⚡ electrical engineering

On Signal Peak Power Constraint of Over-the-Air Federated Learning

This paper identifies that existing Over-the-Air Federated Learning systems often overlook instantaneous peak-power constraints caused by non-linear power amplifiers, and demonstrates that applying standard iterative clipping and filtering to mitigate these constraints leads to significant performance degradation, particularly in multi-carrier OFDM systems due to in-band distortions.

Original authors: Lorenz Bielefeld, Paul Zheng, Oner Hanay, Yao Zhu, Yulin Hu, Anke Schmeink

Published 2026-04-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lorenz Bielefeld, Paul Zheng, Oner Hanay, Yao Zhu, Yulin Hu, Anke Schmeink

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A Group Project with a Broken Microphone

Imagine a group of 40 students (the devices) working on a massive group project (the AI model). They are all in different rooms and cannot share their actual notebooks (data) because of privacy rules. Instead, they only send their "notes on what to improve" (gradients) to a central teacher (the server) once a week.

The Old Way (Digital):
Usually, students would write their notes on paper, put them in envelopes, and mail them one by one. If there are 40 students, the teacher has to open 40 envelopes. This is slow and clogs up the mailbox.

The New Way (Over-the-Air / AirComp):
To speed things up, the students shout their notes into the room at the same time. Because sound waves add up, the teacher hears a single, combined voice that represents the average of all their notes. This is Over-the-Air Computation (AirComp). It's much faster because everyone talks at once.

The Hidden Problem: The "Shouting" Limit

The paper points out a flaw in this "shouting" method that everyone ignored until now.

Imagine the students are using a microphone system connected to a speaker.

The Average Volume: The teacher tells everyone, "Keep your average volume at a comfortable level."
The Peak Volume: However, sometimes a student gets really excited and shouts a specific word very loudly. This is the Peak Power.

The Hardware Glitch:
The speakers (Power Amplifiers) in the system have a limit. They work perfectly when the volume is steady. But if someone shouts too loudly (exceeds the Peak Power Limit), the speaker distorts. It sounds like a crackle or a buzz.

In-Band Distortion: The message itself becomes garbled (the teacher hears "improve" as "im-prove-uh").
Out-of-Band Distortion: The shouting is so loud it bleeds into the next room, disturbing neighbors on other frequencies.

In the world of AI, this "crackle" corrupts the math. If the teacher receives a garbled average, the AI model learns the wrong things and might even fail completely.

The Specific Scenarios

The paper looks at two ways the students send their notes:

Single-Carrier (One continuous stream): Like a single person reading a long sentence. If the sentence has a few very loud words, the speaker distorts.
Multi-Carrier (OFDM): Like a choir where everyone sings a different note at the same time. This is great for speed, but it has a famous problem: if everyone hits their high notes at the exact same moment, the combined volume spikes massively (high PAPR - Peak-to-Average Power Ratio).

The Surprise: The paper found that in this specific AI setup, the "Single-Carrier" method actually had higher peaks than the "Choir" method because the AI gradients (the notes) can sometimes be incredibly large numbers, causing the speaker to blow out.

The "Fix": The Clipping and Filtering Game

To stop the speakers from blowing out, the engineers tried a classic trick called Iterative Clipping and Filtering (ICF).

Clipping: Imagine a bouncer at the door. If a student tries to shout louder than the limit, the bouncer cuts off the top of their voice. "You can't go above 80 decibels!"
The Problem: Cutting off the top of the voice changes the sound wave. It creates a weird "buzz" (distortion).
Filtering: To fix the buzz, they use a filter to smooth out the edges.
The Loop: But smoothing the edges sometimes makes the volume spike again! So they have to repeat the process: Clip, Filter, Clip, Filter... until the volume is safe and the buzz is gone.

What Happened in the Experiments?

The researchers simulated this with a real AI model (LeNet) and a dataset (CIFAR-10). Here is what they found:

The Spikes are Real: Without fixing the peaks, the signal often exceeds the hardware's limit by a huge margin (like trying to shout 100 decibels when the limit is 80).
The Fix Hurts Performance: When they applied the "Clipping and Filtering" fix:
- Single-Carrier: The AI got slightly worse at learning (about 1-2% less accurate), but it still worked.
- Multi-Carrier (The Choir): This was the disaster zone. When the background noise was low (a quiet room), the "fix" actually made the AI diverge (stop learning entirely).

Why did the quiet room make it worse?
In a quiet room, the teacher (server) tries to amplify the students' voices more to hear them clearly. But because the "clipping" already damaged the sound, amplifying it just makes the "crackle" (distortion) louder. The teacher ends up listening to a loud, garbled mess instead of the students' actual notes.

The Takeaway

This paper is a wake-up call. For years, researchers designed these "shouting" AI systems assuming the speakers could handle any volume as long as the average was low.

The Lesson:
Just because the average volume is safe doesn't mean the loud moments won't break the equipment. If we don't account for these "loud moments" (Peak Power Constraints), our AI models might learn the wrong things or crash, especially in complex, high-speed systems.

The authors suggest we need new, smarter ways to control the volume so the AI can learn fast without breaking the hardware.

1. Problem Statement

The paper addresses a critical, yet overlooked, practical constraint in Over-the-Air Federated Learning (AirComp-FL): the instantaneous peak-power constraint imposed by the non-linearity of Radio-Frequency (RF) Power Amplifiers (PAs).

