Prerequisites & Notation

Before You Begin

Chapter 17 closes the technical development of the book by composing AirComp (Chapter 16), federated learning fundamentals (Chapter 9), and secure aggregation (Chapter 10) into a single wireless-FL pipeline. The golden thread — privacy, robustness, communication efficiency — is here recast as a three-way design trade-off over a physical wireless channel. The CommIT contribution on information-theoretically secure federated representation learning appears in §17.4.

  • FedAvg and FL convergence (§9.2)(Review ch09)

    Self-check: State the convergence rate of FedAvg under smoothness and bounded gradient assumptions.

  • Secure aggregation threat model (Chapter 10)(Review ch10)

    Self-check: Recall the honest-but-curious server model and its privacy objective.

  • AirComp MSE bound (Theorem 16.2.1)(Review ch16)

    Self-check: State MSE\mathsf{MSE}^{\star} as a function of minkγk\min_k \gamma_k.

  • Wireless resource allocation basics(Review ch06)

    Self-check: What does water-filling power allocation optimize?

  • Stochastic gradient descent convergence(Review ch28)

    Self-check: How does noisy gradient descent with variance σ2\sigma^2 behave asymptotically for strongly-convex losses?

Notation for This Chapter

Chapter 17 combines FL notation (rounds, learning rate, gradients) with wireless notation (channel, power, MSE).

SymbolMeaningIntroduced
ttFL round index, t=0,1,,T1t = 0, 1, \ldots, T - 1s01
TTTotal FL rounds (budget)s01
θt\boldsymbol{\theta}_tGlobal model parameters at round tts01
gk(t)\mathbf{g}_k^{(t)}Local gradient of user kk at round tts01
ηlr\eta_{\text{lr}}Learning rate (distinguished from receive amplitude)s01
G^t\hat{\mathbf{G}}_tEstimated aggregate gradient at round tts01
\ntnmseagg\ntn{mseagg}Per-round aggregation MSE =E[G^tkgk(t)2]= \mathbb{E}[\|\hat{\mathbf{G}}_t - \sum_k \mathbf{g}_k^{(t)}\|^2]s02
St[n]\mathcal{S}_t \subseteq [n]Scheduled users at round tts03
LLSmoothness constant of the FL losss02
μ\muStrong-convexity constant of the FL losss02
zk\mathbf{z}_kLearned representation of user kk (§17.4)s04