I want to normalize my FFT signal. From this page it says that we can normalize it by dividing the FFT result by the lenght of the signal in time domain.

On the other hand, my supervisor told me that to normalize it, I need to divide the FFT by the sampling frequency. Both methods don't deliver the same results, as you can see in my plots.

EDIT: now I plotted the signal in time domain, in frequency domain and the Power spectrum of my signal to show my point enter image description here

Which one is the right method ? And what is the physical meaning of those methods for normalizing?

Here is my code in python:

def PS_normfreq(time_signal, f_sampling):
    ''' Calculate the Power spectrum of a signal in time domain'''
    fft = np.fft.fft(time_signal) / f_sampling
    PS = np.real(fft * np.conjugate(fft))
    return PS

def PS_normlen(time_signal):
    ''' Calculate the Power spectrum of a signal in time domain'''
    fft = np.fft.fft(time_signal) / len(time_signal)
    PS = np.real(fft * np.conjugate(fft))
    return PS

def FFT_normfreq(time_signal, f_sampling):
    ''' Calculate the Power spectrum of a signal in time domain'''
    fft = np.fft.fft(time_signal) / f_sampling
    return fft

def FFT_normlen(time_signal):
    ''' Calculate the Power spectrum of a signal in time domain'''
    fft = np.fft.fft(time_signal) / len(time_signal)
    return fft

# ensure that the normalization is the same using two different methods (trought sampling frequency and through lenght of the array)
# Define a simple sinusoidal signal
t = np.linspace(0, 2, 500, endpoint=False)  #  start, stop, number
f_signal = 5.0  # Frequency of the sinusoidal signal
signal_test = np.sin(2 * np.pi * f_signal * t)

# Sampling frequency
f_sampling = 1/(t[1]-t[0])  # Example, 50 samples taken evey seconds

# Calculate power spectrum using both versions
ps_version1 = PS_normfreq(signal_test, f_sampling)
ps_version2 = PS_normlen(signal_test)

# Calculate power spectrum using both versions
fft_version1 = FFT_normfreq(signal_test, f_sampling)
fft_version2 = FFT_normlen(signal_test)

# plot the Power spectrum
plt.plot(t, signal_test, label = 'signal')
plt.title('Signal in time domain')
plt.xlabel('Time [s]')

# plot the Fourier transform
plt.plot(np.fft.fftfreq(len(fft_version1), 1/f_sampling), np.real(fft_version1), label = 'norm freq')
plt.plot(np.fft.fftfreq(len(fft_version1), 1/f_sampling), np.real(fft_version2), label = 'norm lenght')
plt.title('Fourier transform')
plt.xlabel('Frequency [Hz]')

# plot the Power spectrum
plt.plot(np.fft.fftshift(np.fft.fftfreq(len(fft_version1), 1/f_sampling)), np.fft.fftshift(ps_version1), label = 'norm freq')
plt.plot(np.fft.fftshift(np.fft.fftfreq(len(fft_version1), 1/f_sampling)), np.fft.fftshift(ps_version2), label = 'norm lenght')
plt.title('Power spectrum')
plt.xlabel('Frequency [Hz]')

# Compare the results
#print("Power Spectrum (Version 1):", ps_version1)
#print("Power Spectrum (Version 2):", ps_version2)
  • 1
    $\begingroup$ f_sampling should be 50, not 1/50. See my answer ;) $\endgroup$
    – Jdip
    Dec 11, 2023 at 18:40
  • $\begingroup$ How are you wanting to normalize? To 0 dB? So that the amplitudes in the time and Fourier domain are equal? $\endgroup$
    – Baddioes
    Dec 11, 2023 at 20:54
  • $\begingroup$ I'm don't know, I just know I need to normalize it but don't understand how to do it and what's the different methods $\endgroup$
    – Apinorr
    Dec 14, 2023 at 9:53
  • 1
    $\begingroup$ Please see my updated answer. Your supervisor is wrong. There is no case where normalizing the DFT by the sampling frequency makes sense. $\endgroup$
    – Jdip
    Dec 14, 2023 at 15:43
  • $\begingroup$ I talked with my supervisor today and he told me that I was indeed computing the PSD $\endgroup$
    – Apinorr
    Dec 15, 2023 at 13:31

2 Answers 2


It's not clear how you are wanting to normalize the function. If you are wanting to normalize the spectrum so that the amplitude of the function in the time domain matches the amplitude of the spectrum in the frequency domain, this depends on if the function you are analyzing is real or complex. I'll show a demonstration for a cosine wave.


