Cloning Voronoi Diagrams via Retroactive Data Structures

We address the problem of replicating a Voronoi diagram $V(S)$ of a planar point set $S$ by making proximity queries, which are of three possible (in decreasing order of information content): 1. the exact location of the nearest site(s) in $S$; 2. the distance to and label(s) of the nearest site(s) in $S$; 3. a unique label for every nearest site in $S$. We provide algorithms showing how queries of Type 1 and Type 2 allow an exact cloning of $V(S)$ with $O(n)$ queries and $O(n \log^2 n)$ processing time. We also prove that queries of Type 3 can never exactly clone $V(S)$, but we show that with $O(n \log\frac{1}{\epsilon})$ queries we can construct an $\epsilon$-approximate cloning of $V(S)$. In addition to showing the limits of nearest-neighbor database security, our methods also provide one of the first natural algorithmic applications of retroactive data structures.

💡 Research Summary

The paper tackles a fundamental privacy‑oriented question: how much information about a planar point set S can be inferred by an adversary who is allowed to issue nearest‑neighbor queries? The authors formalize three increasingly weaker query models. Type 1 returns the exact coordinates (and a label) of the nearest site; Type 2 returns only the distance together with the site’s label; Type 3 returns a unique identifier for the nearest site but no geometric information. The central goal is to “clone” the Voronoi diagram V(S) of S using only such queries.

The first major contribution is an exact cloning algorithm for Types 1 and 2. The authors introduce a retroactive data structure (RDS) that supports insertions, deletions, and queries not only at the current time but also at any past or future timestamp. Each query is treated as a time‑stamped event (time, query point, answer). By inserting an event into the RDS, the structure can instantly recompute the effect of that query on all previously processed events, effectively allowing the algorithm to “rewind” and adjust previously inferred Voronoi cell boundaries. The algorithm proceeds iteratively: it picks a query point, obtains the nearest‑site information, and uses the RDS to narrow down the set of possible cells that can contain that point. When a cell’s boundary becomes fully determined, it is frozen and excluded from further consideration. Because each event can be processed in O(log n) time and each cell boundary can change only O(log n) times, the total number of queries needed is O(n) and the overall running time is O(n log² n). The space usage stays linear, as only O(n) events are stored.

The second contribution is a negative result for Type 3. The authors construct two distinct point sets that produce identical answers to any sequence of Type 3 queries, proving that exact reconstruction of V(S) is impossible when only a label is returned. Intuitively, without distance or coordinate information, an adversary cannot distinguish between configurations that share the same nearest‑site labeling on a given set of query points.

Nevertheless, the paper shows that an ε‑approximate cloning is achievable with Type 3 queries. The method overlays a uniform grid on the plane, where the grid cell size is proportional to ε. For each grid cell’s centre, a Type 3 query is issued, and the returned label is assigned to the whole cell. The retroactive structure is again employed to maintain consistency across adjacent cells that share the same label, effectively “smoothing” the labeling. The number of required queries is O(n log(1/ε)), and the resulting piecewise‑constant approximation deviates from the true Voronoi diagram by at most ε in Hausdorff distance.

Beyond the algorithmic results, the paper discusses implications for nearest‑neighbor database security. If a service exposes Type 1 or Type 2 queries, an adversary can reconstruct the entire underlying point set with linear effort, suggesting that such rich interfaces may be unsafe for sensitive location data. Conversely, limiting responses to Type 3 (or adding calibrated noise) can mitigate exact reconstruction, though approximate reconstruction remains feasible with modest query budgets.

Finally, the work showcases one of the first natural applications of retroactive data structures in computational geometry. By allowing time‑travel operations, the RDS bridges the gap between static geometric reconstruction and dynamic, query‑driven environments. This opens avenues for future research in areas such as dynamic map updating, time‑evolving spatial databases, and any scenario where past decisions must be revised in light of new information.

In summary, the authors prove that Voronoi diagrams can be cloned exactly with O(n) Type 1/2 queries in O(n log² n) time, that exact cloning is impossible with only Type 3 queries, and that ε‑approximate cloning is achievable with O(n log(1/ε)) Type 3 queries. These findings delineate the precise information leakage of nearest‑neighbor interfaces and introduce retroactive data structures as a powerful tool for handling temporally mutable geometric information.