On the existence of shortest directed networks

A directed network connecting a set A to a set B is a digraph containing an a-b path for each a in A and b in B. Vertices in the directed network not in A or B are called Steiner points. We show that in a finitely compact metric space in which geodesics exist, any two finite sets A and B are connected by a shortest directed network. We also bound the number of Steiner points by a function of the sizes of A and B. Previously, such an existence result was known only for the Euclidean plane [M. Alfaro, Pacific J. Math. 167 (1995) 201-214]. The main difficulty is that, unlike the undirected case (Steiner minimal trees), the underlying graphs need not be acyclic. Existence in the undirected case was first shown by E. J. Cockayne [Canad. Math. Bull. 10 (1967) 431-450].

💡 Research Summary

The paper addresses the existence and structural bounds of shortest directed networks (also known as directed Steiner networks) that connect a finite source set A to a finite sink set B in a metric space. While the analogous problem for undirected Steiner trees was solved by Cockayne (1967) for any finitely compact metric space with geodesics, the directed version remained open except for the Euclidean plane, where Alfaro (1995) proved existence using plane‑specific geometry. The main contribution of this work is to extend the existence result to any finitely compact metric space in which every pair of points can be joined by a geodesic, and to provide an explicit polynomial bound on the number of Steiner (intermediate) points required.

Key Definitions.

• An (A,B)‑network is a directed graph G = (V,E) with A∪B ⊆ V ⊆ X such that for every a∈A and b∈B there exists a directed a→b path.
• Vertices in A are called sources, those in B are sinks, and all other vertices are Steiner points.
• The length ℓ(G) is the sum of the metric distances of all arcs.
• A network is simple if every Steiner point has degree at least three (i.e., at least three incident arcs). Simplicity follows from a standard reduction: any Steiner point of degree ≤ 2 can be removed or replaced without increasing total length.

Main Theorem (Structural Bound).
Let m = |A| and n = |B|. Any simple shortest (A,B)‑network contains at most O(m²n + mn²) Steiner points. The proof proceeds by selecting a longest source‑to‑sink directed path P = a₀ x₁ … x_k b₀ in a shortest network G. For each source a, define x(a) as the first vertex of P reachable from a; for each sink b, define y(b) as the last vertex of P that can reach b. Any a→b path can be expressed as a concatenation of three parts: an initial segment from a to x(a), a sub‑path of P (or a “jump” that skips forward along P), and a final segment from y(b) to b. A “jump” (i,j) is a directed sub‑path that goes from vertex x_i to x_j with i > j and is edge‑disjoint from P.

Collect all jumps appearing in all a→b paths into a set J. Choose a minimal subset I ⊆ J such that replacing J by I still covers all required connections. By analyzing the interaction between jumps and the vertices of P, the authors show that every vertex of P must be either an x(a), a y(b), or an endpoint of a jump from I. Consequently, |P| ≤ m + n + 2|I|.

Two combinatorial properties of I are proved: (1) each vertex x_t of P can be “dominated” by at most two jumps, and (2) between any two consecutive jumps (i,j) and (k,ℓ) there exists at least one vertex of P that is dominated by one of them. These properties imply |I| ≤ 4(m + n) + 1, yielding |P| ≤ 9(m + n) + 2. Since each jump contributes at most one Steiner point and each source/sink contributes at most O(mn) jumps, the total number of Steiner points is bounded by O(m²n + mn²).

Existence via Compactness.
A compactness lemma is proved: in a finitely compact metric space with geodesics, among all (A,B)‑networks that use at most s Steiner points there exists a shortest one. The argument uses the finiteness of possible abstract digraph structures (bounded by |A| + |B| + s vertices) and the compactness of X to extract a convergent subsequence of vertex positions, whose limit is a network of minimal length. Combining this lemma with the structural bound (which guarantees a finite s for any optimal network) yields the corollary that a shortest (A,B)‑network always exists for arbitrary finite A and B.

Complexity and Related Problems.
The paper notes that finding a shortest (A,B)‑network is NP‑complete because the Minimal Equivalent Digraph (MED) problem reduces to it. MED asks for a smallest subdigraph preserving reachability, a known NP‑complete problem. Moreover, the point‑to‑point connection problem (where A and B are paired and only a_i→b_i paths are required) inherits the same O(n²) Steiner‑point bound as a special case.

Concluding Remarks.
The authors acknowledge that the O(m²n + mn²) bound is likely not tight; the constant 9 in the path‑length estimate can be reduced, and examples exist where at least m + n Steiner points are necessary. Improving the exponent (currently cubic) is an open challenge and would have direct implications for algorithm design, especially for small‑scale instances. The paper emphasizes that while existence is now settled in great generality, the computational difficulty remains, motivating further research into approximation algorithms or fixed‑parameter tractable methods.

In summary, the paper establishes that in any finitely compact metric space with geodesics, shortest directed networks connecting two finite point sets always exist, and it provides a concrete polynomial bound on the number of auxiliary Steiner points needed. This advances the theory of directed Steiner problems beyond the Euclidean plane and lays groundwork for future algorithmic developments.