# Obtaining power spectrum from ACF, FFT using Matlab and FFTW

I am using Matlab R2012b 64-bit on Windows 7 in order to estimate the power spectrum of a simple signal that is:

$$\cos(10t) + \sin(20t)$$defined in the time interval from 0.0 to 10.0

Here is what I did:

1) There was no predefined value of points so I took 1001 points by forming a time array and sampled the function at these points.

2) I calculated the autocorrelation function of the signal with the built-in function "xcorr".

3) I performed FFT directly on the autocorrelation function with no additional options.

4) The resulting data set had imaginary parts(I think this should not be the case since the autocorrelation function is even).

5) I took the absolute value of the FFT of ACF and plotted it as the power spectrum from the ACF after having it shifted with fftshift.

6) I took the FFT of the sinusoidal signal after padding it with 1000 (n-1) zeros.

7) I calculated the power spectrum from the FFT of the signal as the square of the absolute value of the frequency coefficients.

8) I plotted both of them and compared them by assigning an array to difference between them.

9) They turned out to be the same.

Everything seems to be working correctly but my concerns are as follows:

• How can this data be interpreted? I expected the sum and difference frequencies of the individual sinusoidal functions in the signal.

-How can I normalize the output and represent negative frequencies properly since they have no physical significance?

-Does Matlab use FFTW libraries? My main task is to achieve the same results in a C program using these libraries.

I have benefitted from the following URL: Efficiently calculating autocorrelation using FFTs in organizing the post and achieving these results.

I am also confused about the terms power spectrum, power spectral density and energy spectrum. Are they used interchangeably? I would be grateful for an explanation of this phenomenon and normalization procedure.

P.S : The code snippet I have used is as follows and it can reproduce my results:

t=transpose(0:0.01:10);
n=size(t,1);
f=cos(10*t)+sin(20*t);
acf=xcorr(f,f);

FFTacf=fft(acf);
PSacf=abs(FFTacf);

FFTp=fft([f;zeros(n-1,1)]);
PSf=FFTp.*conj(FFTp);

figure;
subplot(2,1,1);
plot(fftshift(PSf));
title('PS from FFT');
subplot(2,1,2);
plot(fftshift(PSacf));
title('PS from ACF');

d=PSacf-PSf;
fprintf('Max error = %6.2f \n', max(abs(d)));


I am also confused about the terms power spectrum, power spectral density and energy spectrum. Are they used interchangeably? I would be grateful for an explanation of this phenomenon and normalization procedure.

EDIT with @jojek`s helpful comment:

power spectrum is NOT the same as power spectral density despite the fact that terms are often mixed up. Units of power spectral density (PSD) are $x^{2}$/Hz, e.g. $m^{2}s^{-2}$/Hz. Power Spectrum (PS) is equal to PSD*(Equivalent Noise Bandwidth of used window funktion) and has units of $x^{2}$. You often normalise it with N, the number of data points (for discrete time series). For physical processes the PSD and PS are symmetric, thus negative frequencies contain no additional information and only positive frequencies are plotted, negative frequencies are discarded.

energy spectrum is a term of spectroscopy: it describes the intensity of a particle beam as a function of particle energy. What you maybe meant is called energy spectral density, which describes the distribution of energy over frequency; it has units of J/Hz

A good overview is here.

I want to add the term amplitude spectrum/amplitude spectral density, which is the square root of the power spectral density.

• I disagree that Power Spectrum and Power Spectral Density are the same thing. Units of PSD are $x^2/\mathrm{Hz}$, whereas the units of PS are $x^2$ (quantity squared). Moreover, PSD and PS are related to each other: $\mathrm{PS} = \mathrm{PSD} \cdot \mathrm{ENBW}$, where ENBW stands for Equivalent Noise Bandwidth of the window function used. – jojek Nov 19 '16 at 13:33