## Lots of Ingleton Matroids

In 1971, Aubrey Ingleton [1] showed that the rank function of representable matroids satisfies additional linear inequalities that do not follow from the usual rank axioms. The inequality he observed now bears his name. It states that, for all sets $A,B,C,D$ in a representable matroid $M$,
\begin{align*}
&r(A\cup B) + r(A \cup C) + r(A \cup D) + r(B \cup C) + r(B \cup D) \\
\ge & \ r(A) + r(B) + r(A \cup B \cup C) + r(A \cup B \cup D) + r(C \cup D)
\end{align*}

As difficult to typeset as it may be, this is an intriguing fact. The problem of characterizing which matroids are representable is difficult; various authors [2][3] have considered whether it is possible to extend one of the usual axiom systems in a natural way to capture precisely the representable matroids. Ingleton’s inequality suggests the kind of extra axiom that might need to be added to the rank axioms if thinking along these lines. Perhaps more importantly, the Ingleton inequality gives a concise way to certify that a matroid is not representable; if $(A,B,C,D)$ is a $4$-tuple in a matroid $M$ violating the inequality, then $M$ is non-representable. This has been useful to many authors that wish to construct non-representable matroids, in particular in a result of Mayhew, Newman, Welsh and Whittle [4] that constructs a very rich family of excluded minors for the real-representable matroid, all violating Ingleton’s inequality.

Finally, the Ingleton inequality is closely related to everyone’s favourite non-representable matroid, the Vámos matroid $V_8$. This rank-$4$ matroid has eight elements and ground set $E$ that is the disjoint union of four two-element sets $P_1,P_2,P_3,P_4$, in which the dependent four-element sets are precisely the pairs $P_i \cup P_j$ with $i < j$ and $(i,j) \ne (3,4)$. The tuple $(A,B,C,D) = (P_1,P_2,P_3,P_4)$ violates the inequality in this case: the left-hand and right-hand sides are $15$ and $16$ respectively. On the other hand, satisfying Ingleton’s inequality for all choices of $A,B,C,D$ is not enough to imply representability; for example, the non-Desargues matroid satisfies Ingleton’s inequality but is still non-representable, as is the direct sum of the Fano and non-Fano matroids.

Counting

This post is about a recent result [7] I’ve obtained with Jorn van der Pol that answers what we consider to be a natural question: how `close’ is the class of matroids that satisfy Ingleton’s inequality to the class of representable matroids? For convenience, we will call a matroid Ingleton if Ingleton’s inequality is satisfied for all choices of $(A,B,C,D)$. It might not be immediatley clear what the answer should be. Some (weak) positive evidence is the fact that the Ingleton matroids are closed under minors and duality (see [5], Lemmas 3.9 and 4.5) and that $V_8$ is non-Ingleton. However, I think the natural educated guess goes in the other direction; Ingleton’s inequality seems too coarse to capture, even approximately, a notion as intricate as linear representability In fact, Mayhew, Newman and Whittle [3] have in fact shown that it is impossible to define the class of representable matroids by adding any finite list of rank inequalities to the usual rank axioms, let alone a single inequality.

Our result confirms this suspicion.

Theorem 1: For all sufficiently large $n$, the number of Ingleton matroids with ground set $\{1,\dotsc,n\}$ is at least $2^{\frac{1.94 \log n}{n^2}\binom{n}{n/2}}$.

To view this result in context, the lower bound should be compared to the counts of representable matroids and of all matroids. On a fixed ground set $[n] = \{1,\dotsc, n\}$ The number of representable matroids is at most $2^{n^3/4}$ for all $n \ge 12$ [6], and the number of matroids is at least $2^{\tfrac{1}{n}\binom{n}{n/2}}$. (Both these upper and lower bounds are in fact correct counts up to a constant factor in the exponent). This first expression is singly exponential in $n$, having the form $2^{\mathrm{poly}(n)}$, while the second is doubly exponential, having the form $2^{2^n/\mathrm{poly}(n)}$. The lower bound in our theorem is of the second type, showing that the Ingleton matroids asymptoticlaly dwarf the representable matroids in number. In other words, knowing that a matroid is Ingleton tells you essentially nothing about whether it is representable. (In fact, our techniques show that the number of rank-$4$ Ingleton matroids on $[n]$ is asymptotically larger than the class of all representable matroids on $[n]$, which seems surprising.) The ideas in the proof of the above theorem are simple and we obtain a nice excluded minor result on the way; I briefly sketch them below. For the full proof, see https://arxiv.org/abs/1710.01924 .

