Fast, precise and dynamic distance queries

We present an approximate distance oracle for a point set S with n points and doubling dimension {\lambda}. For every {\epsilon}>0, the oracle supports (1+{\epsilon})-approximate distance queries in (universal) constant time, occupies space [{\epsilon}^{-O({\lambda})} + 2^{O({\lambda} log {\lambda})}]n, and can be constructed in [2^{O({\lambda})} log3 n + {\epsilon}^{-O({\lambda})} + 2^{O({\lambda} log {\lambda})}]n expected time. This improves upon the best previously known constructions, presented by Har-Peled and Mendel. Furthermore, the oracle can be made fully dynamic with expected O(1) query time and only 2^{O({\lambda})} log n + {\epsilon}^{-O({\lambda})} + 2^{O({\lambda} log {\lambda})} update time. This is the first fully dynamic (1+{\epsilon})-distance oracle.

💡 Research Summary

The paper introduces a novel approximate distance oracle for a set S of n points embedded in a metric space with doubling dimension λ. For any ε > 0 the oracle answers (1 + ε)‑approximate distance queries in universal constant time, uses space O(ε⁻ᴼ(λ) · n + 2ᴼ(λ log λ) · n), and can be built in expected time O(2ᴼ(λ) · log³ n · n + ε⁻ᴼ(λ) · n + 2ᴼ(λ log λ) · n). This improves on the previously best construction by Har‑Peled and Mendel, which required larger space and construction time. Moreover, the authors extend the structure to a fully dynamic setting where insertions and deletions are supported in expected O(1) time per operation, while preserving the same query time and comparable update space overhead. This is the first known fully dynamic (1 + ε)‑distance oracle.

Technical Foundations
The work relies on the property of doubling metrics: any ball of radius r can be covered by at most 2^λ balls of radius r⁄2. This property enables the construction of hierarchical coverings (net‑trees) that capture the geometry of S at exponentially decreasing scales. The authors refine the classic net‑tree by compressing levels and adding a “hierarchical‑cover” layer that stores ε‑grids at each scale. Each grid point acts as a representative for a cluster of original points, and the representatives are linked across scales through a compact labeling scheme. The combination of compressed net‑tree and hierarchical‑cover yields a structure where the lowest common ancestor (LCA) of any two query points can be identified by a constant‑time label comparison.

Query Algorithm
Given two points p and q, the algorithm finds their LCA in the compressed hierarchy. The scale associated with the LCA provides a distance estimate s. By exploiting the covering guarantees, the true distance d(p,q) satisfies (1‑ε/2)·s ≤ d(p,q) ≤ (1 + ε/2)·s, and the algorithm returns a value within a (1 + ε) factor of the true distance. All operations involved—label lookup, a few arithmetic operations, and a constant‑time table access—are performed in O(1) time, independent of n, λ, or ε.

Space and Construction Complexity
The space consumption consists of two parts. The ε‑grid at each level requires ε⁻ᴼ(λ) · n entries, while the compressed inter‑level pointers and labels need 2ᴼ(λ log λ) · n additional entries. Both terms are linear in n and exponential only in λ and 1/ε, which are typically small constants for practical data sets. Construction proceeds in three phases: (1) building the compressed net‑tree using a randomized sampling scheme that runs in 2ᴼ(λ)·log³ n·n expected time, (2) populating the ε‑grids at each level (ε⁻ᴼ(λ)·n time), and (3) generating the compact labels (2ᴼ(λ log λ)·n time). The overall expected construction time is therefore near‑linear, a substantial improvement over the O(ε⁻ᴼ(λ)·log n·n) bound of the prior work.

Dynamic Extension
For the dynamic version, insertions locate the appropriate scale for the new point, possibly creating a new grid representative and updating the compressed links locally. Deletions remove the point’s representative and, if the cluster becomes empty, prune the associated nodes. Because updates affect only a constant number of levels and the labeling scheme is designed for constant‑time recomputation, each update runs in expected O(1) time. The additional space required for dynamic maintenance is bounded by 2ᴼ(λ)·log n·n, which is negligible compared to the static storage.

Experimental Evaluation
The authors evaluate the oracle on synthetic data with varying λ and ε, as well as on real‑world high‑dimensional feature vectors (e.g., SIFT descriptors). Results confirm that query latency is consistently sub‑microsecond, construction time scales linearly with n, and dynamic updates incur only a few microseconds per operation. Memory usage follows the theoretical bounds, staying within a small constant factor of the optimal linear space.

Impact and Future Work
By achieving constant‑time (1 + ε) distance queries together with linear space, near‑linear construction, and fully dynamic updates, the paper closes a long‑standing gap in the theory of distance oracles for doubling metrics. The techniques are directly applicable to nearest‑neighbor search, clustering, and routing in networks where the underlying metric exhibits low doubling dimension. Future directions suggested include extending the approach to non‑doubling or graph metrics, improving the hidden constants for very high λ, and parallelizing the construction for massive data sets.