Lemma 21.6.1. The subdifferential is the convex hull of the slopes active at x: ∂f(x)=conv{ai:f(x)=⟨ai,x⟩+bi}.
Argument in the manuscript
Proof. Each active affine function supports f at x, so every convex combination of its slopes belongs to ∂f(x). If p lies outside that convex hull, strict separation gives a vector v such that ⟨p,v⟩ exceeds every active ⟨ai,v⟩. For sufficiently small t>0, inactive affine functions remain below the maximum at x+tv. Hence tf(x+tv)−f(x)=i activemax⟨ai,v⟩<⟨p,v⟩, contradicting the subgradient inequality. ◻