--- title: 'Screewidth: Tree-Cut Decomposition in Graph Theory' url: https://www.emergentmind.com/topics/screewidth type: topic --- # Screewidth: Tree-Cut Decomposition in Graph Theory Screewidth is a graph invariant defined from a variation of tree-cut decompositions and introduced to provide a decomposition-theoretic upper bound on scramble number, an invariant arising in chip-firing games and divisorial gonality. In the foundational formulation, screewidth measures how efficiently the vertices of a graph can be organized into a tree-shaped routing structure while controlling both edge traffic across decomposition links and combined bag-size-plus-throughput at decomposition nodes. Its basic role is summarized by the inequality \[ \sn(G)\le \scw(G), \] which places screewidth above scramble number in the chain of width-type and gonality-related invariants [2209.01459]. ## 1. Definition through tree-cut decompositions A tree-cut decomposition of a connected finite multigraph \(G\) is a pair \[ T=(T,\mathcal X), \] where \(T\) is a tree and each node \(b\in V(T)\) is assigned a bag \(X_b\subseteq V(G)\), with \[ \bigcup_{b\in V(T)} X_b = V(G), \qquad X_b\cap X_d=\varnothing \text{ for } b\neq d. \] Thus the bags form a partition of \(V(G)\); unlike ordinary tree decompositions, bags are disjoint, may be empty, and need not induce connected subgraphs. To avoid confusion between the two graphs, vertices and edges of the decomposition tree are called nodes and links [2209.01459]. For a link \(l\in E(T)\), deleting \(l\) disconnects \(T\) into two components. The link adhesion \(\adh(l)\) is the set of edges of \(G\) whose endpoints lie in bags on opposite sides of that cut. For a non-leaf node \(b\in V(T)\), deleting \(b\) disconnects \(T\); the node adhesion \(\adh(b)\) is the set of edges of \(G\) whose endpoints lie in bags belonging to different components of \(T-b\). These notions support two width measures: \[ \lw(T)=\max_{l \in E(T)}\{|\adh(l)|\}, \] \[ \bw(T)=\max_{b \in V(T)}\{|X_b| + |\adh(b)|\}. \] The width of the decomposition is \[ w(T)=\max\{\lw(T), \bw(T)\}, \] and the screewidth of \(G\) is the minimum over all tree-cut decompositions: \[ \scw(G)=\min_T w(T). \] The two terms play different roles: \(\lw(T)\) controls maximum edge traffic through a decomposition link, while \(\bw(T)\) controls the largest sum of bag size and tunneling-edge congestion at a node [2209.01459]. The geometric interpretation is central. One may draw a thickened copy of the decomposition tree, place each graph vertex inside its bag, and route each graph edge along the unique path in the tree joining the bags containing its endpoints. Then link adhesion counts edges passing through a link, and node adhesion counts edges tunneling through a non-leaf node. In this sense, screewidth is a routing-congestion width parameter on a tree [2311.17342]. ## 2. Position among scramble number, treewidth, and gonality Screewidth was introduced to fill a structural gap in scramble theory. Scrambles provide lower bounds on gonality, but prior to screewidth there was no known tree-decomposition-style invariant designed to give upper bounds on scramble number. The defining theorem is \[ \sn(G)\le \scw(G), \] proved by orienting every link of a hypothetical low-width tree-cut decomposition toward a side containing an egg of a scramble, then using a sink node to construct a hitting set of size at most the decomposition width, contradicting the scramble order [2209.01459]. This inequality sits inside a broader chain of invariants. The scramble-number literature recalled in later work gives \[ \tw(G)\le \sn(G)\le \gon(G), \] and therefore \[ \tw(G)\le \sn(G)\le \scw(G). \] The same later paper also recalls the Seymour–Thomas identity \[ \tw(G)=\bn(G)-1 \] for bramble number \(\bn(G)\), and explains that scrambles generalize brambles in the sense that a bramble of order \(k\) becomes a scramble of order either \(k\) or \(k-1\) [2311.17342]. The comparison with gonality is subtler. The original screewidth paper asks whether \[ \scw(G)\le \gon(G) \] holds for all graphs, but leaves the question open. It proves a conditional positive result: if \(G\) has a positive-rank divisor \(D\) that partitions \(V(G)\), then \[ \scw(G)\le \deg(D), \] and in particular, if \(\deg(D)=\gon(G)\), then \(\scw(G)\le \gon(G)\). From this, one obtains that if \(G\) is a \(k\)-edge-connected graph of gonality \(k\), then \[ \sn(G)=\scw(G)=\gon(G)=k \] [2209.01459]. A later paper states that there are no known cases where screewidth is larger than gonality, but it does not prove a general inequality of that form. It also records several families in which screewidth, scramble number, disjoint scramble number, carton number, and gonality coincide, including cycles \(C_n\), complete multipartite graphs \(K_{n_1,\dots,n_k}\), grid graphs \(G_{m,n}\), cylindrical graphs \(Y_{m,n}\), and several Cartesian-product families [2311.17342]. ## 3. Structural properties and extremal behavior The original paper develops a substantial basic theory of screewidth. At the smallest end, \[ \scw(G)=1 \iff G \text{ is a tree}. \] At the dense end, for a simple graph on \(n\) vertices, \[ \scw(G)=n-1 \iff G=K_n. \] More generally, for a graph on \(n\) vertices, \[ \scw(G)\le n, \] with equality iff \(\lambda(G)\ge n\); for simple graphs, this simplifies to \[ \scw(G)\le n-1. \] A sharper simple-graph bound is \[ \scw(G)\le n-\alpha(G), \] where \(\alpha(G)\) is the independence number. In the dense regime \(\delta(G)\ge \lfloor n/2\rfloor+1\), the invariant collapses to \[ \scw(G)=\sn(G)=n-\alpha(G) \] [2209.01459]. Screewidth is monotone under subgraphs: if \(H\) is a subgraph of \(G\), then \[ \scw(H)\le \scw(G). \] It is invariant under subdivision and smoothing of a 2-valent vertex, so it depends on the underlying topological graph shape rather than degree-2 refinements. It also satisfies a bridge formula: if \(e\) is a bridge and \(G-e\) has components \(G_1,G_2\), then \[ \scw(G)=\max\{\scw(G_1),\scw(G_2)\} \] [2209.01459]. By contrast, screewidth is not minor-monotone. The paper gives an explicit example showing that a minor can have larger screewidth than the original graph. This rules out a straightforward Robertson–Seymour-style minor theory for the invariant [2209.01459]. Optimal decompositions can also behave differently from standard treewidth intuition. Empty bags may be necessary: for \(K_3\circ K_n^l\) with \(l>n\), \[ \scw(K_3\circ K_n^l)=n, \] and every width-\(n\) decomposition has at least one empty bag. Still, empty bags can be normalized: any graph has an optimal tree-cut decomposition in which every empty bag has valence \(3\), and if an optimal decomposition has \(m\) leaf nodes, it has at most \(m-2\) empty bags. Nonempty bags need not induce connected subgraphs either; the paper exhibits a graph of screewidth \(2\) for which every optimal decomposition contains a disconnected bag [2209.01459]. The invariant is not simply a repackaging of scramble number. A concrete separation is \[ G=K_3\circ B_{2,3}, \qquad \sn(G)=2,\quad \scw(G)=3. \] Moreover, for every \(n\ge 4\) there exists a graph \(G\) with \[ \sn(G)=n \quad\text{and}\quad \scw(G)=\left\lceil\frac14 n(n+2)\right\rceil-1, \] showing that screewidth can be