Ingleton Sparse Paving Matroids

Our proof actually constructs a large number of Ingleton matroids of a specific sort: sparse paving. These matroids, which have come up in previous matroidunion posts, play a very special role in the landscape of all matroids; they are matroids that, while having a somewhat trivial structure, are conjectured to comprise almost all matroids. For the definition, it is easier to talk about the nonbases of a matroid than its bases. Let $\binom{[n]}{r}$ denote the collection of $r$-element subsets of $[n]$. Given a rank-$r$ matroid on $[n]$, call a dependent set in $\binom{[n]}{r}$ a nonbasis of $M$.

A rank-$r$ matroid is sparse paving if for any two nonbases $U_1,U_2$ of $M$, we have $|U_1 \cap U_2| < r-1$. Equivalently, no two nonbases of $M$ differ by a single exchange, or every dependent set of $M$ is a circuit-hyperplane of $M$. In fact, this condition itself implies that the matroid axioms are satisfied; given any collection $\mathcal{K}$ of $r$-element subsets of $[n]$, if no two sets in $K$ intersect in exactly $r-1$ elements, then $\mathcal{K}$ is the set of nonbases of a sparse paving matroid on $[n]$. Thus, an easy way to guarantee that a set $\mathcal{K}$ is actually the set of nonbases of a matroid is to prove that no two of its members intersect in $r-1$ elements.

Our key lemma gives a simpler way to understand Ingleton’s inequality for sparse paving matroids. In general, it is very hard to mentally juggle the $A$’s, $B$’s, $C$’s and $D$’s while working with the inequality, but for sparse paving matroids, things are much simpler.

Lemma 1: If $M$ is a rank-$r$ sparse paving matroid, then $M$ is Ingleton if and only if there do not exist pairwise disjoint sets $P_1,P_2,P_3,P_4,K$ where $|P_1| = |P_2| = |P_3| = |P_4| = 2$ and $|K|=r-4$, such that the five $r$-element sets $K \cup P_i \cup P_j: i < j, (i,j) \ne (3,4)$ are nonbases, while the set $K \cup P_3 \cup B_4$ is a basis.

This statement may look technical, but it should also look familiar. If $P_1,P_2,P_3,P_4,K$ are sets as above, then the minor $N = (M / K)|(P_1\cup P_2 \cup P_3 \cup P_4)$ is an eight-element sparse paving matroid having a partition $(P_1,P_2,P_3,P_4)$ into two-element sets, where precisely five of the six sets $P_i \cup P_j$ are nonbases, and the last is a basis. This is a structure very similar to that of the Vámos matroid. Call an eight-element, rank-$4$ matroid $N$ having such a property Vámos-like. (Such an $N$ need not be precisely the Vámos matroid, as there may be four-element sets of other forms that are also nonbases of $N$). In any Vámos-like matroid, $(A,B,C,D) = (P_1,P_2,P_3,P_4)$ will violate Ingleton’s inequality. We can restate Lemma 1 as follows.

Lemma 1 (simplified): If $M$ is a sparse paving matroid, then $M$ is Ingleton if and only if $M$ has no Vámos-like minor.

There are evidently only finitely many Vámos-like matroids, since they have eight elements; in fact, thanks to Dillon Mayhew and Gordon Royle’s excellent computational work [8], we know all $39$ of them; as well as the Vámos matroid itself, they include the matroid AG$(3,2)^-$ obtained by relaxing a circuit-hyperplane of the rank-$4$ binary affine geometry. It is easy to show that the sparse paving matroids are themselves a minor-closed class with excluded minors $U_{1,1} \oplus U_{0,2}$ and $U_{0,1} \oplus U_{2,2}$. Combined with Lemma 1, this gives us a nice excluded minor theorem:

Theorem 2: There are precisely $41$ excluded minors for the class of Ingleton sparse paving matroids: the $39$ Vámos-like matroids, as well as $U_{1,1} \oplus U_{0,2}$ and $U_{0,1} \oplus U_{2,2}$.

