[Paper] Federated Learning for Terahertz Wireless Communication

Source: arXiv - 2512.04984v1

Overview

The paper explores how Federated Learning (FL) behaves when the underlying wireless link operates in the Terahertz (THz) band. By marrying a realistic multicarrier THz channel model with the FL optimization loop, the authors reveal why naïve averaging of local updates can stall learning and propose a simple weighting scheme that restores convergence—even in the presence of severe physical impairments like beam‑squint and molecular absorption.

Key Contributions

• Unified THz‑FL framework: A stochastic, multicarrier model that directly ties per‑subcarrier signal‑to‑noise ratios (SNRs) to the variance of local gradient updates.
• “Diversity trap” analysis: Shows that under standard unbiased aggregation the convergence error floor is dictated by the harmonic mean of subcarrier SNRs, making a single deep spectral hole enough to cripple the whole learning process.
• Fundamental bandwidth limit: Demonstrates that beyond a certain bandwidth the added thermal noise and gain roll‑off at band edges increase the convergence error rather than improve it.
• SNR‑weighted aggregation rule: Derives and validates a practical weighting strategy that mitigates the variance singularity caused by spectral holes, enabling reliable FL even with high beam‑squint.
• Extensive simulations: Numerical experiments that map key THz physical parameters (beam‑squint, molecular absorption, jitter) to FL performance metrics such as test accuracy and convergence speed.

Methodology

1. THz Channel Modeling – The authors adopt a wideband, OFDM‑style multicarrier representation where each subcarrier experiences frequency‑selective attenuation from molecular absorption, beam‑squint (angle‑dependent gain loss), and additive thermal noise.
2. Coupling to FL – The local stochastic gradient computed on each device is transmitted over the THz link. The received gradient on the server is modeled as the true gradient plus a zero‑mean noise term whose variance is proportional to the inverse of the subcarrier SNR.
3. Convergence Analysis – Using standard FL theory (smooth, strongly convex objectives) they derive an upper bound on the expected optimality gap. The bound explicitly contains the harmonic mean of the subcarrier SNRs, exposing the “diversity trap”.
4. Weighted Aggregation – To counteract the trap, they propose weighting each device’s update by the average SNR of the subcarriers it used, leading to a modified aggregation formula that restores the usual unbiased convergence rate.
5. Simulation Setup – Realistic THz parameters (e.g., 0.3–1 THz carrier range, molecular absorption lines, antenna array sizes) are plugged into a Python‑based FL simulator (FedAvg on a CNN for MNIST/CIFAR‑10).

Results & Findings

• Convergence error floor: With plain averaging, the test accuracy stalls at ~70 % for a 5‑device FL run when a single subcarrier’s SNR drops below 0 dB (a typical beam‑squint hole).
• Bandwidth sweet spot: Adding more subcarriers improves accuracy up to ~2 GHz total bandwidth; beyond ~5 GHz the error floor rises again due to accumulated noise at the band edges.
• Weighted aggregation wins: Applying the SNR‑weighted rule pushes the accuracy back to >90 % even when multiple deep spectral holes exist, matching the performance of an ideal, noise‑free channel.
• Sensitivity to physical parameters: Molecular absorption peaks cause localized SNR dips; jitter (phase noise) mainly inflates the variance uniformly across subcarriers, both of which are mitigated by the weighting scheme.

Practical Implications

• Design of THz‑enabled FL systems: Engineers should measure or estimate per‑subcarrier SNRs and incorporate them into the aggregation step rather than relying on vanilla FedAvg.
• Spectrum allocation: Simply widening the THz bandwidth is not a silver bullet; system designers must identify the optimal usable bandwidth that balances data rate and learning stability.
• Antenna array tuning: Reducing beam‑squint (e.g., via wider arrays or adaptive beamforming) directly improves the harmonic‑mean SNR, lowering the convergence floor without extra protocol changes.
• Edge‑AI deployments: For applications like distributed sensor fusion or collaborative robotics that may exploit THz links for ultra‑low‑latency updates, the weighted aggregation can be implemented as a lightweight post‑processing step on the server, requiring only SNR metadata from each device.
• Standardization impact: The findings suggest that future FL‑over‑THz protocol specifications should include optional fields for channel quality indicators (CQI) per subcarrier, enabling interoperable weighted aggregation.

Limitations & Future Work

• Model assumptions: The analysis assumes smooth, strongly convex loss functions; extending to non‑convex deep networks (the norm in computer vision) may require tighter bounds.
• Static channel snapshot: Simulations consider quasi‑static THz channels per round; real‑world mobility could introduce rapid SNR fluctuations that challenge the weighting scheme.
• Hardware constraints: The work abstracts away hardware impairments like ADC quantization and power amplifier non‑linearity, which could further affect gradient noise.
• Future directions: The authors propose exploring adaptive bandwidth selection, joint beamforming‑FL optimization, and online SNR estimation techniques to make the approach robust in highly dynamic THz environments.

Authors

• O. Tansel Baydas
• Ozgur B. Akan

Paper Information

• arXiv ID: 2512.04984v1
• Categories: cs.DC
• Published: December 4, 2025
• PDF: Download PDF