Are the edges themselves sampled independently?

In the basic model, positions are sampled and the distance rule determines edges; independent-edge sampling is a different model.

Must the space be a physical map?

No, but its distance must have a defensible meaning for the intended application.

Does a larger radius always improve every business outcome?

It adds potential connections on fixed points, but real cost, interference, and other objectives require additional modeling.

AI Glossary / Graphs and symbolic AI / Intermediate

Random geometric graph

Definition

A random geometric graph is a probabilistic graph model in which nodes occupy randomly chosen positions in a space and nearby nodes are connected. In the basic radius model, two distinct nodes share an edge when their distance is at most a specified threshold. The randomness comes from the positions; the edge rule then follows the geometry.

What it means

The model represents situations where proximity makes a relationship possible. Locations may be physical coordinates, such as sensor positions, or points in another meaningful metric space. A graph realization records one sampled arrangement and the connections produced by the chosen rule.
The phrase describes a family of models, not one universal distribution. To specify a model, state the region, how positions are sampled, the number of nodes, the distance measure, and the connection radius. A common example samples points independently and uniformly in a unit square. Other placement distributions can produce very different networks even with the same node count and radius.

How it works

First sample each node's position. Next compare pairs using the chosen metric and add an undirected edge whenever the distance meets the radius rule. The resulting graph can be analyzed for isolated nodes, connected components, paths, and other properties. Repeating the sampling reveals variation across possible networks.
Geometry links edge patterns: points close to a common neighbor are constrained in where they can lie. Thus this model is generally different from independently choosing every edge with a fixed probability. Boundary conditions matter too. A node near the edge of a square has less surrounding area inside the region than an interior node, which can affect its opportunity to find neighbors.

Three practical examples

A sensor-network study can simulate randomly placed devices and examine whether a simple distance-based communication model creates disconnected groups.
A graph-learning experiment can generate synthetic networks with spatial structure to test whether an algorithm behaves differently from an independent-edge baseline.
A planning exercise can explore how increasing a connection radius changes the number of links and the availability of routes across simulated locations.

Think of it this way

Think of placing pins on a map and connecting each pin to every other pin within a fixed compass radius. Random placement creates different maps, while the compass rule stays the same. The analogy is limited because real connections may depend on walls, terrain, direction, congestion, or equipment rather than on distance alone.

Worked example

Consider an original, deliberately chosen illustrative realization in a unit square. The four points are A = (0.1, 0.1), B = (0.4, 0.1), C = (0.4, 0.5), and D = (0.9, 0.9). Use Euclidean distance and connect distinct points at distance no greater than r = 0.45. The coordinates illustrate a possible configuration; they are not claimed to be a recorded random draw.
The A - B distance is 0.3 and B - C is 0.4, so both pairs are connected. A - C is sqrt(0.3² + 0.4²) = 0.5, so that pair is not connected. C - D is sqrt(0.5² + 0.4²), approximately 0.6403. B - D is approximately 0.9434 and A - D approximately 1.1314. None of these pairs meets the radius.
There are two edges and two connected components: {A, B, C} and {D}. A can reach C through B despite lacking a direct edge. D is isolated. Degrees are 1, 2, 1, and 0, giving average degree 4/4 = 1.
Increase the radius to 0.65 while keeping the same points. A - C and C - D now appear, giving four edges and one connected component. The average degree becomes 8/4 = 2. This comparison holds locations fixed so the change can be attributed to radius. Redrawing positions at the same time would mix two effects.

Illustrative case study

Fictional case: Rowan Field Lab studies a conceptual network of low-cost sensors scattered across a plot. The team first generates many random geometric graphs to explore which radii tend to leave isolated nodes under uniform placement.
It then compares the model with a more realistic placement pattern concentrated near access paths and with measured communication conditions. The simple radius model remains a baseline, while the differences reveal assumptions that matter. This fictional case does not claim that a simulated edge guarantees a working radio link or that the project achieved a particular coverage rate.

Common misconceptions and mistakes

Do not assume that a graph is connected because its average degree seems large. Several dense groups can remain separated. Nor does a path require every pair of nodes along the route's endpoints to be directly connected.
Random geometric graphs are not the same as nearest-neighbor graphs: a radius rule can give different nodes very different degrees. The chosen metric also changes the result. Increasing the radius on fixed points can add edges but cannot remove existing ones under the basic model. Claims about connectivity thresholds or computational cost need their assumptions; they should not be transplanted between dimensions, regions, and placement distributions without checking.

Business use

Businesses may use this model as a transparent starting point for studying spatially constrained networks. Its value is in exposing how placement and connection assumptions affect resilience or reachability before investing in a more detailed simulation.
Record both model settings and the sampled positions when reproducibility matters. Compare many realizations rather than treating one attractive diagram as representative. For applications involving real communication, transport, or service coverage, validate the relationship between distance and actual feasibility. The model can support a decision, but it does not supply that relationship automatically.

Check your understanding

At radius 0.45, can A reach C in the worked example?

Answer: Yes, through B. Their direct distance exceeds the radius, but the two shorter edges create a path.

Related terms

Graph
Metric space
Node degree
connected component
Erdős–Rényi graph
nearest-neighbour graph
spatial network
connectivity

Learn next

Reduce the example radius to 0.35 and list the components: {A, B}, {C}, and {D}. Then keep radius 0.45 but move D closer to C. Separating radius changes from placement changes clarifies what each modeling choice controls.

Sources

Sources: Jung and colleagues, The Aalto Dictionary of Machine Learning, random geometric graph entry: https://aaltodictionaryofml.github.io/ADictML_Main.pdf . NetworkX, random_geometric_graph documentation, model and metric parameters: https://networkx.org/documentation/stable/reference/generated/networkx.generators.geometric.random_geometric_graph.html .

Previous: Random-fuzzy variable Next: Random indexing