---
title: Focused Rank-Width Analysis
url: https://www.emergentmind.com/topics/focused-rank-width
type: topic
---

# Focused Rank-Width Analysis

Searching arXiv for the supplied topic and supporting papers.
{"query":"rank-width stability graph isomorphism bounded rank-width arXiv", "max_results": 10}
“Focused rank-width” is not a standard named graph parameter in the cited literature. The phrase is best understood as an interpretive umbrella for work that takes rank-width as the central structural invariant and, more narrowly, for results that isolate the bounded-rank-width classes that remain tame under additional constraints such as stability. In that narrower sense, the sharpest formulation comes from the equivalence, for graph classes of bounded rank-width, between stable edge relation, stability, monadic stability, and being a first-order transduction of a class with bounded treewidth [2007.07857]. More generally, rank-width measures whether a graph can be recursively decomposed by cuts of low binary rank, and bounded rank-width is equivalent, at the level of graph classes, to bounded clique-width [1505.03737].

## 1. Formal framework

For a finite simple graph \(G=(V,E)\) and \(X\subseteq V\), the cut-rank function is
\[
\rho_G(X)=\operatorname{rk}_{\mathbb{F}_2}(M_{X,\bar X}),
\]
where \(M_{X,\bar X}\) is the \(0\)-\(1\) bipartite adjacency matrix between \(X\) and \(\bar X=V\setminus X\). This function is symmetric, submodular, and satisfies \(\rho_G(\emptyset)=0\). Rank-width is then defined as the branch-width of \(\rho_G\): one takes a cubic tree whose leaves bijectively represent the vertices, each tree edge induces a bipartition of \(V\), and the width of the decomposition is the maximum cut-rank over all induced cuts; \(\mathrm{rw}(G)\) is the minimum such width [1505.03737].

Linear rank-width replaces the decomposition tree by a path-like layout: for a linear ordering \(v_1,\dots,v_n\), its width is \(\max_{1\le i<n}\rho_G(\{v_1,\dots,v_i\})\), and \(\mathrm{lrw}(G)\) is the minimum over all such layouts. In the broader width hierarchy, \(\mathrm{rw}(G)\le \mathrm{lrw}(G)\), bounded tree-width implies bounded rank-width, and bounded rank-width is equivalent to bounded clique-width. The converse to bounded tree-width fails dramatically: the complete graph \(K_n\) has rank-width \(1\) but tree-width \(n-1\). Rank-width is also almost invariant under complementation: the rank-width of \(G\) and \(\overline{G}\) differs by at most \(1\) [1601.03800].

## 2. Stability as a focusing condition

The most explicit “focused” interpretation occurs in the model-theoretic analysis of bounded rank-width. A graph class has a stable edge relation exactly when it excludes some half-graph as a semi-induced subgraph. Within classes of bounded rank-width, this condition is equivalent to stability, equivalent to monadic stability, and equivalent to being a first-order transduction of a class with bounded treewidth. The same paper presents this subclass as “structurally bounded treewidth,” namely bounded rank-width intersected with monadic stability [2007.07857].

This equivalence gives a precise sense in which stability focuses rank-width. Bounded rank-width alone already implies monadic dependence, but not monadic stability: half-graphs have bounded linear rank-width and are unstable. By contrast, once the edge relation is stable, the bounded-rank-width class behaves like a definable image of bounded treewidth. The same framework extends to low rankwidth covers: low rankwidth covers together with stable edge relation are equivalent to being a transduction of a class with bounded expansion, and in particular bounded rank-width plus stable edge relation implies linear \(\chi\)-boundedness [2007.07857].

## 3. Algorithmic role of bounded rank-width

One algorithmic reading of focused rank-width is that low-rank interfaces support canonical decomposition and dynamic programming. Grohe and Schweitzer proved that for every fixed \(k\), graph isomorphism on graphs of rank-width at most \(k\) is decidable in polynomial time, more precisely by computing a canonical treelike decomposition of the cut-rank function and then combining local isomorphism information through group-theoretic dynamic programming. Because clique-width is bounded by a function of rank-width, the same yields polynomial-time isomorphism testing for every fixed clique-width bound [1505.03737].

This structural route was later sharpened by the Weisfeiler–Leman perspective. On graphs of rank-width at most \(k\), the \((3k+4)\)-dimensional Weisfeiler–Leman algorithm identifies every graph in the class, giving an isomorphism test in time \(O(n^{3k+5}\log n)\). The same result yields the first polynomial-time canonisation algorithm for bounded rank-width and implies that fixed-point logic with counting captures polynomial time on every class of graphs of rank-width at most \(k\) [1901.10330].

The same low-rank viewpoint extends beyond isomorphism. For CNF formulas, if one forms the signed incidence graph and measures its signed rank-width, then both \(\#\mathrm{SAT}\) and Max-SAT admit parameterized polynomial algorithms with running time
\[
\mathcal{O}\bigl(t^3\cdot 2^{3t(t+1)/2}\cdot |\phi|\bigr),
\]
where \(t\) is the signed rank-width of the formula. The dynamic programming state space is indexed by subspaces over \(\mathbb{F}_2\), reflecting the fact that rank-width controls the linear-algebraic complexity of clause-variable interaction across decomposition cuts [1006.5621]. This suggests that any focused variant preserving low-rank interfaces should remain compatible with subspace-based dynamic programming.