The Proof

Armed with Lemma 1, we can now take a crack at proving our main theorem. For simplicity, we will prove a slightly weaker result, with a worse constant of $0.2$ and no logarithmic factor in the exponent. The stronger result is obtained by doing some counting tricks a bit more carefully. The good news is that the proof of the weaker theorem is short enough to fit completely in this post.

Theorem 3: For all sufficiently large $n$, there are at least $2^{0.2n^{-2}\binom{n}{n/2}}$ Ingleton sparse paving matroids with ground set $[n]$

Proof
Let $n$ be large and let $r = \left\lfloor n/2 \right\rfloor$ and $N = \binom{n}{r}$. Let $c = 0.4$. We will take a uniformly random subset $\mathcal{X}$ of $\left\lfloor c n^{-2}\binom{n}{r}\right\rfloor$ of the $\binom{n}{r}$ sets in $\binom{[n]}{r}$. We then hope that $\mathcal{X}$ is the set of nonbases of an Ingleton sparse paving matroid. If it is not, we remove some sets in $\mathcal{X}$ so that it is.

We consider two possibilities that, together, encompass both ways $\mathcal{X}$ can fail to be the set of nonbases of a sparse paving matroid. They are

1. $\mathcal{X}$ is not the set of nonbases of a sparse paving matroid. (That is, there are sets $U_1,U_2 \in \mathcal{X}$ whose intersection has size $r-1$.)
2. $\mathcal{X}$ is the set of nonbases of a sparse paving matroid, but this matroid fails to be Ingleton. (That is, there are pairwise disjoint subsets $P_1,P_2,P_3,P_4,K$ of $[n]$ where $|P_1| = |P_2| = |P_3| = |P_4| = 2$ and $|K|=r-4$, such that at least five of the six $r$-element sets $K \cup P_i \cup P_j: i < j, (i,j)$ are nonbases.)

Let $a(\mathcal{X}),b(\mathcal{X})$ denote the number of times each of these types of failure occurs. The condition in (2) is slightly coarser than required, for reasons we will see in a minute. So $a(X)$ is the number of pairs $(U_1,U_2)$ with $|U_1 \cap U_2| = r-1$ and $U_1,U_2 \in X$, and $b(X)$ is the number of $5$-tuples $(P_1,P_2,P_3,P_4,K)$ satisfying the condition in (2).

Claim: If $n$ is large, then $\mathbf{E}(a(\mathcal{X}) + 2b(\mathcal{X})) < \tfrac{1}{2}|\mathcal{X}|$.

Proof of claim: The number of pairs $U_1,U_2$ intersecting in $r-1$ elements is $\binom{n}{r}r(n-r) < n^2\binom{n}{r}$. For each such pair, the probability that $U_1,U_2 \in X$ is at most $(|\mathcal{X}|/\binom{n}{r})^2 \le c^2n^{-4}$. Therefore $\mathbf{E}(a(\mathcal{X})) \le c^2n^{-2}\binom{n}{r} = (1+o(1))c|\mathcal{X}|.$

The number of $5$-tuples $(K,P_1,P_2,P_3,P_4)$ pf disjoint sets where $|K| = r-4$ and $|P_i| = 2$ is at most $\binom{n}{2}^4\binom{n-8}{r-4} < \tfrac{n^8}{16}\binom{n}{r}$. The probability that, for some such tuple, at least five of the sets $K \cup P_i \cup P_j$ are in $X$ is at most $6(|\mathcal{X}|/\binom{n}{r})^5 \le 6c^5n^{-10}$. Therefore
$\mathbf{E}(b(\mathcal{X})) \le \frac{n^8}{16}\binom{n}{r}\cdot 6c^5 n^{-10} = (1+o(1))\tfrac{3}{8}c^4|\mathcal{X}|.$
By linearity of expectation, the claim now holds since $c = 0.4$ gives $c + \tfrac{3}{4}c^4 < \tfrac{1}{2}$.

