Plot FFT as a set of sine waves in python?

Question

I saw someone do this in a presentation but I\'m having a hard time reproducing what he was able to do. Here\'s a slide from his presentation:

Pretty cool.

Answer 1

The discrete Fourier transform gives you the coefficients of complex exponentials that, when summed together, produce the original discrete signal. In particular, the k'th Fourier coefficient gives you information about the amplitude of the sinusoid that has k cycles over the given number of samples.

Note that since your sines don't have integer numbers of cycles in 1000 samples, you actually will not be able to retrieve your original sine waves using an FFT. Instead you'll get a blend of many different sinusoids, including a constant component of ~.4.

You can plot the various component sinusoids and observe that their sum is the original signal using the following code:

freqs = np.fft.fftfreq(len(x),.01)
threshold = 0.0
recomb = np.zeros((len(x),))
for i in range(len(fft3)):
    if abs(fft3[i])/(len(x)) > threshold:
        recomb += 1/(len(x))*(fft3[i].real*np.cos(freqs[i]*2*np.pi*x)-fft3[i].imag*np.sin(freqs[i]*2*np.pi*x))
        plt.plot(x,1/(len(x))*(fft3[i].real*np.cos(freqs[i]*2*np.pi*x)-fft3[i].imag*np.sin(freqs[i]*2*np.pi*x)))
plt.show()

plt.plot(x,recomb,x,sin3)
plt.show()

By changing threshold, you can also choose to exclude sinusoids of low power and see how that affects the final reconstruction.

EDIT: There's a bit of a trap in the above code, although it's not wrong. It hides the inherent symmetry of the DFT for real signals, and plots each of the sinusoids twice at half of their true amplitude. This code is more performant and plots the sinusoids at their correct amplitude:

freqs = np.fft.fftfreq(len(x),.01)
threshold = 0.0
recomb = np.zeros((len(x),))
middle = len(x)//2 + 1
for i in range(middle):
    if abs(fft3[i])/(len(x)) > threshold:
        if i == 0:
            coeff = 2
        else:
            coeff = 1
        sinusoid = 1/(len(x)*coeff/2)*(abs(fft3[i])*np.cos(freqs[i]*2*np.pi*x+cmath.phase(fft3[i])))
        recomb += sinusoid
        plt.plot(x,sinusoid)
plt.show()

plt.plot(x,recomb,x,sin3)
plt.show()

If in the general case you know that the signal was composed of some subset of sinusoids with frequencies that may not line up correctly with the length of the signal, you may be able to identify the frequencies by either zero-padding or extending your signal. You can learn more about that here. If the signals are completely arbitrary and you're just interested in looking at component sinusoids, there's no need for that.