In this post I want to discuss material I have been working on with Daryl Funk, Mike Newman, and Geoff Whittle. In particular, I’m going to discuss a parameter for matroids called decomposition-width. This terminology has been used by Dan Král [Kra12] and Yann Strozecksi [Str10, Str11]. We didn’t discover their work until after we had developed our own notion of decomposition-width, so our definition looks quite different from theirs, although it is equivalent. We have chosen to adopt their terminology.
Decomposition-width has a very natural motivation if you are familiar with matroids representable over finite fields, and matroid branch-width. Consider the following geometric illustration of the binary matroid $AG(3,2)$. The ground set has been partitioned into the sets $U$ and $V$. Let $X$ stand for the set of points coloured purple, and let $X’$ stand for the set of orange points. In the lefthand diagram, $V$ can distinguish between $X$ and $X’$. By this I mean that there is a subset $Z\subseteq V$ (we colour the points in $Z$ green) such that $X\cup Z$ is a circuit while $X’\cup Z$ is independent. However, in the righthand diagram, no subset of $V$ can distinguish $X$ and $X’$ in this way. Geometrically, this is because $X$ and $X’$ span exactly the same subset of the three-point line that lies in the spans of both $U$ and $V$ in the ambient binary space.
In general, let $M$ be a matroid on the ground set $E$, and let $(U,V)$ be a partition of $E$. We define the equivalence relation $\sim_{U}$ on subsets of $U$. We write $X\sim_{U} X’$ to mean that whenever $Z$ is a subset of $V$, both $X\cup Z$ and $X’\cup Z$ are independent, or neither is. This is clearly an equivalence relation.
Now we consider branch-width and decomposition-width. A decomposition of a matroid, $M=(E,\mathcal{I})$, consists of a pair $(T,\varphi)$, where $T$ is a binary tree (by this I mean that every vertex has degree one or three), and $\varphi$ is a bijection from $E$ to the set of leaves of $T$. If $e$ is an edge of $T$ joining vertices $u$ and $v$, then let $U_{e}$ be the subset containing elements $z\in E$ such that the path in $T$ from $\varphi(z)$ to $u$ does not contain $v$. Define $V_{e}$ symmetrically. We say that $U_{e}$ and $V_{e}$ are displayed by the decomposition. Define $\operatorname{bw}(M;T,\varphi)$ to be the maximum of $r(U_{e})+r(V_{e})-r(M)+1$, where the maximum is taken over all edges $e$ with end-vertices $u$ and $v$. Now I will define $\operatorname{dw}(M;T,\varphi)$ to be the maximum number of equivalence classes under the relation $\sim_{U_{e}}$, where we again take the maximum over all displayed sets $U_{e}$. The branch-width of $M$ is the minimum of $\operatorname{bw}(M;T,\varphi)$, where the minimum is taken over all decompositions $(T,\varphi)$. We define the decomposition-width of $M$ in the same way: as the minimum value taken by $\operatorname{dw}(M;T,\varphi)$. We write $\operatorname{bw}(M)$ and $\operatorname{dw}(M)$ for the branch- and decomposition-widths of $M$.
The notion of decomposition-width is clearly motivated by matroids over finite fields, but I won’t discuss those applications now. Instead we will continue to look at more abstract properties of decomposition-width. Král proved this next result for matroids representable over finite fields.
Proposition 1. Let $M$ be a matroid. Then $\operatorname{dw}(M)\geq \operatorname{bw}(M)$.
Proof. Let $E$ be the ground set of $M$, and let $U$ be a subset of $E$. Recall that $\lambda(U)$ is $r(U)+r(E-U)-r(M)$. We will start by proving that $\sim_{U}$ has at least $\lambda(U)+1$ equivalence classes. Define $V$ to be $E-U$. Let $B_{V}$ be a basis of $M|V$, and let $B$ be a basis of $M$ that contains $B_{V}$. Then $B\cap U$ is independent in $M|U$, and
\begin{align*}
r(U)-|B\cap U| &=r(U)-(|B|-|B_{V}|)\\
&=r(U)-(r(M)-r(V))\\
&=r(U)-(r(U)-\lambda(U))\\
&=\lambda(U).
\end{align*}
Therefore we let $(B\cap U)\cup\{x_{1},\ldots, x_{\lambda(U)}\}$ be a basis of $M|U$, where $x_{1},\ldots, x_{\lambda(U)}$ are distinct elements of $U-B$. Next we construct a sequence of distinct elements, $y_{1},\ldots, y_{\lambda(U)}$ from $B_{V}$ such that $(B-\{y_{1},\ldots, y_{i}\})\cup\{x_{1},\ldots, x_{i}\}$ is a basis of $M$ for each $i\in\{0,\ldots, \lambda(U)\}$. We do this recursively. Let $C$ be the unique circuit contained in\[(B-\{y_{1},\ldots, y_{i}\})\cup\{x_{1},\ldots, x_{i}\}\cup x_{i+1}\] and note that $x_{i+1}$ is in $C$. If $C$ contains no elements of $B_{V}$, then it is contained in $(B\cap U)\cup\{x_{1},\ldots, x_{\lambda(U)}\}$, which is impossible. So we simply let $y_{i+1}$ be an arbitrary element in $C\cap B_{V}$.
