Faster Algorithms for Global Minimum Vertex-Cut in Directed Graphs

Abstract: We study the directed global minimum vertex-cut problem: given a directed vertex-weighted graph $G$, compute a vertex-cut $(L,S,R)$ in $G$ of minimum value, which is defined to be the total weight of all vertices in $S$. The problem, together with its edge-based variant, is one of the most basic in graph theory and algorithms, and has been studied extensively. The fastest currently known algorithm for directed global minimum vertex-cut (Henzinger, Rao and Gabow, FOCS 1996 and J. Algorithms 2000) has running time $\tilde{O}(mn)$, where $m$ and $n$ denote the number of edges and vertices in the input graph, respectively. A long line of work over the past decades led to faster algorithms for other main versions of the problem, including the undirected edge-based setting (Karger, STOC 1996 and J. ACM 2000), directed edge-based setting (Cen et al., FOCS 2021), and undirected vertex-based setting (Chuzhoy and Trabelsi, STOC 2025). However, for the vertex-based version in directed graphs, the 29 year-old $\tilde{O}(mn)$-time algorithm of Henzinger, Rao and Gabow remains the state of the art to this day, in all edge-density regimes. In this paper we break the $Θ(mn)$ running time barrier for the first time, by providing a randomized algorithm for directed global minimum vertex-cut, with running time $O\left(mn^{0.976}\cdot\operatorname{polylog} W\right)$ where $W$ is the ratio of largest to smallest vertex weight. Additionally, we provide a randomized $O\left(\min\left{m^{1+o(1)}\cdot k,n^{2+o(1)}\right}\right)$-time algorithm for the unweighted version of directed global minimum vertex-cut, where $k$ is the value of the optimal solution. The best previous algorithm for the problem achieved running time $\tilde O\left(\min\left{k^2 \cdot m, mn^{11/12+o(1)}, n^{2+o(1)}\right}\right)$ (Forster et al., SODA 2020, Li et al., STOC 2021).

• The paper presents novel randomized algorithms that break the Θ(mn) barrier for computing the global minimum vertex-cut in directed graphs.
• It introduces innovative techniques such as random terminal hierarchies and sparsification to effectively handle directed vertex capacities.
• The unweighted case nearly matches the performance of top edge-based methods, marking a significant theoretical and practical advance.

Faster Algorithms for Global Minimum Vertex-Cut in Directed Graphs

Introduction and Motivation

The paper "Faster Algorithms for Global Minimum Vertex-Cut in Directed Graphs" (2512.24355) addresses a longstanding open problem in the domain of graph algorithms, presenting the first significant progress since Henzinger, Rao, and Gabow's $(mn)$-time algorithm from 1996 for computing the global minimum vertex-cut in a directed graph. The minimum vertex-cut problem, in both its edge and vertex variants, is foundational in graph theory and algorithm design due to its connection to network reliability, connectivity, and flow computations. While substantial improvements have occurred for edge-based and undirected versions, the directed vertex-cut problem has resisted better-than-$(mn)$ solutions even as edge-centric approaches push towards near-linear time.

This work introduces new randomized algorithms that break the $\Theta(mn)$ barrier for the directed vertex-cut, with further contributions for the unweighted case. These results not only improve the asymptotic performance but also highlight new structural and algorithmic ideas for handling the difficulties unique to directed vertex-separators, for which classic techniques such as tree packing (Tutte/Nash-Williams) and isolating cuts (in the undirected regime) do not generalize.

Problem Statement and Background

Given a directed, vertex-weighted graph $G=(V,E)$ with weights $w(v) \geq 0$ on $v \in V$, the global minimum vertex-cut is the tripartition $(L, S, R)$ of $V$ minimizing $w(S)$, where no edge exists from $L$ to $R$ and $L,R \neq \emptyset$. In the unweighted case $w(v) = 1$ for all $v$. The problem generalizes vertex-connectivity computation. Prior to this work, the fast algorithm for undirected edge-min-cut (Karger, $\tilde{O}(m)$) and various directed/undirected edge- and vertex-cut algorithms achieved near-optimal complexity, but directed vertex-cut lagged behind with only $(mn)$-time solutions.

A major bottleneck is that many tools—such as contraction, tree packing, and the isolating cuts lemma—do not extend to the directed-vertex setting. Furthermore, naive reduction from vertex to edge capacities through split graphs does not preserve the structure of global vertex-cuts in the directed case. Thus, improving upon existing approaches required algorithmic innovations tailored to the directed vertex-cut landscape.