Now, let $\mathcal{X}_0$ be a set of size $\left\lfloor c n^{-2}\binom{n}{r}\right\rfloor$ for which $a(\mathcal{X}) + 2b(\mathcal{X}) < \tfrac{1}{2}|\mathcal{X}_0|$. We now remove from $\mathcal{X}_0$ one of the two sets $U_1,U_2$ for each pair contributing to $a(\mathcal{X}_0)$, and two of the sets $K \cup P_i \cup P_j$ for each tuple $(K,P_1,P_2,P_3,P_4)$ contributing to $b(\mathcal{X}_0)$. This leaves a subset $\mathcal{X}_0’$ of $\mathcal{X}_0$ of size $\tfrac{1}{2}|\mathcal{X}_0| \approx \tfrac{1}{2}cn^{-2}\binom{n}{r} = 0.2n^{-2}\binom{n}{r}$. By construction and Lemma 1, $\mathcal{X}_0’$ is the set of nonbases of a rank-$r$ Ingleton sparse paving matroid on $[n]$. However (and this is where condition (2) needed to be strengthened slightly), so are all subsets of $\mathcal{X}_0’$. This puts the size of $\mathcal{X}_0’$ in the exponent; there are therefore at least $2^{0.2n^{-2}\binom{n}{n/2}}$ Ingleton sparse paving matroids on $[n]$, as required.

References

1. A.W. Ingleton. Representation of matroids. In D.J.A Welsh, editor, Combinatorial mathematics and its applications (Proceedings of a conference held at the Mathematical Institute, Oxford, from 7-10 July, 1969). Academic Press, 1971.
2. P. Vámos. The missing axiom of matroid theory is lost forever. J. London Math. Soc. 18 (1978), 403-408
3. D. Mayhew, M. Newman and G. Whittle. Yes, the “missing axiom” of matroid theory is lost forever, arXiv:1412.8399
4. D. Mayhew, M. Newman and G. Whittle. On excluded minors for real-representability. J. Combin. Theory Ser. B 66 (2009), 685-689.
5. A. Cameron. Kinser inequalities and related matroids. Master’s Thesis, Victoria University of Wellington. Also available at arXiv:1401.0500
6. P. Nelson. Almost all matroids are non-representable. arXiv:1605.04288
7. P. Nelson and J. van der Pol. Doubly exponentially many Ingleton matroids. arXiv:1710.01924
8. D. Mayhew and G.F. Royle. Matroids with nine elements. J. Combin. Theory Ser. B 98 (2008), 882-890.

# Clutters III

A long time ago I started a series of posts abut clutters. The most recent post followed the textbook by Gérards Cornuéjols in defining several important classes, and introduced a Venn diagram, showing the relationships between them.

In this post we will spend a little more time discussing this diagram.

We let $H=(S,\mathcal{A})$ be a clutter. This means that $S$ is a finite set, and the members of $\mathcal{A}$ are subsets of $S$, none of which is properly contained in another. We let $M$ stand for the incidence matrix of $H$. This means that the columns of $M$ are labelled by the elements of $S$, and the rows are labelled by members of $\mathcal{A}$, where an entry of $M$ is one if and only if the corresponding element of $S$ is contained in the corresponding member of $\mathcal{A}$. Any entry of $M$ that is not one is zero. Let $w$ be a vector in $\mathbb{R}^{S}$ with non-negative values. We have two fundamental linear programs:

(1) Find $x\in \mathbb{R}^{S}$ that minimises $w^{T}x$ subject to the constraints $x\geq \mathbf{0}$ and $Mx\geq \mathbf{1}$.

The vectors $\mathbf{0}$ and $\mathbf{1}$ have all entries equal to zero and one, respectively. When we write that real vectors $a$ and $b$ with the same number of entries satisfy $a\geq b$, we mean that each entry of $a$ is at least equal to the corresponding entry in $b$.

(2) Find $y\in\mathbb{R}^{\mathcal{A}}$ that maximises $y^{T}\mathbf{1}$ subject to the constraints $y\geq \mathbf{0}$ and $y^{T}M\leq w$.

