Graph-Based Nearest-Neighbor Search without the Spread

$\renewcommand{\Re}{\mathbb{R}}$Recent work showed how to construct nearest-neighbor graphs of linear size, on a given set $P$ of $n$ points in $\Re^d$, such that one can answer approximate nearest-neighbor queries in logarithmic time in the spread. Unfortunately, the spread might be unbounded in $n$, and an interesting theoretical question is how to remove the dependency on the spread. Here, we show how to construct an external linear-size data structure that, combined with the linear-size graph, allows us to answer ANN queries in logarithmic time in $n$.

💡 Research Summary

The paper addresses a fundamental limitation of recent graph‑based approximate nearest‑neighbor (ANN) data structures: their query time depends logarithmically on the spread Ψ of the data set (the ratio between the farthest and closest pair distances). Since Ψ can be exponential in the number of points n, this dependence is undesirable from a theoretical standpoint. The authors present a complete solution that eliminates the spread dependence while preserving linear space and achieving logarithmic query time in n.

The core of the approach is a two‑stage reduction. First, the authors show how to reduce the general, unbounded‑spread case to a bounded‑spread case using three auxiliary components: (i) a constant‑approximation ANN data structure that answers O(1)‑approximate queries in constant time, (ii) a low‑quality hierarchical well‑separated tree (HST) built on the point set P, and (iii) a collection of (1+ε)‑ANN graphs constructed on subsets of P whose spread is polynomially bounded. The HST provides a hierarchy of resolutions; at each level the points are clustered into subgraphs whose spread is small enough to apply existing ANN graph constructions (e.g., the greedy‑permutation graph of Har‑Peled et al.). This yields a family of “multi‑resolution” graphs, each supporting fast ANN queries but only within its own resolution.

The second stage introduces an external linear‑size data structure that allows the search to skip directly to the appropriate resolution. A constant‑approximation ANN query supplies a rough distance estimate p to the true nearest neighbor. Using the HST, the algorithm identifies the first point in the greedy permutation that lies within the resolution indicated by p. A “reverse tree” then moves a few steps backward in the permutation to ensure the search radius is sufficiently large. From this entry point, a slightly modified greedy‑permutation search (as in Har‑Peled et al.) proceeds, and the number of steps depends only on ε and the dimension d, not on the spread. Consequently, the overall query time becomes O(log n + ε⁻ᵈ·log (1/ε)), and the total space is linear, O(n).

The paper formalizes these ideas in two main theorems. Theorem 4.15 (Multi‑resolution graphs + Rough ANN) shows that with O(n·ε⁻ᵈ) space one can answer (1+ε)‑ANN queries in O(log n) time when the spread is polynomially bounded. Theorem 5.13 (Greedy permutation + Rough ANN) combines the greedy‑permutation graph with the external structure to achieve O(n) space and O(log n + ε⁻ᵈ·log (1/ε)) query time for arbitrary data sets, completely removing the log Ψ factor. An additional bootstrapping technique first obtains a coarse (ε = ½) ANN, then uses its result as a starting vertex for a finer‑grained search, further improving practical performance.

The authors also discuss how their reduction differs from the classic Indyk‑Motwani reduction, which relies on an approximate near‑neighbor decision structure that is not readily implementable with graph search. Their new reduction works directly with ANN structures and HSTs, making it applicable to Euclidean spaces and, more generally, to metrics with bounded doubling dimension.

In summary, the paper delivers a theoretically optimal solution for graph‑based ANN search: linear‑size space, logarithmic query time in the number of points, and no dependence on the spread. This advances both the theoretical understanding of graph‑based ANN methods and provides a practical framework for large‑scale, high‑dimensional nearest‑neighbor retrieval.