Main Results

1. General Weighted Directed Vertex-Cut Algorithm:

The authors present a randomized algorithm for the directed, weighted global minimum vertex-cut problem with running time

$O\left(m n^{0.976} \cdot \mathrm{polylog} W\right)$

where $m$ and $n$ are the edge and vertex counts, and $W$ the weight ratio. This is the first sub-$(mn)$ algorithm for the problem, leveraging novel techniques rather than reductions.

2. Unweighted Directed Vertex-Cut Improvement:

For the unweighted case, the paper provides a randomized

$O\left(\min\,\{ m^{1+o(1)}k, n^{2+o(1)} \}\right)$

algorithm, with $k$ the value of the optimal min-cut. This closely matches the best known directed edge-cut bounds $\tilde{O}(\min \{km, n^2\})$, a previously unattainable milestone for vertex separators.

3. Simpler/Improved Undirected Algorithm:

The algorithmic structure also leads to simplifications and improvements over the recent undirected vertex-cut algorithms (e.g., Chuzhoy & Trabelsi, STOC 2025), offering a technical leap and more elegant proofs.

Algorithmic Techniques and Structural Insights

Sublinear Vertex-Cut via Random Terminal Hierarchies

A core contribution is the design of a random sampling/partitioning framework over potential separators that incrementally tracks promising terminal sets. The algorithm constructs hierarchical partitions of sampled terminals and anti-terminals (for negative coverage) that, with high probability, are "hit" by some minimum cut. The process closely mimics recursive ball-growing but with careful handling of the directionality and vertex restrictions.

For the unweighted case, the construction enables efficient handling of potentially large collections of terminal pairs without incurring a direct $O(mn)$ cost for each one. The approach is extended and refined in the weighted regime.

Sparsification and Gradual Commitment

The method avoids processing all pairs by gradually 'committing' to subsets of the terminals in a hierarchical manner, focusing on preserving and propagating minimum cuts through carefully structured subgraphs. This reduces redundant min-cut computation and cleverly exploits flow sparsifications and anti-terminal constraints to maintain correctness.

Decomposition by Density and Balance

The approach distinguishes cases based on whether the input graph or the minimum cut is "balanced" (i.e., large $|L|$ or large total out-degree of $L$) or "unbalanced". Special-case routines efficiently handle balanced minimum cuts using random sampling, whereas for unbalanced/dense regimes, more intricate partitioning is applied.

Local Flow and Split Graph Augmentation

Weighted/dense cases utilize local flow techniques and adaptations of split graphs to capture capacity scaling, augmented by shortcuts and modified adjacency representations tailored to vertex-based capacities.

Numerical Impact and Comparative Performance

• Weighted algorithm's $O(m n^{0.976} \cdot \mathrm{polylog} W)$ runtime represents a significant improvement over prior $O(mn)$ or $O(m n^{1+o(1)})$ (deterministic) bounds.
• Unweighted case: nearly matches the best known edge-based algorithms, previously unattainable for vertex cuts.
• Algorithmic structure substantially simplifies and in some cases accelerates previous undirected or approximate vertex-cut routines.

Contrasting Claims and Implications

• The paper demonstrates that the edge-centric approaches (e.g., tree packings) fundamentally fail for directed vertex-cuts. Their new framework is both technically distinct and more generally applicable.
• While succeeding in significantly breaking the $(mn)$ barrier, the algorithm is randomized. Recent independent works (Frandsen et al.) have shown that for dense graphs, reductions to undirected versions are sufficient for $n^{2+o(1)}$-time, but no prior work achieved a similar improvement in both sparse/dense and weighted/unweighted settings within a unified framework.

Future Directions and Theoretical Implications

• The methodology has potential to seed further advances in directed separator, flow, and cut problems, possibly enabling deterministic analogues or even finer complexity.
• The paradigm may adapt for other hard-to-approximate graph partitioning problems where directionality undermines classical techniques.
• The approach may inspire novel parameterizations (e.g., for fixed-parameter tractability wrt cut size $k$) and finer-grained complexity categorizations.

Conclusion

This paper breaks a near-three-decade complexity barrier for the global minimum directed vertex-cut problem, introducing randomized algorithms with sub-$(mn)$ runtime and connections to advances in undirected and edge-based cut computations (2512.24355). The work opens novel algorithmic territory for directed graphs, extends sparsification and partitioning frameworks to new domains, and suggests a path for further unification and generalization of separator algorithms in graph theory and combinatorial optimization.