Tree-Cut Decomposition in Graph Theory

• Tree-cut decomposition is an edge-based graph partitioning that organizes disjoint vertex bags into a tree structure, using adhesions to encode finite edge-cuts.
• It serves as the immersion-theoretic analogue of treewidth, linking local torso complexity with dual concepts like brambles, tangles, and lean decompositions.
• Recent algorithmic advances involve FPT approximations, dynamic programming on nice decompositions, and variants such as slim tree-cut width for tailored applications.

Searching arXiv for recent and foundational papers on tree-cut decomposition and related variants. Tree-cut decomposition is an edge-based graph decomposition in which a graph $G$ is represented by a tree $T$ whose nodes carry pairwise disjoint bags partitioning $V(G)$, while the edges of $T$ encode finite edge-separations of $G$. In Wollan’s formulation, and in later equivalent reformulations, the associated width parameter controls both the sizes of these adhesions and the local complexity of contracted “torsos,” making tree-cut width the immersion-theoretic analogue of treewidth for minors (Kim et al., 2015, Giannopoulou et al., 2018, Bożyk et al., 2021). Subsequent work developed constructive approximation algorithms, linked and lean normal forms, duality with brambles and tangles, variants such as slim tree-cut width, and applications to parameterized algorithms and infinite graph structure (Kim et al., 2015, Giannopoulou et al., 2018, Ganian et al., 2022, Ganian et al., 2022, Pitz et al., 18 Jun 2026).

1. Formal definition and width notions

A tree-cut decomposition of a graph $G$ is a pair $(T,\mathcal X)$ where $T$ is a tree and $\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$ is a near-partition of $V(G)$: the bags are pairwise disjoint, may be empty, and satisfy $T$0 (Kim et al., 2015, Bożyk et al., 2021). This disjoint-bag condition is the primary formal difference from tree decompositions, where bags overlap.

For a tree edge $T$1, deleting $T$2 splits $T$3 into two components. The corresponding adhesion is the set of edges of $T$4 with endpoints in the unions of bags on the two sides. In one notation,

$T$5

where $T$6 and $T$7 are the components of $T$8 containing $T$9 and $V(G)$0 (Kim et al., 2015). This is the edge cut represented by $V(G)$1.

The second ingredient is the torso at a node $V(G)$2. If $V(G)$3 are the components of $V(G)$4, and $V(G)$5, then the torso $V(G)$6 is obtained from $V(G)$7 by contracting each non-empty $V(G)$8 to a single vertex $V(G)$9 (Kim et al., 2015). Wollan’s width uses the 3-center of $T$0, obtained by repeatedly deleting degree-1 vertices outside $T$1 and dissolving degree-2 vertices outside $T$2. If $T$3 denotes this 3-center, then the width of $T$4 is

$T$5

and the tree-cut width $T$6 is the minimum such width over all tree-cut decompositions of $T$7 (Kim et al., 2015).

A later equivalent formulation, due to Giannopoulou, Pilipczuk, Raymond, Thilikos, and Wrochna, replaces torso 3-centers by a local quantity

$T$8

and defines width as the maximum of the edge adhesions and the values $T$9. For every graph, the optimized value of this GPRTW width equals Wollan’s original tree-cut width (Bożyk et al., 2021). In 3-edge-connected graphs, every adhesion is bold, so the width simplifies to

$G$0

(Giannopoulou et al., 2018).

2. Position within structural graph theory

Tree-cut width was introduced as an analogue of treewidth for the immersion order. Treewidth is vertex-centric: it controls overlapping bags and small vertex separators. Tree-cut width is edge-centric: it uses disjoint bags, measures adhesions by edge cuts, and controls local branching structure through torsos (Kim et al., 2015, Giannopoulou et al., 2018). This change aligns the parameter with edge-disjoint routing and immersions rather than minor structure.

The parameter is closely tied to immersed walls. Wollan’s theorem states that excluding a fixed wall as an immersion forces bounded tree-cut width, and the 2021 duality work summarizes this by saying that tree-cut width is functionally equivalent to the largest wall that can be found as an immersion (Bożyk et al., 2021). This is the immersion analogue of the grid-minor role played by treewidth.

Tree-cut width also sits between several classical width measures. Wollan showed $G$1, while Ganian, Kim, and Saurabh proved $G$2; the quadratic upper bound is asymptotically tight (Kim et al., 2015). Bounded carving-width is equivalent to bounded treewidth together with bounded maximum degree, whereas tree-cut width can remain bounded under unbounded degree and is therefore strictly more flexible for edge-connectivity phenomena (Kim et al., 2015).