Lemma 1. Any clutter with the Max Flow Min Cut property also has the packing property.

Proof. Let $H$ be a clutter with the Max Flow Min Cut property. This means that for any choice of vector $w$ with non-negative integer entries, the programs (1) and (2) both have optimal solutions with integer values.

We will show that $H$ has the packing property. According to the definition in the literature, this means that for any choice of vector $w$ with entries equal to $0$, $1$, or $+\infty$, there are optimal solutions to (1) and (2) with integer entries. I think there is a problem with this definition. Assume that $r$ is a row of $M$, and every member of the support of $r$ receives a weight of $+\infty$ in the vector $w$. Then (2) cannot have an optimal solution. If $y$ is a purported optimal solution, then we can improve it by adding $1$ to the entry of $y$ that corresponds to row $r$. We are instructed that if entry $i$ in $w$ is $+\infty$, then this means when $x$ is a solution to (1), then entry $i$ of $x$ must be $0$. Again, we have a problem, for if $r$ is a row of $M$, and the entire support of $r$ is weighted $+\infty$, then (1) has no solution: if $x$ were a solution, then it would be zero in every entry in the support of $r$, meaning that $Mx$ has a zero entry.

The literature is unanimous in saying that $H$ has the packing property if (1) and (2) both have integral optimal solutions for any choice of vector $w$ with entries $0$, $1$, and $+\infty$. As far as I can see, this means that any clutter with $\mathcal{A}$ non-empty does not have the packing property: simple declare $w$ to be the vector with all entries equal to $+\infty$. Then neither (1) nor (2) has an optimal solution at all. I think the way to recover the definition is to say that whenever $w$ with entries $0$, $1$, or $+\infty$ is chosen in such a way that (1) and (2) have solutions, they both have optimal solutions that are integral. This is the definition that I will use here.

After this detour, we return to our task, and assume that $H$ has the Max Flow Min Cut property. Assume that $w$ is a vector with entries equal to $0$, $1$, or $+\infty$, and that (1) and (2) both have solutions. This means that any row in $M$ has a member of its support which is not weighted $+\infty$ by $w$. We obtain the vector $u$ by replacing each $+\infty$ entry in $w$ with an integer that is greater than $|S|$. Now because $H$ has the Max Flow Min Cut property, it follows that there are optimal integral solutions, $x$ and $y$, to (1) and (2) (relative to the vector $u$). We will show that $x$ and $y$ are also optimal solutions to (1) and (2) relative to the vector $w$.

We partition $S$ into $S_{0}$, $S_{1}$, and $S_{+\infty}$ according to whether an element of $S$ receives a weight of $0$, $1$, or $+\infty$ in $w$. We have assumed that no member of $\mathcal{A}$ is contained in $S_{+\infty}$. We note that if $z\in \mathbb{Z}^{S}$ is a vector which is equal to zero for each element of $S_{+\infty}$, and one everywhere else, then $z$ is a solution to (1), by this assumption. Moreover, $w^{T}z=u^{T}z\leq |S|$. Now it follows that $x$ must be zero in every entry in $S_{+\infty}$, for otherwise $u^{T}x>|S|$, and therefore $x$ is not an optimal solution to (1) relative to $u$. Since $x$ is integral and optimal, it follows that we can assume every entry is either one or zero. If $x$ is not an optimal solution to (1) relative to $w$, then we let $z$ be an optimal solution with $w^{T}z < w^{T}x$. But by convention, $z$ must be zero in every entry of $S_{+\infty}$. Therefore $u^{T}z=w^{T}z < w^{T}x=u^{T}z$, and we have a contradiction to the optimality of $x$. Thus $x$ is an optimal solution to (1) relative to the $\{0,1,+\infty\}$-vector $w$.

Now for problem (2). Since $y$ is integral and non-negative, and $y^{T}M\leq w$, where every member of $\mathcal{A}$ contains an element of $S_{0}$ or $S_{1}$, it follows that each entry of $y$ must be either one or zero. Let $z$ be any solution of (2) relative to $w$. Exactly the same argument shows that each entry of $z$ is between zero and one. Therefore $z^{T}\mathbf{1}\leq y^{T}\mathbf{1}$ so $y$ is an optimal solution to (2).

