The image is taken from DSP Guide, where plot [a] is the time domain to synthesize [b] is the frequency domain where the amplitude denotes spectral density and [c] shows the actual magnitude of the sinusoidal amplitude.

To convert spectral density to sinusoidal amplitude (i.e. from [b] to [a]), we multiply the spectral density by 2/N bandwidth. For instance, frequency sample number 4 has spectral density of 32, multiplying 32 by 2/N (or 2/32) gives us 2 - the actual sinusoidal amplitude. But why?

The spectral density counts the number of times the signal falls within a unit bandwidth (i.e. the value 32 for frequency number 4 means there are 32 samples that have frequency between 3.5 and 4.5 ). What the spectral density tells us is the count or how probable a specific range of frequency embedded within a signal. I don't see how it has any correlation to the sinusoidal amplitude, or why multiplying by the bandwidth can magically converts a probability density or a count into a physical sinusoidal amplitude such as voltage.

enter image description here


1 Answer 1


That's not spectral density, it's just real DFT.

If you have a $60 Hz$ sine that lasts for 1 sec, its DFT will be half as large as for the same sine that lasts for 2 secs, since there's half the number of samples and thus half the (unnormalized) correlation: sum(product). To get the same value for both, we divide out the sample count.

Why multiplying by 2/N results in sinusoidal amplitude?

Because any (non-Nyqist/DC; see "special cases" in the article) bin will correlate with integer periods as $N/2$:

$$ \sum_{n=0}^{N-1} (A \cos(2\pi k n)) \cdot \cos(2\pi k n) = A (N/2) \tag{1} $$

This holds for real and complete DFT. The units are carried over with $A$.

The article doesn't discuss finding amplitude of a sinusoid in a general signal from DFT (for which DFT is ill-suited), however - it's rather about normalizing DFT: real has half the number of bins of complete DFT, so we double the normalization.

What the spectral density tells us is ... how probable a specific range of frequency embedded within a signal.

There's no need for probabilities in interpreting results for deterministic signals - but if we insist, the normalizations remain accurate.

If this still doesn't make sense, then it's about understanding how DFT works fundamentally. In addition to a closer look at DSP guide, I recommend this clip.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.