Analysis · Concept hubDeep story

Universal Approximation Theorem

1989 CE20th-century United States (Cybenko)

Concept

A neural network with one hidden layer can approximate any continuous function.

Understand it in one breath

"A single hidden-layer neural network can approximate any continuous function to arbitrary precision." Cybenko's theorem (1989). The first mathematical guarantee for why deep learning works. Given enough width, expressiveness is unbounded — but how wide you actually need is still open (one of deep learning theory's biggest gaps).

At a glance

-4-202400.20.40.60.81
σ(x) = 1/(1+e^(-x))

Key formula

fC([0,1]n),  ϕ(x)=iciσ(wix+bi):  fϕ<ε\forall f \in C([0,1]^n),\; \exists\, \phi(x) = \sum_i c_i \sigma(w_i \cdot x + b_i):\; \|f - \phi\|_\infty < \varepsilon

Key moments

1989 CE

Cybenko — the theorem for sigmoidal functions

George Cybenko proved that a neural network with one hidden layer and a sigmoidal activation can approximate any continuous function on a compact domain arbitrarily well.

1991 CE

Hornik — broader activation functions

Kurt Hornik showed that universal approximation depends less on a particular sigmoid and holds for broad classes of activation functions.

2017 CE

Why depth helps

A growing body of theory demonstrated that some functions can be represented exponentially more efficiently by deep networks than by shallow ones.

Modern applications

A theoretical foundation for neural networks, general function approximation, value-function estimation in reinforcement learning, and physics-informed neural networks for differential equations.

Beyond MathVoyage

Loading…