑒 = 𝑊
𝑒

However this attempt fails because adding the constraint 𝑒 𝑥 𝑒 𝑤 𝑒 = 𝑊 makes the polytope non-
Í
integral. Consider the following example (see fig. 6) where values of 𝑤 𝑒 are indicated in black and
those in 𝑥 𝑒 are red, with target weight 𝑊 = 9

21
 5 7
0.5 0.5

1
3

Figure 6: Counterexample for obvious LP formulation

We can see that no spanning tree of weight 9 exists, however corresponding LP gives a feasible
non-integral solution.

3.7.2 Exact or odd spanning tree

Since the case with weights is not approachable, an alternative to be considered is the case with no
weights on edges and choosing spanning trees with exactly 𝑘 red edges, in a graph colored with
red or blue edges. We know (due to Balas) that for polytopes 𝑃𝑖 , 𝑖 ∈ {1, · · · , 𝑛},

Theorem 3.38. 𝑥𝑐( 𝑃𝑖 ) ≤ 𝑥𝑐(𝑃𝑖 )

Ð Í
𝑖 𝑖

This result is simple to obtain by taking points 𝑥 𝑖 in polytopes 𝑃𝑖 and taking their convex
combination to get new point 𝑥. Now, consider the polytope of odd spanning trees. Let 𝑃𝑖 denote
polytope of spanning trees with exactly 𝑖 red edges.
Õ
𝑥𝑐(𝑃𝑜𝑑𝑑 ) ≤ 𝑃𝑖
odd 𝑖∈ℕ

A related question was tackled by Svennson et. al. [JSY], where it was shown that extension
complexity of odd perfect matchings in bipartite graphs is exponential, hence the same result holds
for exact matchings as well. This paper has a 2-step approach to solving this problem,

1. Giving an inequality description for odd perfect matchings (albeit with exponentially many
constraints).

2. Relax the polytope and lower bound its extension complexity.

The approach used for lower bounding the extension complexity is similar to the one used by
Rothvoss [Rot17] in lower bounding extension complexity of the perfect matching polytope.
Back to tackling this question, the problem of finding exactly 𝑘 red edges, odd red edges has a
polynomial time algorithm. So, it is natural to expect a small extended formulation for this problem.
However, this remains open.

22
4 Discussion
In conclusion the set of open problems relevant to this survey are,

1. Extension complexity of dilworth truncation in general matroids. As a first step, consider

modified constraints as 𝑥(𝑆) ≤ 2𝑟(𝑆) − 1. For graphic matroids, we see that 2|𝑆| − 3 is a 𝑘 − 𝑙
sparsity matroid, hence it has a small extension complexity (One possible approach can be
using the hitting set).

2. Extension complexity of odd red spanning free polytope. First, we need to find an LP
formulation for this followed in order to get started.

3. Alternative method for extension of regular matroid using hitting set.

5 Acknowledgements
We are thankful to Prof. Rohit Gurjar for helpful discussions and providing direction to this work.

References
[ACF18] Manuel Aprile, Alfonso Cevallos, and Yuri Faenza. On 2-level polytopes arising in
combinatorial settings. SIAM Journal on Discrete Mathematics, 32(3):1857–1886, 2018.

[AF19] Manuel Aprile and Samuel Fiorini. Regular matroids have polynomial extension
complexity, 2019.

[Apr22] Manuel Aprile. Extended formulations for matroid polytopes through randomized
protocols. Operations Research Letters, 50(2):145–149, 2022.

[Edm03] Jack Edmonds. Submodular Functions, Matroids, and Certain Polyhedra, page 11–26.
Springer-Verlag, Berlin, Heidelberg, 2003.

[FFGT15] Yuri Faenza, Samuel Fiorini, Roland Grappe, and Hans Raj Tiwary. Extended formu-
lations, nonnegative factorizations, and randomized communication protocols. Math.
Program., 153(1):75–94, oct 2015.

[FMP+ 12] Samuel Fiorini, Serge Massar, Sebastian Pokutta, Hans Raj Tiwary, and Ronald de Wolf.
Linear vs. semidefinite extended formulations: Exponential separation and strong lower
bounds. In Proceedings of the Forty-Fourth Annual ACM Symposium on Theory of Computing,
STOC ’12, page 95–106, New York, NY, USA, 2012. Association for Computing Machinery.

[FRT12] Samuel Fiorini, Thomas Rothvo, and Hans Raj Tiwary. Extended formulations for
polygons. Discrete Comput. Geom., 48(3):658–668, oct 2012.

[Hak65] S.L. Hakimi. On the degrees of the vertices of a directed graph. Journal of the Franklin
Institute, 279(4):290–308, 1965.

23
[HIH21] Keiko Imai Hiroshi Imai and Hidefumi Hiraishi. Extended formulations of lower-
truncated transversal polymatroids. Optimization Methods and Software, 36(2-3):326–331,
2021.

[IKK+ 16] Satoru Iwata, Naoyuki Kamiyama, Naoki Katoh, Shuji Kijima, and Yoshio Okamoto.
Extended formulations for sparsity matroids. Math. Program., 158(1–2):565–574, jul 2016.

[JSY] Xinrui Jia, Ola Svensson, and Weiqiang Yuan. The Exact Bipartite Matching Polytope Has
Exponential Extension Complexity, pages 1635–1654.

[KLdW15] Jedrzej Kaniewski, Troy Lee, and Ronald de Wolf. Query complexity in expectation.
In Magnús M. Halldórsson, Kazuo Iwama, Naoki Kobayashi, and Bettina Speckmann,
editors, Automata, Languages, and Programming, pages 761–772, Berlin, Heidelberg, 2015.
Springer Berlin Heidelberg.

[KPT10] Volker Kaibel, Kanstantsin Pashkovich, and Dirk Oliver Theis. Symmetry matters for the
sizes of extended formulations. In Integer Programming and Combinatorial Optimization,
pages 135–148. Springer Berlin Heidelberg, 2010.

[Li18] Yaqiao Li. Lecture notes in extension complexity, February 2018.

[Oxl06] James G. Oxley. Matroid Theory (Oxford Graduate Texts in Mathematics). Oxford University
Press, Inc., USA, 2006.

[PV13] Sebastian Pokutta and Mathieu Van Vyve. A note on the extension complexity of the
knapsack polytope. Operations Research Letters, 41(4):347–350, 2013.

[Rot17] Thomas Rothvoss. The matching polytope has exponential extension complexity. J.
ACM, 64(6), sep 2017.

[S+ 03] Alexander Schrijver et al. Combinatorial optimization: polyhedra and efficiency, volume 24.
Springer, 2003.

[Shi20] Yaroslav Shitov. Sublinear extensions of polygons, 2020.

[Swa87] E. R. Swart. P = NP. Technical Report, University of Guelph, 1986; revision 1987.

[Yan88] Mihalis Yannakakis. Expressing combinatorial optimization problems by linear programs.

In Proceedings of the twentieth annual ACM symposium on Theory of computing, pages 223–228,
1988.

24