I finally got an answer, which of course is derived from a model (inspired by Engineering).
But at least here is an answer, which I could not get before, and I got it reading the Neural Engineering book.
Basically an estimate of the number of neurons is a function of:
- the noise leve of the neurons
- the required precision of the representation
- the required range of the representation
There is also a formula, which is derived using calculus.