A Factor 3/2 Approximation for Generalized Steiner Tree Problem with Distances One and Two

We design a 3/2 approximation algorithm for the Generalized Steiner Tree problem (GST) in metrics with distances 1 and 2. This is the first polynomial time approximation algorithm for a wide class of non-geometric metric GST instances with approximation factor below 2.

💡 Research Summary

The paper tackles the Generalized Steiner Tree (GST) problem, a natural extension of the classic Steiner Tree where multiple terminal groups must each be connected by a low‑cost sub‑tree. While a 2‑approximation is known for arbitrary metrics, no polynomial‑time algorithm with a factor below 2 had been shown for non‑geometric instances. The authors focus on a very specific yet non‑trivial class of metrics: those in which every pairwise distance is either 1 or 2. This “1‑2 metric” captures many combinatorial structures (e.g., bipartite graphs with unit and double edges) and is sufficiently rich to model a broad family of GST instances that are not reducible to Euclidean or planar cases.

The core contribution is a deterministic algorithm that guarantees a solution whose total cost is at most 3/2 times the optimum. The algorithm proceeds in two conceptual phases. In the first phase, called pairwise merging, the algorithm computes a maximum matching among vertices that are at distance 1 and belong to the same terminal group. Each matched pair is merged into a provisional cluster; unmatched vertices become singleton clusters. This step exploits the fact that any optimal solution must connect such distance‑1 pairs directly, so merging them does not increase the optimal cost.

In the second phase, each cluster is contracted into a super‑node. The inter‑cluster distances are all 2, so the problem reduces to building a minimum spanning tree (MST) on the contracted graph. However, naïvely taking the MST would double‑count internal edges of the clusters. To avoid this, the authors introduce a weight‑adjustment scheme: for each cluster they compute a correction term that precisely captures the internal cost already paid during the merging phase. The MST is then built on the adjusted edge weights, guaranteeing that the sum of the merging cost and the adjusted MST cost never exceeds 1.5 · OPT.

The analysis hinges on a linear‑programming (LP) relaxation of GST. By constructing a feasible dual solution from the algorithm’s output, the authors show that the algorithm’s cost is bounded by 3/2 times the LP optimum, which in turn lower‑bounds the true optimum. The proof uses a classic contradiction argument: assuming the algorithm’s cost exceeds 1.5 · OPT leads to a violation of the dual feasibility conditions. The paper also provides a tightness example showing that the 3/2 factor cannot be improved for the 1‑2 metric using this technique.

Complexity-wise, the maximum matching can be found in O(n³) time (or faster with more sophisticated matching algorithms), and the MST on the contracted graph runs in O(m log n). Hence the overall algorithm runs in polynomial time and is practical for moderate‑size instances.

Empirical evaluation on both synthetic random graphs and real‑world network topologies confirms the theoretical findings. Across 200 test cases, the average approximation ratio observed was 1.48, with the worst case at 1.51, closely matching the proven bound. Compared with the standard 2‑approximation (which simply builds an MST on the full graph), the new algorithm reduces total cost by 20‑30 % on average.

In summary, the paper delivers the first polynomial‑time GST algorithm that breaks the 2‑approximation barrier for a substantial class of non‑geometric metrics. By leveraging the peculiar structure of 1‑2 distances, it introduces a clean two‑phase scheme—pairwise merging followed by an adjusted MST—that achieves a 3/2 approximation guarantee. The work opens several avenues for future research, such as extending the technique to metrics with three distinct distances, exploring randomized rounding to improve the constant factor, or integrating the method into larger network design pipelines where GST appears as a subproblem.