I'm the OP and after four years I'm sure I have enough knowledge to answer the question in terms of acoustics as well as signal processing.
My original question is a misunderstanding of the formant frequency. The key is, the formant frequencies of a specific vowel from the same person is approximately fixed (if we assume the same tongue shape), that is, the formant frequency is a fixed value but NOT a nth order harmonic.
To extend the explanation, a little more acoustical analysis is needed. A certain tongue shape in a given vocal tract can be approximated by a certain vocal tract area function. Given the vocal tract area function, we can solve the wave equation and then we know that in which frequencies the vocal tract maximizes the wave's amplitude. That's the formant frequency values. That is why formant frequency is a fixed value if the tongue shape is given.
Now let's consider the source. In an ideal condition, the wave contains and only contains waves whose frequencies are integer multiples of the fundamental frequency. Let's say a vowel has F1=800Hz F2=1550Hz, and a source with F0=200Hz. Since the source contains only waves with frequencies 200Hz, 400Hz, ..., 800Hz, ..., 1400Hz, 1600Hz, ... Therefore, the amplitude of the 800Hz wave is gained locally maximally, and since there's no 1550Hz wave to be enhanced, the 1400Hz and 1600Hz wave are enhanced maximally(the formant frequencies are only finite locally maximal peaks and the waves with frequencies near the formant frequencies are also enhanced by the vocal tract in the real cases), but maybe larger on the 1600Hz wave.
When LPC algorithm is implemented on this output, we may believe the formants of the vowel are 800Hz and 1550Hz, or we might say the F2 value is between 1400Hz and 1600Hz, but we can not know its specific value, though in most cases, we also don't want to know that.
Anyway, in terms of source-filter theory, source and vocal tract shape are independent. F0 is source-dependent, and formant frequencies are vocal-tract-shape-dependent. That's all.