A lower bound for metric 1-median selection

Consider the problem of finding a point in an n-point metric space with the minimum average distance to all points. We show that this problem has no deterministic $o(n^2)$-query $(4-\Omega(1))$-approximation algorithms.

💡 Research Summary

The paper investigates the metric 1‑median problem—finding a point in an n‑point metric space that minimizes the average distance to all other points—under an oracle query model. In this model an algorithm may query the distance d(x, y) for any unordered pair (x, y) but may not query a point with itself and may not repeat a query. The authors consider any deterministic algorithm A that makes only o(n²) distance queries and prove that such an algorithm cannot achieve a (4 − ε)‑approximation for any constant ε > 0.

The proof proceeds by constructing an adversarial metric space that is revealed gradually as the algorithm asks queries. The construction uses a small set S of size ⌈δn⌉ (with a constant δ ∈ (0,0.1)) and a set B of vertices whose degree (the number of incident queried edges) exceeds δn. For each query (x_i, y_i) the adversary “freezes’’ the distance according to a case analysis (Equation 5). The frozen value is chosen from {1,2,3,4} based on whether the endpoints lie in S, whether they are already highly connected, and the current degrees α_i(·). After all q(n)=o(n²) queries have been answered, the remaining undefined distances are set by another case analysis (Equation 8), which ensures symmetry and that every distance is still in {1,2,3,4}.

A series of lemmas (2‑14) verify that the resulting function d satisfies all metric axioms, especially the triangle inequality. The point (\hat p) is defined as the vertex in S with the smallest degree after all queries; it turns out to be an excellent 1‑median: most vertices are at distance 1 from (\hat p), and the total cost (\sum_x d(\hat p,x)) is at most (n + O(δn)). In contrast, the algorithm’s output p must be a vertex whose degree eventually becomes n − 1 (since the algorithm queries all edges incident to p). Lemma 16 shows that for almost all vertices z_k(t) that become adjacent to p after a certain point, the distance d(p, z_k(t)) equals 4. Consequently Lemma 17 yields (\sum_x d(p,x) \ge 4 (n - 2⌈δn⌉ - 2)).

Comparing the two costs gives
\