Heterogeneous Mean Field Game Framework for LEO Satellite-Assisted V2X Networks

Abstract: Coordinating mixed fleets of massive vehicles under stringent delay constraints is a central scalability bottleneck in next-generation mobile computing networks, especially when passenger cars, freight trucks, and autonomous vehicles share the same radio and multi-access edge computing (MEC) infrastructure. Heterogeneous mean field games (HMFG) are a principled framework for this setting, but a fundamental design question remains open: how many agent types should be used for a fleet of size $N$? The difficulty is a two-sided trade-off that existing theory does not resolve: using more types improves heterogeneity representation, but it reduces per-class sample size and weakens the mean-field approximation accuracy. This paper resolves that trade-off through an explicit $\varepsilon$-Nash error decomposition, a closed-form type-selection law, a heterogeneity-aware equilibrium solver, and a robust extension to time-varying LEO backhaul dynamics. For the 1D queue state space, the optimal type count satisfies $K^*(N)=Θ(N^{1/3})$; for the joint queue-channel model ($d=2$), the scaling becomes $K^*(N)=Θ(N^{1/5})$ with logarithmic correction. The unified formula $K^*(N)=Θ(N^{α/(α+β)})$ provides dimension-dependent design guidance, reducing type granularity to a principled, set-once system parameter rather than a per-deployment tuning burden. Experiments validate the 1D scaling law with empirical slope $0.334 \pm 0.004$, achieve $2.3\times$ faster PDHG convergence at $K=5$, and deliver up to $29.5\%$ lower delay and $60\%$ higher throughput than homogeneous baselines. Unlike model-free DRL methods whose training complexity scales with the state-action space, the proposed HMFG solver has per-iteration complexity $O(K^2 N_q N_t)$ independent of fleet size $N$, making it suitable for large-scale mobile edge computing deployment.

• The paper introduces a closed‐form scaling law (cube-root law) for selecting the optimal type count in heterogeneous mean field games, enhancing modeling fidelity.
• It presents a heterogeneity-aware PDHG algorithm that achieves up to 2.3× faster convergence and improved delay, throughput, and energy performance.
• Empirical results validate the framework’s ability to guarantee QoS with 100% satisfaction in large-scale vehicular networks under dynamic LEO backhaul conditions.

Heterogeneous Mean Field Game Framework for Scalable LEO Satellite-Assisted V2X Networks

Introduction and Motivation

The paper "Heterogeneous Mean Field Game Framework for LEO Satellite-Assisted V2X Networks" (2604.00621) addresses the question of scalable and heterogeneity-aware resource allocation in massive vehicular networks with LEO satellite backhaul. Future V2X deployments will coordinate fleets of $10^4$–$10^5$ vehicles exhibiting diverse behaviors—passenger cars, trucks, and autonomous vehicles—each imposing distinct data and QoS requirements. In this context, homogeneous mean field game (MFG) models are inadequate; they overlook meaningful physical and economic distinctions across vehicle classes. Heterogeneous MFGs (HMFG) provide the necessary type-level granularity but introduce a delicate model selection problem: finer class granularity improves modeling fidelity but reduces the representative sample size per type, undermining mean-field approximation accuracy.

In large $N$ regimes, coordination based on exact Nash equilibria becomes computationally intractable ($O(N^2)$ complexity). Prior work has developed scalable MFG-based optimizations, e.g., via primal-dual methods with complexity independent of $N$ [kang2023time], but the agent-type resolution (i.e., number $K$ of types) has remained a heuristic, unprincipled selection—fixed arbitrarily, often $K=1,2,3$ [zhang2022hmfmarl, xu2025joint]. The core contribution of this paper is a closed-form, theoretically justified scaling law for optimal type-count $K^*(N)$, a heterogeneity-aware equilibrium solver, and robustness guarantees under realistic LEO backhaul dynamics.

Problem Formulation

Consider a V2X system where an RSU equipped with MEC must allocate communication and compute resources to a mixed fleet ($N$ vehicles) over a time horizon $[0, T]$. Each vehicle's state is a queue length process with type-dependent arrival rates, energy-delay trade-off weights, and hardware/action constraints. The population is partitioned into $10^5$0 types, and each type may have distinctive queue dynamics, cost preferences, and control spaces. State evolution is governed by SDEs, and resource sharing induces cross-type coupling via interference and computational surcharge terms.

The HMFG setting leads to $10^5$1 coupled HJB-FPK systems with congestion feedback, where each type's best-response is parameterized by the mixed fleet mean field. The LEO satellite augmentation introduces additional nonstationarity via piecewise-constant backhaul rates and time-varying computation surcharge.

Figure 1: V2X–MEC scenario with K vehicle types, cross-type mean field interaction, and LEO backhaul for modeling dynamic network conditions.

Type Granularity: Theoretical Foundation

Error Decomposition

The central technical advance is an explicit decomposition of the $10^5$2-Nash approximation error for the $10^5$3-type HMFG:

$10^5$4

Here, $10^5$5 is the type-discretization bias (how well $10^5$6 types represent true fleet heterogeneity) and $10^5$7 is the sample variance term due to finite class size. The exponents $10^5$8 (empirical measure convergence rate) and $10^5$9 (drift regularity in type) depend on the state-space and model class; $N$0 in 1D, $N$1 in 2D settings (with logarithmic corrections), and $N$2 for Lipschitz class parametrizations.

