Operating Regimes of Decentralized Learning Under Mobility and Bandwidth Constraints

Samuele Sabella, Chiara Boldrini, Lorenzo Valerio, Marco Conti, Andrea PassarellaarXiv 2026 · arXiv

TL;DR

Decentralized learning holds immense promise for privacy-preserving, serverless AI on mobile devices, but standard algorithms assume perfect, static network connections. This paper stress-tests decentralized learning under realistic wireless constraints—modeling node mobility, limited bandwidth, and network contention. The authors identify three distinct operating regimes, revealing that frequent but partial model updates are surprisingly effective, while overly dense node crowding can actually cripple learning due to network congestion.

What Problem Does This Paper Address?

In classical Federated Learning (FL), mobile clients sync directly with a central coordinator server. In fully decentralized learning, there is no server; devices (nodes) train models locally and share updates directly with nearby peers. While this eliminates single points of failure and massive server bandwidth bottlenecks, it introduces a massive wildcard: the physical world.

Most theoretical analyses of decentralized learning assume a static or smoothly varying network graph with infinite bandwidth. In reality, devices move, connections drop mid-transfer, and multiple devices competing for the same wireless channel cause severe network contention. This paper addresses the gap between idealized algorithmic theory and the messy reality of deploying decentralized learning over actual wireless technologies like Bluetooth LE, LTE, and Wi-Fi.

Baseline performance comparison between a fully connected network with infinite bandwidth (LOCALHOST) and no communication (BROKEN_LINK).

Methods

The authors implement an asynchronous decentralized averaging protocol (DecAvg) designed to handle real-world network disruptions. Key components of their system model include:

  • Asynchronous Training & Dissemination: Nodes continuously train on local data and attempt to broadcast their latest model parameters to any available neighbors in parallel, rather than waiting for rigid global communication rounds.
  • Partial (Truncated) Transfers: If two mobile nodes move out of range before a model transfer completes, the receiving node doesn't discard the data. Instead, it keeps the fully received layer-wise parameter tensors and integrates them into its model aggregation.
  • Contention Modeling: The system accounts for channel congestion using a shared-medium policy where a node's available bandwidth is dynamically scaled down based on the number of simultaneous transfers it is handling.
  • Mobility Control: Using the Random Waypoint (RWP) mobility model, the authors systematically vary the simulation area to control two critical network statistics: Contact Time (CT) (how long nodes stay in range) and Inter-Contact Time (ICT) (how long they spend apart).

Datasets

The primary evaluation is conducted on the CIFAR-10 dataset distributed across 50 mobile clients in a highly non-IID (heterogeneous) fashion using a Zipf distribution ($\alpha = 3.5$). The authors utilize a lightweight convolutional neural network with residual blocks. MNIST is also referenced in the paper's preliminary analysis to contrast model sizes and transmission times.

Key Insights and Ideas

The paper's core contribution is the categorization of decentralized learning behavior into three system-level operating regimes:

  • The Connectivity-Limited Regime: When nodes rarely meet (high ICT), learning is bottlenecked purely by how slowly information diffuses across the physical space. Even if nodes have ultra-fast Wi-Fi that guarantees 100% complete model transfers during a brief encounter, the overall convergence remains slow.
  • The Bandwidth-Tolerant Regime: When nodes meet frequently (low ICT), the network becomes highly robust to truncated updates. Even if low-bandwidth links (like LTE or Bluetooth) mean that only a small fraction of the model tensors are successfully transmitted per contact, the high frequency of interactions compensates for the partial data.
  • The Contention-Limited Regime: Counter-intuitively, more connectivity isn't always better. When node density is extremely high, the sheer volume of concurrent transfers triggers severe wireless channel contention. This drops the effective throughput so drastically that intermediate mobility patterns actually outperform hyper-dense mixing.

Key Results

  • The Contention Dip: Under LTE connectivity, decreasing the Inter-Contact Time (ICT) from 100 seconds to 25 seconds actually degraded final test accuracy. This happened because the increased node density caused severe channel contention, dropping the fraction of successfully completed updates. An intermediate ICT of 100 seconds yielded the best performance.
Fig. 3. Impact of inter-contact time (ICT) on convergence for fixed wireless technologies (LTE and Wi-Fi 4), highlighting the transition from sparse mixing to dense mixing where channel contention can emerge. Top-left: CIFAR-10 test accuracy vs. time under LTE for different ICT values. Top-right: CIFAR-10 test accuracy vs. time under Wi-Fi 4. Bottom-left: distribution of successful transmission fractions under LTE. Bottom-right: distribution under Wi-Fi 4.
  • High-Throughput Resilience: Wi-Fi 4 proved robust enough to avoid severe contention bottlenecks. At an ICT of 25 seconds, Wi-Fi 4 achieved a global test accuracy close to 70%, matching the ideal, unlimited-bandwidth benchmark (LOCALHOST).
  • Tensor Priority Matters: When network congestion is severe, the order in which model layers are transmitted becomes crucial. Prioritizing the transmission of the classifier head tensors rather than the deeper feature-extraction layers yielded noticeably better accuracy under high-contention scenarios. The authors hypothesize that faster alignment on decision boundaries helps stabilize learning when updates are frequently cut short.
Comparison of tensor transmission order under LTE, demonstrating that prioritizing the classifier head yields better accuracy during high network congestion.

Limitations and Future Work

While this work provides a vital reality check for decentralized learning, the authors note a few limitations. The evaluation relies on the idealized Random Waypoint mobility model, which features uniform node distribution and lacks the social or spatial clustering seen in real human movement. Additionally, the experiments were bounded to a moderate-scale convolutional model on CIFAR-10.

Future work will focus on validating these operating regimes using large-scale model architectures (like Transformers) and evaluating them against real-world human and vehicular mobility traces.

Summary updated

Abstract

Decentralized learning is a promising paradigm for collaborative training in mobile and pervasive systems, as it avoids a central coordinator and does not require sharing raw data. Yet, most analyses rely on idealized communication assumptions that break down in wireless settings, where connectivity is intermittent, topology changes due to mobility, and bandwidth is limited. We study decentralized averaging under client asynchrony, time-varying contact graphs, and technology-dependent throughput constraints. We implement a fully decentralized protocol that overlaps synchronization with local training and supports partial tensor-level transfers when contacts end early. Using Random Waypoint mobility and multiple wireless technologies (Bluetooth LE, LTE, and Wi-Fi), we quantify how network dynamics and link capacity impact convergence. We identify three operating regimes: (i) inter-contact time largely dictates convergence via mixing, (ii) partial updates are often well tolerated when contacts are frequent, and (iii) very dense contact patterns can trigger contention, reducing effective throughput. These findings provide a practical lens to reason about decentralized learning deployments over realistic wireless systems, highlighting when improving connectivity, increasing bandwidth, or mitigating contention is most impactful.
BibTeX
@misc{sabella2026operating,
  author = {Samuele Sabella and Chiara Boldrini and Lorenzo Valerio and Marco Conti and Andrea Passarella},
  title = {{Operating Regimes of Decentralized Learning Under Mobility and Bandwidth Constraints}},
  year = {2026},
  eprint = {2606.28342},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.28342},
}
Figure