Linear Clique-Width: Structure and Obstructions

• Linear clique-width is a sequential graph parameter defined by constructing graphs one vertex at a time using a fixed set of label-based operations.
• It leverages modular decomposition and universal obstructions, such as quasi-threshold graphs, to differentiate bounded from unbounded classes.
• The framework connects with CMSO transductions, dynamic programming, and well-quasi-order theories to enable practical algorithm design.

Linear clique-width is a graph width parameter measuring the complexity of building a graph in a strictly sequential fashion using a finite palette of labels and a fixed repertoire of label-based operations. It is a linearized version of clique-width: the underlying construction remains label-based, but the parse tree is constrained to be path-like, so vertices are introduced one-by-one rather than by combining two previously built subgraphs. Across hereditary graph classes, the parameter sits at the intersection of modular decomposition, obstruction theory, logical transductions, and algorithmic metatheorems. Recent work gives a sharp lifting theorem through modular decomposition, identifies quasi-threshold graphs and their complements as the canonical hereditary obstructions to lifting boundedness from prime graphs, and places unbounded linear clique-width at a precise threshold in the CMSO transduction hierarchy (Brignall et al., 25 Feb 2026, Bojanczyk et al., 29 Jan 2025).

1. Definitions, formalisms, and basic inequalities

Fix $k \in \mathbb{N}$ and a label set $[k] = \{1,2,\dots,k\}$. A $k$-expression builds a labeled graph by repeatedly applying single-vertex creation, disjoint union $G \oplus H$, edge insertion $\eta_{i,j}$ for $i \neq j$, and relabeling $\rho_{i\to j}$ for $i \neq j$. A graph $G$ has clique-width at most $k$ if it can be constructed by a $[k] = \{1,2,\dots,k\}$0-expression, and

$[k] = \{1,2,\dots,k\}$1

Linear clique-width restricts this construction to be sequential: vertices are introduced one-by-one, and there is no binary $[k] = \{1,2,\dots,k\}$2 combining two previously built graphs. Equivalently, one uses a linear $[k] = \{1,2,\dots,k\}$3-expression consisting of steps that introduce a new vertex, apply $[k] = \{1,2,\dots,k\}$4, or apply $[k] = \{1,2,\dots,k\}$5. The linear clique-width is

$[k] = \{1,2,\dots,k\}$6

These equivalent formalisms appear both in the standard expression language and in algebraic presentations of clique-width (Brignall et al., 25 Feb 2026, Bojanczyk et al., 29 Jan 2025).

The sequential restriction is strict. By design, $[k] = \{1,2,\dots,k\}$7, and for many classes the inequality is strict. A standard example is the class of cographs: every cograph has $[k] = \{1,2,\dots,k\}$8, but cographs have unbounded linear clique-width (Brignall et al., 25 Feb 2026). A complementary estimate also plays a recurrent role:

$[k] = \{1,2,\dots,k\}$9

which makes complement-closed obstruction statements particularly natural (Brignall et al., 2013).

Several technical notions are specific to linear expressions. A label is a sink if it is assigned to vertices but never used in subsequent edge insertions or relabelings; sink labels are useful in inflation arguments and in sharp upper bounds for complete or anti-complete modular skeletons (Brignall et al., 25 Feb 2026). In algorithmic work on linear expressions, one also tracks live labels: a label is live at time $k$0 if its current vertices can still participate in a future join. This notion yields state-space reductions in dynamic programming over linear expressions (Dell et al., 25 Jun 2026).

2. Modular decomposition and the sequential inflation viewpoint

A module in a graph $k$1 is a set $k$2 such that every vertex outside $k$3 is adjacent either to all of $k$4 or to none of $k$5. The trivial modules are $k$6, singletons, and $k$7. A graph on at least two vertices is prime if all of its modules are trivial. Modular decomposition asserts that every graph is uniquely an inflation of a prime skeleton: if $k$8 is a graph and $k$9 is a family of nonempty graphs, then

$G \oplus H$0

is obtained by replacing each vertex $G \oplus H$1 of $G \oplus H$2 by $G \oplus H$3 and making the bipartite adjacency between $G \oplus H$4 and $G \oplus H$5 complete if and only if $G \oplus H$6 is adjacent to $G \oplus H$7 in $G \oplus H$8 (Brignall et al., 25 Feb 2026).

