The Generalized Shannon Theorem

Sampling a narrowband signal with sampling frequency $F_s$ lower than twice the maximum frequency of the signal does not always lead to aliasing.
In this example, you can adjust the sampling frequency to see how it affects the spectral content of the sampled signal (in red). The original signal (in blue) is a narrowband signal with central frequency $F_0=1000\ \mathrm{Hz}$ and bandwidth $B=200\ \mathrm{Hz}$.
Find which sampling frequencies are acceptable, and why.

4000
300 4000

Original signal

Sampled signal

Open for comments

In this example, aliasing occurs when $F_s$ lies in:

$$ [0,440] \cup [450,550] \cup [600,733] \cup [900,1100] \cup [1800,2200] $$

The generalized Shannon theorem states that to avoid aliasing when sampling a narrowband signal (with central frequency $F_0$ and bandwidth $B$), the sampling frequency must obviously be at least twice the bandwidth of the signal:

$$ F_s \geq 2B $$

Yet, not all such frequencies are acceptable, given the possible overlap between positive and negative spectral images. This excludes the following sampling frequencies:

$$ \left[\frac{2F_0-B}{k}, \frac{2F_0+B}{k}\right], $$

with $k$ integer but not 0. As a matter of fact, for $k$ odd, values of $F_s$ in these intervals put the Nyquist frequency inside a spectral image; for $k$ even, they put the zero frequency inside a spectral image. In both cases, this causes aliasing.

In our example, this clearly leads to excluding $F_s$ from:

$$ [0,400] \cup [333,440] \cup [450,550] \cup [600,733] \cup [900,1100] \cup [1800,2200] $$

So, in this example, the smallest possible sampling frequency is $440\ \mathrm{Hz}$.
Notice, by examining the spectral content of the sampled signal, and by listening to it, that using sampling frequencies $F_s<2(F_0+B/2)$ leads to signals in the $[0,F_s/2]$ band that are different from the original signal. However, since no aliasing occurs, it is still possible to recover the original signal by further digital upsampling and band-pass filtering.

Nota Bene: for convenience, the effect of the sampling frequency on the magnitude of the sampled signal spectrum has been compensated for in the plot.