PSD,

Autocorrelation, and Noise in MATLAB

Aaron Scher

Energy and power of a signal
Consider a continuous time deterministic signal 𝑥 𝑡 . We are specifically interested in analyzing
the characteristics of the signal over the time range [Tbegin to Tend]. To do this, let us defined a
“windowed” version of 𝑥 𝑡 :

𝑥 𝑡 for 𝑇*+,-. ≤ 𝑡 ≤ 𝑇+.0
𝑥# 𝑡 = (1)
0 for all other times

Suppose further that we sample the signal with a sampling frequency 𝑓; . We can write 𝑥# 𝑛 =
𝑥[𝑛] = 𝑥 𝑡? , where 𝑡? are the sampling times:

𝑡? = 𝑇*+,-. + 𝑛 − 1 𝑇; , 𝑛 = 1 … 𝑁F (2)

The last time 𝑡 GH = 𝑇I?J .

Instantaneous power [W] at each discrete time tn.
𝑝[𝑛] = 𝑥[𝑛]L (3.a)
p=x.^2 (MATLAB) (3.b)
Average power Pg [W]:
O L
𝑃N = ? 𝑥[𝑛] (4.a)
GH
Pg=(1/N0).*sum(x.^2) (MATLAB) (4.b)

Total energy Eg [J]:
𝐸N = 𝑇; ? 𝑥[𝑛]L (5.a)
Eg=1/fs*sum(x.^2) (MATLAB) (5.b)


Spectrum of a signal
The two-sided discrete Fourier transform of a signal can be found in MATLAB can be found with
the following command:
X=1/fs*fftshift(fft(x,N)) (MATLAB) (6)

where N is the number of frequency points in the FFT. The discrete Fourier transform X has
length 𝑁 = length(X), and is defined at discrete frequency points 𝑓Q :

R R
𝑓Q = − S + 𝑘 − 1 S , 𝑛 = 1 … 𝑁 (7)
L GUO

Recall that 𝑁F = length(x) is the number of discrete time points of the original signal 𝑥[𝑛].
In general, 𝑁F ≠ 𝑁, and you are free to choose N as large as you want (so long as your
computer can handle it). In fact, you will generally set N>>N0. A good value for N is N=215.

1
Power spectral density
In two lines of code you can compute the power spectral density (PSD):

X=1/fs*fftshift(fft(x,N)) (MATLAB) (8.a)

PSD=(abs(X).^2)/((N0-1)*Ts) (MATLAB) (8.b)

Alternatively, you can also use MATLAB’s built in “periodogram” function:

PSD=periodogram(x,[],N,fs,’centered’) (MATLAB) (9)

Average Power
Once you have the PSD, you can compute the average power (Pg):

Pg=fs/N*sum(PSD) (MATLAB) (10)


Autocorrelation
The autocorrelation function can be calculated with the following MATLAB command.

Rxx=1/fs*xcorr(x,x) (MATLAB) (11)

Below are the equations to calculate the signal’s energy and power spectral density using the
autocorrelation function:

Total energy Eg [J]:

Eg=max(Rxx) (MATLAB) (12)

PSD [W/Hz]:

PSD=1/(N0-1)*abs(fftshift(fft(Rxx,N))) (MATLAB) (13)


Noise
You can create white Gaussian noise in MATLAB with the following commands:


noise_PSD = ... %choose noise PSD [W/Hz] (MATLAB) (14.a)

variance = ... noise_PSD*fs (MATLAB) (14.b)

sigma = ... sqrt(variance) (MATLAB) (14.c)

2

noise = transpose(sigma*randn(N0,1)) (MATLAB) (14.d)


In the MATLAB code below, I create a noisy sinusoid and compute/generate the following
information:
• Plot histogram of the noise
• Plot of noisy sinusoid in the time domain
• Signal-to-noise ratio (SNR)
• Plots of PSD computing using methods presented above
• Average PSD of noise
• Average value of noise


%Clear variables. clear command window, close all figures:
clc;
clear all;
close all;
%%%Setup and define variables
f0=10; %frequency of sinusoidal signal (Hz)
fs=100; %sampling frequency (Hz)
Ts=1/fs; %sampling period (seconds)
N0=3000; %number of samples
t=[0:Ts:Ts*(N0-1)]; %Sample times
noise_PSD=.5; %This is the desired noise power spectral density in W/Hz.
variance=noise_PSD*fs;% Variance = sigma^2
sigma=sqrt(variance);
noise=transpose(sigma*randn(N0,1));%create sampled white Gaussian noise.
xsignal=20*sin(2*pi*f0*t); %create sampled sinusoidal signal
x=xsignal+noise; %Add signal to noise

figure(1)
histogram(noise,30) %plot histogram
set(gca,'FontSize',14) %set font size of axis tick labels to 18
xlabel('Noise amplitude','fontsize',14)
ylabel('Frequency of occurance','fontsize',14)
title('Simulated histogram of white Gaussian noise','fontsize',14)

SNR_try1=snr(xsignal,noise) %calculate SNR using built in "snr" function.