This suggests a recurring structural interpretation: treewidth captures graphs that can be recursively separated by small vertex sets, while tree-cut width captures graphs that can be recursively separated by small edge cuts without losing the essential local branching encoded in torsos. A plausible implication is that tree-cut width is the more natural parameter when the target obstruction theory or algorithmic task is governed by edge-disjoint paths rather than vertex separators.

3. Linkedness, leanness, and duality

A central refinement is the leanness property, an edge-analogue of Thomas’s Menger-like condition for tree decompositions. A tree-cut decomposition $G$3 is lean if for every $G$4, every links $G$5, and every sets $G$6, $G$7 with $G$8, either there exist $G$9 edge-disjoint paths in $G$0 linking $G$1 to $G$2, or some link $G$3 on the path from $G$4 to $G$5 in $G$6 satisfies $G$7 (Giannopoulou et al., 2018). Every graph admits a lean tree-cut decomposition of width $G$8, so minimum width can be achieved simultaneously with this Menger-type visibility of small edge cuts (Giannopoulou et al., 2018).

A complementary perspective is duality. The 2021 work introduced ab-tree-cut width by separating bag-width and adhesion-width: $G$9 The resulting parameter is functionally equivalent to Wollan’s tree-cut width, but it supports a clean duality theorem: for integers $(T,\mathcal X)$0, a graph has no tree-cut decomposition of adhesion-width $(T,\mathcal X)$1 and bag-width $(T,\mathcal X)$2 if and only if it has an appropriately defined bramble of adhesion-order $(T,\mathcal X)$3 and bag-order $(T,\mathcal X)$4, and equivalently an $(T,\mathcal X)$5-tangle of edge separations (Bożyk et al., 2021). The same work also gives a cops–dogs–robber game characterization of this biparametric form (Bożyk et al., 2021).

For infinite graphs, linkedness takes a more explicit decomposition-theoretic form. A rooted tree-cut decomposition is linked if along every finite upward path $(T,\mathcal X)$6 in the decomposition tree there are

$(T,\mathcal X)$7

pairwise edge-disjoint paths in $(T,\mathcal X)$8 linking the outer adhesion sets (Pitz et al., 18 Jun 2026). The 2026 result proves that every graph admits a linked, componental, rooted tree-cut decomposition of finite adhesion that displays all undominated edge-ends, and moreover the edge-degree of each displayed end equals the $(T,\mathcal X)$9 of the boundary sizes along its representing ray in the decomposition tree (Pitz et al., 18 Jun 2026). In locally finite graphs, this yields a linked finite-adhesion tree-cut decomposition into finite parts that displays all ends and their edge-degrees (Pitz et al., 18 Jun 2026).

4. Algorithmic theory and nice decompositions

Computing tree-cut width exactly is NP-complete (Kim et al., 2015). A foundational constructive result is a parameterized 2-approximation: given $T$0 and $T$1, there is an algorithm that either reports that no tree-cut decomposition of width at most $T$2 exists, or returns one of width at most $T$3, in time

$T$4

(Kim et al., 2015). This was the first constructive FPT algorithm, even approximate, for tree-cut width (Kim et al., 2015).

For dynamic programming, the decisive structural tool is the notion of a nice tree-cut decomposition. In such a rooted decomposition, every thin node $T$5 satisfies

$T$6

where thin means adhesion at most $T$7 and $T$8 is the union of bags in the subtree rooted at $T$9 (Ganian et al., 2022). Any width-$\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$0 tree-cut decomposition can be transformed into a nice one of the same width, and if $\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$1, there exists a nice width-$\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$2 decomposition with at most $\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$3 nodes, computable in time $\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$4 from a given decomposition (Ganian et al., 2022). A key combinatorial consequence is that for every node $\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$5, the set $\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$6 of “complex” children satisfies

$\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$7

(Ganian et al., 2022). This bounded branching drives the paper’s dynamic-programming-and-ILP schemes.

Using that framework, the first fixed-parameter algorithmic applications of tree-cut width were obtained for Capacitated Vertex Cover, Capacitated Dominating Set, and Imbalance. Given a width-$\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$8 decomposition, the paper proves runtimes

$\mathcal X=\{X_t\subseteq V(G)\mid t\in V(T)\}$9

for these three problems, respectively (Ganian et al., 2022). By contrast, List Coloring, Precoloring Extension, and Boolean CSP remain W[1]-hard when parameterized by tree-cut width (for Boolean CSP, by tree-cut width of the incidence graph), even though these problems are FPT under $V(G)$0 (Ganian et al., 2022). This establishes that tree-cut width is not a uniformly stronger algorithmic surrogate for treewidth-plus-degree; rather, it is selective, favoring problems whose structure interacts directly with edge separations and capacities.