For linear clique-width, this decomposition is structurally decisive. The one-step modular decomposition theorem used in the 2026 lifting result states that every graph $G \oplus H$9 on at least two vertices admits a unique prime graph $\eta_{i,j}$0 on at least two vertices and nonempty graphs $\eta_{i,j}$1 such that $\eta_{i,j}$2; when $\eta_{i,j}$3 has at least four vertices, the graphs $\eta_{i,j}$4 are also unique. When $\eta_{i,j}$5 or its complement is disconnected, the decomposition is refined so that $\eta_{i,j}$6 is taken to be as large an anti-complete or complete skeleton as possible (Brignall et al., 25 Feb 2026).

The inflation formalism yields direct linear clique-width bounds. If

$\eta_{i,j}$7

then

$\eta_{i,j}$8

The construction simulates a linear expression for the skeleton and replaces the insertion of each skeleton vertex by a linear construction of the corresponding module, using a reserved block of labels and then relabeling the module to the skeleton label. When $\eta_{i,j}$9 is complete or anti-complete, the bound improves to

$i \neq j$0

using a sink label to collapse finished modules (Brignall et al., 25 Feb 2026).

Two further lemmas are central in sequential settings. First, for any vertex $i \neq j$1, there is a linear expression for $i \neq j$2 with at most $i \neq j$3 labels that begins by inserting $i \neq j$4. Second, if $i \neq j$5 with $i \neq j$6, then either some module already has full complexity, $i \neq j$7, or there are distinct modules $i \neq j$8 with

$i \neq j$9

These statements enable the induction driving the lifting theorem (Brignall et al., 25 Feb 2026).

3. The lifting theorem and the quasi-threshold obstructions

For clique-width, boundedness lifts cleanly through modular decomposition: a hereditary class has bounded clique-width if and only if its prime members do. For linear clique-width, this lifting property fails in general. The simplest counterexample is the class of cographs: its only prime members are $\rho_{i\to j}$0 and $\rho_{i\to j}$1, each of linear clique-width at most $\rho_{i\to j}$2, yet the class itself has unbounded linear clique-width (Brignall et al., 25 Feb 2026).

The sharp repair is the main theorem of "Linear clique-width and modular decomposition" (Brignall et al., 25 Feb 2026). Let $\rho_{i\to j}$3 be hereditary. Then

$\rho_{i\to j}$4

holds if and only if both of the following conditions hold:

1. the prime members of $\rho_{i\to j}$5 have bounded linear clique-width; and
2. $\rho_{i\to j}$6 contains neither all quasi-threshold graphs nor all complements of quasi-threshold graphs.

A quasi-threshold graph, also called a trivially perfect graph, is defined inductively from $\rho_{i\to j}$7 by disjoint union and join with $\rho_{i\to j}$8. Equivalently, quasi-threshold graphs are exactly the $\rho_{i\to j}$9-free graphs. Their complements form the co-quasi-threshold graphs (Brignall et al., 25 Feb 2026).

The proof uses explicit universal families. Define $i \neq j$0, and for $i \neq j$1,

$i \neq j$2

Dually, let $i \neq j$3, and for $i \neq j$4,

$i \neq j$5

Every quasi-threshold graph embeds as an induced subgraph of some $i \neq j$6, and every co-quasi-threshold graph embeds into some $i \neq j$7. Since quasi-threshold graphs and their complements have unbounded linear clique-width, these universal graphs provide canonical obstructions (Brignall et al., 25 Feb 2026).

The resulting quantitative statement is explicit. Suppose all prime induced subgraphs of $i \neq j$8 have linear clique-width at most $i \neq j$9. Proposition 4.1 of the paper states:

If $G$0, then $G$1 contains $G$2 or $G$3 as an induced subgraph.

Contrapositively, if a hereditary class excludes $G$4 and $G$5 and its prime members have $G$6, then every graph in the class satisfies

$G$7

Thus one obtains the explicit bound

$G$8

This generalizes the earlier theorem of Brignall, Korpelainen, and Vatter for hereditary classes of cographs, which showed that bounded linear clique-width is equivalent to excluding all quasi-threshold graphs and all complements of quasi-threshold graphs inside the cographs (Brignall et al., 2013, Brignall et al., 25 Feb 2026).

A common misconception is that prime control alone should suffice because it does for clique-width. The modern theorem shows precisely where that intuition fails: quasi-threshold and co-quasi-threshold self-embedding patterns survive modular composition in a way that the strictly sequential formalism cannot absorb (Brignall et al., 25 Feb 2026).

4. Obstruction families, minimal unbounded classes, and finite minimal obstructions

The quasi-threshold families are the canonical obstructions for the modular lifting theorem, but they are not the whole obstruction landscape for linear clique-width. Independent lines of work show that hereditary classes of unbounded linear clique-width admit many minimal forms.

