Greedy Routing in Decentralized Networks

Updated 6 February 2026

Greedy routing networks are decentralized systems where nodes make local forwarding decisions based on proximity to a destination, ensuring efficient packet delivery.
They employ geometric and combinatorial frameworks—such as Delaunay graphs and Schnyder drawings—to guarantee navigability and scalability across diverse domains.
Dynamic models of throughput, congestion, and epidemic spread, along with game-theoretic incentives, enhance network robustness and practical applicability.

Greedy routing networks are decentralized communication substrates in which all forwarding decisions are made locally and myopically: each node, given a packet and the target’s coordinates (or virtual identifier), transmits the packet to the neighboring node closest to the destination according to a specified metric. This paradigm underpins efficient navigation, scalability, and robustness in large-scale, mobile, or ad hoc networks across diverse domains, including wireless sensor fields, vehicular networks, Internet overlays, and even brain connectomics. The core structural property is a greedy-supporting topology (a “greedy embedding” or combinatorial equivalent), guaranteeing that local, one-hop-at-a-time choices suffice to deliver packets efficiently and reliably to their targets. This article surveys fundamental definitions, the geometric and combinatorial conditions for perfect greedy routing, analytical models of emergent throughput and congestion dynamics, and incentive-compatible frameworks for self-organizing navigable systems.

1. Formal Model and Structural Conditions

Let $G=(V,E)$ be a graph, with each node $v \in V$ assigned a geometric coordinate $p(v)\in\mathbb{R}^2$ , and $d(u,v)$ the metric distance. The canonical greedy routing protocol forwards any packet at $u$ for destination $x$ to the neighbor $w\in N(u)$ minimizing $d(w,x)$ . Greedy routing succeeds if, for any source–target pair $(s,x)$ , repeated local minimization eventually delivers the packet to the node $v^*=\arg\min_{v\in V} d(p(v),x)$ .

Necessary and sufficient conditions for perfect geometric greedy routing (i.e., success for all targets, never getting stuck at local minima away from $v^*$ ) are captured by the containment of the undirected Delaunay graph $\mathrm{DT}(V)$ as a subgraph. Specifically, greedy routing perfectly succeeds on $G$ if and only if, for every $v$ , the vertex region

$\mathrm{VR}_G(v) = \{ x \in \mathbb{R}^2 : d(x, v) \leq d(x, w)\ \forall w \in N(v) \}$

is exactly the Voronoi cell $\mathrm{VC}(v)$ . This holds precisely when $E$ contains every non-degenerate Delaunay edge, i.e., greedy forwarding always has an “escape” towards the true closest node unless already at the target (0903.5208).

The construction and maintenance of greedy-supporting substrates is algorithmically tractable for planar graphs and can be dynamically maintained in mobile agents with geometric position tracking.

2. Dynamical Traffic and Epidemic Models

Greedy routing on mobile agent networks exhibits intricate transport and congestion phenomena. In spatial models with $N$ agents moving at speed $v$ in a domain of area $L^2$ (density $\rho$ ) and communication radius $r$ , each agent forwards up to $C$ packets per time step; packets are generated at rate $R$ with random source–destination pairs. Greedy forwarding proceeds as follows:

If the destination is within range ( $d(i,j)\leq r$ ), direct delivery occurs.
Otherwise, the next hop $k^*\in \mathcal N_i$ is selected to minimize $d(k,j)$ .

The order parameter for congestion is

$\eta(R) = \lim_{t\to\infty} \frac{C}{R}\cdot\frac{\langle \Delta N_p \rangle}{\Delta t}\ ,$

with a critical injection rate $R_c$ signifying the onset of congestion (above which packet accumulation diverges). Notably, $R_c$ increases monotonically with communication radius $r$ , and is maximized at intermediate agent speeds ( $v\approx 0.1$ for specific units), as slow or very fast topology updates reduce capacity. Travel time $\langle T \rangle$ decreases with $r$ and increases rapidly if $v\gg 0.1$ (Yang et al., 2011).

In coupled epidemic-traffic dynamics (SIS model, infinite delivery), the infection threshold is

$\beta_c = \frac{N}{R\,\langle T \rangle}\ .$

Increasing $r$ or operating at optimal $v$ increases $\beta_c$ , raising robustness to infection spread. In congested regimes ( $C$ finite, $R\gg R_c$ ), $\beta_c$ saturates at $1/C$, showing that overload bottlenecks can suppress epidemic propagation.

3. Greedy Embeddings and Planar Graph Models

Not all graphs admit geometric greedy embeddings, but several combinatorial conditions guarantee such embeddings for wide classes. Maximal planar (triangulated) graphs can always be represented by Schnyder drawings—assigning to each node barycentric coordinates in $\mathbb{R}^3$ based on directed tree paths—that guarantee local progress under greedy routing in the induced virtual coordinate system (Leone et al., 2016). For 3-connected planar graphs, succinct (i.e., $O(\log n)$ -bit) coordinate assignments exist for which local distance comparisons suffice for greedy forwarding, with linear preprocessing time and minimal per-hop computation (0812.3893, Leone et al., 2015).

Combinatorial greedy routing can be realized without explicit geometric distances. By exploiting Schnyder’s standard $3$-dimensional representations (three compatible total orders), greedy paths can be fully described and computed using local rank-majorization and sector decomposition, supporting compact routing with guaranteed delivery in $O(\log n)$ space per node (Leone et al., 2015).

