I am working with the Nyquist criterion for zero ISI when I find something that makes me think that I misunderstood this criterion.

What Nyquist says is that to avoid ISI in the sampling we must ensure that the sample frecuency is less than 2 times the bandwith of the signal, so if the channel filter is something like a rise cosine the ISI in the sampling will be $0$.

At the same time, we have (another) Nyquist criterion which says that to be able to perfect reconstruct a signal, the sample frequency must be at least $2$ times the bandwith of the signal.

So... taken into account that the case of $F_s=2B$ is not implementable, how can we assure no ISI at the same time that we have perfect reconstruction? Aren't this two criteria contradictory? Am I missing something?


  • $\begingroup$ The Nyquist ISI criterion is not directly related to Nyquist sampling rate. Also, the "twice the bandwidth" condition is only true for real-valued sampling. If you simultaneously sample I and Q of an equivalent baseband signal, one time the bandwidth is sufficient as sample rate. $\endgroup$ – Marcus Müller Dec 16 '16 at 21:38
  • $\begingroup$ What Nyquist says is that to avoid ISI in the sampling we must ensure that the sample frecuency is less than 2 times the bandwith of the signal, <-- where do you take that from? $\endgroup$ – Marcus Müller Dec 16 '16 at 21:39

My answer is related to this question Why root raised cosine filter can eliminate intersymbol interference (ISI) ?

A classical system with Nyquist pulse $p(t)$ is:

enter image description here

The equivalent baseband signal has bandwidth limited in $[-1/2T,1/2T]$ and a sampling rate $F_s > 1/T$ is enough to avoid aliasing.

We normally sample at rate $t=kT$. I wonder what would happen if suddenly I sample faster than the normal sampling rate, for example at $t = kT/2$ instead of $t = kT$ ?

  • In the frequency domain, the composite filter $g(t) = p(t) \star p^*(-t)$ has Fourier transform $G(f)$ satisfies $\sum_k G(f-k/T) = T, \forall f$. Double the sampling rate $\sum_k G(f-2k/T)$ is not constant any more. This is true at least for typical pulses such as sinc or raised cosine.
  • In the time domain, thank Dan Boschen for the visualization below, the sample at impair $k$ of $kT/2$ is contributions of two adjacent symbols. enter image description here

Thus, is it true that for free-ISI-Nyquist-pulse-shaped signal, the largest sampling rate is the value of nominal (Nyquist) bandwidth $1/T$ ?

One can say that we still have correct data by taking only one per two samples (downsampling). What if I sample at rate $F_s > 1/T$ that satisfies Shannon-Nyquist sampling theorem to avoid aliasing ? We still have ISI.

OK another one can say that the criterion $F_s > 1/T$ is in the sense that the entire information after being sampled is intact, just take a baseband filter to get back the continous time spectrum and then sample it at the correct sampling rate at $t=kT$. Have I answered my questions by myself ?


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.