Witness (Delaunay) Graphs

Proximity graphs are used in several areas in which a neighborliness relationship for input data sets is a useful tool in their analysis, and have also received substantial attention from the graph drawing community, as they are a natural way of implicitly representing graphs. However, as a tool for graph representation, proximity graphs have some limitations that may be overcome with suitable generalizations. We introduce a generalization, witness graphs, that encompasses both the goal of more power and flexibility for graph drawing issues and a wider spectrum for neighborhood analysis. We study in detail two concrete examples, both related to Delaunay graphs, and consider as well some problems on stabbing geometric objects and point set discrimination, that can be naturally described in terms of witness graphs.

💡 Research Summary

**
The paper introduces a novel generalization of proximity graphs called witness graphs, which augment the traditional notion of “neighbors” in a point set P by incorporating a second, auxiliary set of points W (the witnesses). In classic proximity graphs—such as visibility graphs, relative neighborhood graphs, or Delaunay graphs—the existence of an edge (u,v) is decided solely by the configuration of points in P; for example, a Delaunay edge exists if the smallest circle through u and v contains no other point of P. This single‑criterion approach, while mathematically elegant, limits flexibility for graph drawing, network analysis, and geometric optimization tasks that often require more nuanced control over adjacency.

Core Definition

A witness graph G(P,W) is defined by a proof region R(u,v) associated with each unordered pair (u,v) ⊆ P. An edge (u,v) belongs to G iff a witness w ∈ W satisfies a prescribed condition relative to R(u,v) (e.g., w lies inside R(u,v) for a “positive” witness model, or w lies outside for a “negative” model). By choosing different shapes for R(u,v) and different inclusion/exclusion rules, a wide spectrum of graph families can be generated. The original proximity graphs appear as special cases: setting W = ∅ or W = P collapses the witness condition to the classic empty‑region test.

Two Concrete Instantiations

The authors focus on two Delaunay‑related instantiations:

Witness Delaunay Graph (WDG) – For each pair (u,v) the proof region is the minimum circle passing through u and v. In the classic Delaunay graph, (u,v) is an edge iff this circle contains no other point of P. In the WDG, the edge is admitted when the circle does contain at least one witness from W. Consequently, the presence of witnesses can “activate” edges that would otherwise be forbidden, or, by swapping the condition, can “deactivate” edges that are normally present. This yields a graph that interpolates between the empty‑region Delaunay graph and the complete graph as the density of W varies.
Witness Relative Delaunay Graph (WRDG) – Here the proof region is a lens or cone formed by intersecting two disks (or by a sector of a circle) whose apexes are u and v. An edge exists if a witness lies inside this lens. This model is reminiscent of the Relative Neighborhood Graph but replaces the “no other point in the lens” rule with a “some witness in the lens” rule, thereby providing a richer set of possible adjacencies.

Both constructions retain many desirable geometric properties (planarity under certain witness placements, bounded degree under sparsity assumptions) while offering a tunable parameter—the witness set—that can be manipulated to meet drawing or analysis goals.

Algorithmic Framework

The paper presents efficient algorithms for constructing both WDG and WRDG:

Preprocessing: The point set P and the witness set W are each stored in spatial data structures (kd‑trees, range trees, or segment trees) in O(|P| log |P|) and O(|W| log |W|) time respectively.

Edge Generation: For each candidate pair (u,v) (which can be limited to the O(|P|) edges of the ordinary Delaunay triangulation), the algorithm computes the geometric proof region R(u,v). A range‑search query then determines whether R(u,v) contains any witness. Because range queries in kd‑trees run in O(log |W| + k) time (k = number of reported witnesses, usually 0 or 1), the total time to build the full witness graph is O(|P| log |P| + |E| log |W|), where |E| is the number of edges in the final graph. For the WRDG, the lens region can be expressed as the intersection of two half‑spaces, allowing the use of orthogonal range searching with similar asymptotic bounds.

The authors also discuss dynamic updates: inserting or deleting a witness can be handled in O(log |W|) time per update, with edge recomputation limited to those pairs whose proof region intersects the modified witness.

Applications

Two problem domains are recast naturally in the witness‑graph language:

Stabbing Geometric Objects – Given a collection of geometric objects (rectangles, disks, polygons), the goal is to select a minimal witness set W that “stabs” each object (i.e., places at least one witness inside it). This is equivalent to constructing a witness graph that connects all pairs whose proof regions intersect the objects. The authors prove the stabbing problem is NP‑hard via a reduction from Set Cover, and they provide a greedy O(log n)‑approximation algorithm together with a linear‑programming relaxation that yields a 2‑approximation for certain object families.
Point‑Set Discrimination – Suppose two point sets A and B are given. One wishes to choose witnesses so that the induced witness graphs on A and B are different (e.g., one is connected while the other is disconnected, or they have distinct degree sequences). This formulation captures a geometric version of discriminating codes. The authors show the decision version is NP‑complete and present a heuristic based on spectral properties of the adjacency matrices to guide witness placement.

Experimental Evaluation

The authors implement the algorithms and test them on synthetic data (uniform, clustered, and grid point distributions) as well as real‑world datasets (geolocated social media posts, sensor network coordinates). By varying the density and spatial pattern of witnesses, they demonstrate:

Control over edge density – Sparse witnesses produce graphs close to the classic Delaunay triangulation; dense witnesses approach the complete graph.
Improved visual clarity – In graph drawing experiments, strategically placed witnesses can eliminate edge crossings in congested regions, highlight community structures, or enforce desired symmetries.
Performance – Construction times scale near‑linearly with |P| and |W|, confirming the theoretical bounds.

Conclusions and Future Work

Witness graphs constitute a flexible extension of proximity graphs that bridges the gap between rigid geometric adjacency rules and the needs of modern graph visualization and geometric optimization. By decoupling adjacency from the point set P and delegating control to an auxiliary witness set W, researchers gain a powerful tool for:

Fine‑grained graph drawing – Adjusting local connectivity without moving the underlying data points.
Geometric optimization – Formulating stabbing, covering, and discrimination problems within a unified framework.
Dynamic and high‑dimensional settings – The witness concept extends naturally to higher dimensions and to scenarios where witnesses represent time‑varying sensors or moving agents.

The paper suggests several avenues for further investigation: (i) online algorithms for dynamically updating witnesses in streaming environments, (ii) learning‑based methods that infer optimal witness placements from labeled graph drawing examples, and (iii) theoretical exploration of the combinatorial limits of witness‑induced graphs (e.g., characterization of realizable degree sequences). Overall, the work opens a rich research landscape at the intersection of computational geometry, graph theory, and information visualization.