Bandwidth of an entire song

Question

My question has to do with the difference between the frequencies of a single note, and the frequencies of an entire song.

If I have a 5 second signal of the form: $x(t)=\sin(8\pi t)$, here is the frequency response with zero-padding:

For a signal of the form $x(t)=\sin(8\pi t)\sin^2(\pi t/5)$, here is how this looks:

Here is the frequency response with zero-padding:

My intuition about the second signal is that it is a 4Hz tone that gets louder and then quieter again. In both cases, the highest frequency contained in the signal is higher than 4Hz, even though it was a 4Hz tone. This indicates to me that the frequencies that we hear are not the same as the frequencies in the Fourier transform. This further indicates that just because the highest note of a song may be at 16kHz does not mean that the bandwidth of the song is at -16kHz to +16kHz and that 32kHz sampling rate is sufficient. Music is typically recorded at 44.1kHz, but is the bandwidth of a song really -22.05kHz to +22.05kHz, even if every individual note is? I took the FFT of Bad Habits by Ed Sheeran out of curiosity. The highest frequency component actually appears to be at 22.05 kHz. Doesn't this indicate that if we had a higher sampling rate than 44.1kHz, we would have seen higher frequency components of the music? In other words, the FFT looks like higher frequency components got "cut off" by a low sampling rate.

My second question is about understanding how a note played during a song will affect the FT of the song. Without zero-padding, the first signal is purely 4Hz and the second signal has nonzero components at 4Hz and the two adjacent bins. With zero-padding, it appears there are a large number of nonzero bins in each, and the second signal actually appears as the first with sidelobe suppression. This seems significant to me because if an 8kHz tone, played for one second, appears in a 3 minute song, it would not affect the FT of the song by adding a pure 8kHz tone in. I think it would appear as an 8kHz tone of one second duration, zero-padded to a 3 minute duration (and therefore including the sidelobes), since the sidelobes are important to destructively interfere with the note outside of the timeframe it is supposed to be played at. Is this correct?

Edit: I just remembered something probably critical. Any signal of finite time necessarily has infinite bandwidth. If the highest tone in a song is 16kHz, then the highest frequency component of the whole song would be a "smeared" 16kHz, and some sidelobes will be cut off when sampling at 44.1kHz. Therefore the DFT is lossy. Part of my confusion is probably because I read elsewhere on the internet that the DFT is lossless, but I am thinking now that must be wrong since all real signals are of finite time/infinite bandwidth, therefore all real signals must have an infinite sampling rate to be truly lossless. Is this correct?

Edit #2: Envidia pointed out that I had forgotten to fftshift Bad Habits. It definitely looks better now.

Answer 1

First of all, kudos to you: I appreciate the effort and thinking you've managed to articulate in your question.

The DFT is a mathematical tool. As such, the parameters used to compute it can hide or reveal information that is there or not there. For example, zero-padding will reveal side-lobes for a single tone but that is just an artifact of the DFT of a finite-length sequence. You don't "hear" these side lobes when listening to a single tone. You only hear that single tone (a perfect frequency peak if the parameters of the DFT are chosen appropriately). As another example, if the parameters aren't set correctly, giving poor frequency resolution, the DFT will hide information from you.

That being said, case-in-point:

This indicates to me that the frequencies that we hear are not the same as the frequencies in the Fourier transform.

$5$ seconds is too short to see the expected peaks in the frequency domain your signal has: recall that $$\sin(A)\sin(B) = \frac{1}{2}\big(\cos(A-B)-\cos(A+B)\big)$$ In your case, the FFT should therefore show peaks at $3.8\,\tt{Hz}$, $4\,\tt{Hz}$ and $4.2\,\tt{Hz}$.
Set your parameters appropriately: increase the duration to $10$ seconds and you should see:

Music is typically recorded at 44.1kHz, but is the bandwidth of a song really -22.05kHz to +22.05kHz?

I think you already know the answer: in the analog world, no. Some details:
Harmonics are integer multiples of a fundamental frequency. For example, if a pianist bangs on a middle $A$, which has fundamental frequency of $440 \,\tt{Hz}$, its harmonics would be at $880\,\tt{Hz}$ (2nd harmonic), $1320\,\tt{Hz}$ (3rd harmonic), $1760\,\tt{Hz}$ (4th harmonic), and so on. There's technically no upper limit to how high harmonics can go in terms of frequency. However, the amplitude (~loudness) of these harmonics generally decreases as the frequency increases, making them less significant in the overall sound, especially as they move out of the range of human hearing.

Which brings us to:

The highest frequency component actually appears to be at 22.05 kHz. Doesn't this indicate that if we had a higher sampling rate than 44.1kHz, we would have seen higher frequency components of the music?

Correct! And some audio systems claim to use $96\,\tt{kHz}$ for that exact purpose (although I have a hard time believing anyone that claims they can hear the difference with standard $44.1\,\tt{kHz}$ or $48\,\tt{kHz}$, unless they'd want their dogs to be able to?) - edit: there are other purposes to higher sampling rates such as post-processing considerations and less stringent anti-aliasing/reconstruction filter requirements, but that's outside the scope of your question, see comments -. As far as quality of playback is concerned, there's no point in going higher because: 1. Humans can not hear past $\approx 20\,\tt{kHz}$, and 2. as previously mentioned, the harmonics will generally be well below our hearing threshold anyways.

if an 8kHz tone, played for one second, appears in a 3 minute song, [...] it would appear as an 8kHz tone of one second duration, zero-padded to a 3 minute duration (and therefore including the sidelobes), since the sidelobes are important to destructively interfere with the note outside of the timeframe it is supposed to be played at. Is this correct?

You are correct that it would be interpolated because of the length of your DFT. But again, this is only the result of a mathematical operation. The side-lobes would interfere with other frequencies side-lobes, but this is not typically referred to as "destructive interference" in the context of Fourier analysis. Instead, it's an aspect of the spectral leakage caused by the finite duration of the signal.

Finally:

Part of my confusion is probably because I read elsewhere on the internet that the DFT is lossless, but I am thinking now that must be wrong since all real signals are of finite time/infinite bandwidth, therefore all real signals must have an infinite sampling rate to be truly lossless. Is this correct?

Yes and no. The term "lossless" in the context of DFT refers to the fact that, theoretically, the DFT of a signal and its inverse can reconstruct the original signal exactly, without any information loss. But like I said earlier, in real-world scenarios, signals are band-limited, and practical sampling rates are chosen to capture the essential frequency content of these signals. So while it's theoretically true that an infinite sampling rate is required to capture all frequency components of a finite-duration signal, in practice, a sufficiently high sampling rate (like $44.1 \, \tt{kHz}$ for audio) is adequate to capture all the significant components within the human hearing range.