The Kolmogorov-Arnold representation theorem says any function with N inputs can be created using a two layer network of at most 2 N + 1 nodes where each noed's output is a non-linear function of the sum of its inputs. However, the non-linear functions at each node may all be different and very complex including discontinuities which are effectively, meaning that the result at first appears to only have very abstract thoeretical interest. However, Kolmogorov-Arnold Networks are showing promise at being able to create usable neural networks using this result.
Used in Chap. 7: page 99
Links:
Wikipedia:
Kolmogorov–Arnold representation theorem