We have shown that relative to the vector $w$, both (1) and (2) have optimal solutions that are integral. Hence $H$ has the packing property. $\square$

Lemma 2. A clutter with the packing property packs.

Proof. This one is easy. In order to prove that $H$ packs, we merely need to show that (1) and (2) have optimal integral solutions when $w$ is the vector with all entries equal to one. But this follows immediately from our revised definition of clutters with the packing property. $\square$

The final containment we should show is that clutters with the packing property are ideal. Idealness means that (1) has an optimal integral solution for all vectors $w\in \mathbb{R}^{S}$. This proof is difficult, so I will leave it for a future post. Usually we prove it by using a theorem due to Lehman [Leh].

Theorem (Lehman). The clutter $H$ is ideal if and only if (1) has an optimal integral solution for all choices of vector $w\in\{0,1,+\infty\}^{S}$.

Question. Is there a short proof that clutters with the packing property are ideal? One that does not rely on Lehman’s (quite difficult) theorem?

We will conclude with some examples showing that various containments are proper.

Let $C_{3}^{2}$ and $C_{3}^{2+}$ be the clutters with incidence matrices
$\begin{bmatrix} 1&1&0\\ 0&1&1\\ 1&0&1 \end{bmatrix} \quad\text{and}\quad \begin{bmatrix} 1&1&0&1\\ 0&1&1&1\\ 1&0&1&1 \end{bmatrix}.$
Let $Q_{6}$ and $Q_{6}^{+}$ be the clutters with incidence matrices
$\begin{bmatrix} 1&1&0&1&0&0\\ 1&0&1&0&1&0\\ 0&1&1&0&0&1\\ 0&0&0&1&1&1 \end{bmatrix} \quad\text{and}\quad \begin{bmatrix} 1&1&0&1&0&0&1\\ 1&0&1&0&1&0&1\\ 0&1&1&0&0&1&1\\ 0&0&0&1&1&1&1 \end{bmatrix}$

Exercise. Check that:

1. $C_{3}^{2}$ is not ideal and does not pack,
2. $C_{3}^{2+}$ packs, but is not ideal,
3. $Q_{6}$ is ideal, but does not pack,
4. $Q_{6}^{+}$ is ideal and packs, but does not have the packing property.

This leaves one cell in the Venn diagram without a clutter: the clutters with the packing property that do not have the Max Flow Min Cut property. In fact, Conforti and Cornuéjols [CC] have speculated that no such clutter exists.

Conjecture (Conforti and Cornuéjols). A clutter has the packing property if and only if it has the Max Flow Min Cut property.

[CC] M. Conforti and G. Cornuéjols, Clutters that Pack and the Max Flow Min Cut Property: A Conjecture, The Fourth Bellairs Workshop on Combinatorial Optimization, W.R. Pulleyblank and F.B. Shepherd eds. (1993).

[Leh] A. Lehman, On the width-length inequality. Mathematical Programming December 1979, Volume 16, Issue 1, pp 245–259.

# SiGMa 2017 Recap

The Workshop on Structure in Graphs and Matroids (SiGMa 2017) just wrapped up a few weeks ago in Waterloo.  This was a continuation of similar workshops organized by Bert Gerards in 20082010, and 2012 and by Stefan van Zwam and Rudi Pendavingh in 2014 and 2016.

I would like to thank Jim Geelen, Peter Nelson, Luke Postle, and Stefan van Zwam (the organizers of this year’s workshop) who did an excellent job in choosing the program and making sure everything ran smoothly.  They even helped fellow Matroid Union blogger Nathan Bowler overcome Canadian Visa issues in time to give his (very nice) plenary talk (thanks also go to Alan).

This year the workshop was held in commemoration of William T. Tutte, who would have turned 100 this year.  See here for a short biography of Professor Tutte.  Some matroid theorists may not be aware of Tutte’s extremely important contributions as a code breaker at Bletchley Park during the Second World War.  The C&O Department at Waterloo has also been hosting a Distinguished Lecture Series in honour of Tutte this summer.  Click on this link for the list of speakers and to watch the videos on YouTube. SiGMa 2017 also featured a rare conference dinner speech by Jim Geelen on Tutte’s work.

