Cycle rank

In graph theory, the cycle rank of a directed graph is a digraph connectivity measure proposed first by Eggan and Büchi (Eggan 1963). Intuitively, this concept measures how close a digraph is to a directed acyclic graph (DAG), in the sense that a DAG has cycle rank zero, while a complete digraph of order n with a self-loop at each vertex has cycle rank n. When applied to undirected graphs, the concept of cycle rank bears many different names in the research literature, including vertex ranking number, ordered chromatic number, minimum elimination tree height and tree-depth (Bodlaender et al. 1998, Rossman 2008)

Besides its original application in the studying the star height of formal languages, the measure has found use in sparse matrix computations (see Bodlaender et al. 1995) and logic (Rossman 2008).

Contents

Definition

The cycle rank r(G) of a digraph G = (V, E) is inductively defined as follows:

r(G) = 1 %2B \min_{v\in V} r(G-v).\,

History

In the special case of undirected graphs, the cycle rank was rediscovered about twenty years later in the context of sparse matrix computations (Schreiber 1982). Apparently being unaware of the original definition from the 1960s, later authors generalized the concept once again to digraphs (Eisenstat & Liu 2005).

Bounds

Any n-vertex forest has cycle rank O(log n). For, in a forest, one can always find a constant number of vertices the removal of which leaves a forest that can be partitioned into two smaller subforests with at most 2n/3 vertices each. By recursively partitioning each of these two subforests, we can easily derive a logarithmic upper bound on the cycle rank. The same technique, applied to a tree decomposition of a graph, shows that, if the treewidth of an n-vertex graph G is t, then the pathwidth of G is O(t log n), see (Bodlaender et al. 1995). Since outerplanar graphs, series-parallel graphs, and Halin graphs all have bounded treewidth, they all also have at most logarithmic cycle rank.

Cycle rank formulas for some digraphs

As mentioned in the introduction, the cycle rank of a directed acyclic graph is 0, while a complete digraph of order n with a self-loop at each vertex has cycle rank n. Apart from these, the cycle rank of a few other digraphs is known: the undirected path P_n of order n, which possesses a symmetric edge relation and no self-loops, has cycle rank \lfloor \log n\rfloor (McNaughton 1969). For the directed (m\times n)-torus T_{m,n}, i.e., the cartesian product of two directed circuits of lengths m and n, we have r(T_{n,n}) = n and r(T_{m,n}) = \min\{m,n\} %2B 1 for m ≠ n (Eggan 1963, Gruber & Holzer 2008).

Computing the cycle rank

Computing the cycle rank is computationally hard: already in the case of undirected graphs, the corresponding decision problem is NP-complete (Pothen 1988). The same holds true in the case of digraphs (Gruber 2008). For undirected graphs, problem remains NP-complete for cobipartite graphs (Pothen 1988), that is, complements of bipartite graphs, for bipartite graphs (Bodlaender et al. 1998), as well as for chordal graphs (Dereniowski & Nadolski 2006). On the positive side, the problem is solvable in polynomial time on interval graphs (Aspvall & Heggernes 1994), as well as on permutation, trapezoid, circular-arc, circular permutation graphs, and cocomparability graphs of bounded dimension (Deogun et al. 1999). For undirected trees, the problem is even solvable in linear time (Iyer, Ratliff & Vijayan 1988, Schäffer 1989). Concerning the approximation complexity of the problem, Bodlaender et al. (1995) give an O((\log n)^2)-approximation algorithm for the undirected case.

Applications

Star height of regular languages

The very first application of cycle rank was in formal language theory, for studying the star height of regular languages. Eggan (1963) established a relation between the theories of regular expressions, finite automata, and of directed graphs. In subsequent years, this relation became known as Eggan's theorem, cf. Sakarovitch (2009). In automata theory, a nondeterministic finite automaton with ε-moves (ε-NFA) is defined as a 5-tuple, (Q, Σ, δ, q0, F), consisting of

A word w ∈ Σ* is accepted by the ε-NFA if there exists a directed path from the initial state q0 to some final state in F using edges from δ, such that the concatenation of all labels visited along the path yields the word w. The set of all words over Σ* accepted by the automaton is the language accepted by the automaton A.

When speaking of digraph properties of a nondeterministic finite automaton A with state set Q, we naturally address the digraph with vertex set Q induced by its transition relation. Now the theorem is stated as follows.

Eggan's Theorem: The star height of a regular language L equals the minimum cycle rank among all nondeterministic finite automata with ε-moves accepting L.

Proofs of this theorem are given by Eggan (1963), and more recently by Sakarovitch (2009).

Cholesky factorization in sparse matrix computations

Another application of this concept lies in sparse matrix computations, namely for computing the Cholesky factorization of a (symmetric) matrix in parallel. A given sparse (n\times n)-matrix M may be interpreted as the adjacency matrix of some symmetric digraph G on n vertices, in a way such that the non-zero entries of the matrix are in one-to-one correspondence with the edges of G. If the cycle rank of the digraph G is at most k, then the Cholesky factorization of M can be computed in at most k steps on a parallel computer with n processors (Dereniowski & Kubale 2003).

See also

References