4. Large-Scale, Small-World, and Growth-Driven Navigability

Greedy routing achieves ultra-short paths in specific random and small-world topologies, provided the distribution of long-range links is tuned to match the spatial growth properties.

For fixed-growth graphs of dimension $\alpha$ , where balls of radius $\ell$ about any node $u$ scale as $\Theta(\ell^\alpha)$ , efficiency is maximized by adding “highway” nodes (density $k = \Theta(\log n)$ ) and assigning long-range links weighted according to $d_G(u,v)^{-s}$ with $s\approx\alpha$ . Greedy routing then achieves $\Theta(\log n)$ expected hop-count, and the augmented network’s diameter matches this up to subpolynomial factors. Empirical work on U.S. road networks reveals that real graphs have optimal clustering exponents $s$ closely matching their measured dimension $\alpha$ , and tuning $s$ to $\alpha$ improves routing efficiency by 10–30% over lattice-based heuristics (Gila et al., 5 Feb 2025).

Small-world random graphs and geometric inhomogeneous random graphs (GIRGs) with power-law degree distributions and geometric proximity exhibit greedy routing path lengths of $O(\log\log n)$ , with constant success probability and stretch, provided edge probabilities are appropriately tuned. This result formally explains Milgram’s letter experiments and supports decentralized navigation protocols in inhomogeneous settings (Bringmann et al., 2016).

5. Self-Organization, Incentives, and Game-Theoretic Construction

Greedy routing is robust to decentralized implementation, but in autonomous environments, agents must be incentivized to construct and maintain greedy-supporting overlays, especially under cost constraints. Game-theoretic models (the “greedy routing reachability game”) offer the following characterization:

Agents (nodes) select incident links to maximize navigability (guaranteeing greedy reachability to all destinations) while minimizing link cost. For undirected networks in 2D Euclidean space, equilibria (pure Nash) do not always exist, but approximate equilibria can be achieved via geometric spanners (specifically, $\Theta_k$ -graphs with $k\ge6$ ), yielding at most 1.8 times the minimum possible edge count (price of anarchy between 1.75 and 1.8) (Lenzner et al., 30 Jan 2026, Berger et al., 2024).
The minimum necessary set of links for 100% greedy navigability—termed the “Greedy Navigational Core” (GNC)—is, in general, an NP-hard set cover problem, but high-precision approximations can be found using local set covering heuristics (heszberger et al., 2021).
Games in the hidden-hyperbolic metric set reward-sharing mechanisms to bootstrap global cooperation: when all agents cooperate, high delivery success ( $\sim90\%$ for O( $\log N$ ) paths) is achieved; otherwise, the non-functional (no routing) state can result. Critical-mass thresholds for cooperation are sharply defined and depend sensitively on the benefit per delivered packet, degree heterogeneity, and initial spatial clustering (Kleineberg et al., 2016).

6. Algorithmic and Practical Implementations

Practical greedy routing schemes are deployed in energy-limited and resource-constrained networks:

In sensor networks, grid-based greedy hierarchical chain protocols can use spread spectrum code-masks (e.g., Persian Greedy Chain), exploiting local energy-aware adaptation and spread-spectrum transmission for robustness and energy efficiency, markedly outperforming classical LEACH clustering (Noghabi et al., 2012).
Reactive local detour algorithms (e.g., reactive deflection, backtracking) can rescue routing from voids (topological holes), providing near-constant route stretch and reducing packet loss to below 4% in realistic mesh densities, indicating applicability for non-planar, heterogeneous wireless networks (0902.4157, Mahmood et al., 2018).
Secure variants (S-GPSR) integrate lightweight, direct-observation trust metrics to mitigate sinkhole and selective-forwarding attacks, boosting delivery ratios by up to 25–30 percentage points over the baseline, with minimal added delay and low computational overhead (Samundiswary et al., 2010).
In highly dynamic networks such as UAV swarms, distance-based greedy forwarding with local state and prediction achieves semi-analytically bounded hop counts and energy consumption; predictive variants outperform global route recomputation under high mobility (Khaledi et al., 2018).

7. Measures of Navigability and Structural Implications

Greedy routing efficiency is rigorously quantified by integrated measures:

The GR-score (or GR-efficiency) is the mean over all node pairs of the shortest-path distance divided by the greedy path length, with failed pairs contributing zero. This quantity synthesizes both path-stretch and success ratio, enabling unambiguous comparison across networks and embeddings (Muscoloni et al., 2019).
Greedy spatial navigation centrality and essentiality measures reveal paradoxical effects (Braess’s paradox): in some graphs, removing specific edges increases overall greedy navigability by steering navigators away from traps (Lee et al., 2011).

The unique structural signature of a network’s “greedy navigability” (success rates, path length ratios) is not determined solely by conventional topological metrics. For instance, even in human brain structural networks, the minimal GNC subnetwork achieves full greedy navigability (~4–5 outgoing edges per node at mesoscopic scales), with high overlap with the empirical connectome, suggesting that greedy-efficient wiring is a natural emergent design principle (heszberger et al., 2021).

Greedy routing networks synthesize geometric structure, decentralized operation, and navigability constraints to yield scalable architectures for communication and computation. The interplay between local rules, global properties (such as Delaunay containment or hyperbolic geometry), dynamical throughput limits, and agent incentives determines the real-world utility and theoretical limits of these systems across physical, biological, and engineered domains.