Minimum-Weight Steiner Triangulation of Convex Polygons Requires Interior Steiner Points

Abstract: We construct a convex polygon for which the minimum-weight Steiner triangulation requires an interior Steiner point. This provides a counterexample to a 1994 conjecture of Eppstein that minimum-weight Steiner triangulation of convex polygons needs only Steiner points on the boundary of the polygon.

• The paper constructs a 28-vertex convex polygon where the optimal boundary-only triangulation costs 49,425.5605, while one interior Steiner point reduces the cost to 49,425.4212.
• The authors combine exact dynamic programming, exponentially spaced polygon vertices, geometric rerouting lemmas, and finite case analysis to exclude every boundary Steiner-point configuration.
• The result disproves Eppstein’s 1994 conjecture that boundary Steiner points always suffice and shows that tree-dual dynamic programming cannot solve the general convex-polygon problem without a new framework.

The minimum-weight Steiner triangulation problem asks for a triangulation of a geometric input, with Steiner points permitted as additional vertices, minimizing total edge length. In "Minimum-Weight Steiner Triangulation of Convex Polygons Requires Interior Steiner Points" (2606.25302), Eppstein and Hadizadeh construct an explicit convex polygon whose optimal Steiner triangulation strictly requires at least one interior Steiner point. This disproves a 1994 conjecture of Eppstein that boundary Steiner points suffice for optimality on convex inputs [Eppstein94], and it eliminates the tree-dual dynamic-programming approach that the conjecture would have enabled.

Background and motivation

For convex polygons without Steiner points, minimum-weight triangulation is solvable in polynomial time by classical dynamic programming [Gil-RR-79, Kli-ADM-80]. With Steiner points allowed, Clarkson gave a logarithmic approximation [Cla-SODA-91], and Eppstein developed quadtree-based constant-factor approximations for both point sets and convex polygons [Eppstein94], later extended to polygons with holes [ChengLee02] and to three-dimensional minimum-area variants [ChengDey99]. Eppstein's polygon algorithm uses only boundary Steiner points, and he conjectured this restriction is without loss of optimality. The conjecture mattered structurally: triangulations using only boundary vertices have weak duals that are trees, which would potentially admit polynomial-time exact optimization via dynamic programming over the dual.

A necessary feature of any single-Steiner-point counterexample is high degree: a Steiner point of degree at most six can be contracted to its nearest neighbor, reducing total weight by the triangle inequality. The paper's strategy is therefore to build a polygon whose best non-Steiner triangulation has a high-degree vertex with neighbors at exponentially increasing distances, so that relocating those adjacencies through a nearby interior point shortens many edges simultaneously.

The counterexample

The polygon $P$ has 28 vertices: $v_0=(0,0)$, the extreme vertex $e=(16000,0)$, and 13 pairs of symmetric vertices $v_i=((2^i-i)/5,\,\pm(2^i+1))$ for $1\le i\le 13$. Two features drive the construction. On the left side, the $y$-coordinates grow exponentially while $x$ grows only linearly in $i$, producing a steep convex chain on which nonconsecutive diagonals are expensive; consequently the cheapest non-Steiner triangulation is the fan centered at $v_0$. Meanwhile, the distant extreme vertex $e$ creates long, shallow hull edges whose poor geometry prevents boundary Steiner points from exploiting the improvement available to an interior point.

The main theorem states three facts: (i) $v_0=(0,0)$0; (ii) $v_0=(0,0)$1; and (iii) $v_0=(0,0)$2. Part (i) is certified exactly by the standard convex-polygon dynamic program, which confirms the fan at $v_0=(0,0)$3 is optimal. Part (iii) is established by the explicit interior point $v_0=(0,0)$4: connecting $v_0=(0,0)$5 to all 28 vertices yields total length $v_0=(0,0)$6, strictly below the fan's cost. The margin between these values is small (roughly $v_0=(0,0)$7), so the separation rests on exact arithmetic rather than robust inequalities — a point worth noting when assessing numerical stability of the example.

