I'm currently computing the spectrogram with the matplotlib. I specify NFFT=512 but the resulting image has a height of 257. I then tried to just do a STFT (short time fourier transform) which gives me 512 dimensional vectors (as expected). If I plot the result of the STFT I can see that half of the 512 values are just mirrored so really I only get 257 values (like the matplotlib). Can somebody explain to me why that is the case? I always thought of the FT as a basis transform, why would it introduce this redundancy?
Thank you.
The redundancy is because you input a strictly real signal to your FFT, thus the DFT result is complex conjugate (Hermitian) symmetric. This redundancy is due to the fact that all the imaginary components of strictly real input are zero. But the output of this DFT can include non-zero imaginary components to indicate phase. Thus, the this DFT result has to be conjugate symmetric so that all the imaginary components in the result will cancel out between the two DFT result halves (same magnitudes, but opposite phases), indicating strictly real input. Also, the lower 257 bins of the basis transform will have 512 degrees of (scaler)freedom, just like the input. However, a spectrogram throws away all phase information, so it can only display 257 unique values (magnitude-only).
If you input a complex (quadrature, for instance) signal to a DFT, then there would likely not be Hermitian redundancy, and you would have 1024 degrees of freedom from a 512 length DFT.
If you want an image height of 512 (given real input), try an FFT size of 1024.