The Universal Approximation Theorem states that a feedforward neural network with even a single hidden layer can approximate any continuous function defined on a bounded domain to arbitrarily small error, provided it has a nonlinear activation function and enough neurons. George Cybenko proved it for sigmoid activations in 1989, and Kurt Hornik extended it to general nonlinear activation functions in 1991.
The theorem provides the theoretical basis for the claim that neural networks can learn complex relationships that linear models cannot capture. It is an existence result only: it says nothing about how many neurons are needed or whether gradient descent will actually find the required weights. Deep networks are used in practice because they achieve the same expressive power with far fewer parameters.