\begin{equation}X[k] = \sum_{n=0}^{N-1}A\cos(\frac{2\pi f_{0}n}{N})e^{-j2\pi\frac{kn}{N}}\end{equation}

Using Euler's formula, we can rewrite this as

\begin{equation}X[k] = A\sum_{n=0}^{N-1}(\frac{e^{j2\pi\frac{f_{0}n}{N}}+e^{-j2\pi\frac{f_{0}n}{N}}}{2})e^{-j2\pi\frac{kn}{N}}\end{equation}

Expanding this, we get

\begin{equation}X[k] = \frac{A}{2}\sum_{n=0}^{N-1}e^{j2\pi\frac{f_{0}n}{N}}e^{-j2\pi\frac{kn}{N}}+\frac{A}{2}\sum_{n=0}^{N-1}e^{-j2\pi\frac{f_{0}n}{N}}e^{-j2\pi\frac{kn}{N}}\end{equation}

Simplifying, this becomes

\begin{equation}X[k] = \frac{A}{2}\sum_{n=0}^{N-1}e^{-j2\pi\frac{(k-f_{0})n}{N}}+\frac{A}{2}\sum_{n=0}^{N-1}e^{-j2\pi\frac{(k+f_{0})n}{N}}\end{equation}

Each complex exponential has a magnitude of one, so

\begin{equation}\sum_{n=0}^{N-1}1 = N\end{equation}

This means the amplitude of each delta after the Fourier transform is applied is $\frac{AN}{2}$, so to normalize back to A, you need to multiply by $\frac{2}{N}$. A similar proof shows that for complex signals you multiply by $\frac{1}{N}$.

This leads into a discussion of the power spectral density (PSD) as mentioned by the other commenter. That definition of a PSD seems to be incorrect, assuming $f_{s}$ is the sample rate which it appears to be. There are two definitions given by Stoica and Moses in "Spectral Analysis of Signals". They are

\begin{equation} \phi_{1}(\omega) = \sum_{k=-\infty}^{\infty}r(k)e^{-j\omega k}\end{equation}


\begin{equation} \phi_{2}(\omega) = \lim_{N \to \infty}E\{{\frac{1}{N}\lvert\sum_{t=0}^{N-1}y(t)e^{-j\omega t}\rvert^{2}}\}\end{equation}

where $r(k)$ is the autocorrelation sequence (see eqs 1.3.7 and 1.3.10). These are shown to be equivalent under the assumption that $r(k)$ decays sufficiently quickly, ie

\begin{equation}\lim_{N \to \infty}\sum_{k=-N}^{N}|k|r(k)e^{-j\omega k} = 0\end{equation}

(see equations 1.3.11-1.3.17).

The definition of the PSD listed by the other user (a) assumes the signal being analyzed is real (ie sines and cosines not complex exponentials), and (b) makes the common misconception of how to normalize the PSD. Dividing by the sample rate squashes everything. The correct way to normalize the PSD is by the bin width.

Notice that both definitions of the PSD are defined for continuous spectra, ie, the PSD is the DTFT of the autocorrelation sequence, not the DFT, which means the PSD is a continuous spectrum. However, it is obviously impossible to compute a true continuous spectrum, so we approximate it using the DFT. If our sample rate is $f_{s}$ and our signal is length $N$, then our sample points in the spectrum will lie at integer multiples of the fundamental frequency $\delta_{f} = \frac{f_{s}}{N}$. For a frequency $f_{0}$ being one of the integer multiples of the fundamental frequency, the power spectrum accumulates all of the power in the range $[f_{0}-\frac{\delta_{f}}{2},f_{0}-\frac{\delta_{f}}{2})$ into a single frequency bin, much like an FFT accumulates the amplitude into a single frequency bin. This gives the total power of the spectrum across a finite frequency band.

To normalize to get a PSD, we have to divide by the bandwidth (ie the bin width $\delta_{f}$) across which the power has been accumulated. This gives us an average power spectral density, ie the average power contained across a finite bandwidth. This is because the assumption is that random signals have finite average power. Random signals do not have finite energy as they are assumed second order stationary, and therefore do not have DTFTs that converge. However, second order stationary random signals have finite average power, hence the need for a PSD estimate via the DTFT of the autocorrelation sequence. This is also the reason for the limit in the Periodogram estimate (see the discussion at the beginning of section 1.3 in the Stoica and Moses book mentioned above).

