Digraph Decompositions and Monotonicity in Digraph Searching

We consider monotonicity problems for graph searching games. Variants of these games - defined by the type of moves allowed for the players - have been found to be closely connected to graph decompositions and associated width measures such as path- or tree-width. Of particular interest is the question whether these games are monotone, i.e. whether the cops can catch a robber without ever allowing the robber to reach positions that have been cleared before. The monotonicity problem for graph searching games has intensely been studied in the literature, but for two types of games the problem was left unresolved. These are the games on digraphs where the robber is invisible and lazy or visible and fast. In this paper, we solve the problems by giving examples showing that both types of games are non-monotone. Graph searching games on digraphs are closely related to recent proposals for digraph decompositions generalising tree-width to directed graphs. These proposals have partly been motivated by attempts to develop a structure theory for digraphs similar to the graph minor theory developed by Robertson and Seymour for undirected graphs, and partly by the immense number of algorithmic results using tree-width of undirected graphs and the hope that part of this success might be reproducible on digraphs using a directed tree-width. Unfortunately the number of applications for the digraphs measures introduced so far is still small. We therefore explore the limits of the algorithmic applicability of digraph decompositions. In particular, we show that various natural candidates for problems that might benefit from digraphs having small directed tree-width remain NP-complete even on almost acyclic graphs.

💡 Research Summary

The paper investigates monotonicity in graph searching games played on directed graphs (digraphs) and explores the algorithmic limits of several directed graph decomposition measures. Graph searching games, also known as Cops‑and‑Robber games, involve a set of cops that try to capture a robber moving on the vertices of a graph. Depending on whether the robber is visible or invisible, fast (dynamic) or lazy (inert), different variants arise, each giving rise to a width parameter: for undirected graphs the visible‑fast variant corresponds to tree‑width, the invisible‑fast variant to path‑width, etc. In the undirected setting all these variants are known to be monotone – the cops can capture the robber without ever having to re‑occupy a vertex that has already been cleared, and the minimum number of cops needed does not increase when monotonicity is required.

For directed graphs the situation is more complex. Four natural variants have been studied: (i) SCC‑visible‑dynamic (scc‑vis), (ii) reachability‑visible‑dynamic (reach‑vis), (iii) reachability‑invisible‑dynamic (reach‑invis) which defines DAG‑width, and (iv) reachability‑invisible‑inert (reach‑inert) which defines Kelly‑width. While the reach‑invis (DAG‑width) and reach‑inert (Kelly‑width) games were conjectured to be monotone, the monotonicity of the visible‑fast and invisible‑lazy variants on digraphs remained open.

The authors resolve this long‑standing open problem by constructing explicit counter‑examples. For every integer p ≥ 2 they build a digraph Dₚ consisting of three large cliques (C₀, C₂, C₂₁) and an independent set C₁₁, with all possible directed edges from one part to the next. They prove that the cop‑width (the minimum number of cops needed to win) of Dₚ in the DAG‑game is 3p − 1, but any monotone strategy requires at least 4p − 2 cops. Hence the DAG‑game is non‑monotone. A similar construction shows that the Kelly‑game (reach‑inert) is also non‑monotone. These results demonstrate that, unlike the undirected case, the directed variants with a visible fast robber or an invisible lazy robber cannot be forced to be monotone without a substantial increase in resources.

Having established non‑monotonicity, the paper turns to the broader question of how useful directed width measures are for algorithmic graph theory. Existing notions—directed tree‑width (dtw), DAG‑width, Kelly‑width, directed path‑width—are all motivated by the hope that many NP‑hard problems become tractable on digraphs of bounded width, mirroring the success of tree‑width for undirected graphs. The authors systematically examine a suite of classic NP‑complete problems (disjoint paths, Hamiltonian cycle, Minimum Equivalent Subgraph, Feedback Vertex Set, Feedback Arc Set, Graph Grundy Numbering, etc.) and ask whether any of them become polynomial‑time solvable on digraphs whose directed width is bounded by a constant.

Their findings are striking: apart from the already known tractable cases (disjoint paths and Hamiltonicity, which are solvable on bounded directed tree‑width), all other examined problems remain NP‑complete even on digraphs that are “almost acyclic” – i.e., graphs with very low global connectivity and consequently very small directed path‑width or tree‑width. For example, Feedback Vertex Set and Feedback Arc Set are NP‑complete on digraphs with directed tree‑width = 2, and Minimum Equivalent Subgraph is NP‑complete on graphs of directed path‑width = 1. This demonstrates that bounded directed width does not, by itself, guarantee algorithmic tractability for a wide class of problems.

The paper concludes by discussing the implications of these results. The non‑monotonicity of DAG‑ and Kelly‑games indicates that the natural game‑theoretic characterisations of directed width measures cannot be simplified to monotone strategies, limiting the design of clean, monotone algorithms. Moreover, the persistence of NP‑hardness on low‑width digraphs suggests that current width parameters capture only a narrow structural aspect of digraphs, insufficient for a general “graph minor theory” analogue in the directed setting. The authors propose several avenues for future work: developing new structural parameters that combine width with additional constraints (e.g., bounded feedback sets, limited cycle length), tightening the bounds on monotone versus non‑monotone cop‑width, and investigating specific graph classes (such as DAGs that are close to trees) where monotonicity might hold. Overall, the paper resolves a key open problem in directed graph searching, clarifies the limits of existing directed width measures, and sets a research agenda for extending the algorithmic toolbox for digraphs.