Over-the-Air Computation (AirComp)

AirComp exploits the natural summation property of the wireless multiple-access channel (MAC). Instead of avoiding interference through orthogonal scheduling, all devices transmit simultaneously using analog modulation. The receiver's antenna directly observes the sum of the modulated signals, which corresponds to a desired mathematical function—typically the weighted sum or arithmetic mean—of the pre-processed sensor readings. This collapses communication and computation into a single physical layer operation.

To compute arbitrary functions beyond simple summation, AirComp relies on nomographic decomposition. A target function f(x₁, ..., xₖ) is decomposed into:

• A pre-processing function φ(x) applied at each device before transmission.
• A post-processing function ψ(y) applied at the fusion center after receiving the superimposed signal y. This ensures that f(x₁, ..., xₖ) = ψ( Σ φ(xᵢ) ), enabling the computation of geometric means, Euclidean norms, and maximum values over the air.

To ensure amplitude alignment at the receiver, AirComp employs channel inversion (CI) precoding at the transmitter. Each device multiplies its signal by the inverse of its uplink channel coefficient hₖ. This forces all signals to arrive with equalized magnitude and phase coherence, ensuring the receiver observes a true arithmetic sum. The technique is power-limited; devices in deep fades require excessive transmit power, necessitating truncated channel inversion to maintain energy efficiency.

Unlike digital multiple-access schemes that require K orthogonal time slots for K devices, AirComp achieves latency that is independent of network size. All devices transmit in a single time slot, making it ideal for ultra-reliable low-latency communication (URLLC) and real-time control in Industry 4.0. The communication delay is bounded by a single symbol duration, regardless of whether 10 or 10,000 sensors participate.

AirComp provides a physical layer privacy guarantee absent in digital federated learning. The fusion center never receives individual data points; it only observes the aggregate sum corrupted by receiver noise. Reconstructing any single device's raw data from the superimposed analog signal is mathematically ill-posed, offering a form of privacy by design that complements differential privacy mechanisms without adding computational overhead.

AirComp operates in two regimes:

• Uncoded AirComp: Direct analog transmission of modulated values. Optimal for computing linear functions under Gaussian MAC assumptions but vulnerable to noise.
• Coded AirComp: Introduces structured redundancy using lattice codes or nested linear codes to combat channel noise while preserving the linear computability of the sum. This extends AirComp to noisy, fading channels with provable computation rate guarantees.

The canonical algorithm for federated learning where clients compute local model updates and a central server aggregates them by computing a weighted average. In the context of AirComp, the waveform superposition property of the multiple-access channel directly computes this sum in the analog domain, eliminating the need for a digital aggregation step. This physical-layer computation of the mean is the mathematical foundation of Over-the-Air Federated Learning.

The fundamental challenge in federated systems where local datasets are non-IID (not independently and identically distributed). In AirComp, this heterogeneity is compounded by channel fading and receiver noise. A client with a statistically divergent dataset and a poor channel condition can introduce both a biased update and a high error floor. Robust AirComp designs must jointly address data heterogeneity and communication unreliability to prevent model divergence.

A cryptographic protocol that allows a server to compute the sum of client vectors without inspecting individual contributions. AirComp provides a form of physical-layer secure aggregation by default: the receiver observes only the superimposed analog waveform, not the individual pre-processed signals. However, this is not cryptographically secure against a malicious observer. True privacy requires combining AirComp with differential privacy noise injection or homomorphic encryption over the air.

Techniques like sparsification and quantization reduce the communication payload in distributed learning. In AirComp, these methods serve a dual purpose:

• Sparsification reduces the number of active subcarriers, mitigating interference.
• Stochastic quantization aligns gradient precision with the modulation order. The goal is to match the gradient information content to the instantaneous channel capacity for maximum communication efficiency.

The property of a distributed system to converge correctly despite a fraction of arbitrary or malicious participants. In AirComp, a Byzantine client can transmit a high-power, deliberately corrupted waveform to poison the aggregated sum. Defenses include robust aggregation rules (e.g., geometric median, trimmed mean) that are computed over the air, or reputation-weighted power control that attenuates the transmission of untrusted nodes at the physical layer.

Accurate CSI at the transmitter is critical for AirComp. Each client must apply a pre-equalization factor (channel inversion) to its signal so that the superposition at the receiver directly represents the desired mathematical function. Imperfect CSI leads to misalignment errors that corrupt the computation. Modern approaches use neural network-based channel prediction to provide the precise, low-latency CSI required for coherent AirComp.