[Paper Review] A Labeling Approach to Incremental Cycle Detection

Edith Cohen, Amos Fiat+2 more|arXiv (Cornell University)|Oct 31, 2013

Complexity and Algorithms in Graphs15 references3 citations

TL;DR

This paper introduces a novel labeling technique for incremental cycle detection in directed acyclic graphs (DAGs), using recursive random ranks to maintain weak topological order. The method enables efficient cycle detection with total time complexity matching prior bests—$O(n^2/log n)$ and $O(m \cdot \min\{m^{1/2}, n^{2/3}\})$—while offering greater simplicity and better suitability for distributed implementation through localized label comparisons and pruning.

ABSTRACT

In the \emph{incremental cycle detection} problem arcs are added to a directed acyclic graph and the algorithm has to report if the new arc closes a cycle. One seeks to minimize the total time to process the entire sequence of arc insertions, or until a cycle appears. In a recent breakthrough, Bender, Fineman, Gilbert and Tarjan \cite{BeFiGiTa11} presented two different algorithms, with time complexity $O(n^2 \log n)$ and $O(m \cdot \min \{m^{1/2}, n^{2/3} \})$, respectively. In this paper we introduce a new technique for incremental cycle detection that allows us to obtain both bounds (up to a logarithmic factor). Furthermore, our approach seems more amiable for distributed implementation.

Motivation & Objective

• To address the incremental cycle detection problem in directed acyclic graphs (DAGs) where edges are added one by one and cycles must be detected as soon as they form.
• To unify previous algorithms under a single, more intuitive framework based on recursive vertex labeling with random ranks.
• To improve the algorithm's suitability for distributed systems by leveraging local, label-based pruning in forward and backward searches.
• To achieve time complexity matching the best-known centralized bounds, up to poly-logarithmic factors, without sacrificing correctness or efficiency.

Proposed method

• Assign random ranks from $1$ to $qn$ to a subset of $qn$ vertices in the DAG, where $0 < q \leq 1$.
• Define a recursive label $\ell(u)$ for each vertex $u$ as a sequence of ranked vertices, starting with the minimum-rank vertex that can reach $u$.
• Extend labels recursively: each subsequent vertex in $\ell(u)$ is the minimum-rank vertex reachable from the previous label vertex and able to reach $u$.
• Use lexicographic ordering of label sequences to rule out paths: if $\ell(y) \succ \ell(x)$, then no path from $y$ to $x$ exists.
• Maintain labels dynamically during edge insertions using message-passing schedules that propagate label updates efficiently.
• Prune forward and backward searches using label comparisons, reducing the number of vertices explored during cycle detection.

Experimental results

Research questions

• RQ1 Can a unified framework be developed for incremental cycle detection that simplifies prior algorithms while maintaining optimal time complexity?
• RQ2 How can recursive random labeling be used to maintain weak topological order and enable efficient path existence checks in dynamic DAGs?
• RQ3 To what extent do label-based pruning techniques reduce the cost of forward and backward searches in cycle detection?
• RQ4 Can the proposed labeling method be efficiently distributed, and how does it compare to prior approaches in terms of message complexity and synchrony assumptions?
• RQ5 What is the impact of parameter $q$—the fraction of ranked vertices—on the trade-off between forward and backward search costs?

Key findings

• The expected length of the recursive labels is logarithmic in the number of vertices, ensuring compact representation and efficient comparison.
• Label sequences are lexicographically descending along any path in a DAG, enabling efficient path existence checks via label comparison.
• The set of vertices sharing the same label as a given vertex $u$, from which $u$ is reachable, is of size $O(\log n / q)$ with high probability.
• By choosing $q = 1$, the algorithm achieves $O(n^2 \log^2 n)$ total time with no backward search, due to unique labels.
• For sparse graphs ($m = O(n)$), setting $q = 1/\sqrt{n}$ yields a total time of $O(n^{3/2} \log n)$, matching known bounds.
• With $q = \sqrt[3]{\log n / n}$, the forward and backward search times balance at $O(m \cdot n^{2/3} \log^{4/3} n)$, achieving the best trade-off for dense graphs.