Context: AirComp leverages the superposition property of wireless channels to aggregate local model updates from edge devices directly over the air, solving the communication bottleneck of digital transmission.
The Gap: Existing AirComp designs optimize for average power constraints to minimize Mean Squared Error (MSE). However, they ignore the Peak-to-Average Power Ratio (PAPR).
The Issue:
- Analog transmission in AirComp depends on gradient value distributions, which can result in high instantaneous peaks.
- Multi-carrier OFDM systems inherently suffer from high PAPR.
- When the instantaneous signal power exceeds the PA's linear region (saturation point), it causes in-band distortion (corrupting the aggregated gradient) and out-of-band emissions (interfering with adjacent frequencies).
Core Question: Does enforcing instantaneous peak-power constraints degrade FL performance, and if so, how?

2. Methodology

The authors propose a system model and a mitigation strategy to analyze the impact of peak power constraints on both Single-Carrier and Multi-Carrier (OFDM) AirComp-FL systems.

A. System Model

Federated Learning Setup: A base station (BS) coordinates $K$ user equipments (UEs). UEs compute local gradients, normalize them (zero mean, unit variance), and transmit them via analog signals.
Transmission Schemes:
1. Single-Carrier: Gradient values are transmitted as real-valued samples in time slots.
2. Multi-Carrier (OFDM): Gradient values are mapped to subcarriers. The paper assumes equal power allocation per subcarrier to simplify the problem, decoupling the optimization per subcarrier.
Power Control: The standard approach minimizes MSE under an average power budget ( $P_{avg,max}$ ). The optimal solution involves channel inversion for strong channels and maximum power for weak channels.

B. The Constraint & Mitigation (ICF)

The paper introduces an instantaneous amplitude limit $A_{max}$ (corresponding to $P_{inst,max}$ ). To satisfy this without changing the modulation scheme, the authors apply Iterative Clipping and Filtering (ICF):

Clipping: The time-domain signal amplitude is limited to $A_{max}$ . This introduces non-linear distortion.
Filtering: A low-pass filter is applied in the frequency domain to suppress out-of-band emissions caused by clipping.
Iteration: Since filtering can cause "peak regrowth" (increasing PAPR again), the clipping and filtering steps are repeated until out-of-band emissions fall below a specific threshold (e.g., -10 dBm).

C. Simulation Setup

Dataset/Model: LeNet CNN trained on CIFAR-10.
Network: 40 UEs, 62,006 parameters.
Channel: Free-space path loss + Rayleigh fading.
Power Constraints: Average power = 23 dBm; Instantaneous peak power = 26 dBm.
Comparison: Performance is compared between "Unclipped" (theoretical, violating PA limits) and "Clipped" (ICF applied) scenarios across varying noise levels.

3. Key Contributions

Identification of a Practical Blind Spot: This is the first work to explicitly highlight that standard AirComp-FL designs violate instantaneous peak-power constraints due to high PAPR in both single-carrier and OFDM systems.
Characterization of PAPR in AirComp: The authors demonstrate that AirComp signals exhibit high PAPR. Interestingly, they found that single-carrier transmission can exhibit even higher PAPR than multi-carrier OFDM in this context because the power is directly proportional to the gradient values, which can theoretically be unbounded.
Evaluation of ICF: The paper provides a detailed implementation and analysis of ICF as the standard solution for PAPR reduction in this context.
Performance Impact Analysis: The study quantifies the trade-off between adhering to hardware constraints and maintaining FL convergence accuracy.

4. Simulation Results

The simulations reveal several counter-intuitive and critical findings:

PAPR Reduction: Without clipping, PAPR is extremely high (~35 dB). ICF successfully reduces this to ~15 dB, keeping the signal within the PA's linear region.
Error Metrics (MSE vs. TSE):
- Clipping increases the True Squared Error (TSE) compared to the unclipped baseline.
- Crucial Finding for Multi-Carrier: In low-noise environments, the clipped system performs worse than the unclipped system.
- Reasoning: In low-noise scenarios, the optimal power control algorithm assigns higher scaling factors ( $\alpha$ ) to maximize signal strength. These high scaling factors amplify the in-band distortion introduced by the clipping process, leading to significant aggregation errors.
FL Accuracy Impact:
- Single-Carrier: Clipping causes a consistent but small accuracy drop (1.5–2% at later stages). A noticeable drop occurs around 50 communication rounds.
- Multi-Carrier (OFDM): The impact is severe in low-noise conditions. At $\sigma^2 = -110$ dBm/Hz, the training diverges completely due to amplified in-band distortion.
- Peak Drop: Accuracy can drop by up to 8% at specific training stages (e.g., round 50) due to the clipping artifacts.

5. Significance and Conclusion

Practical Necessity: The paper proves that ignoring instantaneous peak-power constraints renders theoretical AirComp-FL designs impractical for real-world hardware, as PAs will inevitably distort the signal.
Paradoxical Behavior: The finding that lower noise leads to worse performance under peak-power constraints is significant. It implies that in high-SNR regimes (typical of modern networks), the distortion from necessary clipping becomes the dominant error source, potentially causing model divergence.
Future Directions: The work calls for:
1. Peak-Power-Aware Power Control: New optimization algorithms that account for instantaneous limits rather than just average limits.
2. Better PAPR Reduction: Developing techniques specifically tailored for AirComp that minimize in-band distortion, as standard ICF is insufficient for high-precision FL in low-noise environments.

In summary, this paper serves as a vital reality check for the AirComp-FL community, demonstrating that hardware non-linearities can fundamentally undermine the convergence of distributed learning, particularly in multi-carrier systems with high signal-to-noise ratios.