Clip Harder When You Sum: Amplifier Back-Off as the Hidden Straggler in Over-the-Air Federated Learning

In over-the-air computation (AirComp), many phones or sensors transmit their model updates at the same time on the same frequencies, and the radio channel adds them up. The server receives the sum of all updates in one transmission, no matter how many devices take part. The catch is that every device has to pre-scale its signal to undo its own channel, and all devices must share one common scale. The device with the worst channel therefore sets the received signal strength for everyone. This device is a wireless straggler: it slows or degrades the whole round, as a slow worker does in a synchronous GPU cluster. Prior work treats the straggler's limit as a fixed average-power budget. On real multi-carrier hardware, however, the binding limit is the power amplifier's peak. OFDM signals (orthogonal frequency-division multiplexing, which spreads data across many parallel subcarriers) have a high peak-to-average power ratio (PAPR), meaning occasional spikes far above their mean power. To keep those spikes out of the amplifier's distorting region, devices "back off" their average power by several decibels. That back-off shrinks the budget the straggler has to work with. Undoing a deep fade on a single subcarrier makes the spikes worse, which forces even more back-off.

This note proposes that the right amount of back-off is very different when signals are summed than when one signal is sent point-to-point. When a single device clips its peaks, the receiver sees that distortion directly. When K devices clip independently and their signals are summed, each device's clipping distortion is uncorrelated with its own gradient and with the other devices' distortion. The distortion therefore averages down in the aggregate while the gradients add up coherently. AirComp devices should thus drive their amplifiers much closer to saturation than point-to-point links do. The weakest device should clip hardest, because it gains the most average power by doing so. With a simple gain correction, the aggregated gradient stays unbiased, which stochastic gradient descent requires.

Core Claim

Let device $k$ send an OFDM signal $x_k$ whose subcarriers carry its pre-scaled gradient entries. The amplifier clips $x_k$ at a level set by clipping ratio $\gamma_k$, the clip level divided by the root-mean-square amplitude. The Bussgang decomposition is the standard result that a memoryless nonlinearity applied to a Gaussian-like input equals a scaled copy of the input plus uncorrelated distortion. It gives

$$\mathrm{clip}_{\gamma_k}(x_k) = \alpha(\gamma_k)\, x_k + d_k, \qquad \mathbb{E}[x_k d_k^*] = 0,$$

where $\alpha(\gamma_k) \le 1$ is a known deterministic gain and $d_k$ is distortion whose power $\sigma_d^2(\gamma_k)$ grows as the clip level drops. If each device divides by $\alpha(\gamma_k)$ before transmitting, the received sum carries the true gradient sum at a common scale $\eta$. The error in the averaged gradient on one subcarrier then behaves as

$$\mathrm{MSE}(\eta,\gamma) \approx \frac{1}{K^2}\sum_{k=1}^{K} |h_k|^2\,\sigma_d^2(\gamma_k)\,\frac{\eta^2}{|h_k|^2\alpha(\gamma_k)^2}\cdot\frac{1}{P_{\text{avg},k}} \;+\; \frac{\sigma_n^2}{K^2\eta^2},$$

where $h_k$ is the channel gain, $\sigma_n^2$ is receiver noise, and $\eta$ is capped by the weakest device's usable average power $P_{\text{avg},k}(\gamma_k)$. That usable power rises as $\gamma_k$ falls, because less back-off is needed. The distortion term is a sum of $K$ independent contributions divided by $K^2$, so it shrinks roughly as $1/K$. The noise term shrinks as the weakest device's power rises. The prediction that makes this falsifiable is that the MSE-optimal clipping ratio falls as $K$ grows, and falls further for weak-channel devices. A point-to-point link, where $K=1$, is the most conservative case. Standard practice fixes back-off at the point-to-point value, so it should leave several decibels of straggler power unused.

