5
$\begingroup$

in trying to understand the convolution theorem for DTFT, I'm faced with the following problem which I can't get my head around.

First, let me state the convolution theorem for the DTFT as follows:

\begin{equation} y[n] = (x*h)[n] \iff Y(\omega) = X(\omega)H(\omega), \quad n \in \mathbb{Z}, \omega \in \mathbb{R} \end{equation} where $x, h, y$ are signals in the time domain and $X,H,Y$ their spectra in the frequency domain.

According to Wikipedia, we also have the converse for the DTFT, i.e.:

\begin{equation} y[n] = h[n]\cdot x[n] \iff Y(\omega) = X(\omega) * H(\omega) \tag{1} \end{equation} where \begin{equation} X(\omega) * H(\omega) = \int_{-\pi}^{\pi} X(\xi)H(\omega - \xi)\,d\xi \end{equation}

Now assume $x[n]$ and $h[n]$ are non-zero within the range $-(N-1)/2 \leq n \leq (N-1)/2$, where $N$ is odd. Moreover, let $h[n] = 1$ within this range such that $h[n]$ can be thought of as a rectangular window.

Now, this implies \begin{equation} y[n] = h[n] \cdot x[n] = x[n] \end{equation} and by equation (1), \begin{equation} Y(\omega) = X(\omega) * H(\omega) = X(\omega) \end{equation} However, it is known that the Fourier transform for a rectangular window (i.e. $H(\omega)$) is an aliased sinc function. It's hard to imagine that the cyclic convolution of $H(\omega)$ with any function $X(\omega)$ leaves $X(\omega)$ unchanged. So, I believe there must be a fault in logic somewhere. Please help illuminate!

$\endgroup$

1 Answer 1

2
$\begingroup$

It may be hard to imagine but there's no fault in your logic. Look at the dual problem: assume that $x(t)$ is a lowpass signal with no frequency components above $\omega_x$. If you filter $x(t)$ with an ideal lowpass filter with cut-off frequency $\omega_c>\omega_x$, the signal $x(t)$ will remain unchanged. Mathematically this means

$$\int_{-\infty}^{\infty}x(t-\tau)\frac{\sin(\omega_c\tau)}{\pi\tau}d\tau = x(t)\tag{1}$$

The validity of Eq. $(1)$ might also be hard to imagine, but we can accept it more easily by considering that its implication in the frequency domain is trivial.

$\endgroup$
1
  • 1
    $\begingroup$ Thanks for this. I was thinking about the analogy between my problem and the DFT case where the length of the signal is finite, and $h[n]=1$. It appears in this case, then $H[0]=N$, and $H[k]=0, k\neq 0$. However, the fact that in the DTFT case, $H(\omega)$ is an aliased sinc function rather than an impulse highlights the fact that what I have there is not quite as general as it seems. In particular, the form of $X(\omega)$ must be restricted by the fact that $x[n]=0, |n|>N/2$. It's this restriction that makes the convolution behaves like an identity operator. $\endgroup$
    – Tim Mak
    Oct 21, 2022 at 2:43

Your Answer

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

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