SNR_try2=10*log10(sum(xsignal.^2)/sum(noise.^2)) %manually calculate SNR.
%If everything is correct, the two SNR calculations above should agree.

%Plot noise in time-domain

figure(2)
plot(t,x)
set(gca,'FontSize',14) %set font size of axis tick labels to 18
xlabel('Time (s)','fontsize',14)
ylabel('Amplitude','fontsize',14)
title('Noisey sinusoid','fontsize',14)
grid on

%Plot power spectral density (PSD) of noise using three different methods:
%

3
%Method 1. Calcululate PSD from amplitude spectrum
N=2^16; %Number of discrete points in the FFT.)
y=fft(x,N)/fs; %fft of noise
z=fftshift(y);%center noise spectrum
f_vec=[0:1:N-1]*fs/N-fs/2; %designate sample frequencies
amplitude_spectrum=abs(z); %compute two-sided amplitude spectrum
ESD1=amplitude_spectrum.^2; %ESD = |F(w)|^2;
PSD1=ESD1/((N0-1)*Ts);% PSD=ESD/T where T = total time of sample
figure(3)
plot(f_vec,10*log10(PSD1));
xlabel('Frequency [Hz]','fontsize',14)
ylabel('dB/Hz','fontsize',14)
title('Power spectral density - method 1','fontsize',14)
grid on
set(gcf,'color','w'); %set background color from grey (default) to white
axis tight
%calculate average power using PSD calclated from method 1:
Average_power_method_1=sum(PSD1)*fs/N% Pav=sum(PSD)*delta_f where
delta_f=fs/N
%
%Method 2 - Calculate PSD from autocorrelation
time_lag=((-length(x)+1):1:(length(x)-1))*Ts;
auto_cor=xcorr(x,x)/fs; %Use xcorr function to find PSD
y=1/fs*fft(auto_cor,N); %fft of auto correlation function
PSD2=abs(1/(N0-1)*fftshift(fft(auto_cor,N)));
figure(4)
plot(f_vec,10*log10(PSD2));%use convolution
xlabel('Frequency [Hz]','fontsize',14)
ylabel('dB/Hz','fontsize',14)
title('Power spectral density - method 2','fontsize',14)
grid on
set(gcf,'color','w'); %set background color from grey (default) to white
axis tight
%calculate average power using PSD calclated from method 1:
Average_power_method_2=sum(PSD2)*fs/N %Pav=sum(PSD)*delta_f where
delta_f=fs/N

%
%Method 3 - Calculate PSD using built in pwelch function
figure(5)
PSD3=periodogram(x,[],N,fs,'centered');
plot(10*log10(PSD3))
xlabel('Frequency [Hz]','fontsize',14)
ylabel('dB/Hz','fontsize',14)
title('Power spectral density - method 3','fontsize',14)
grid on
set(gcf,'color','w'); %set background color from grey (default) to white
axis tight
Average_power_method_3=sum(PSD3)*fs/N %Pav=sum(PSD)*delta_f where
delta_f=fs/N

%
%Calculate mean and average PSD of noise:
PSD_noise=periodogram(noise,[],N,fs,'centered');
Average_noise_PSD=mean(PSD_noise)
Mean_noise=mean(noise)

4
Results of running MATLAB code:

SNR_try1 =

6.3039


SNR_try2 =

6.3039


Average_power_method_1 =

245.5948


Average_power_method_2 =

245.5948


Average_power_method_3 =

245.5129


Average_noise_PSD =

0.4684


Mean_noise =

0.0983

5
 Simulated histogram of white Gaussian noise
300

250
Frequency of occurance

200

150

100

50

0
-30 -20 -10 0 10 20 30
Noise amplitude


Noisey sinusoid
40

30

20

10
Amplitude

-10

-20

-30

-40
0 5 10 15 20 25 30
Time (s)

6
 Power spectral density - method 1
30

20

10

-10
dB/Hz

-20

-30

-40

-50

-60

-50 -40 -30 -20 -10 0 10 20 30 40

Frequency [Hz]


The power spectral density plots for methods 2 and 3 exactly match that for method 1 (shown
above).