5. Variants and extensions

Several later variants modify the local torso notion while preserving the basic tree-cut paradigm. Slim tree-cut width changes the suppression threshold in torsos from the 3-center to the 2-center: outside the bag, only degree-$V(G)$1 vertices are deleted, while degree-2 vertices are retained (Ganian et al., 2022). If $V(G)$2 denotes slim tree-cut width and $V(G)$3 the even stricter 0-tree-cut width based on the 1-center, then for every graph

$V(G)$4

(Ganian et al., 2022). Slim tree-cut width is designed to retain the structural advantages of tree-cut width while recovering the algorithmic behavior expected from an edge-cut analogue of treewidth (Ganian et al., 2022).

This variant admits several strong characterizations. It is monotone under weak immersions, closed under bounded edge sums, and asymptotically equivalent to super edge-cut width, a spanning-tree-based parameter on supergraphs (Ganian et al., 2022). It also has a forbidden-immersion characterization in terms of walls and windmills: large slim tree-cut width forces either a large wall immersion or a large windmill immersion, and conversely excluding both bounds the parameter by a function (Ganian et al., 2022). Algorithmically, there is an FPT approximation algorithm that, given $V(G)$5 and $V(G)$6, either reports that no decomposition of slim width at most $V(G)$7 exists or returns one of slim width at most $V(G)$8 in time

$V(G)$9

(Ganian et al., 2022).

A different modification is screewidth, defined from a variation of tree-cut decompositions by combining link adhesions with the quantity $T$00 at nodes. Screewidth $T$01 always upper-bounds scramble number $T$02,

$T$03

but the two are not always equal (Cenek et al., 2022). This places tree-cut-type decompositions in the orbit of chip-firing, scramble number, and divisorial gonality rather than immersion theory alone (Cenek et al., 2022).

Beyond width parameters in the Wollan lineage, there are broader tree-like cut decompositions. Madry’s hierarchical $T$04-tree decomposition represents cuts of a graph by a collection of tree-like graphs consisting of a forest plus a small core. A 2026 dynamic version maintains a hierarchical $T$05-tree decomposition under edge updates, preserving cuts within a polylogarithmic factor and supporting amortized update time $T$06 for any fixed $T$07 (Goranci et al., 14 Jan 2026). This is not Wollan’s tree-cut decomposition, but it belongs to the same broader family of tree-shaped representations of edge cuts.

6. Terminological boundaries and distinct uses of the phrase

The phrase “tree-cut decomposition” is not used uniformly across the literature. In structural graph theory, the standard meaning is Wollan’s edge-based decomposition with disjoint bags, adhesions, and torso-based width (Kim et al., 2015, Giannopoulou et al., 2018). Several adjacent constructions are conceptually related but formally distinct.

One nearby object is the tree of decomposition of a $T$08-connected graph by a set of pairwise independent $T$09-vertex cutsets. In the biconnected case, this yields a bipartite tree whose vertices are parts and single $T$10-vertex cutsets, generalizing the block–cutpoint tree (Karpov, 2014). This decomposition is vertex-cutset-based rather than edge-cut-based, and its applications in that paper concern planarity, coloring, and critical biconnected graphs rather than tree-cut width (Karpov, 2014).

Another important distinction concerns work on balanced cuts in graphs equipped with a tree decomposition. The paper on minimum bisection in trees and tree-like graphs develops bounds in terms of tree decompositions and path-like structure, and although its viewpoint is “close in spirit” to edge-cut decompositions, it does not use tree-cut decomposition in Wollan’s sense (Fernandes et al., 2017). The same caution applies to canonical tree-decomposition theory for $T$11-blocks, which is based on nested vertex separations rather than edge-cut adhesions (Carmesin et al., 2011).

Outside graph structure theory, “tree-cut” can denote a literal cut in a hierarchical tree. In probabilistic image segmentation, a region tree induces valid segmentations by choosing one active node per root-to-leaf path; the resulting tree-cut model defines a probability distribution over tree-consistent segmentations and supports exact dynamic programming for MAP inference and sampling (Hu et al., 2015). This usage is algorithmically tree-based, but it is unrelated to tree-cut width and immersion theory. Likewise, the cut-tree of a randomly destroyed tree is a genealogical tree of connected components under random edge deletions, not a width-minimizing graph decomposition (Berzunza, 2015).

The term therefore has a stable technical meaning in immersion-based structural graph theory, but it also supports a broader family resemblance: tree-indexed representations of cut structure, balanced separation, or hierarchical partitioning. The literature since 2015 shows both the value of keeping these meanings distinct and the productivity of moving ideas between them.