This is not the idea behind the closest existing work. Semantic Communications with Discrete-time Analog Transmission: A PAPR Perspective learns a clipping threshold end to end for a single image link. It cuts PAPR from about 21 dB to 6 dB at a 2–3 dB PSNR cost. It has no multiple-access channel, no straggler, and no averaging across transmitters. Relay-Assisted Cooperative Federated Learning models the weakest-channel bottleneck and rescues it with relays, but uses a scalar power constraint with no envelope statistics or amplifier nonlinearity. PAR-Aware Large-Scale Multi-User MIMO-OFDM Downlink is the work that tells us pre-equalizing frequency-selective channels inflates PAPR. It studies a downlink carrying independent data streams, not summed computation. Our contribution is a closed-form, per-device choice of clipping level for the aggregation task, derived from how distortion averages across devices.

Research Program

The first piece is analytical. We will derive $\alpha(\gamma)$ and $\sigma_d^2(\gamma)$ for pre-equalized OFDM gradient signals. These signals are not Gaussian when a deep-faded subcarrier dominates the envelope, so we will check where the Bussgang approximation breaks. We then solve the joint problem over the common scale $\eta$ and the per-device ratios $\gamma_k$. Given $\eta$, the problem separates across devices, so a one-dimensional search over $\eta$ with closed-form $\gamma_k(\eta)$ should suffice. We will also bound the bias that remains when the gain correction uses an estimated rather than true $\alpha$, following the unbiasedness requirement $\mathbb{E}[\hat g]=g$ stressed in Approximate Gradient Coding for Distributed Learning with Heterogeneous Stragglers.

The second piece is a hard constraint on out-of-band emission. Clipping spreads power into neighboring channels, and spectrum masks forbid that. Every operating point we report must therefore meet an adjacent-channel leakage limit, enforced with one clip-and-filter pass whose in-band effect is folded into $\sigma_d^2$. A method that wins only by breaking the mask does not count.

The third piece checks the scheme against a realistic amplifier. We will use a smooth saturating amplifier model rather than an ideal clipper, and an energy model that credits the efficiency gained from lower back-off. This lets us report energy per training round alongside accuracy. It also connects to the idle-energy accounting of Reducing Energy Bloat in Large Model Training, whose Perseus system slows fast GPUs to match stragglers in datacenters. Here the corresponding saving comes from the radio amplifier instead.

Baselines

The do-nothing reference is coherent OFDM AirComp with per-subcarrier channel inversion and a fixed back-off chosen, as in standard practice, so that the unclipped envelope rarely saturates. It is identical to our system except for the clipping policy, so any gain comes from that policy alone.

The first strong competitor is truncated channel inversion. Devices whose channel falls below a threshold stay silent, which removes the straggler instead of accommodating it. This is the OBDA-style baseline used by Bayesian AirComp with Sign-Alignment Precoding for Wireless Federated Learning. We tune its threshold on the same grid budget as our $\eta$ search.

The second competitor is sign-alignment precoding with Bayesian aggregation from that same paper. It sidesteps channel inversion by sending 1-bit gradient signs precoded with only the sign of the channel. It reports 3.0% (MNIST) and 3.7% (CIFAR-10) higher accuracy than truncated inversion in heterogeneous settings, and over 95% on MNIST and about 70% on CIFAR-10 overall. It is the strongest published straggler-robust scheme in the file, but it was evaluated on flat fading without OFDM. We will re-run it over the same OFDM channel and amplifier, so its PAPR cost appears.

The third competitor is non-coherent balanced-numeral AirComp from Over-the-Air Computation over Balanced Numerals. It encodes gradient digits as which subcarriers are active and detects them by energy, avoiding inversion entirely. It reports approximately 98% MNIST accuracy with 25 non-IID devices, while a 1-bit frequency-shift-keying majority vote struggles to reach 80%. We use its 25-device non-IID MNIST split directly. We measure its PAPR, which that paper did not analyze, under the same amplifier.

We also include DFT-spread OFDM with pre-equalization as a waveform baseline. DFT-spread OFDM first mixes the data with a Fourier transform so the transmitted signal behaves more like a single low-PAPR carrier.