## 4. Structural bridges to tree-width, path-width, and minimal layers

Rank-width is not tree-width, but it admits a strong tree-width shadow. Every graph of rank-width \(k\) is a pivot-minor of a graph of tree-width at most \(2k\), and every graph of linear rank-width \(k\) is a pivot-minor of a graph of path-width at most \(k+1\). In the extremal case \(k=1\), graphs of rank-width at most \(1\) are exactly distance-hereditary graphs, exactly the vertex-minors of trees, while graphs of linear rank-width at most \(1\) are exactly the vertex-minors of paths; in the bipartite setting these become pivot-minors of trees and paths, respectively [1203.3606].

The rank-width-\(1\) stratum is especially revealing. For distance-hereditary graphs, linear rank-width can be computed in time \(\mathcal{O}(n^2\log^2 n)\), together with a witnessing linear layout. The key structural tool is the canonical split decomposition, whose bags are cliques or stars, together with the notion of limbs, which correspond to vertex-minors and support a recursive characterization of linear rank-width similar to the classical characterization of path-width for forests [1403.1081]. This suggests that one productive meaning of focus is restriction to particularly transparent layers of the rank-width hierarchy, where the decomposition anatomy becomes explicit.

These results also clarify a frequent misconception. Bounded rank-width does not mean sparse, and it does not collapse to bounded tree-width; rather, it describes graphs that admit tree-like recursive cuts of low binary rank. The tree-width and path-width host graphs obtained through pivot-minors therefore do not identify rank-width with sparse structure, but show that low-rank cut structure can be encoded inside sparse hosts [1203.3606].

## 5. Obstructions, hardness, and class boundaries

The obstruction theory for linear rank-width is explicit but large. For every fixed \(p\), the forbidden pivot-minors for \(\sigma\)-symmetric matrices over a fixed finite field of linear rank-width at most \(p\) have size bounded doubly exponentially in \(p\); as a corollary, forbidden vertex-minors for graphs of linear rank-width at most \(p\) have size at most \(2^{2^{O(p)}}\). The proof adapts Lagergren’s pseudo-minor order from path-width to the pivot-minor world through boundaried \(s\)-labelled graphs and linear \(s\)-profiles [1412.6201].

From the algorithmic side, bounded rank-width supports fixed-parameter algorithms, but for several classical problems the dependence on the parameter is provably quadratic in the exponent. Assuming ETH, there is no \(2^{o(k^2)}n^{O(1)}\) algorithm parameterized by linear rank-width for Independent Set, Weighted Dominating Set, Maximum Induced Matching, or Feedback Vertex Set. These lower bounds match the known \(2^{O(k^2)}n^{O(1)}\) algorithms and show that, for these problems, rank-width-based dynamic programming is already asymptotically tight in its parameter dependence [2210.02117].

At the level of graph classes, bounded rank-width is also a genuine restriction. Even-hole-free graphs do not have bounded rank-width in general, and the same remains true even for a class of \((\)diamond, even hole\()\)-free graphs with no clique cutset. This gives a negative answer to the question whether forbidding clique cutsets within even-hole-free graphs suffices to bound rank-width, and it blocks direct application of Courcelle–Makowsky style meta-theorems to that class [1611.09907].

## 6. Modern extensions and broader significance

Recent work has pushed the algorithmics of rank-width close to linear time. Given an \(n\)-vertex \(m\)-edge graph \(G\) and an integer \(k\), one can now, in time
\[
O_k(n^{1+o(1)})+O(m),
\]
either output a rank-decomposition of width at most \(k\) or conclude that \(\mathrm{rw}(G)>k\). The same algorithm also returns a \((2^{k+1}-1)\)-expression for clique-width. Its main ingredient is a fully dynamic data structure that maintains a rank-decomposition of width at most \(4k\) under edge insertions and deletions, under the promise that rank-width never exceeds \(k\); the same structure can maintain fixed \({\sf CMSO}_1\) properties with the same amortized update time [2402.12364].

A different line of work recasts rank-width categorically. In a prop of open graphs, monoidal width captures rank-width up to a constant factor: for a graph \(G\) encoded as a morphism \(g\),
\[
\frac{1}{2}\operatorname{rw}(G)\le \operatorname{mwd}(g)\le 2\operatorname{rw}(G).
\]
This does not change the graph-theoretic invariant, but it reinterprets rank-width as the minimal peak interface size in a compositional construction, thereby providing an algebraic template for rank-width-like parameters on other kinds of structures [2205.08916].

At the opposite extreme from tame focused subclasses, recent deterministic lower-bound constructions show how large rank-width can become. By deriving cut-rank lower bounds from edge expansion, one obtains bounded-degree deterministic graph families with provably maximum rank-width \(\Theta(n)\), including explicit expander-based families. In the corresponding graph states, entanglement width equals rank-width, so these families also yield deterministic graph states with entanglement width \(\Theta(n)\) that remain constant-depth preparable under all-to-all connectivity [2606.07110]. This suggests a final, broader reading of focused rank-width: the subject is not only about isolating tame subclasses, but also about understanding exactly which structural mechanisms—stability, low-rank interfaces, expansion, or compositionality—govern the full range of rank-width behaviour.

Source: https://www.emergentmind.com/topics/focused-rank-width