Imagine two points on a map. If you measure the distance between them using a straight line, you are calculating Euclidean distance—the "as-the-crow-flies" metric. This geometric simplification assumes a perfectly flat, unobstructed world where agents can move freely in any direction. However, humans are not crows. We do not fly over buildings, rivers, or highway dividers; our movement is strictly bound by physical infrastructure.
When spatial models rely on Euclidean distance, they introduce a fundamental bias. For example, two locations separated by a river might be geographically close but logically distant if the nearest bridge is miles away. As shown by Shaokun Zhang et al. in A Hybrid Model for Forecasting Traffic Flow: Using Layerwise Structure and Markov Transition Matrix, ignoring real-world constraints like one-way streets and geographical barriers leads to highly inaccurate spatial correlations. To build models that reflect reality, researchers across urban planning, logistics, and network science are shifting from coordinate-based geometry to network-constrained topology.
Projecting Space onto Graphs
To capture true proximity, spatial analysis must map entities directly onto network graphs where nodes represent intersections or decision points, and edges represent navigable paths.
This topological shift fundamentally changes how we evaluate similarity and interest in urban environments. In trajectory analysis, for instance, two vehicles might drive parallel to each other on opposite sides of a divided highway barrier. A Euclidean metric would classify their paths as nearly identical. However, Ziquan Fang et al. demonstrate in Spatio-Temporal Trajectory Similarity Learning in Road Networks that by integrating road network topology into trajectory embeddings, models can respect physical constraints like one-way restrictions, preventing disconnected paths from being incorrectly clustered together.
Similarly, when identifying urban features, evaluating points of interest (POIs) based on a simple radius can be misleading. Dimitrios Skoutas addresses this in Identifying and Describing Streets of Interest by mapping POIs directly to road network edges, calculating their density along actual navigable paths rather than across arbitrary 2D space. This ensures that urban discovery tools reflect what a pedestrian or driver actually experiences on the street.
The Human and Economic Cost of Geometric Simplification
The distinction between geometric distance and network distance is not just an academic nuance; it has profound implications for public health, resource allocation, and social equity.
Consider healthcare accessibility. Traditional metrics often use provider-to-population ratios or straight-line buffers to determine if a community is well-served. But as Bin Zhu et al. point out in Beyond distance: integrating economic burden into large-scale primary healthcare accessibility analysis, these methods fail to account for rugged terrain, infrastructure quality, and the actual time required to reach a doctor. When evaluating access to critical resources like pharmacies or vaccine centers, Inmaculada Hernandez et al. in Role of independent versus chain pharmacies in providing pharmacy access: a nationwide, individual-level geographic information systems analysis and Da Qi Chen et al. in Efficient and Equitable Deployment of Mobile Vaccine Distribution Centers emphasize that driving time and network topology provide a far more realistic measure of travel burdens than simple proximity.
When these network constraints are ignored, resource deserts become invisible. Jingchuan Guo et al. utilized network-based geographic information systems in Income disparities in driving distance to health care infrastructure in the United States: a geographic information systems analysis and Racial disparities in access to health care infrastructure across US counties: A geographic information systems analysis to uncover stark income and racial disparities in healthcare access across the United States, demonstrating that low-income and minority populations often face significantly longer real-world driving distances.
In emergency scenarios, this gap can be fatal. Wenxi Huang demonstrated a clear link between longer network-based driving distances to healthcare facilities and increased COVID-19 mortality rates in Association between distance to community health care facilities and COVID-19–related mortality across U.S. counties in the COVID-19–vaccine era. By switching to network-constrained modeling, planners can optimize emergency infrastructure—such as the placement of public-access naloxone kits modeled by Kwan Leung et al. in Optimizing placement of public-access naloxone kits using geospatial analytics: a modelling study—to maximize actual life-saving coverage along pedestrian and road paths.
Overcoming Computational Friction
If network-constrained distance is universally superior for spatial modeling, why has Euclidean distance persisted for so long? The answer comes down to computational complexity.
Calculating the Euclidean distance between two points requires a single, constant-time ($O(1)$) operation. In contrast, finding the shortest path along a complex road network requires graph traversal algorithms like Dijkstra's, which scale poorly when evaluating millions of pairs. This friction is particularly acute in spatial partitioning problems, such as generating Voronoi diagrams. Mehdi Sharifzadeh notes in Additively Weighted Voronoi Diagrams for Optimal Sequenced Route Queries that extending spatial partitioning logic to graph-based distances remains a significant challenge, requiring advanced geometric pre-computation to handle sequenced route queries efficiently. Furthermore, urban transit is rarely uniform; Andreas Gemsa et al. show in Higher Order City Voronoi Diagrams that modeling modern cities requires higher-order diagrams that blend grid-based walking paths with high-speed, node-restricted transit networks.
To bypass this computational bottleneck, researchers are developing hybrid spatiotemporal frameworks and deep learning heuristics. For example, Bi Yu Chen et al. introduced space-time buffering techniques in Toward space-time buffering for spatiotemporal proximity analysis of movement data to dramatically accelerate proximity analysis for moving objects restricted to road networks. In the domain of combinatorial optimization, Danhua Meng et al. developed EFormer in EFormer: An Effective Edge-based Transformer for Vehicle Routing Problems, an edge-based Transformer that solves vehicle routing problems by operating directly on non-Euclidean edge weights (such as traffic delays and one-way constraints) rather than raw node coordinates.
Generalizing Beyond Geography
The intuition behind network-constrained distance extends far beyond urban planning and geography; it applies to any system where interaction is governed by structural pathways rather than spatial proximity.
In neuroanatomy, for instance, estimating the functional connectivity of neural networks faces a similar challenge. Cells that are physically adjacent may not communicate if they lack direct synaptic wiring. Simona Ullo et al. demonstrated in Functional connectivity estimation over large networks at cellular resolution based on electrophysiological recordings and structural prior that measuring distance via heat propagation along physical axonal pathways provides a significantly better predictor of functional interaction than straight-line Euclidean distance.
Even in macro-level infrastructure design, such as generating synthetic power grids, Rounak Meyur et al. showed in Creating Realistic Power Distribution Networks using Interdependent Road Infrastructure that constraining power line optimization to follow existing road networks yields highly accurate "digital twins" of utility infrastructure, because utility lines naturally follow physical rights-of-way.
Whether mapping the synapses of the brain, the distribution of emergency medical supplies, or the flow of urban traffic, the core lesson of this research direction is clear: space is defined not by coordinates, but by connections. By replacing the simplification of Euclidean distance with network-constrained topology, models gain the fidelity required to solve real-world problems.
Key Foundational Work
Computational data modeling for network-constrained moving objects (2003) — contributed 5 concept elements to this direction.