Hopefully this clarifies some things!

EDIT: Here is some Matlab code to show the normalization in the first part of my post.

% Define Frequency location
f1 = 1;
f2 = 3;

% Sample rate
fs = 10;

% Signal length
N = 100;

% Frequency resolution
df = fs/N;

% Define frequency locations
w = -fs/2:df:fs/2-df;

% Define signals
t = 0:N-1;
A = 10*cos(2*pi*f1/fs*t);
B = 10*exp(1i*2*pi*f2/fs*t);

% Plot signals
figure; plot(w,abs(fftshift(fft(A))))
hold on;
hold off;
legend('Cosine','Complex Exponential')

figure; plot(w,abs(fftshift(fft(A*2/N))))
hold on;
hold off;
legend('Normalized Cosine','Normalized Complex Exponential')

Un-normalized vs. Normalized Sinusoids

Additionally, if you want to normalize to 0 dB, calculate the FFT, divide by the maximum absolute value, and then take the log and plot.

EDIT 2: I had to go back and review some notation for this. It turns out I had a couple misunderstandings as well. As previously, "Spectral Analysis of Signals" will be the reference for all the derivations.

TL;DR: The PSD does not (in general) give the power content of a signal at frequency $\omega$. It instead gives the power at $\omega$ in the signal's autocorrelation sequence.

Full breakdown: It's not a super long proof but an important one. Let's define the autocorrelation as

\begin{equation}r(k) = E\{y(t)y^{*}(t-k)\}\end{equation}

The PSD is then defined (as noted previously) as

\begin{equation}\phi(\omega) = \sum_{k=-\infty}^{\infty}r(k)e^{-j\omega k}\end{equation}

Now let's define a new autocorrelation sequence as follows

\begin{equation}r'(k) = E\{y(t)y^{*}(t+k)\}\end{equation}

This is obviously very similar to the definition of $r(k)$, however, it is an important difference to note. Now, if we define a new PSD based on this estimate of the autocorrelation sequence, we get

\begin{equation}\phi'(\omega) = \sum_{k=-\infty}^{\infty}r’(k)e^{-j\omega k} = \sum_{k=-\infty}^{\infty}r(k)e^{j\omega k} = \phi(-\omega)\end{equation}

If the PSD represented the power in the signal, we would have gotten $\phi(\omega) = \phi'(\omega)$, because the signal's spectral content shouldn't be dependent on it's autocorrelation sequence. However, the PSD does clearly depend on the definition of the signal's autocorrelation sequence, so it cannot represent the power in the signal. It may be enough in the case that the original signal is real to say that the PSD relates to the power in the signal, and therefore has a relationship with the power spectrum. However, this is not generally true, as $\phi(\omega) \neq \phi(-\omega)$ for complex valued signals. I also noted before that we wanted to compute an average power spectral density as random signals have finite average power. This is given by the second definition of the PSD originally written.

Therefore, the frequency averaged squared magnitude spectrum is a PSD estimate, not a power spectrum estimate. The proof as I see it is as follows. Let the frequency averaged squared magnitude spectrum be

\begin{equation} \phi(\omega) = \frac{1}{N}\lvert Y(\omega)\rvert^{2} = \frac{1}{N}Y(\omega)Y^{*}(\omega)\end{equation}

Taking the inverse Fourier transform, we get something proportional to


This is a simplified definition of the autocorrelation sequence. Therefore, the average squared magnitude spectrum is a scaled PSD, that again describes the power contained in the autocorrelation sequence, not the signal itself. I'm not sure if there is a proof for the relation between the power spectrum and the power spectral density for real valued signals, but it might only need to be that $\phi(\omega) = \phi(-\omega)$ for real spectra.

I edited my earlier response in light of this new information.

1 Stoica, P., & Moses, R. L. (2005). Spectral analysis of signals (Vol. 452, pp. 25-26). Upper Saddle River, NJ: Pearson Prentice Hall.

