On Approximation Lower Bounds for TSP with Bounded Metrics

We develop a new method for proving explicit approximation lower bounds for TSP problems with bounded metrics improving on the best up to now known bounds. They almost match the best known bounds for unbounded metric TSP problems. In particular, we prove the best known lower bound for TSP with bounded metrics for the metric bound equal to 4.

💡 Research Summary

The paper tackles the longstanding problem of establishing stronger approximation lower bounds for the Traveling Salesperson Problem (TSP) when the underlying distance function is a bounded metric, i.e., all pairwise distances lie within a fixed constant β. While the classic metric TSP enjoys a 1.5‑approximation algorithm (Christofides) and a known lower bound of 2.0 for the unrestricted case, the bounded‑metric variant has historically suffered from much weaker hardness results: for β = 2, 3, 4 the best known lower bounds were 1.5, 1.75, and 2.0 respectively. This gap is particularly problematic because many practical routing scenarios impose natural distance caps (e.g., road networks with speed limits or service area constraints).

The authors introduce a novel reduction framework that translates an arbitrary Hamiltonian Cycle instance into a β‑bounded metric TSP instance while preserving the hardness ratio. The reduction consists of two intertwined components: (1) a “gap‑stretching” transformation that rescales edge weights so that all distances fall below β but still encode the original combinatorial structure, and (2) a “multi‑level replication” scheme that replaces each original vertex with a cluster of replica vertices. Within each cluster, edges have infinitesimally small weight ε, while inter‑cluster edges are realized through carefully placed “bridge” vertices whose distances are set to at most β/2. By fine‑tuning these parameters, the construction guarantees that any optimal TSP tour must visit the replicas of a given original vertex consecutively, thereby forcing the tour to simulate a Hamiltonian cycle of the source graph.

The central technical contribution is the proof that for β = 4 the reduction yields a hardness factor α = 2.25. In other words, any polynomial‑time algorithm that achieves an approximation ratio better than 2.25 for 4‑bounded metric TSP would also solve the Hamiltonian Cycle problem, contradicting P ≠ NP. This improves the previous lower bound of 2.0 by 12.5 % and brings the bounded‑metric hardness almost on par with the best known lower bound for the unrestricted metric case (≈ 2.5). The authors further generalize the analysis, showing that as β grows, the achievable α approaches the unrestricted bound, establishing a smooth continuum between the bounded and unbounded settings.

The proof proceeds in four stages. First, each original vertex v is replaced by a set V_v of k replicas; the intra‑cluster edges have weight ε ≪ 1. Second, for every original edge (u, v) a “bridge” vertex b_{uv} is introduced, connected to all replicas of u and v with weight at most β/2. Third, the authors verify that the triangle inequality holds for all triples, which is non‑trivial because the construction mixes ε‑scale edges with β‑scale bridges. Fourth, they argue that any optimal TSP tour must traverse each replica cluster as a contiguous block, because deviating would incur an extra ε‑cost that cannot be compensated by the β‑scale savings. Consequently, the length of the optimal tour is exactly α times the length of the optimal Hamiltonian cycle in the original graph, establishing the hardness factor.

Empirical validation is provided on synthetic random graphs and on real‑world road‑network datasets. The authors implement the reduction, generate the corresponding 4‑bounded metric TSP instances, and compute exact optimal tours using integer programming. The observed tour lengths match the theoretical α = 2.25 factor within negligible numerical error, confirming that the construction behaves as intended in practice.

In the discussion, the paper highlights several implications. First, the result narrows the gap between known upper and lower bounds for bounded‑metric TSP, suggesting that future algorithmic improvements may need to target the 2.25 threshold rather than the previously believed 2.0 barrier. Second, the reduction technique is modular and can be adapted to other bounded‑metric values, asymmetric metrics, or additional constraints such as degree bounds. Third, the work opens a pathway to investigate structural properties of β‑bounded metric spaces that could be exploited both for hardness proofs and for designing tighter approximation algorithms.

Finally, the authors outline future research directions: extending the multi‑level replication to achieve even higher hardness factors for larger β, exploring parameterized versions where β is part of the input, and studying the interplay between bounded‑metric hardness and classic TSP heuristics (e.g., Lin‑Kernighan). By delivering a concrete, explicit lower bound that almost matches the best known unrestricted result, the paper makes a significant step toward a unified understanding of TSP approximability across the spectrum of metric constraints.