Hierarchical Routing Structure

• Hierarchical routing structure is a multi-level network overlay that organizes spatial graphs into distinct regions with hubs, corridors, feeder paths, and express metro links to realistically model mobility flows.
• It employs spatial discretization, radial organization from a reference center, and weighted, bidirectional transition dynamics to simulate movement from local neighborhoods to central hubs.
• This framework supports privacy-preserving synthetic trajectory generation and aggregate flow modeling, providing analytical invariants and practical insights for urban mobility research.

A hierarchical routing structure is a network overlay designed to impose multi-level organization onto a base spatial graph, enabling efficient, structured, and realistic modeling of flows and trajectories in human mobility systems. Such structures introduce concepts like regional hubs, corridor backbones, feeder paths, and express metro links, each serving distinct functional and topological roles in mediating movement across scales from local neighborhoods to city centers. This multiscale partitioning of the movement space is central to network-driven generative mobility models and underpins current methods for synthetic trajectory generation, aggregate flow modeling, and privacy-preserving simulations.

1. Construction of Hierarchical Routing Overlays

The construction of a hierarchical routing structure begins with spatial discretization, typically using a uniform grid (e.g., H3 hexagonal tessellation), where each cell centroid becomes a node in the base graph. Immediate neighbors are linked by undirected edges, establishing local contiguity. The hierarchical overlay is then built through the following procedure (Meyer et al., 9 Jan 2026):

• Reference Center: An absolute (e.g., geometric centroid) or functional (e.g., city center) node $c$ is chosen as a root for radial organization.
• Hubs: Nodes are partitioned into radial bands around $c$. Within each band, a small number of well-separated nodes are selected as hubs, acting as aggregators or transfer points.
• Corridor Backbone: For every hub $h$, the unique shortest path to $c$ is computed and all such paths collectively form the backbone corridor structure, defining main arterials.
• Feeder Paths: Non-hub nodes $i$ are mapped to their nearest hub via shortest path; these paths, labeled as feeders, serve to collect and distribute movement between peripheral areas and their corresponding hubs.
• Metro Links: Each hub is additionally connected by an express edge to the nearest hub that is closer to $c$, yielding a network of long-range, high-capacity bypasses.
• Symmetrization: All overlay edges are bidirectional to guarantee strong connectivity.

Each node is assigned a scalar potential $\Psi_i$ equal to its graph distance from $c$, which enables definition of edge directionality (inward, outward, neutral) via $\sigma_{ij}=\mathrm{sign}(\Psi_j-\Psi_i)$ for each overlay edge $j\to i$.

2. Integration With Network-Driven Dynamics

The routing hierarchy is directly integrated into the mobility model through weighted, time-dependent transition matrices $P(t)$, which describe the probability of moving between any pair of nodes at each discrete time step. The hierarchical structure enters the kernel via:

• Overlay Weights $\omega_{ij}$: Edges are attributed static weights according to their type (e.g., backbone, feeder, metro), typically with backbone/metro links given higher probability mass to reflect their role as main routes.
• Distance Decay: The transition kernel is modulated by a gravity-type factor inversely proportional to distance $d_{ij}^\beta$, with $\beta>0$, penalizing long-range moves except along high-weighted corridors or metro links.
• Directional Bias: Edge orientation $\sigma_{ij}$ is used to apply time-varying bias schedules $\phi_{ij}(t)$, enforcing, for example, strong centripetal flows during morning (commuting) intervals and centrifugal patterns in the evening.
• Population Mass and Temporal Schedules: Site-specific destination and origin “masses” $m_i(t), m_j(t)$ allow time-of-day dependent boosting of central, hub, or peripheral nodes.

Normalization ensures each column of $P(t)$ sums to one, and stay probabilities $s_j(t)$ permit agents to remain in place.

3. Markovian Trajectory Generation on Hierarchical Networks

Given the hierarchical routing structure and time-dependent transition dynamics, synthetic mobility trajectories are generated under a memoryless Markovian protocol:

• A finite population $K$ of indistinguishable “privacy-enhanced persons” (PEPs), initialized according to a population vector $\mathbf p_0$, is propagated over the network by sequentially sampling the next location for each PEP from the appropriate column of $P(t)$.
• At each time step, only the current node influences the transition — there is no explicit memory or agent-level behavioral rule.
• Aggregated origin-destination (OD) matrices over any interval reflect the empirical counts of these categorical jumps.

