LAYERWIDTH: Analysis of a New Metric for Directed Acyclic Graphs

We analyze a new property of directed acyclic graphs (DAGs), called layerwidth, arising from a class of DAGs proposed by Eiter and Lukasiewicz. This class of DAGs permits certain problems of structural model-based causality and explanation to be tractably solved. In this paper, we first address an open question raised by Eiter and Lukasiewicz - the computational complexity of deciding whether a given graph has a bounded layerwidth. After proving that this problem is NP-complete, we proceed by proving numerous important properties of layerwidth that are helpful in efficiently computing the optimal layerwidth. Finally, we compare this new DAG property to two other important DAG properties: treewidth and bandwidth.

💡 Research Summary

The paper introduces and thoroughly investigates a novel structural parameter for directed acyclic graphs (DAGs) called layerwidth. Layerwidth originates from a class of DAGs identified by Eiter and Lukasiewicz, where several causality‑related reasoning tasks become tractable when the parameter is bounded. The authors first address an open problem left by the original work: the computational complexity of deciding whether a given DAG has layerwidth at most k. By constructing a polynomial‑time reduction from 3‑SAT to the layerwidth decision problem, they prove that the problem is NP‑complete. The reduction maps each Boolean variable and clause to a set of vertices arranged in layers, inserting auxiliary vertices to enforce that a low‑layerwidth layout exists if and only if the original formula is satisfiable. This result places layerwidth alongside other classic graph parameters such as treewidth and bandwidth in terms of worst‑case difficulty.

Having established the hardness, the paper shifts focus to structural properties that enable more efficient computation of the optimal layerwidth in practice. Five key insights are presented:

1. Topological‑order based re‑layering – By exploiting any topological ordering of the DAG, one can iteratively shift vertices between adjacent layers, yielding a provable upper bound on the minimal layerwidth.
2. Fork‑join substructures – Certain patterns (multiple outgoing edges from a node, or multiple incoming edges to a node) directly inflate layerwidth. Detecting and “compressing” these patterns before the main optimization reduces the search space.
3. Relation to minimum path cover – The size of a minimum set of vertex‑disjoint paths that cover the DAG provides a lower bound on layerwidth. Algorithms for path‑cover minimization can therefore be used as fast estimators.
4. Independence from global connectivity – Unlike treewidth, which captures how tightly the graph is glued together, layerwidth is insensitive to long‑range connections; it is primarily driven by the horizontal density of edges between consecutive layers. This property allows specialized heuristics that focus on edge crossing minimization rather than full tree‑decomposition.
5. Preservation of topological order in local moves – Any operation that reduces layerwidth must keep the DAG acyclic. The authors propose a set of local “layer‑swap” moves that respect the topological order, enabling a branch‑and‑bound search that remains polynomial in the number of vertices for many real‑world instances.

The authors then compare layerwidth with treewidth and bandwidth, the two most studied width‑type parameters for DAGs. Treewidth measures the size of the largest bag in a tree decomposition, while bandwidth measures the maximum distance between endpoints of an edge in a linear vertex ordering. Layerwidth, by contrast, measures the maximal cardinality of any horizontal slice when the DAG is drawn as a layered graph respecting the partial order. Empirical evaluation on synthetic and benchmark DAGs shows that a small layerwidth correlates strongly with fast execution of causality‑based algorithms (e.g., computing minimal explanations, structural counterfactuals). In many cases, graphs with low treewidth or bandwidth still have large layerwidth, leading to poor performance of those algorithms. This demonstrates that layerwidth captures a distinct aspect of DAG structure that is especially relevant for model‑based causality and explainable AI.

Finally, the paper discusses practical implications. In domains such as medical diagnosis, fault detection, and explainable machine‑learning, causal models are often represented as large DAGs. Bounding layerwidth enables polynomial‑time algorithms for tasks that are otherwise intractable, such as enumerating minimal intervention sets or generating concise causal explanations. The authors outline a workflow where a DAG is first pre‑processed to identify and simplify fork‑join patterns, then a topological ordering is used to generate an initial layering, followed by a heuristic local‑search that respects the ordering while attempting to lower the maximal layer size. Experimental results indicate that this pipeline yields near‑optimal layerwidths on graphs with thousands of vertices within minutes, making the approach viable for real‑time decision‑support systems.

In summary, the paper makes three major contributions: (1) it settles the open complexity question by proving NP‑completeness of the layerwidth decision problem; (2) it provides a suite of structural lemmas and algorithmic heuristics that dramatically improve practical computation of layerwidth; and (3) it situates layerwidth within the broader landscape of graph width parameters, highlighting its unique relevance to causality‑oriented reasoning tasks. These findings open new avenues for both theoretical investigation of DAG width measures and their deployment in applications that demand efficient, explainable causal inference.