Very briefly, Nyquist's theorem states that, when sampling at a given rate, the highest frequency that can appear in the sampled signal is half the sampling frequency.

If the sampled signal contains frequencies higher than half the sampling frequency (higher than 4 KHz when sampling at 8 KHz as is the case for µ-law), these higher frequencies will appear folded down to below half the sampling frequency when the signal is reconstructed. This is the aliasing problem.

A visual example of the same phenomenon is when you see wheels turning backwards (e.g. train in movie).

In the present applet, the highest harmonic I use is the 5th. So I have to keep the fundamentals below 750 Hz (5x750 is getting close to 4 KHz).