Neural Network Universal Approximation — Blueprint

7 Input-convex networks — Amos et al.

Definition 41 Fully input-convex network
#

A FICNN threads the input \(y\) into every layer, \(z_{i+1}=\rho _i(W^z_i z_i + W^y_i y + b_i)\), with nonnegative propagation weights \(W^z_i\ge 0\), unconstrained input skip \(W^y_i\), and convex nondecreasing activations \(\rho _i\).

Theorem 42 ICNN soundness

Such a network denotes a convex function of its input.

Proof

By induction over the layer chain: each layer maps convex coordinates to convex coordinates — a nonnegative combination of convex functions plus the affine input skip is convex, and a convex nondecreasing activation composed with a convex function is convex.

Theorem 43 Convex universal approximation

Let \(f\) be convex and differentiable on \(\mathbb {R}^d\), and let \(K\) be compact. For every \(\varepsilon \gt 0\) there is a fully input-convex network, with nonnegative propagation weights and convex nondecreasing (ReLU and identity) activations, that is within \(\varepsilon \) of \(f\) uniformly on \(K\).

Proof

Cover \(K\) by finitely many points \(x_1,\dots ,x_n\) close enough that \(f\)’s modulus of continuity is below \(\varepsilon /2\) at that scale. At each \(x_i\), the tangent plane \(y \mapsto f(x_i) + \nabla f(x_i)\cdot (y-x_i)\) minorizes \(f\) by convexity, and equals \(f\) at \(x_i\); taking the pointwise maximum of these tangent planes over \(i\) therefore stays below \(f\) everywhere and comes within \(\varepsilon \) of \(f\) on \(K\). This max-of-affine-functions is realized exactly by a nonnegative-weight ICNN with ReLU and identity activations (a max of two affine functions is \(\mathrm{ReLU}(u-v)+v\), iterated), giving the network witness.