Optimal Planar $k$ Nearest Neighbors Queries for Additively Weighted Euclidean Distances and Convex Metrics

Abstract

Given a set $P$ of $n$ point sites in the plane, the $k$ nearest neighbors problem is to build a data structure that, for a query point $v$ and a query integer $k$, reports the $k$ nearest point sites of $v$ in $P$ efficiently. For the Euclidean metric, after four decades of efforts, an optimal data structure has been developed with $O(n)$ space, $O(\log n+k)$ query time and $O(n\log n)$ preprocessing time [Mat92b, Cha00, AC09, CT16]. In contrast, for a family of distance functions, the state-of-the-art data structure [AES99, KMR+20, Liu22] attains $O(n\log\log n)$ space, $O(\log n+k)$ query time and expected $O(n\log n)$ preprocessing time. An open question is the existence of an optimal data structure with $O(n)$ space and $O(\log n + k)$ query time for distance functions more general than the Euclidean metric. Such an optimal data structure also remains unknown for non-point sites in the Euclidean metric, even for disk sites or, more generally, additively weighted Euclidean distances. We develop an optimal $k$-nearest-neighbors data structure for additively weighted Euclidean distances with $O(n)$ space, $O(\log n + k)$ query time and expected $O(n\log n)$ preprocessing time. Since this data structure can be adjusted to disk sites, it serves as an initial result toward optimality for non-point sites. We also adapt our data structure to point sites under a family of convex metrics, including commonly used $L_p$ metrics and uniform orientation metrics. The data structure can be further extended to distance functions that satisfy certain axioms. Thus, we demonstrate the existence of optimal $k$-nearest-neighbors data structures for distance functions more general than point sites in the Euclidean metric. Our main technical advance is to generalize Matoušek's shallow partition theorem for hyperplanes [Mat92b] to these settings, while achieving a logarithmic crossing number.