I am analyzing a recording from a microelectrode array on a neuronal cell culture and I found something that is disturbing me a lot. Through python, I have taken different chunks of the same signal profile from the original recording, calculated the FFT of each of them and then, the mean among them.

In theory, this makes sense since I am just amplifying common patterns on the wave, however, I am getting a huge peak at the 0 component but if you calculate the mean of the signal, they have zero DC component. I don't know if this issue might be a consequence of a bad use of the FFT on python or maybe I am missing something. Do any of you have any idea of what is going on?

enter image description here

enter image description here

The code is very long but the specific section of the chunks and the fft is here:

chunk = np.zeros((16, 300000))
chunk[0,:] = signal1[320000:620000]
chunk[1,:] = signal1[950000:1250000]
chunk[2,:] = signal1[1640000:1940000]
chunk[3,:] = signal1[2311000:2611000]
chunk[4,:] = signal2[320000:620000]
chunk[5,:] = signal2[950000:1250000]
chunk[6,:] = signal2[1640000:1940000]
chunk[7,:] = signal2[2311000:2611000]
chunk[8,:] = signal3[320000:620000]
chunk[9,:] = signal3[950000:1250000]
chunk[10,:] = signal3[1640000:1940000]
chunk[11,:] = signal3[2311000:2611000]
chunk[12,:] = signal4[320000:620000]
chunk[13,:] = signal4[950000:1250000]
chunk[14,:] = signal4[1640000:1940000]
chunk[15,:] = signal4[2340000:2640000]

myfftchunk0 = np.fft.fft(chunk[0,:])
myfftchunk1 = np.fft.fft(chunk[1,:])
myfftchunk2 = np.fft.fft(chunk[2,:])
myfftchunk3 = np.fft.fft(chunk[3,:])
myfftchunk4 = np.fft.fft(chunk[4,:])
myfftchunk5 = np.fft.fft(chunk[5,:])
myfftchunk6 = np.fft.fft(chunk[6,:])
myfftchunk7 = np.fft.fft(chunk[7,:])
myfftchunk8 = np.fft.fft(chunk[8,:])
myfftchunk9 = np.fft.fft(chunk[9,:])
myfftchunk10 = np.fft.fft(chunk[10,:])
myfftchunk11 = np.fft.fft(chunk[11,:])
myfftchunk12 = np.fft.fft(chunk[12,:])
myfftchunk13 = np.fft.fft(chunk[13,:])
myfftchunk14 = np.fft.fft(chunk[14,:])
myfftchunk15 = np.fft.fft(chunk[15,:])

Meansignal = np.zeros(len(chunk[1, :]))

for i in range(len(chunk[1, :])):
    Meansignal[i] = myfftchunk0[i] + myfftchunk1[i] + myfftchunk2[i] + myfftchunk3[i] + myfftchunk4[i] + myfftchunk5[i] + myfftchunk6[i] + myfftchunk7[i] + myfftchunk8[i] + myfftchunk9[i] + myfftchunk10[i] + myfftchunk11[i] + myfftchunk12[i] + myfftchunk13[i] + myfftchunk14[i] + myfftchunk15[i]
    Meansignal[i] = Meansignal[i] / 16

mynewffthalfMean = Meansignal[:len(Meansignal) // 2]
fnewHalfMean = np.linspace(0, fs / 2, len(mynewffthalfMean))

# Plot figure 2 - FFT
plt.plot(fnewHalfMean, np.abs(abs(mynewffthalfMean))**2)
plt.xlabel('Frequency (Hz)')
plt.ylabel('Magnitude (dB)')

1 Answer 1


You cannot simultaneously have a non-zero value at the first bin of the DFT (the 0 component) and a zero-mean since the first component is by definition of the DFT the (scaled) mean of the signal:

from the DFT:

$$X[k] = \sum_{n=0}^{N-1}x[n]e^{-j2\pi nk/N}$$

$$X[0] = \sum_{n=0}^{N-1}x[n]e^{j0}= \sum_{n=0}^{N-1}x[n] = N\mu_x$$

So there is a mean in the digital data, and perhaps the OP's confusion is there is not a mean of any significance in the analog waveform being sampled. There are several opportunities to introduce offsets notably in the ADC converter itself, or it may be as simple as the mapped amplitude range of the digital waveform includes an offset. But taking the mean of the data that is being FFT'd should be consistent with what is above, with further scaling by any windowing that may be done according to $\sum_{n=0}^{N-1}w_n$.

  • $\begingroup$ Thanks for you answer. First, I have check out the mean of the signal in real time, it is around -3.56 which is not significant. I have check out the other channels they also have an offset around -3, but again, I don't know how this tiny offset value gets that amplified on the FFT 0 component. Would it be a good step to remove the offset of the signal and then, compute the DFT? $\endgroup$ Sep 25, 2020 at 7:16
  • $\begingroup$ Just to close this thread: I tried to remove again the DC component and even that it was very little, that was the caused of the high DC component in the frequency spectrum. Thanks a lot! $\endgroup$ Sep 25, 2020 at 12:21
  • $\begingroup$ @\Mufasa thanks for the confirmation. To close, please mark this answer as "correct" unless you feel it is incomplete. $\endgroup$ Sep 25, 2020 at 12:28

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.