quadratically larger than scramble number [2209.01459]. ## 4. Vertex congestion, line graphs, and upper bounds A later paper places screewidth in a congestion-theoretic framework. It defines an embedding of \(G\) as an injection \(\pi\) from \(V(G)\) to the leaves of a sub-cubic tree \(T\). For each edge \(vw\in E(G)\), let \(P_{vw}\) be the path in \(T\) from \(\pi(v)\) to \(\pi(w)\). The vertex congestion of \(\pi\) is \[ \max_{u\in V(T)} \left|\{vw\in E(G)\mid u\in P_{vw}\}\right|, \] and \(\vcon(G)\) is the minimum of this quantity over all such embeddings [2311.17342]. The key theorem is \[ \scw(G)\le \vcon(G) \qquad (|V(G)|\ge 3). \] The proof starts from a minimum-congestion embedding into a subcubic tree, places each graph vertex into the leaf where it is embedded, and interprets the resulting structure as a tree-cut decomposition. Because graph vertices lie only in leaves, non-leaf bags are empty, their contribution is purely tunneling-edge congestion, and link adhesion is bounded by the contribution of an adjacent non-leaf node. Hence the decomposition width is at most the vertex congestion [2311.17342]. This connects screewidth to line-graph treewidth via the recalled identity \[ \vcon(G)=\tw(L(G))+1. \] Combining this with \(\sn(G)\le \scw(G)\) yields \[ \sn(G)\le \scw(G)\le \vcon(G)=\tw(L(G))+1. \] The same paper states the resulting scramble-number bound as \[ \sn(G)\le (\tw(G)+1)\Delta(G)-1, \] where \(\Delta(G)\) is maximum degree. It also uses the chain \[ \tw(G)\le \sn(G)\le \scw(G)\le \vcon(G)=\tw(L(G))+1 \] to give a new proof of the best known bound \[ \tw(L(G))\ge \tw(G)-1 \] [2311.17342]. The congestion viewpoint also yields asymptotic consequences. Using excluded-minor treewidth bounds, the paper concludes that if \(G\) belongs to a family with no \(K_r\) minor and bounded maximum degree, then \[ \sn(G)=O(\sqrt n). \] In particular, planar graphs of bounded degree have scramble number \(O(\sqrt n)\). Since screewidth bounds scramble number from above, this does not directly determine \(\scw(G)\), but it shows how decomposition and congestion machinery can control scramble-related invariants on minor-closed families [2311.17342]. ## 5. Exact formulas for banana trees For banana trees, screewidth admits a complete combinatorial formula. A banana tree is a multigraph whose underlying simple graph is a tree; the parallel edges joining adjacent vertices form an edge bunch. The 2025 paper proves: > For a banana tree \(G\), let \(k\) be the maximum positive integer such that there is a connected subgraph with at least \(k\) vertices with at least \(k\) edges between each adjacent pair. Then, \[ \sn(G)=\scw(G)=k. \] Equivalently, \[ \scw(G)=\max_{H\subseteq G\text{ connected}} \min\Bigl\{|V(H)|,\ \min_{uv\in E(H)} |E(u,v)|\Bigr\}. \] Thus screewidth is determined by the largest connected region whose size and minimum edge-bunch multiplicity are simultaneously at least \(k\) [2510.03499]. The lower bound is obtained by taking singleton eggs on such a thick connected subgraph: if the subgraph has \(k\) vertices and every adjacent pair is joined by at least \(k\) edges, then the scramble has hitting number \(k\) and egg-cut number at least \(k\). The upper bound comes from an explicit tree-cut decomposition: vertices joined by more than \(k\) parallel edges are merged into a bag, while crossing an edge bunch of size at most \(k\) creates a new bag. By construction, maximum link adhesion is at most \(k\), there is no bag adhesion, and bag size is at most \(k\) [2510.03499]. For banana paths \(B_{(a_1,\dots,a_n)}\), this becomes \[ \scw(B_{(a_1,\dots,a_n)}) = \max_{0\le i