This protocol naturally aligns realized trajectory ensembles with the prescribed Markov dynamics, with any discrepancy (e.g., in multi-step transfer matrices) attributable solely to finite sampling noise.

4. Analytical Properties: Invariants and Consistency

The main analytical features imposed by the hierarchical routing structure are:

• Periodic Invariant Distribution: The daily dynamical operator $Q = P(T)\cdots P(1)$ is column-stochastic, irreducible, and aperiodic, ensuring a unique strictly positive invariant vector $\pi$ via Perron–Frobenius theory, satisfying $Q\pi=\pi$.
• Structural Stationarity: The invariant $\pi$ is entirely determined by the routing hierarchy, the transition kernel, and the imposed schedule biases, and is independent of agent-level heterogeneity or historical path dependencies, reflecting a stable macroscopic reference state (Meyer et al., 9 Jan 2026).
• Validation: Direct comparison between the analytical product of transition matrices and the empirically estimated aggregates demonstrates convergence rate $\propto K^{-1/2}$ — finite-population sampling error — and no systemic model bias.

5. Privacy and Modeling Advantages

The hierarchical routing framework possesses intrinsic privacy-preserving properties:

• There is no retention of agent identity, no stored paths, and no individual-level model calibration; the system relies purely on exogenously prescribed network and macro-level schedule inputs.
• Only aggregate flows and transition kernels are ever imposed or reproduced; individual trajectories are uniquely synthetic and cannot be used for trajectory-recovery or re-identification attacks.
• The model is thus immune to attacks prevalent in data-driven mobility simulation frameworks, while still supporting rigorous calibration and empirical realism at the population level.

6. Significance in Relation to Classical and Contemporary Mobility Models

The hierarchical routing structure diverges meaningfully from both classical unconstrained (e.g., fully connected, gravity-based) and purely opportunity-driven models:

• Compared to Gravity Models: Standard gravity models represent OD flows as $T_{ij} \propto N_i N_j / d_{ij}^\gamma$ over a flat connectivity landscape. The hierarchical overlay introduces restricted high-conductance corridors and multilevel paths, capturing real-world constraints such as multi-modal transfers, express connections, and differentiated arterial/backroad usage (Barbosa-Filho et al., 2017).
• Opposed to Purely Local Exploration/Preference Models: The absence of individual memory in the Markovian system sidesteps the exploration–preferential return (EPR) paradigm (Pappalardo et al., 2016), yet, via network structure and schedule, reproduces characteristic flows and hubs.
• Contrast to Deep Learning or GAN-based Approaches: Hierarchically structured kernels enable interpretability and exogenous control absent from black-box neural generators (e.g., MoGAN (Mauro et al., 2022), PMT (Wu et al., 2024)), while preserving privacy by not leveraging agent histories.

This approach is thus uniquely suited to large-scale, privacy-sensitive trajectory synthesis, policy counterfactuals, and the study of spatial flow patterns under exogenously defined network constraints.

7. Extensions and Future Directions

Potential directions for extending hierarchical routing structures in human mobility modeling include:

• Integration with Deep Generative Priors: Incorporation of transformer-based or diffusion models conditioned on routing hierarchies for high-dimensional behavioral fidelity beyond memoryless transitions (Hong et al., 7 Oct 2025).
• Adaptive Schedules: Real-time modulation of corridor capacities and bias schedules responding to observed or simulated congestion patterns.
• Multilayer and Multimodal Networks: Embedding additional routing logic for multimodal integration (e.g., nested layers for bus, rail, pedestrian) and cross-layer transfers, reflecting true hierarchical urban mobility.
• Urban Design and Epidemics: Application of hierarchical constraints in the simulation of evacuation, public health interventions, or the design of sustainable urban transportation systems.
• Analytical Study of Hierarchy Effects: Theoretical and empirical assessment of how hierarchy depth, hub density, and corridor configuration affect macroscopic observables such as mean travel time, flow heterogeneity, and system resilience.

In summary, hierarchical routing structures offer a formally grounded, rigorously analyzable, and empirically validated framework for modeling, synthesizing, and understanding the organization of mobility flows on complex spatial networks, while maintaining strong privacy guarantees and interpretability (Meyer et al., 9 Jan 2026).