In case you were not able to attend, all the talks from SiGMa 2017 were recorded and are now viewable on the C&O Department’s YouTube Channel.  The overall quality of talks was very high.  I won’t bias you with all of my personal favourites, but for example, all the plenaries and Paul Seymour were excellent.

# The Template Theorem

Geelen, Gerards, and Whittle are still working on the write-up of their Herculean effort to build a matroid minor structure theory generalizing the work that Robertson and Seymour did for graphs. One of the consequences of their work has been published, namely a discussion of the highly connected members of a minor-closed class of matroids [1]. This work was mentioned by Peter in passing here and in more detail here. What Peter didn’t talk about was the most detailed version of that structure theorem. Today I will do just that. To keep things light, I will focus on binary matroids, but [1] has results for matroids representable over any finite field.

## 1. Low-rank perturbations

The key concept is that of a perturbation of a representable matroid. If $M = M[A]$, and $T$ is a matrix with the same dimensions as $A$, then $M[A+T]$ is a perturbation of $M$. The hope is that, if we know a lot about $M$ and $T$ has low rank, then the resulting matroid still resembles $M$ to a large extent. We will study perturbations of graphic matroids, and start with a few examples. It will be convenient to drop the restriction that the perturbed matroid has the same size as $M$, and therefore we will allow the addition of a bounded number of elements.

Let $A$ be the vertex-edge incidence matrix of a graph, and let $A’$ be the matrix obtained from $A$ by adding an arbitrary row. The matroid $M[A’]$ is known as an even-cycle matroid. It is an instance of the class of lift matroids frequently discussed on this blog, such as by Irene (here,  here, and here) and two weeks ago by Daryl (here). Note that it can be obtained from $M$ by coextending the matroid by one element, and then deleting that element. They can be visualized by coloring the edges of the graph, calling an edge even if its corresponding column in the matrix has a 0 in the new row, and odd if it has a 1. This class of matroids is closed under minors.

Let $A$ again be the vertex-edge incidence matrix of a graph, and this time let $A’$ be the matrix obtained from $A$ by adding an arbitrary column. The matroid $M[A’]$ will have one extra element, and is known as a graft. Grafts played a crucial role in the proofs of many fundamental results in matroid theory, including Seymour’s Decomposition Theorem for regular matroids [2]. They can be visualized by coloring the vertices of the graph whose corresponding matrix row has a 1 in the new column.

Note that the class of grafts is not closed under minors (contracting the graft element destroys the graphic structure). However, a closely related class, where we add an element to the graphic matroid and immediately contract that element, is closed under minors. The duals of these matroids are known as the even cut matroids. They have been extensively studied by Guenin, Pivotto, and Wollan (see, for instance, Irene’s PhD thesis [3]).

### 1.3 Adding a few elements to a low-rank flat

Let $G$ be a graph with $t$ distinguished vertices. We take the vertex-edge incidence matrix, and add a few columns, but now these columns are only allowed to have nonzero entries corresponding to the $t$ distinguished vertices. The resulting class of matroids will resemble a graphic matroid everywhere but in a low-rank set. Note that this can be seen as a variation of the “adding a column” construction, but sometimes it is useful to consider the operation separately.

## 2. Frame templates

Next, we ask ourselves what happens if we allow several of the above operations in a row. Can there be fundamentally different ways to perturb a graphic matroid? Geelen, Gerards, and Whittle answered this question in full generality through the introduction of templates.

Definition. A (binary) frame template is a tuple $\Phi = (C, X, Y_0, Y_1, A_1, \Delta, \Lambda)$ with the following elements:

1. Finite pairwise disjoint sets $C, X, Y_0, Y_1$;
2. A matrix $A_1$ with rows labeled by $X$ and columns labeled by $Y_0\cup Y_1 \cup C$;
3. A set of row vectors $\Delta$ closed under addition, with entries labeled by $Y_0 \cup Y_1 \cup C$;
4. A set of column vectors $\Lambda$ closed under addition, with entries labeled by $X$.