One source is the family $G$9 built from an infinite graph $k$0 determined by an infinite word $k$1. If $k$2 is periodic and contains at least one $k$3, then $k$4 is a minimal hereditary class of graphs of unbounded clique-width and linear clique-width. In particular, the sequence $k$5 yields infinitely many pairwise incomparable minimal hereditary classes of unbounded clique-width and linear clique-width (Collins et al., 2017). This places linear clique-width in sharp contrast with the tree-width situation, where planar graphs form the unique minimal minor-closed class of unbounded tree-width (Collins et al., 2017).

A broader grid framework is developed in "A framework for minimal hereditary classes of graphs of unbounded clique-width" (Brignall et al., 2022). Here a hereditary class $k$6 is generated from a triple $k$7 controlling consecutive-column edges, bonds between non-consecutive columns, and within-column edges. The parameter $k$8, defined from distinct neighborhoods in a two-row auxiliary graph, characterizes unbounded clique-width:

$k$9

For a large family $[k] = \{1,2,\dots,k\}$00 of recurrent triples with bounded bond-complexity parameter $[k] = \{1,2,\dots,k\}$01, the corresponding classes are minimal hereditary classes of both unbounded clique-width and unbounded linear clique-width (Brignall et al., 2022).

The same framework captures several previously known minimal classes, including bipartite permutation graphs, unit interval graphs, bichain graphs, split permutation graphs, and infinite families defined by periodic or recurrent words (Brignall et al., 2022). The proofs are constructive: proper hereditary subclasses admit explicit bounded-label linear expressions, with an overall bound of

$[k] = \{1,2,\dots,k\}$02

in the panel construction for subclasses obtained by forbidding a finite $[k] = \{1,2,\dots,k\}$03 block (Brignall et al., 2022).

At the level of finite obstructions for fixed bounds, "Minimal forbidden induced subgraphs of graphs of bounded clique-width and bounded linear clique-width" identifies explicit families minimal for classes of the form $[k] = \{1,2,\dots,k\}$04 or $[k] = \{1,2,\dots,k\}$05 (Meister et al., 2013). Among them are:

• the path-power family $[k] = \{1,2,\dots,k\}$06, minimal for both bounded clique-width and bounded linear clique-width at level $[k] = \{1,2,\dots,k\}$07;
• the family $[k] = \{1,2,\dots,k\}$08, minimal for bounded linear clique-width, and witnessing strict separation because $[k] = \{1,2,\dots,k\}$09 for $[k] = \{1,2,\dots,k\}$10;
• the family $[k] = \{1,2,\dots,k\}$11, minimal obstructions for $[k] = \{1,2,\dots,k\}$12;
• the graphs $[k] = \{1,2,\dots,k\}$13, minimal obstructions for $[k] = \{1,2,\dots,k\}$14 (Meister et al., 2013).

These results clarify an important distinction. Quasi-threshold and co-quasi-threshold graphs are the universal hereditary obstructions to lifting boundedness from prime graphs, whereas minimal unbounded hereditary classes and minimal forbidden induced subgraphs describe a wider obstruction theory for the parameter itself (Brignall et al., 25 Feb 2026, Collins et al., 2017, Meister et al., 2013).

5. Logical transductions, well-quasi-order, and algorithmic consequences

A major structural development is the dense analogue of the Pathwidth Theorem. If a class of graphs has unbounded linear clique-width, then it can produce all trees via some fixed CMSO transduction (Bojanczyk et al., 29 Jan 2025). More precisely, for a class $[k] = \{1,2,\dots,k\}$15 of bounded clique-width, either $[k] = \{1,2,\dots,k\}$16 has bounded linear clique-width, or there is a surjective MSO transduction from $[k] = \{1,2,\dots,k\}$17 onto the class of trees. Combined with the case of unbounded clique-width, this yields the corollary stated in the abstract: if a class has unbounded linear clique-width, then there exists a fixed CMSO transduction producing all trees from the class (Bojanczyk et al., 29 Jan 2025).

This result positions linear clique-width in the adjacency/CMSO transduction hierarchy:

$[k] = \{1,2,\dots,k\}$18

Up to and including Trees, the order does not change if one uses MSO rather than CMSO (Bojanczyk et al., 29 Jan 2025). In this hierarchy, bounded linear clique-width corresponds to the path-like side, while escaping bounded linearity inside bounded clique-width is already sufficient to CMSO-transduce all trees.

