There was a similar question asked here, however I would like to focus on some specifics of this problem. Let me present the Python code sample to illustrate the situation when only a part of the period of the input signal is analyzed using FFT:

N = 1200
H = 4
fig = plt.figure()
sig = np.zeros(N)
for h in range(H):
        sig += [(h+1)*np.sin(2*np.pi*(h+1)*x/N) for x in (range(N))]
ax1, ax2 = (None, None)
for h in range(H):
        if (h):
                fig.add_subplot(H, 2, 2*h+1, sharex=ax1)
                ax1 = fig.add_subplot(H, 2, 2*h+1)
        if (h):
                fig.add_subplot(H, 2, 2*h+2, sharex=ax2)
                ax2 = fig.add_subplot(H, 2, 2*h+2)
        fft = pd.DataFrame()
        fft['coef'] = np.fft.fft(sig[:int(N/(h+1))])
        fft['freq'] = np.fft.fftfreq(int(N/(h+1)))
        fft['per']  = 1/fft['freq']
        fft['mag']  = fft['coef'].apply(lambda z: np.sqrt(z.real**2 + z.imag**2))
        rside = fft['freq'] > 0
        plt.bar(fft[rside]['per'], fft[rside]['mag'])

The parameter H specifies the number of composite sines, their magnitude and the number of divisions of the original signal. For each division we get 1 row of subplots: the input signal in the left column and the period spectrum in the right column. The X axes in each column are shared, so the plots are aligned. Only positive frequencies are considered. The example with H = 4 and the signal length at N = 1200 should produce the set of plots like on the picture below.

Period spectrum of fragmented input signal

The FFT results in the top row are spot on, but on other plots from the right column we can see that the spikes representing the correct dominant cycles are missing, but some additional smaller spikes appear at periods much smaller than the smallest one used to generate the test signal.

In a general case when given a sample of discrete observations or a time series data I would like to have a way to test how would some bigger cycles fit in the spectrum, ie. if we consider the Mercury's distance from the barycenter in that other thread someone suggested that orbital cycles of other planets should be considered, all of which are well known and some are longer than 8 years (the size of the input data). I was searching online, but due to lacks in theoretical (and practical) knowledge on the matter I probably wasn't using the right keywords, so I couldn't find any helpful information. These are the questions I would like to ask:

  • Can the small cycles that FFT 'found' in the period slices provide clues about the bigger ones that were actually in effect, ie. could there be a correlation in frequency or phase between the small fake cycles and the real big ones?
  • Is there a way to expand the period domain or in other words assume there are bigger cycles at play and include them in the spectrum? Can Fourier Transform be somehow manipulated to do so?
  • Lastly is Fourier Transform the right tool or at least a starting point to deal with such issue?
  • $\begingroup$ Have you looked into the continuous Fourier transform? Harder to compute, but allows for fractional periods, including periods less than 1. $\endgroup$
    – user7409
    Commented Apr 7, 2017 at 16:44
  • $\begingroup$ I'm not looking for periods smaller than 1, but for periods larger that the input signal length. Or did you mean frequencies? $\endgroup$
    – mac13k
    Commented Apr 7, 2017 at 18:43
  • $\begingroup$ Yes, that's probably what I meant. Any case where you have less than a full cosine wave. $\endgroup$
    – user7409
    Commented Apr 7, 2017 at 18:44

1 Answer 1


If you plot frequency instead of period, you might be able to see that some of the increasing artifacts are due to the shrinking rectangular window on your signal.

An FFT decomposes the entire length of its input into sinusoids. You have created a flat spot in your data input, and it takes a linear combination of a lot of sinusoids to represent those extended flat spots. So the FFT of your latter shorter signal segments becomes more and more about the flat spot than your signal.

Added: There are other methods for estimating the parameters of a known small number of sinusoids within a known distribution of noise, such as MUSIC and ESPIRIT, that may be more suitable for estimation of sine waves not equal to or possibly shorter than the lowest DFT basis vectors.

  • $\begingroup$ Of course what you wrote makes sense and it explains the example, but you didn't address any of the questions regarding finding or testing the input signal for the possible presence of larger cycles... $\endgroup$
    – mac13k
    Commented Apr 8, 2017 at 21:27
  • $\begingroup$ When it comes to very low frequencies, those whose periods are longer than your data, there are some techniques that are possible and were mentioned but acquiring those low frequencies is more a matter of the methods and instrumentation used in acquisition than the signal processing used to search for them. Very low frequencies requires very stable low drift amplification and high levels of mechanical isolation. The question that typically occurs, is if the low frequencies are important, why didn't you design your experiment to collect them properly. Garbage in garbage out is true $\endgroup$
    – user28715
    Commented Jul 8, 2017 at 18:01

Your Answer

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

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