Definition. A matrix $A’$ is said to respect the template $\Phi$ if it is of this form:

Here, a unit column is a column with exactly one nonzero. The incidence matrix and the columns labeled by $Z$ can be of arbitrary size.

Definition. Let $A’$ be a matrix respecting the template $\Phi$. Let $A$ be obtained from $A’$ by

1. Taking each column from $Z$ and adding to it some column labeled by an element of $Y_1$;
2. Deleting the columns labeled by $Y_1$;
3. Contracting the elements labeled by $C$ in the corresponding matroid.

Then $A$ and $M[A]$ are said to conform to the template $\Phi$.

One should think about templates as recipes for constructing families of matroids.

Exercise. Describe the examples from sections 1.1 – 1.3 using templates.

The main result from [1] is that templates can be used to describe all sufficiently large, sufficiently highly connected matroids in any minor-closed class. Highly connected means the following:

Definition. A matroid $M$ is vertically $k$-connected if, for all separations $(X,Y)$ with $\lambda(X) < k$, either $r(X) = r(M)$ or $r(Y) = r(M)$. In other words, one side of the separation is spanning.

The precise result is:

Theorem (Geelen, Gerards, Whittle [1]). Let $\mathcal{M}$ be a proper minor-closed class of binary matroids. There exist constants $k, l$ and frame templates $\Phi_1, \ldots, \Phi_t, \Psi_1, \ldots, \Psi_s$ such that:

1. Every matroid conforming to $\Phi_i$ is in $\mathcal{M}$ for $i = 1, \ldots, t$;
2. Every matroid whose dual conforms to $\Psi_j$ is in $\mathcal{M}$ for $j = 1, \ldots, s$;
3. For every vertically $k$ connected matroid $M \in \mathcal{M}$ with at least $l$ elements, there either exists an $i$ such that $M$ conforms to $\Phi_i$ or a $j$ such that $M^*$ conforms to $\Psi_j$.

The third property says that the structure of the highly connected matroids in the class is completely described by a finite list of templates. The first two properties put some quite strict constraints on those templates: any matroid we can build using the template must be a member of the class!

The third property is reminiscent of the result by Robertson and Seymour that the “torsos” of a tree-decomposition are nearly-embeddable in some surface of bounded genus (see [4, Theorem 12.6.6]). But in the graph minors theorem there is no analog of the converse: most graphs nearly-embeddable on that surface won’t be members of the class being studied.

The first two properties can be used to determine the full set of templates $\Phi_1, \ldots, \Phi_t, \Psi_1, \ldots, \Psi_s$ for a given class of matroids. I have worked on several such results with my PhD student Kevin Grace [5, 6]. For instance, let’s consider Seymour’s 1-flowing conjecture, discussed by Dillon here. With Kevin I proved the following:

Theorem (Grace, vZ [5]). There exist constants $k, l$ such that every vertically $k$-connected 1-flowing matroid on at least $l$ elements is either graphic or cographic.

In other words, any counterexample to Seymour’s conjecture will have to be small, or have a low-order vertical separation. The proof proceeds by considering an arbitrary frame template, and showing that it either has to be trivial, or it can be used to build an excluded minor for the class of 1-flowing matroids. In the process we develop some tools to help with proofs by induction on templates, and to clean up the matrix $A_1$, but those will have to wait for another day.

Other applications of templates include growth rate results, and finding sufficient sets of excluded minors to characterize the highly connected members of a minor-closed class of matroids.

#### References

1. J. Geelen, B. Gerards and G. Whittle, The highly connected matroids in minor-closed classes, Ann. Comb. 19 (2015), 107–123.
2. P. D. Seymour, Decomposition of regular matroids, J. Combin. Theory Ser. B 28(3) (1980), 305-359.
3. I. Pivotto, Even Cycle and Even Cut Matroids. PhD Thesis, University of Waterloo (2011).
4. R. Diestel, Graph Theory, Springer GTM 173, 5th edition (2016).
5. K. Grace, S. H. M. van Zwam, Templates for Binary Matroids, SIAM J. Disc. Mathem. 31(1) (2017), 254 — 282.
6. K. Grace, S. H. M. van Zwam, The highly connected even-cycle and even-cut matroids, Submitted (2016).