  • $\begingroup$ "To normalize to get a PSD, we have to divide by the bandwidth $d_f$". To go from Power Spectrum to PSD, we have to divide by the effective noise-equivalent bandwidth $\text{ENBW} = \text{NENBW}\cdot d_f$, with $\text{NENBW} = N\frac{S_2}{S_1}$. See Heinzel et al $\endgroup$
    – Jdip
    Dec 11, 2023 at 22:53
  • $\begingroup$ My discussion was not including windowing. Note that for no window being used, ie $w_{j} = 1 \forall j \in [0,N-1]$, then $NENBW*f_{res} = \frac{f_{s}}{N}*N\frac{S_{2}}{(S_{1})^{2}} = \frac{f_{s}}{N}$. But, generally speaking yes, if you use windowing, that is the equation. $\endgroup$
    – Baddioes
    Dec 11, 2023 at 22:57
  • $\begingroup$ You discussion is including windowing. It's a rectangular window. The normalization mentioned in the reference is to go from a normalized Power Spectrum $\text{NPS} = \frac{2|X|^2}{N^2}$ to a PSD, in which case $\text{PSD} = \frac{\text{NPS}}{\text{ENBW}}$ $\endgroup$
    – Jdip
    Dec 11, 2023 at 23:02
  • 1
    $\begingroup$ See my updated answer under the "Edit 2" section. $\endgroup$
    – Baddioes
    Dec 12, 2023 at 7:40
  • 1
    $\begingroup$ Interesting! I'll take some time and go through it when I have the bandwidth. Great to have you as a new contributor :) $\endgroup$
    – Jdip
    Dec 12, 2023 at 7:43

You should look at this answer for more detail, but I'll summarize it here for you:

Denote by $|X|^2$ the frequency averaged Squared Magnitude Spectrum and $w$ the window applied.

  • The PSD is: $$\frac{2|X|^2}{f_s \times S_2} \quad \texttt{with} \quad S_2 = \sum_{i = 0}^{N - 1} w_i^2$$
  • The Power Spectrum is: $$\frac{2|X|^2}{(S_1)^2} \quad \texttt{with} \quad S_1 = \sum_{i = 0}^{N - 1} w_i$$

For your purposes, $w$ is a rectangular window. Here would be the updated code depending on which method you want to use. Please note that your sampling frequency was wrong.

import numpy as np
import matplotlib.pyplot as plt

# Define a simple sinusoidal signal
t = np.linspace(0, 2, 100, endpoint=False)  #  start, stop, number, from 0 to 2 seconds
f_signal = 5.0  # Frequency of the sinusoidal signal
signal_test = np.sin(2 * np.pi * f_signal * t)

# Sampling frequency
f_sampling = 1/(t[1]-t[0]) 

def PS(time_signal, f_sampling, method='ps'):
    fft = np.fft.fft(time_signal)
    mag_squared = np.real(fft * np.conjugate(fft))
    f = np.fft.fftfreq(len(time_signal), 1/f_sampling)

    if method == 'psd':
        scaling_factor = 2 / (f_sampling * len(time_signal))
        scaling_factor = 2 / (len(time_signal)**2) 

    PS = scaling_factor * mag_squared
    return f, PS

# Calculate power spectrum using both versions
f, ps_version1 = PS(signal_test, f_sampling, 'psd') # Power Spectral Density scaling
f, ps_version2 = PS(signal_test, f_sampling, 'ps')  # Power Spectrum scaling

plt.plot(f[0:len(f)//2-1], ps_version1[0:len(f)//2-1])
plt.plot(f[0:len(f)//2-1], ps_version2[0:len(f)//2-1])
plt.legend(['PSD', 'Power Spectrum'])
plt.ylabel('Power (dB)')
plt.xlabel('Frequency (Hz)')

enter image description here

Edit: fft normalization

To normalize the DFT, in general you would use the Power Spectrum normalization, except remove the $.^2$: $$\frac{2|X|}{S_1}$$

  • $\begingroup$ This definition of the PSD is not quite correct. See my answer. $\endgroup$
    – Baddioes
    Dec 11, 2023 at 22:34
  • $\begingroup$ These scaling factors are used by Matlab, Scipy, etc. Now that does not mean they can't be wrong, but for practical purposes that is what they use. See my comment on your answer for more detail. $\endgroup$
    – Jdip
    Dec 11, 2023 at 22:42
  • $\begingroup$ It's not showing up for me yet, I'll take a look when it does though. $\endgroup$
    – Baddioes
    Dec 11, 2023 at 22:47
  • $\begingroup$ Oh thanks I corrected my sampling frequency indeed. However, I don't see how this is answering my question :/ I didn't ask about the power spectrum nor the power spectral density but the normalization of the Fourier transform $\endgroup$
    – Apinorr
    Dec 12, 2023 at 10:35
  • $\begingroup$ In your code you are computing the power spectrum… so I gave you the normalization for the power spectrum. $\endgroup$
    – Jdip
    Dec 12, 2023 at 17:16

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.