Excluding boundary Steiner points

Part (ii) is the technical core. The authors must show no optimal boundary-only triangulation places any Steiner point anywhere on the hull. The argument proceeds by a sequence of exclusion lemmas, each exhibiting a local rerouting that strictly decreases cost whenever a boundary Steiner point exists in a given configuration.

Extreme edges. Three observations supply the quantitative backbone: distances between points on non-extreme hull edges satisfy $v_0=(0,0)$8 (a consequence of the $v_0=(0,0)$9 slope bound on the left chain); distances from such points to the upper extreme line are bounded below by $e=(16000,0)$0 (for $e=(16000,0)$1) or $e=(16000,0)$2 (for $e=(16000,0)$3); and every point on a non-extreme hull edge lies at distance between $e=(16000,0)$4 and $e=(16000,0)$5 from $e=(16000,0)$6, sharpening to at most $e=(16000,0)$7 when $e=(16000,0)$8. Using these bounds, the paper first rules out Steiner points on the short edge adjacent to each extreme edge, then rules out one-sided configurations on the extreme edges themselves, and finally eliminates all extreme-edge Steiner points via a finite case analysis. This last step relies on five structural reductions (bounding points per hull edge, extremal placement of neighbors, discarding low-$e=(16000,0)$9 auxiliary points, symmetry, and monotone location arguments) that reduce the analysis to finitely many threshold configurations, each verified numerically — for instance, case $v_i=((2^i-i)/5,\,\pm(2^i+1))$0 gives new cost $v_i=((2^i-i)/5,\,\pm(2^i+1))$1 versus old cost $v_i=((2^i-i)/5,\,\pm(2^i+1))$2, and case $v_i=((2^i-i)/5,\,\pm(2^i+1))$3 gives $v_i=((2^i-i)/5,\,\pm(2^i+1))$4 versus $v_i=((2^i-i)/5,\,\pm(2^i+1))$5.

Non-extreme edges. A key geometric lemma shows that rerouting an upper neighbor from a Steiner point $v_i=((2^i-i)/5,\,\pm(2^i+1))$6 to the upper endpoint $v_i=((2^i-i)/5,\,\pm(2^i+1))$7 gains at least $v_i=((2^i-i)/5,\,\pm(2^i+1))$8, derived from a law-of-cosines argument showing $v_i=((2^i-i)/5,\,\pm(2^i+1))$9 for the relevant angle. Combined with counting claims (at most one Steiner point per non-extreme edge, at most 13 neighbors per side, balanced neighbor counts when the extreme vertex is not adjacent), the paper proves that deleting any non-extreme-edge Steiner point and rerouting toward one endpoint always strictly decreases cost, in both balanced ($1\le i\le 13$0) and unbalanced cases, with a height dichotomy at $1\le i\le 13$1 handling the interaction with the extreme vertex. Additional lemmas exclude consecutive-edge placements and enforce a minimum degree of three beyond endpoints.

One methodological caveat applies throughout part (ii): several final comparisons are finite numerical verifications of specific extremal configurations rather than closed-form symbolic inequalities, so the correctness of part (ii) depends on the accuracy of those computed constants and on the reductions genuinely capturing the worst case.

Limitations and open questions

The result is existential and specific: it exhibits one 28-vertex polygon with a small but strict gap, and does not address how common such examples are, whether the gap can be made large, or whether interior Steiner points help by more than a single point in general. The authors note directly that the computational complexity of minimum-weight Steiner triangulation of convex polygons remains unresolved — it is unknown whether the problem is NP-hard or admits a polynomial-time algorithm once interior Steiner points are permitted. The disproof of the conjecture removes the tree-dual dynamic programming route, but no replacement algorithmic framework is proposed.

Conclusion

This paper settles a thirty-year-old structural question by explicit counterexample: there exists a convex polygon for which boundary Steiner points provide no improvement over the ordinary optimum, while a single well-placed interior Steiner point strictly reduces total edge length. The proof combines exact dynamic programming, a carefully engineered exponential-chain geometry, and an extensive case analysis excluding all boundary Steiner placements. Its principal consequence is negative but clarifying — optimal Steiner triangulations of convex polygons are inherently non-tree-structured — leaving the precise complexity status of the problem as its central open question.