Sorry if I ask a basic question but I am a bit lost.

I have a data set in frequential domain, I need to go to time domain to perform time-gating and then going back to frequencies. In order to do this, I compute an inverse Discrete Fourier Transform.

Shannon theorem tells us that two times the difference between the greatest and the least frequency, respectively noted as $f_N$ and $f_0 ( \ne 0)$, contained in the signal must be less then the sample-rate in time domain ($F_t$). Do I need to do some zero padding to avoid aliasing ?

Without zero padding on frequential domain, what is the relation between $F_t$ and $f_N, f_0$ ? My guess is that $F_t=f_N-f_0$. Is that correct ?

If I must do some zero padding, should I do it this way :

$$(0,\dots,0,f_0,\dots,f_N,0 \dots,0)$$

With enough zeros such that the first frequency contained in the padded signal is $0$ and the last is $2 f_N$ ? If my guess above is correct, then the new sample rate is high enough to avoid aliasing, because it will be $F_t=2f_N>2(f_N-f_0)$.

Thanks for your help.

  • 1
    $\begingroup$ Can you define what "time-gating" means, mathematically, for the time-sample vector that you get out of your idft? $\endgroup$ Jul 30, 2018 at 9:39
  • $\begingroup$ Sure. Let $(x_0,\dots,x_N)$ the corresponding time vector (without zero padding), what I am doing is the Hadamard product (or termwise product) with another vector $p=(p_0,\dots,p_N)$. That vector $p$ has a lot of zeros at the beginning and at the end, and where it is non-null, I compute a Kaiser-Bessel window (accordingly to papers I have read, this is the most efficient window to use in my situation). And finally I obtain something like this : $(0,\dots,0,p_i x_i, \dots, p_{i+l} x_{i+l},0\dots,0)$ @MarcusMüller $\endgroup$
    – nicomezi
    Jul 30, 2018 at 10:22
  • $\begingroup$ did your process work correctly when you simply adhere to using powers of 2 number of samples so you can avoid needing to do any padding ? only after you have this simple approach working would I worry about using non powers of 2 $\endgroup$ Jul 30, 2018 at 13:35
  • $\begingroup$ @ScottStensland you're conflating things! this has nothing to do with zero-padding to achieve an efficient-to-calculate transform length (by the way, FFTs exist for far more sizes than just powers of 2). . $\endgroup$ Jul 30, 2018 at 13:38

1 Answer 1


You'll not experience aliasing – in the end, the IDFT result per definition only contains frequency components that can be represented by the original frequency domain.

If you will, just stop worrying about aliasing: the DFT and inverse DFT are just base change matrices in $\mathbb C^N$, and are invertible.

Aliasing can only happen when something is non-invertible. This is not the case here; so, no problem.

What you might see is leakage, but you're already taking care of that by windowing.

  • $\begingroup$ Sorry but after reflection I am having another problem. You said that aliasing can only happen if something is non invertible, and windowing is not invertible. So I am unsure that I am not introducing new frequencies that will produce aliasing when going back to frequency domain. Am I mistaken ? @MarcusMüller $\endgroup$
    – nicomezi
    Jul 31, 2018 at 7:00
  • $\begingroup$ yes, you are. Aliasing doesn't happen. $\endgroup$ Jul 31, 2018 at 7:24

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.