Suppose we are given an input signal s[m,n] with DTFT $S(\omega_1, \omega_2)$.

We sample it at $\omega_1 = \frac{2 \pi k}{256}$ and $\omega_2 = \frac{2 \pi l}{256}$ to get a 256 point DFT S[k,l]. Now suppose we take the IDFT of S[k,l] to get $\tilde{s}[m,n]$

I am trying to understand the relationship between $\tilde{s}[m,n]$ and s[m,n]. My understanding is that I should be able to express $\tilde{s}[m,n]$ as a summation of every 256th sample of the original signal.

So far I have been advised to try expressing the Sampled DTFT as a dirac comb, i.e.:

$S[k,l] = \sum_{k = -\infty}^{\infty} \sum_{l = -\infty}^{\infty} S( \frac{2 \pi k}{256}, \frac{2 \pi l}{256}) \delta(\omega_1 - \frac{2 \pi k}{256}, \omega_2-\frac{2 \pi l}{256}) $

and use the definition of the IDFT:

$\tilde{s}[m,n] = \frac{1}{256^2}\sum_{k=0}^{255} \sum_{l=0}^{255} S[k,l] e^{j(\frac{2 \pi km}{256}+\frac{2 \pi ln}{256})}$

Using these two pieces of information I can get the expression:

$\tilde{s}[m,n] = \frac{1}{256^2}\sum_{k=0}^{255} \sum_{l=0}^{255} [\sum_{k = -\infty}^{\infty} \sum_{l = -\infty}^{\infty} S( \frac{2 \pi k}{256}, \frac{2 \pi l}{256}) \delta(\omega_1 - \frac{2 \pi k}{256}, \omega_2-\frac{2 \pi l}{256})] e^{j(\frac{2 \pi km}{256}+\frac{2 \pi ln}{256})}$

but I am not sure how to proceed from here. Any advice would be helpful I have been racking my brain over this for a couple hours now.


For simplicity, I prefer a 1D notation without losing generality.

The IDFT associated with the properly taken uniform (in frequency) samples of a given valid DTFT can be shown to be the following.

Case-1: Assume a finite length signal $x[n]$ of length $N$, and consider its DTFT $X(\omega)$. Now if you take $M \geq N$ uniform (frequency) samples of $X(\omega)$ according to $$ X[k] = X( \frac{2\pi}{M}k) ~~~, k=0,1,...,M-1 $$ then the M-point inverse DFT of $X[k]$ will exactly be $x[n]$ padded with $M-N$ zeros; i.e., $$ y[n] = IDFT_M\{ DFT_M\{ x[n] \} \} = x_M[n] $$ So the sequence $x[n]$ and $y[n]$ have the same values for the first $N$ samples (i.e., all samples of $x[n]$ are retained in $y[n]$), but $y[n]$ have further (effectively padded) zero samples.

If $ M < N$ , then time-aliasing in $x[n]$ will happen and the first $N-M$ samples of $y[n]$ will be in error. But the remaning $2M-N$ samples of $y[n]$ will be identical to the original $x[n]$ at the corresponding index positions.

Case-2: $x[n]$ is infinite length (and stable) sequence and $X(\omega)$ is its DTFT. Then according to $M < N$ subcondition of case-1, all samples of $y[n]$ will be aliased. No samples of $y[n]$ will be equal to $x[n]$, except other than (pure) coincidence.

  • $\begingroup$ Correct, I am expecting aliasing in y[n], but I am trying to figure out how to prove derive the expression which relates the aliased $\tilde{s}[m,n]$ values to those of the original signal. $\endgroup$ – Filip Oct 15 '18 at 0:25
  • $\begingroup$ aliased values of y[n] are addition of last $N-M$ samples of $x[n]$ to first $N-M$ samples of $x[n]$ ; i.e. $y[0] = x[0] + x[M]$, $y[1] = x[1]+x[M+1]$,...,$y[N-M-1] = x[N-M+1]+x[N-1]$. You can see those indices by a simple plot of the periodic extension of the circularly alised inverse DFT. $\endgroup$ – Fat32 Oct 15 '18 at 0:30
  • $\begingroup$ or stated in other words: $$ \tilde{y}[n] = \sum_k x[n-kM] $$ $\endgroup$ – Fat32 Oct 15 '18 at 0:32
  • $\begingroup$ But what is the math behind this fact, to me it is not obvious why there is aliasing in the discrete time domain. $\endgroup$ – Filip Oct 15 '18 at 0:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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