Ferkans — Interactive Telecom Tutor

Why Compute Over the Air?

Federated learning (Chapter 9) and secure aggregation (Chapter 10) both reduce to one fundamental operation: the server wants the sum $\sum_k \mathbf{g}_k$ of the users' gradients. Classical digital-uplink FL transmits one bit at a time, giving the server all $n$ gradients and letting it add them. The per-user communication cost scales linearly in $n$ — the uplink quickly becomes the bottleneck.

AirComp (over-the-air computation) turns the multiple-access channel's superposition from a nuisance into a feature. All users transmit simultaneously on the same frequency; the wireless channel physically adds the signals. The access point receives $\sum_k h_k x_k + \mathbf{w}$ — the aggregate directly, in one channel use — with MSE dominated by the noise $\mathbf{w}$ , not by the per-user payload.

The point is that the sum is computed in the analog domain, with communication cost independent of $n$ . AirComp reframes the bottleneck: the limit is no longer bandwidth per user, but the MSE floor imposed by channel heterogeneity and noise. The rest of this chapter develops the model, power-control strategy, function class, and privacy implications.

,

Definition:
The AirComp Signal Model

There are $n$ single-antenna users, each holding a source value $s_k \in \mathbb{C}$ (for example, one scalar entry of a gradient). The access point wants to estimate $y = \sum_{k=1}^{n} s_k$ .

Pre-processing. User $k$ maps $s_k$ to a transmit symbol $x_k = b_k s_k$ , where $b_k \in \mathbb{C}$ is a scaling chosen to (i) match the receive target and (ii) respect the power constraint $\mathbb{E}[|x_k|^2] \leq P_k$ .

MAC superposition. All users transmit in the same channel use. The access point observes $r \;=\; \sum_{k=1}^{n} h_k \, x_k \;+\; \mathbf{w} \;=\; \sum_{k=1}^{n} h_k b_k s_k \;+\; \mathbf{w}, \qquad \mathbf{w} \sim \mathcal{CN}(0, \sigma^2).$ The channel gains $h_k \in \mathbb{C}$ are assumed known at the transmitters (CSIT, typical for TDD reciprocity).

Post-processing. The receiver forms $\hat{y} = r / \eta$ for a common receive amplitude $\eta$ . When $b_k h_k = \eta$ for all $k$ (magnitude alignment, §16.2): $\hat{y} \;=\; \sum_{k=1}^{n} s_k \;+\; \frac{\mathbf{w}}{\eta}.$ The aggregate is recovered up to an additive noise term whose variance scales with $1/|\eta|^2$ . No digital decoding, quantization, or per-user bandwidth — the entire aggregation is a single analog channel use.

,

AirComp (Over-the-Air Computation)

A physical-layer aggregation scheme where users transmit analog pre-processed values simultaneously; the wireless channel's natural superposition computes the aggregate. The receiver estimates the result from the superimposed signal. The per-user communication cost is $O(1)$ symbols, independent of the number of users.

Magnitude Alignment

The power-control condition $b_k h_k = \eta$ (a common receive amplitude for every user). Alignment is necessary so that the superposition $\sum_k h_k b_k s_k$ equals $\eta \sum_k s_k$ , up to noise.

Aggregation MSE

The mean-squared error $\mathbb{E}[|\hat{y} - \sum_k s_k|^2]$ between the AirComp estimate and the true sum. The core performance metric of AirComp. Under magnitude alignment, $\mathsf{MSE} = \sigma^2 / |\eta|^2$ .

Example: Two-User AirComp Over an AWGN-Free MAC

Two users hold $s_1, s_2 \in \mathbb{R}$ and want the access point to learn $y = s_1 + s_2$ . The channel gains are $h_1 = 1$ , $h_2 = 2$ . The noise variance is $\sigma^2 = 0$ (ideal). Design transmit scalings $b_1, b_2$ that recover $y$ from a single channel use, and compute the receive amplitude $\eta$ .

Solution

Magnitude-alignment condition

Require $b_1 h_1 = b_2 h_2 = \eta$ . With $h_1 = 1, h_2 = 2$ , this gives $b_1 = \eta$ and $b_2 = \eta / 2$ .

Choose $\eta$

Any $\eta > 0$ works; pick $\eta = 1$ for simplicity. Then $b_1 = 1, b_2 = 0.5$ .

Received signal

$r = h_1 b_1 s_1 + h_2 b_2 s_2 + 0 = s_1 + s_2 = y$ . The sum is recovered exactly.