The key ablations remove one ingredient at a time. The first uses a common clipping ratio for all devices rather than per-device ratios. The second clips aggressively without the $1/\alpha$ gain correction, which exposes the bias. The third uses our clipping policy but plugs in the $K=1$ optimal ratio, which tests whether the averaging argument drives the gain.

For all methods, device count, channel realizations, amplifier model, leakage limit, learning rate, and number of rounds are shared. Beating the competitors means one of two things at an equal leakage limit: higher final test accuracy at equal energy per round, or equal accuracy at measurably lower energy. The relay paper gives no numerical result in the file ("no reported result found on Lacuna for relay-assisted AirComp accuracy"), so it is discussed but not reproduced.

Why not just…?

Why not just drop the stragglers? Truncation removes the weakest device and lets everyone else transmit louder. Under non-IID data, though, it removes that device's data from the model. Bayesian AirComp with Sign-Alignment Precoding measures a 3.0–3.7% accuracy loss from doing so. Our experiment sweeps the non-IID skew. If truncation matches us when data is nearly IID, we will say so; the claim is about heterogeneous data.

Why not just send 1-bit signs, which have constant envelope? Constant-envelope symbols on each subcarrier do not stay constant-envelope after the inverse Fourier transform that builds the OFDM signal. Summing hundreds of subcarriers makes the time-domain envelope roughly Gaussian again, with high PAPR, so sign schemes still need back-off. They also lose magnitude information: the balanced-numeral paper shows 1-bit voting stalling near 80% where multi-level coding reaches about 98%. Sign precoding could still hold up well because it avoids inversion-driven spikes. Re-running it on our OFDM-plus-amplifier channel settles this directly.

Why not just use DFT-spread OFDM to lower PAPR? This is the most serious threat. DFT-spread OFDM cuts PAPR by several decibels for ordinary data. We expect the advantage to vanish once each device multiplies its subcarriers by $1/h_k(f)$, because a deep fade then dominates the spectrum and breaks the single-carrier structure. The PAR-aware MIMO-OFDM work flags this mechanism for downlinks. If the PAPR CCDF curve (the probability that PAPR exceeds each level) of pre-equalized DFT-spread OFDM stays low under our channels, the simpler fix wins. In that case we will recast the contribution as per-device clipping on top of DFT-spread OFDM and test whether the averaging gain still adds anything.

Why not just learn the clipping end to end, as in DeepJSCC? Trainable clipping works for one link. In federated learning, however, the encoder would have to be retrained whenever the device set or channels change, and learned clipping carries no unbiasedness guarantee. Our policy is closed form and recomputed each round from channel estimates. We include a learned-threshold variant as a secondary comparison to see how much a learned policy gains beyond the analytic one.

Proposed Experiments

The simulation uses $K \in \{8, 25, 100\}$ devices, 256 or 1024 subcarriers, and frequency-selective Rayleigh multipath with independent device channels, plus a saturating amplifier model. Training uses the 25-device non-IID MNIST split of the balanced-numeral paper, the CIFAR-10 heterogeneous setting of the Bayesian AirComp paper, and Fashion-MNIST ( Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms) as a harder small benchmark. Metrics are per-round aggregation MSE against the ideal gradient average, test accuracy over rounds, PAPR CCDF, adjacent-channel leakage, and amplifier energy per round.

The first decisive check measures the MSE-optimal clipping ratio as a function of $K$. If it does not fall with $K$, the averaging argument is wrong and the proposal fails at its core. The second checks bias by comparing corrected and uncorrected aggressive clipping over long training runs. A drift in the uncorrected run's accuracy that the corrected run avoids would confirm that the gain correction matters. The third checks for a straggler-specific gain. We plot improvement against the spread of channel strengths across devices. The claim predicts that gains grow with that spread and vanish when all channels are equal. If gains are flat in the spread, the benefit is generic PAPR reduction rather than straggler relief, and we would report it as such.