System optimality versus self-organization in metro networks: An optimal transport analysis

Technical Summary: System Optimality versus Self-Organization in Metro Networks

Problem Statement
Urban transportation network design is traditionally approached through top-down planning grounded in engineering and economics, aiming for system-wide optimality (e.g., minimizing travel time or maximizing accessibility). However, cities are complex systems where infrastructure and mobility demand interact through feedback loops: network structure shapes origin-destination (OD) flows, while OD flows adapt and influence network development. Despite this interdependence, there is a lack of computational tools to benchmark real-world networks against alternative generative principles. A key open question remains whether real-world metro networks structurally resemble those implied by global system optimization or those emerging from bottom-up, self-organizing travel behaviors.

Methodology
The authors introduce a data-driven framework to compare three distinct metro network structures using geometric comparisons and optimal transport (OT) theory:

1. Empirical Network: The real-world Mass Rapid Transit (MRT) system.
2. Self-Organized (Desire-Path) Network: A minimal model derived from local rules reflecting human navigation tendencies. It assumes a preference for short travel times and the emergence of shared high-use corridors through trajectory bundling.
3. System-Optimal Network: A network derived from global optimization that minimizes aggregate travel time subject to an infrastructure cost constraint, treating link width (speed) as the design variable.

Model Construction Details:

• Data Source: The study utilizes anonymized mobile phone GPS data (CityData.ai, 2019) to infer general urban OD demand patterns across all modes, and aggregated MRT trip data (LTA, 2023) for the empirical network.
• Desire-Path Model: OD paths are initially modeled as Euclidean straight lines. These trajectories are then partitioned and subjected to a point-advection process using weighted Kernel Density Estimation (KDE). Intermediate points are iteratively displaced toward regions of higher trajectory density until convergence. Network nodes are extracted as local maxima of the density field, and edges are formed based on the frequency of transitions between nodes.
• System-Optimal Model: The study area is discretized into a regular grid. The model solves a non-linear optimization problem to assign link speeds (weights) that minimize total travel time for all OD pairs, constrained by a total infrastructure cost budget. This is solved using simulated annealing coupled with repeated shortest-path computations.
• Comparison Metrics:
  • Hausdorff Distance: Measures geometric similarity between the spatial sets of network nodes/edges.
  • Optimal Transport (OT): Quantifies the global discrepancy between networks by treating them as probability distributions of "network mass" (derived from node strength/weighted degree). The OT distance represents the minimal cost to transform the distribution of one network into another.

Case Study Application
The framework is applied to Singapore's MRT system.

• Geometric Results: The empirical MRT network exhibits a mean Hausdorff distance of 5.49 km from the desire-path benchmark, compared to 6.97 km from the system-optimal benchmark.
• Optimal Transport Results: The OT cost to align the empirical network with the desire-path benchmark is 1.39 km, whereas the cost to align it with the system-optimal benchmark is 2.12 km.
• Structural Observations: The empirical network is substantially closer to the desire-path benchmark. Notable deviations occur in the Central Catchment Nature Reserve, where environmental constraints prevent direct rail alignment. The system-optimal benchmark suggests direct connections across this reserve to reduce aggregate travel time, while the desire-path benchmark and the empirical network both route around it, reflecting the concentration of actual travel flows.

Key Contributions

1. Computational Framework: The paper provides a novel framework to benchmark empirical metro networks against self-organization and system optimality principles using OD flows derived from mobile phone data.
2. Quantification of Discrepancies: It quantifies structural differences that are not captured by conventional performance indicators alone, revealing that Singapore's MRT is structurally closer to a self-organized, demand-driven configuration than a globally optimized one.
3. Generalizability: The framework is designed to be transferable across regions and applicable to other infrastructure networks, enabling the diagnosis of design-use misalignments.

Significance and Claims
The authors claim that their findings highlight that system-wide optimality alone may be inadequate for guiding practical interventions. The empirical network's proximity to the desire-path benchmark suggests that real-world systems evolve through adaptive processes driven by local service needs and bottom-up dynamics rather than purely centralized optimization.

The paper posits that planning approaches should explicitly incorporate local service needs. While the system-optimal benchmark identifies potential efficiency gains (e.g., the Cross Island Line crossing the nature reserve), the desire-path approach offers a lower-risk strategy for network extensions as it aligns with existing travel patterns, reducing the likelihood of post-construction demand shifts. The framework serves as a diagnostic tool to identify connectivity gaps supported by both benchmarks (high-impact opportunities) versus those supported by only one, aiding in the prioritization of adaptive infrastructure planning.

Limitations and Future Work
The authors note that the framework is descriptive rather than causal; both benchmarks are "reverse-engineered" from observed OD patterns which are themselves shaped by the existing infrastructure. Future work could incorporate practical constraints (e.g., protected areas) directly into the models, expand the system-optimal formulation to include objectives like energy consumption and robustness, and explicitly model the co-evolution of demand and infrastructure through iterative counterfactual updates.