Wireless Aggregation Latency in Edge Learning with Fractional Power Control

Technical Summary: Wireless Aggregation Latency in Edge Learning with Fractional Power Control

Problem Statement
In hierarchical federated learning (HFL) systems, edge servers aggregate client updates locally before transmitting them to a central core server. While this architecture improves scalability, the uplink transmission from edge servers to the core server creates a multiple-access bottleneck. This "core aggregation latency" (CAL) is heavily influenced by heterogeneous wireless channels, characterized by varying path loss and fading. The paper addresses the challenge of minimizing the cumulative core aggregation latency (C-CAL) over the duration of a learning task, specifically investigating how fractional power control (FPC) can mitigate these wireless bottlenecks without requiring complex model-specific adaptations.

Methodology
The authors develop an analytical framework that decouples learning dynamics from communication latency through the following steps:

1. Stochastic System Modeling:

   • Network: Edge server (ES) locations are modeled as random variables within a bounded region, and wireless channels are modeled using flat-fading, path loss, and lognormal shadowing.
   • Latency: Transmission latency is defined based on Shannon capacity, where the Signal-to-Noise Ratio (SNR) is adjusted via a fractional power control (FPC) rule. The rule adapts transmit power based on channel gain $h_n$ using an exponent $b \in [0, 1]$.
   • Protocol: The system employs Time Division Multiple Access (TDMA) for the uplink aggregation phase.

2. Renewal Reward Formulation:

   • The paper models the HFL process as a renewal reward process. Each learning round constitutes a renewal cycle, and the task completion (stopping time $S$) defines the end of the process.
   • Under the assumption of independent and identically distributed (iid) server selection and channel realizations across rounds, the authors apply Wald's equality. This allows the Mean Cumulative Core Aggregation Latency (MC-CAL) to be decomposed into the product of two independent components:
     $$\bar{T}_{c-cal} = \bar{S} \times \bar{T}_{cal}$$
     Where $\bar{S}$ is the expected number of learning rounds (determined by learning dynamics, data, and stopping criteria) and $\bar{T}_{cal}$ is the expected per-round aggregation latency (determined by communication system parameters).

3. Analytical Derivation:

   • The authors derive analytical upper bounds for $\bar{T}_{cal}$ under the FPC policy. They establish necessary and sufficient conditions for the finiteness of the expected latency, linking it to the moments of the fading distribution and the spatial distribution of edge servers.
   • Specifically, they show that the bound depends on the term $E[h_n^{b-1}]$, which captures the joint effect of fading statistics and server geometry.

Key Contributions

• Decomposition of Latency: The paper provides a rigorous theoretical decomposition of cumulative aggregation latency, separating the influence of learning convergence (stopping time) from communication efficiency (per-round latency). This allows for the independent analysis of communication-layer improvements.
• Analytical Bounds under FPC: The study derives exact upper bounds on the mean per-round aggregation latency for HFL systems utilizing fractional power control. It proves that even modest FPC exponents are sufficient to ensure finite latency and significantly reduce the cumulative delay.
• Model-Agnostic Mechanism: The proposed approach is described as "model-agnostic," meaning the benefits of FPC apply regardless of the specific learning algorithm, dataset distribution, or model architecture, provided the communication assumptions hold.

Results
Numerical simulations were conducted using MNIST and CIFAR-10 datasets with varying data heterogeneity (IID vs. Label Skew) and network parameters (Nakagami-m fading, lognormal shadowing).

• Impact of FPC: The results demonstrate that even small FPC exponents (e.g., $b = 0.05$ or $b = 0.25$) yield substantial reductions in C-CAL. For instance, at a deployment radius of 300m, a modest exponent of $b=0.05$ reduced the per-round latency by approximately 65%.
• Weak Channel Mitigation: The efficacy of FPC is particularly pronounced in weak channel conditions (e.g., Non-Line-of-Sight or large deployment ranges). In these scenarios, $b=0.25$ reduced the cumulative latency by roughly two orders of magnitude compared to no power control ($b=0$).
• Tightness of Bounds: The derived analytical upper bounds were found to be tight, closely matching simulation results as the deployment radius increased.
• Data Heterogeneity: While data heterogeneity (skew) significantly increased the number of required learning rounds ($\bar{S}$), the reduction in per-round latency ($\bar{T}_{cal}$) via FPC remained consistent across different data distributions.

Significance and Claims
The paper claims that fractional power control serves as a simple, effective, and model-agnostic mechanism to mitigate wireless aggregation bottlenecks in hierarchical edge learning. By isolating the communication latency component, the authors show that network operators can accelerate large-scale HFL systems by adjusting power control parameters without modifying the underlying learning algorithms. The work highlights that modest adjustments to power control exponents can substantially reduce the total time required for distributed learning tasks, particularly in challenging wireless environments characterized by high path loss and fading. The authors conclude that this framework provides a foundational tool for analyzing and optimizing communication efficiency in edge learning, with future work planned to extend these findings to more diverse communication environments and heterogeneous datasets.