Operational interpretation

The user with the stronger channel ( $h_2 = 2$ ) must transmit at half the amplitude. Channel inversion: the worst channel determines the shared receive level. With noise, this becomes the key source of the MSE floor (§16.2).

AirComp over the MAC: Analog Aggregation in One Channel Use

Animation of the AirComp aggregation:

n = 4

users pre-process their source values, transmit simultaneously, and the wireless channel physically adds the signals. The receiver divides by

\eta

to recover

\sum_k s_k

plus noise. Visual emphasis on the single channel use and the fact that the receiver never sees individual

s_k

— a privacy property made explicit in §16.4.

AirComp vs. Digital Uplink Aggregation

Property	Digital uplink (Ch. 10)	AirComp (Ch. 16)
Channel uses per aggregation	$\Theta(n)$ — orthogonal per-user slots	$\Theta(1)$ — single MAC use
Bandwidth scaling	Linear in $n$	Independent of $n$
Aggregation accuracy	Quantization + noise per user	MSE $= \sigma^2/\|\eta\|^2$ (channel-limited)
Individual-gradient leakage	Server decodes each $\mathbf{g}_k$	Server sees only $\sum_k s_k$ + noise
CSIT requirement	None (orthogonal)	Yes (pre-equalization)
Synchronization requirement	Symbol-level	Symbol and carrier-phase

The AirComp–Secure-Aggregation Synergy

AirComp is natively privacy-preserving for the sum: the receiver observes the superposition $\sum_k h_k b_k s_k + \mathbf{w}$ and cannot separate individual contributions. This is a structural property of the MAC — no cryptographic protocol required. The cryptographic pairwise masking of Chapter 10 (Bonawitz et al.) solves the same problem in the digital domain at $\Theta(n^2)$ key exchanges; AirComp achieves the aggregate at $\Theta(1)$ communication cost and $0$ key exchanges.

Two caveats temper the claim. First, "server learns only the sum" presumes an honest-but-curious server that cannot deploy multiple receive antennas to separate users via beamforming — the non-colluding-antennas assumption that §16.4 scrutinizes. Second, AirComp demands tight synchronization and CSIT, which may be unavailable in some deployments. The golden thread — privacy vs. communication efficiency — is visible here: AirComp buys $O(n)$ efficiency and sum-privacy at the cost of stricter physical-layer requirements.

,

Common Mistake: AirComp Is Not 'Free'

Mistake:

Conclude from the $\Theta(1)$ channel-use count that AirComp replaces digital aggregation at no cost.

Correction:

AirComp requires: (1) channel-state information at the transmitter (CSIT) for magnitude alignment; (2) tight symbol and carrier-phase synchronization across all $n$ users (harder than digital, which tolerates per-user offsets); (3) an analog front end that transmits real-valued pre-processed samples (not the standard digital modem); and (4) a known common power-control target $\eta$ . Real deployments must budget for these. The $\Theta(1)$ bandwidth saving is real, but so is the increase in physical-layer coordination complexity.

Power Cost of Magnitude Alignment

Explore how the required transmit power $|b_k|^2 = |\eta|^2 / |h_k|^2$ depends on the per-user channel gain $|h_k|$ . Users with weak channels pay a large multiplicative penalty — the bottleneck user dominates the shared power budget. The plot displays the per-user required power against channel gain for a common receive target $|\eta|^2 = 1$ .

Parameters

|\eta|^2

— receive amplitude target1

|h|_{\min}

— minimum channel gain0.2

Key Takeaway

AirComp turns MAC superposition into a one-shot analog aggregator. With $n$ synchronized users and CSIT, the access point recovers $\sum_k s_k$ in a single channel use, with MSE bounded by the noise-to-alignment ratio $\sigma^2/|\eta|^2$ . The cost is analog-front-end and tight synchronization — offset against the $\Theta(n)$ -to- $\Theta(1)$ bandwidth gain and the native sum-privacy. The rest of this chapter turns "what is $\eta$ ?" into a concrete optimization (§16.2), broadens the function class (§16.3), and quantifies the privacy (§16.4).

Quick Check

In the AirComp signal model with $n$ users and a Gaussian MAC, which of the following is an inherent requirement (not a design choice)?

Every user must transmit the same source value $s_k$ .

The receiver must separately decode each $x_k$ before combining.

The channel gains $h_k$ must be known at the transmitters (CSIT).

All users must be at the same physical distance from the receiver.