By understanding the Shannon-Nyquist theorem. If you want to run a pointless experiment, feel free, but you haven't shown any sign of wanting to either actually experiment or actually understand the very basics of PCM. Don't give me homework to labor over whatever piece of lint flies off your keyboard.
This assumes that all of this is sincere, but I'll be pretty clear that I don't believe that's the case.
This was sincere. And very concrete. The exercise shows that the Whittaker–Nyquist–Kotelnikov–Shannon theorem doesn't literally apply to PCM encoding, because the theorem operates on
infinitely precise values, defined at discrete time points.
Infinitely precise value is a mathematical abstraction. It doesn't really exist. If it did, we could encode a description of the entire universe in just one number between 0.0 and 1.0, using digits from 0 to 9, placed in a sequence after the decimal point. PCM, unless it has infinite bit depth, can't encode a value with infinite precision.
With PCM, you have to take into account the precision of the encoding of the dimensions: such as frequency, amplitude, and phase. Assuming that the three PCM samples you encoded contain 24 x 3 = 72 bits, that's the core amount of information the decoder would have to work with. Additional information provided to the decoder is that there are two (two bits worth if information) band-limited sinusoids (sets the context for the Math operational in this case).
It is those 72 + 2 = 74 bits of information, plus the context of applicable decoding algorithms, that the decoder has to work with. Using DSP tricks, the decoder can trade the precision of determining amplitudes for the precision of determining the frequencies, trade both or one by one for the precision of determining the phases, yet it can't get something for nothing: the overall amount of information will not increase.
You can apply a straight discrete Fourier transform to the three samples, which will keep the amplitude precision, yet you'll end up with just three frequency bins. If your original sinusoids were, say of 100 Hz and 10,000 Hz, we could deduce from the transform graph, with decent precision, the energies that would go to tweeter and woofer of a three-way speaker.
Let's say we are not that interested in the energies, and want to know instead whether the signal contains musically consonant or musically dissonant sinusoids. In this case, we would apply a discrete Fourier transform after padding the samples, with, say, 4,093 zero-valued samples. That would allow us to resolve the frequencies much better, yet the precision of amplitude's determination would go down.
The human hearing system doesn't operate exactly like the Fourier transform, yet it is not excepted from the fundamental laws of Information Theory. Based on just those three samples, a human would also have trouble determining what that signal represents. Which brings us to the evolutionary value of the transients detection, which employs specialized neural structures, such as the Octopus Cells.
Even just one sample with sufficient amplitude, presented in the absence of any other audio information, could quickly tell the hearing system, and ultimately the decision-making part of the brain, that
something is happening. This one bit of timely information could be just enough for an animal to stop doing whatever it was doing, and start listening, looking, and sniffing more attentively, or simply quickly leave the scene without further ado.
If you want to increase the amount of information transferred via a sequence of samples of fixed bit depth, you have to supply more samples. This was my original point. The paper by Stuart and Craven makes it more concrete, by showing how many more samples, on average over a set of quasi-realistic music signals, you'd need to provide in order to achieve a desired amplitude precision.
The required number of samples, for a target amplitude precision, depends on the particulars of the PCM encoding (such as dithering algorithm). It also depends on the sampling rate. Higher sampling rate, even if the number of samples needed to achieve the target precision was exactly the same, results in a shorter required physical time.