1
$\begingroup$

Initially I asked this question on mathoverflow. I however thought physicists may face this sort of problem more than mathematicians (I am an engineer). Due to that, I decided to ask here as well.

Consider a function $f(x)$ that is numerically defined in $-1 \leq x \leq 1$ interval (assume $N$ samples). I am trying to compute the action of $f(d/dx)$ on a function $g(x)$ using the Fourier transform as follows:

\begin{equation} f(d/dx) g(x)=\int_{-\infty}^{\infty} f(i\omega)\mathcal{G}(\omega)e^{i\omega x}dw, \end{equation}

where $\mathcal{G}$ is the Fourier transform of $g(x)$. Say for instance, if $f(x)=x$, the above integral leads to $\frac{d}{dx}g(x)$ as expected. I am however stuck in calculating $f(i\omega)$ when $f$ is known numerically.

Question: How to evaluate $f(i\omega)$ given $N$ available samples?

Edit: Is this the right approach to compute $f(d/dx)$ numerically?

I am unsure about the selected tags below. Please help to add proper tags as well.

$\endgroup$
1
  • 1
    $\begingroup$ You should look for analytic continuation. You have a function defined on a real interval and you are trying to obtain it's value on the imaginary axis. In general this is impossible (the function could be whatever on the imaginary axis) but if your function is analytic then you're good. In general a method known as maximum entropy is used to compute this sort of analytic continuation for functions which are given numerically $\endgroup$
    – lcv
    Sep 28, 2022 at 9:34

1 Answer 1

2
$\begingroup$

Very crudely, I'd suggest the following. Let's represent the numerical function as $$ f(x) = \sum_n f_n \delta(x-n\Delta x) , $$ where $f_n$ represents the numerical value. Then we represent the Dirac deltas as $$ \delta(x)=\int \exp(i2\pi xa)\ da $$ so that $$ f(x) = \sum_n f_n \int \exp(i2\pi xa-i2\pi na\Delta x)\ da . $$ It then means that $$ f(i\omega) = \sum_n f_n \int \exp(-2\pi a\omega-i2\pi na\Delta x)\ da . $$ It may be possible to solve the required integrals, assuming they are convergent.

$\endgroup$
2
  • $\begingroup$ Thank you for your response. I am afraid that this solution is numerically unstable because $e^{-2\pi a \omega}$ for negative $\omega$ s quickly explods. $\endgroup$
    – Mirar
    Sep 28, 2022 at 7:34
  • $\begingroup$ Yes, but fortunately it is multiplied by $\mathcal{G}(\omega)$ that would hopefully suppress it on the negative side. That is why I made the comment about the convergence. $\endgroup$ Sep 28, 2022 at 8:41

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.