Autcorrelation Via FFT and Zero Padding

Hi,

I would like to understand conceptually why when calculating the Autocorrelation Function of a vector x via fft, the best way is to zero padding in the way

nFFT = 2^(nextpow2(length(x))+1); % nFFT = 2*length(x) will do it as well

F = fft(x-mean(x),nFFT);

and no, for instance

nFFT = 2^(nextpow2(length(y)));

Why do the way we pad the vector influence so much the approximation of the Autocorrelation?

Thank you.