2
$\begingroup$

Fourier decomposition of a mixed signal can straightforwardly give me the frequencies of the different components and their relative amplitudes, but how can I extract the components of a mixed signal if they have the same frequency and differ only in phase?

I have a data series consisting of a metric of global activity in a certain domain (e.g. volume of Google searches). There is a strong diurnal rhythm in the data, i.e. activity is high during certain hours and low during other hours, on a 24-hour cycle. The peaks and troughs of the activity occur respectively during the day and night periods in North America.

I suspect that, while the North American contribution dominates, there are smaller contributions to the overall global activity from other time-zones. These would follow the same 24-hour cycle but would be shifted by a certain number of hours and have smaller amplitudes.

My problem is to identify the amplitude of the signal centred on each time-zone. I know that the relative amplitudes of the different phased signals affect the balance between real and imaginary components of the complex Fourier coefficient, but can I go backwards from the Fourier analysis to extract the amplitudes of the different-phased contributions? Or is there something similar to Fourier analysis that will do this?

To put it another way, I have a signal of the form $S(t)=\sum_{i=1}^{24}a_i cos⁡(ωt+2πi/24)+\epsilon(t)$ (where $\epsilon(t)$ is noise), and my problem is to find the $a_i$. I have thousands of data points covering multiple cycles, so if there is nothing like Fourier analysis that can do it, instead I could treat this as an overdetermined set of simultaneous equations in 24 unknowns. Is that the best or only way to approach the problem?

$\endgroup$
3
  • 1
    $\begingroup$ That's pretty much impossible without additional assumptions even if the noise is absent since $\cos(x+a)+\cos(x-a)=2\cos a\cos x$, say, so how on Earth can you hope to distinguish the combination of two signals from North America and Asia from one from Europe? $\endgroup$
    – fedja
    Sep 25, 2016 at 14:29
  • 1
    $\begingroup$ $\cos(\omega t + 2\pi i / 24) = \cos(\omega t) \cos(2\pi i / 24) - \sin(\omega t) \sin(2\pi i / 24)$. Your Fourier transform should give you both the cosine series and the sine series. You have now 24 variables ($a_1 \ldots a_{24}$) and 2 equations, which is vastly underdetermined. Now: it is absolutely necessary that the equations are underdetermined, no sort of signal analysis can ever reveal to you how to separate two signals that are exactly 180-degrees out of phase. $\endgroup$ Sep 25, 2016 at 14:29
  • 1
    $\begingroup$ Right, so you seem to be saying that, given a waveform known to be composed of 24 phase-shifted signals, there is no unique solution for the amplitudes of the 24 signals, and in fact there are 22 degrees of freedom available. $\endgroup$
    – Parsifal
    Sep 26, 2016 at 15:28

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.