Let $x[n] = \cos(wn)\ \ $ be our signal. Let the other signal be $x_2[n] = \cos(2wn)\ \ $ . The spectrum of $x[n]$ and $x_2[n]$ are well defined and consist of impulses at the relevant frequencies with weight $\pi$. However, if we interpret $x_2[n]$ as the decimated version of $x[n]$, the spectrum of $x_2[n]$ must have impulses with weight $\frac{\pi}{2}$. I have the explanation but I would like to hear the alternatives, what is the best way to resolve these seemingly contradicting facts?

  • $\begingroup$ How do you determine that the impulses at $w$ have weight $\pi$? $\endgroup$ – Jim Clay Mar 23 '13 at 20:34
  • $\begingroup$ @JimClay: Jim, the DTFT of a discrete-time cosine function with frequency w is a periodic train of impulses at frequencies -w+2\pi k and w+2\pi k, k=-\inf to \inf. All impulses have weight \pi. This comes from the fact that the DTFT of a complex exponential e^jw0t is 2\pi \delta(w-w0). 2\pi is there to cancel the 1 / 2\pi factor in the IDTFT. $\endgroup$ – Dost Mar 24 '13 at 19:03

The following correction of Jason R's answer completely resolves the problem: The frequency spectrum of $X_2(ω)$ consists of two components, which are scaled versions of $X(ω)$. Those are $4\pi$-periodic instead of being $2\pi$-periodic. When those $4\pi$-periodic versions are added up, the sum becomes $2π$-periodic again. THERE IS NO ALIASING or IMPULSES SUPERIMPOSED ON EACH OTHER. The derivation from $$ X_2(\omega) = \frac{1}{2} \pi\left(\delta\left(\frac{\omega}{2}-\omega_0\right) + \delta\left(\frac{\omega}{2}+\omega_0\right) + \delta\left(\frac{\omega}{2}-\omega_0-\pi\right) + \delta\left(\frac{\omega}{2}+\omega_0-\pi\right)\right) $$ to $$ X_2(\omega) = \frac{1}{2} \pi\left(\delta\left(\omega-2\omega_0\right) + \delta\left(\omega+2\omega_0\right) + \delta\left(\omega-2\omega_0-2\pi\right) + \delta\left(\omega+2\omega_0-2\pi\right)\right) $$ is not correct. Instead it must read $$ X_2(ω)=π(δ(ω−2ω_0 )+δ(ω+2ω_0)+δ(ω−2ω_0−2π)+δ(ω+2ω_0−2π)) \qquad (∗) $$ This is due to the scaling property of impulses $$ δ(aω)=\frac{δ(ω)}{|a|} $$ Then, observing that (*) is $2π$-periodic, two of the impulses are redundant and hence the DTFT of $x_2[n]=x[2n]$ becomes $$ X_2(ω)=π(δ(ω−2ω_0)+δ(ω+2ω_0)) $$

| improve this answer | |

I'm not sure where you get the idea that $x_2[n]$ must have weight $\frac{\pi}{2}$. You can conclude that the weights are the same for both cases using a simple change of varible in the DTFT equation:

$$ x[n] = \cos(\omega_0n) $$

$$ \begin{align} X(\omega) &= \sum_{n=-\infty}^{\infty} x[n] e^{-j\omega n} \\ &= \sum_{n=-\infty}^{\infty} \cos(\omega_0 n) e^{-j\omega n} \\ &= \pi(\delta(\omega-\omega_0) + \delta(\omega+\omega_0)) \end{align} $$

Now consider the other signal:

$$ x_2[n] = \cos(2\omega_0 n) $$

$$ \begin{align} X_2(\omega) &= \sum_{n=-\infty}^{\infty} x_2[n] e^{-j\omega n} \\ &= \sum_{n=-\infty}^{\infty} \cos(2 \omega_0 n) e^{-j\omega n} \end{align} $$

Perform a change of variable:

$$ n' = 2n $$

Since the limits of the summation over $n$ extend to infinity, we also sum from $-\infty$ to $\infty$ over $n'$.

$$ X_2(\omega) = \sum_{n'=-\infty}^{\infty} \cos(\omega_0 n') e^{-j\omega \frac{n'}{2}} $$

Now, regroup the exponent inside the sum to yield:

$$ X_2(\omega) = \sum_{n=-\infty}^{\infty} \cos(\omega_0 n') e^{-j\frac{\omega}{2} n'} $$

