English

Approximate k-flat Nearest Neighbor Search

Computational Geometry 2014-11-07 v1 Data Structures and Algorithms

Abstract

Let kk be a nonnegative integer. In the approximate kk-flat nearest neighbor (kk-ANN) problem, we are given a set PRdP \subset \mathbb{R}^d of nn points in dd-dimensional space and a fixed approximation factor c>1c > 1. Our goal is to preprocess PP so that we can efficiently answer approximate kk-flat nearest neighbor queries: given a kk-flat FF, find a point in PP whose distance to FF is within a factor cc of the distance between FF and the closest point in PP. The case k=0k = 0 corresponds to the well-studied approximate nearest neighbor problem, for which a plethora of results are known, both in low and high dimensions. The case k=1k = 1 is called approximate line nearest neighbor. In this case, we are aware of only one provably efficient data structure, due to Andoni, Indyk, Krauthgamer, and Nguyen. For k2k \geq 2, we know of no previous results. We present the first efficient data structure that can handle approximate nearest neighbor queries for arbitrary kk. We use a data structure for 00-ANN-queries as a black box, and the performance depends on the parameters of the 00-ANN solution: suppose we have an 00-ANN structure with query time O(nρ)O(n^{\rho}) and space requirement O(n1+σ)O(n^{1+\sigma}), for ρ,σ>0\rho, \sigma > 0. Then we can answer kk-ANN queries in time O(nk/(k+1ρ)+t)O(n^{k/(k + 1 - \rho) + t}) and space O(n1+σk/(k+1ρ)+nlogO(1/t)n)O(n^{1+\sigma k/(k + 1 - \rho)} + n\log^{O(1/t)} n). Here, t>0t > 0 is an arbitrary constant and the OO-notation hides exponential factors in kk, 1/t1/t, and cc and polynomials in dd. Our new data structures also give an improvement in the space requirement over the previous result for 11-ANN: we can achieve near-linear space and sublinear query time, a further step towards practical applications where space constitutes the bottleneck.

Keywords

Cite

@article{arxiv.1411.1519,
  title  = {Approximate k-flat Nearest Neighbor Search},
  author = {Wolfgang Mulzer and Huy L. Nguyen and Paul Seiferth and Yannik Stein},
  journal= {arXiv preprint arXiv:1411.1519},
  year   = {2014}
}

Comments

22 pages, 1 figure

R2 v1 2026-06-22T06:49:40.874Z