Universal Approximation Theorem vs Fourier transform

654 Views Asked by At

According to my understanding of the Universal Approximation Theorem (one of the mathematical foundations of neural networks) when applied to functions $f:\mathbb{R} \rightarrow \mathbb{C}$ it says that any function $f(t)$ can be approximated as:

$$f(t) = \sum_j c_j\phi(a_jt+b_j) $$

where $\phi$ can be any non-polynomial function and $a_i$,$b_i$,$c_i$ are finite sequences of real numbers.

If we choice $\phi(t)=e^{it}$ we obtain:

$$f(t) = \sum_j c_je^{i\left(a_jt+b_j\right)} $$

Comparing with the inverse Fourier transform:

$$f(t) = \int F(w)e^{iwt}dw $$

both expressions agrees if:

$$ F(w) = \sum_j c_j e^{ib_j}\delta(a_j-w) $$

that means that any function has a discrete spectrum, something not true.

I do not know if my interpretation of the Universal Approximation Theorem is incorrect and/or these calculus contains an error.

1

There are 1 best solutions below

0
On

I think I found an argument which should make your arguments invalid.

According to Wikipedia, the universal approximation theorem says that $f$ can be approximated (arbitrarily well) on every compact subset of $\Bbb R$.

This is not the same as $f(t) = \sum_j c_j\phi(a_jt+b_j) $ for all $t\in\Bbb R$!

In the case of $\phi(t)=e^{it}$, one can show that such an approximation is possible on each closed interval (using theory for Fourier Series), but in general, one cannot make such a claim for all $t\in\Bbb R$ at once.