A complementary perspective comes from definability over words. A class of graphs has bounded linear clique-width if and only if it is contained in the image of some MSO-interpretation of finite words (Lopez, 2024). This connection underlies recent results on induced-subgraph well-quasi-order. For a class $[k] = \{1,2,\dots,k\}$19 given as the image of an MSO-interpretation of words, there exists a computable $[k] = \{1,2,\dots,k\}$20 such that the following are equivalent: $[k] = \{1,2,\dots,k\}$21 is $[k] = \{1,2,\dots,k\}$22-well-quasi-ordered by induced subgraphs, $[k] = \{1,2,\dots,k\}$23 is $[k] = \{1,2,\dots,k\}$24-well-quasi-ordered, and $[k] = \{1,2,\dots,k\}$25 is labelled-well-quasi-ordered. These properties are decidable (Lopez, 2024). As a corollary, Pouzet’s second conjecture holds for bounded linear clique-width classes: $[k] = \{1,2,\dots,k\}$26-wqo and labelled-wqo coincide in this regime (Lopez, 2024).

Algorithmically, the sequential nature of linear clique-width permits refined dynamic programs when a linear expression is given. For equitable coloring, there exists an algorithm that, given an integer $[k] = \{1,2,\dots,k\}$27 and an $[k] = \{1,2,\dots,k\}$28-vertex graph $[k] = \{1,2,\dots,k\}$29 together with a linear $[k] = \{1,2,\dots,k\}$30-expression constructing $[k] = \{1,2,\dots,k\}$31, computes the number of equitable $[k] = \{1,2,\dots,k\}$32-colorings of $[k] = \{1,2,\dots,k\}$33 in time

$[k] = \{1,2,\dots,k\}$34

The improvement over the general clique-width algorithm comes from tracking color-sets only for live labels and excluding both $[k] = \{1,2,\dots,k\}$35 and $[k] = \{1,2,\dots,k\}$36 from the state alphabet on live labels (Dell et al., 25 Jun 2026). More generally, the constructive inflation bounds from modular decomposition suggest practical synthesis of linear expressions along modular decomposition trees whenever bounds on prime graphs and excluded universal obstructions are available (Brignall et al., 25 Feb 2026).

6. Related parameters, methodological contrasts, and open directions

Linear clique-width is closely related to, but distinct from, several neighboring width notions. Every graph satisfies $[k] = \{1,2,\dots,k\}$37 (Brignall et al., 25 Feb 2026). The transduction-theoretic analysis of unbounded linear clique-width also uses a notion of rank of a set that is functionally related to the parity-matrix rank used in rank-width over $[k] = \{1,2,\dots,k\}$38 (Bojanczyk et al., 29 Jan 2025). In the study of minimal hereditary classes, rank-width enters through the inequalities

$[k] = \{1,2,\dots,k\}$39

together with monotonicity of rank-width under vertex-minors (Collins et al., 2017). These connections explain why vertex-minor methods and local complementations appear naturally in lower-bound constructions.

Methodologically, the 2026 modular-decomposition theorem differs sharply from the older cograph proof of Brignall, Korpelainen, and Vatter. The earlier proof followed a blueprint from permutation classes relying on well-quasi-order arguments; the new proof avoids well-quasi-order altogether, using only modular decomposition and the explicit universal graphs $[k] = \{1,2,\dots,k\}$40 and $[k] = \{1,2,\dots,k\}$41 (Brignall et al., 2013, Brignall et al., 25 Feb 2026). This suggests a broader role for explicit universal obstruction families in dense width theory.

Several open directions remain. The bound $[k] = \{1,2,\dots,k\}$42 in the main lifting proposition is linearly tight in $[k] = \{1,2,\dots,k\}$43, since $[k] = \{1,2,\dots,k\}$44 grows linearly in $[k] = \{1,2,\dots,k\}$45, but it is likely not optimal; refining constants and sharpening the dependence on $[k] = \{1,2,\dots,k\}$46 is open (Brignall et al., 25 Feb 2026). Further classification of minimal hereditary classes of unbounded linear clique-width, beyond the quasi-threshold universals and the currently known grid- or word-based families, remains active (Collins et al., 2017, Brignall et al., 2022). On the logical side, a conjectural vertex-minor analogue asks whether unbounded linear rank-width classes admit all trees as vertex-minors, while another direction asks whether classes of unbounded linear clique-width FO-transduce classes containing subdivisions of all trees (Bojanczyk et al., 29 Jan 2025). On the well-quasi-order side, extending the effective MSO-word techniques beyond bounded linear clique-width toward bounded clique-width more generally would require tree-based analogues of the factorization-forest machinery currently available for words (Lopez, 2024).

Taken together, these results place linear clique-width at a precise structural threshold. It is the parameter governing when dense graph constructions remain genuinely path-like, when modular decomposition can be lifted from primes to whole hereditary classes, and when bounded clique-width classes cross from linear discipline into the logical strength required to produce all trees (Brignall et al., 25 Feb 2026, Bojanczyk et al., 29 Jan 2025).