Closed-Form Optimal Type Count

Minimizing $N$3 analytically yields a unique optimal type count:

$N$4

In canonical 1D queue-state settings, this delivers the cube-root law, $N$5. For modern fleets ($N$6), only $N$7 is necessary, a dramatic reduction relative to the vehicle count.

Figure 2: Validation of scaling-law—empirical optimal $N$8 closely matches $N$9; adaptive selection achieves orders-of-magnitude reduction in approximation error compared to fixed-$O(N^2)$0 Heterogeneous/Homogeneous baselines.

The methodology also incorporates unbalanced fleet populations. When class sizes are heavily skewed (e.g., 10% autonomous vehicles), the optimal $O(N^2)$1 is further reduced by a factor depending on the smallest class fraction, accommodating practical deployment realities.

Figure 3: $O(N^2)$2 and $O(N^2)$3-Nash error for balanced (uniform) versus heavily skewed fleets; optimal type count decreases with minimal class proportion as predicted theoretically.

Heterogeneity-Aware HMFG Solver Design

Algorithmic Pipeline

The authors present a modular solution workflow:

1. Type Selection Algorithm: Uses closed-form formula for $O(N^2)$4, adjusting for heterogeneity (quantified via Wasserstein distance), composition skew, and LEO topology variation.
2. Heterogeneity-Aware Step Size Adaptation: Primal-dual hybrid gradient (PDHG) steps are tuned according to observed instantaneous heterogeneity, with explicit correction terms to guarantee convergence in the presence of strong cross-type coupling.
3. Efficient Mean Field Equilibrium Solver: Parallelized, vehicle-count-independent complexity $O(N^2)$5, with pipeline compatibility for LEO-augmented, time-varying optimal controls.

   Figure 4: PDHG residuals—heterogeneity-aware adaptive step-sizes enable $O(N^2)$6 convergence speedup compared to fixed step-size protocols.

Complexity and Scalability

Empirically, the per-iteration complexity grows as $O(N^2)$7 (since $O(N^2)$8), equivalently $O(N^2)$9 independent of $N$0. This is a clear advantage over direct-multiagent (MARL) or model-free methods, whose costs scale with the size of the state-action space or $N$1.

Figure 5: Wall-clock iteration runtime—proposed method exhibits sublinear scaling in $N$2, maintaining efficiency in massively large fleet regimes.

Empirical Results

Comprehensive experiments confirm the theoretical findings and exhibit strong numerical improvements over homogeneous and fixed-type baselines:

• Delay Reduction: Up to $N$3 lower average transmission delay; at $N$4 all vehicles satisfy the V2X latency threshold (100ms), while the homogeneous solver achieves only $N$5 satisfaction.
• Throughput and Reliability: $N$6 higher per-vehicle throughput and $N$7 lower packet loss compared to G-prox.
• Energy Efficiency: $N$8 reduction in energy per transmitted bit.
• Robustness: Under LEO backhaul dynamics, only a negligible error penalty is incurred, and order-optimal $N$9 scaling is preserved provided topology variation is not pathologically severe.

Figure 6: Six-KPI performance profile—proposed HMFG achieves lowest transmission delay, highest throughput, and lowest energy/bit, outperforming all fixed-type baselines.

Figure 7: Transmission delay CDF—proposed method guarantees 100% QoS below 100ms at $K$0; G-prox achieves only 48.3%. Mean-field averaging at $K$1 gives universal satisfaction, but mean-delay ordering remains optimal for type-adaptive solutions.

Practical and Theoretical Implications

The principal implication is that type-granularity selection in HMFGs can be principled, computable, and fixed system-design choice. Type count should scale sublinearly with $K$2; excessive partitioning worsens mean-field approximation due to undersampled classes, while excessively coarse models fail to capture critical operational heterogeneity. The provided law is dimension- and heterogeneity-dependent and is robust to moderate non-stationarities introduced by dynamic LEO satellite backhaul.

Additionally, adaptive step-sizing enables safe, rapid solver convergence even as heterogeneity and non-stationarity increase. The entire pipeline is compatible with RSU–MEC architectures and practical for online plug-and-play deployment.

Future Directions

Potential extensions include:

• Global (as opposed to local) well-posedness for HMFGs under unbounded time horizons or severe LEO topology perturbation.
• Joint optimization for higher-dimensional joint queue-channel models and semantic-aware codebook integration.
• Hybridization with optimization-driven DRL to handle partial/abrupt knowledge or adversarial uncertainty regimes.

Conclusion

This work rigorously quantifies the optimal model granularity needed for scalable, robust V2X coordination under strong agent heterogeneity and dynamic LEO backhaul. The cube-root law for agent types, per-iteration complexity guarantees, and strong practical gains demote type granularity from a heuristic to a controlled, system-level design parameter. These results establish a new baseline for analytical and deployable large-scale vehicular network control via mean field games.