We complete the claim by showing that $(B\cap U)\cup\{x_{1},\ldots,x_{i}\}$ and $(B\cap U)\cup\{x_{1},\ldots, x_{j}\}$ are inequivalent under $\sim_{U}$ whenever $i< j$. Indeed, if $Z=B_{V}-\{y_{1},\ldots, y_{i}\}$, then $(B\cap U)\cup\{x_{1},\ldots, x_{i}\}\cup Z$ is a basis of $M$, and is properly contained in $(B\cap U)\cup\{x_{1},\ldots, x_{j}\}\cup Z$, so the last set is dependent, and we are done.
Now assume for a contradiction that $\operatorname{bw}(M)>\operatorname{dw}(M)$. Let $(T,\varphi)$ be a decomposition of $M$ such that if $U$ is any set displayed by an edge of $T$, then $\sim_{U}$ has at most $\operatorname{dw}(M)$ equivalence classes. There is some edge $e$ of $T$ displaying a set $U_{e}$ such that $\lambda(U_{e})+1>\operatorname{dw}(M)$, for otherwise this decomposition of $M$ certifies that
$\operatorname{bw}(M)\leq \operatorname{dw}(M)$. But $\sim_{U_{e}}$ has at least $\lambda_{M}(U_{e})+1$ equivalence classes by the previous claim. As $\lambda_{M}(U_{e})+1>\operatorname{dw}(M)$, this contradicts our choice of $(T,\varphi)$. $\square$
Král states the next result without proof.
Proposition 2. Let $x$ be an element of the matroid $M$. Then $\operatorname{dw}(M\backslash x) \leq \operatorname{dw}(M)$ and
$\operatorname{dw}(M/x) \leq \operatorname{dw}(M)$.
Proof. Let $(T,\varphi)$ be a decomposition of $M$ and assume that whenever $U$ is a displayed set, then $\sim_{U}$ has no more than $\operatorname{dw}(M)$ equivalence classes. Let $T’$ be the tree obtained from $T$ by deleting $\varphi(x)$ and contracting an edge so that every vertex in $T’$ has degree one or three. Let $U$ be any subset of $E(M)-x$ displayed by $T’$. Then either $U$ or $U\cup x$ is displayed by $T$. Let $M’$ be either $M\backslash x$ or $M/x$. We will show that in $M’$, the number of equivalence classes under $\sim_{U}$ is no greater than the number of classes under $\sim_{U}$ or $\sim_{U\cup x}$ in $M$. Let $X$ and $X’$ be representatives of distinct classes under $\sim_{U}$ in $M’$. We will be done if we can show that these representatives correspond to distinct classes in $M$. Without loss of generality, we can assume that $Z$ is a subset of $E(M)-(U\cup x)$ such that $X\cup Z$ is independent in $M’$, but $X’\cup Z$ is dependent. If $M’=M\backslash x$, then $X\cup Z$ is independent in $M$ and $X’\cup Z$ is dependent, and thus we are done. So we assume that $M’=M/x$. If $U$ is displayed by $T$, then we observe that $X\cup (Z\cup x)$ is independent in $M$, while $X’\cup (Z\cup x)$ is dependent. On the other hand, if $U\cup x$ is displayed, then $(X\cup x)\cup Z$ is independent in $M$ and $(X’\cup x)\cup Z$ is dependent. $\square$
When we combine Propositions 1 and 2, we see that the class of matroids with decomposition-width at most $k$ is a minor-closed subclass of the matroids with branch-width at most $k$. The class of matroids with branch-width at most $k$ has finitely many excluded minors [GGRW03]. In contrast to this, Mike and I convinced ourselves that there are classes of the form $\{M\colon \operatorname{dw}(M) \leq k\}$ with infinitely many excluded minors. I guess we’d had a couple of beers by that point, but I think our argument holds up. I’ll eventually add that argument to this post. If anyone presents a proof in the comments before I do then I will buy them a drink at the next SIGMA meeting.
References
[GGRW03] J. F. Geelen, A. M. H. Gerards, N. Robertson, and G. P. Whittle. On the excluded minors for the matroids of branch-width $k$. J. Combin. Theory Ser. B 88 (2003), no. 2, 261–265.
[Kra12] D. Král. Decomposition width of matroids. Discrete Appl. Math. 160 (2012), no. 6, 913–923.
[Str10] Y. Strozecki. Enumeration complexity and matroid decomposition. Ph.D. thesis, Université Paris Diderot (2010).
[Str11] Y. Strozecki. Monadic second-order model-checking on decomposable matroids. Discrete Appl. Math. 159 (2011), no. 10, 1022–1039.