Then via inspection we can observe that $X_2(\omega) = X\left(\frac{\omega}{2}\right)$. Then, in terms of the equation that we found previously, we conclude that:

$$ \begin{align} X_2(\omega) &= X\left(\frac{\omega}{2}\right) \\ &= \pi\left(\delta\left(\frac{\omega}{2}-\omega_0\right) + \delta\left(\frac{\omega}{2}+\omega_0\right)\right) \\ &= \pi\left(\delta\left(\omega-2\omega_0\right) + \delta\left(\omega+2\omega_0\right)\right) \end{align} $$

Edit: After thinking about it a bit more and looking at the notation that you posted in the below comment, I assert that the two are equivalent, although that may not be readily apparent. While the equation given in your comment has a typo, you can find a derivation on slide 22 in this presentation. That structure makes explicit the aliasing that can happen when decimating a signal. The correct form is:

$$ x[an] \Leftrightarrow \frac{1}{a} \sum_{k=0}^{a-1} X\left(\frac{\omega}{a} - \frac{2\pi k}{a}\right) $$

So, for the case you posed, we can write:

$$ \begin{align} x_2[n] = x[2n] \Leftrightarrow X_2(\omega) &= \frac{1}{2} \sum_{k=0}^{1} X\left(\frac{\omega}{2} - \frac{2\pi k}{2}\right) \\ &= \frac{1}{2} \sum_{k=0}^{1} X\left(\frac{\omega - 2 \pi k}{2}\right) \\ &= \frac{1}{2} \left(X\left(\frac{\omega}{2}\right) + X\left(\frac{\omega}{2} - \pi\right)\right) \end{align} $$

We then substitute the known expression for $X(\omega)$ to yield:

$$ X_2(\omega) = \frac{1}{2} \pi\left(\delta\left(\frac{\omega}{2}-\omega_0\right) + \delta\left(\frac{\omega}{2}+\omega_0\right) + \delta\left(\frac{\omega}{2}-\omega_0-\pi\right) + \delta\left(\frac{\omega}{2}+\omega_0-\pi\right)\right) $$

We then manipulate the arguments to the impulses, noting that they are nonzero when their argument is equal to zero:

$$ X_2(\omega) = \frac{1}{2} \pi\left(\delta\left(\omega-2\omega_0\right) + \delta\left(\omega+2\omega_0\right) + \delta\left(\omega-2\omega_0-2\pi\right) + \delta\left(\omega+2\omega_0-2\pi\right)\right) $$

Which, noting the $2\pi$-periodicity of $X(\omega)$, is equivalent to:

$$ \begin{align} X_2(\omega) &= \frac{1}{2} \pi\left(\delta\left(\omega-2\omega_0\right) + \delta\left(\omega+2\omega_0\right) + \delta\left(\omega-2\omega_0\right) + \delta\left(\omega+2\omega_0\right)\right) \\ &= \pi\left(\delta\left(\omega-2\omega_0\right) + \delta\left(\omega+2\omega_0\right)\right) \\ &= X\left(\frac{\omega}{2}\right) \end{align} $$

| improve this answer | |
  • $\begingroup$ Jason, what you say is absolutely correct. However, the contradiction is that; $x_2[n] = x[2n]$ which is the downsampled version of $x[n]$ by a factor of 2. And the downsampling operation by a factor of 2 brings a 1/2 multiplicative factor in the frequency domain. $\endgroup$ – Dost Mar 27 '13 at 19:07
  • $\begingroup$ Do you have a reference for your claim? Details like extra multiplicative constants in the result are dependent upon the exact transform definition that is used. $\endgroup$ – Jason R Mar 27 '13 at 19:38
  • $\begingroup$ The following is textbook knowledge: The frequency response of the decimated-by-M signal $x_d[n]$ of $x[n]$ is $$X_d(\omega)=\frac{1}{M}\sum_{i=0}^{M-1}X\left(\frac{\omega}{M}-\frac{2\pi j}{M}\right)$$ $\endgroup$ – Dost Mar 27 '13 at 22:09
  • $\begingroup$ @JasonR I don't think your change of variable works. If $n' = 2n$, then the summation in $n'$ will have to be zero for $n'$ odd....which is a different summation. $\endgroup$ – Peter K. Apr 2 '13 at 16:22

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.