version 1.0.0.0 (5.83 KB) by
John D'Errico

Noise variance estimation from a signal vector or array

Some curve fitting or smoothing tools can benefit from knowledge of the noise variance to expect on your data. Kalman filters use this information, also some spline fitting tools. So I wrote a function to extract the noise variance from a signal vector. It also works on any specified dimension of an array.

A few examples of this code in use:

Simple linear data, with purely additive N(0,1) gaussian noise:

t = 0:10000;

x = t + randn(size(t));

mv = estimatenoise(x)

mv =

1.0166

Gaussian noise added to a sine wave (Nominal variance = 0.01)

t = linspace(0,1,1000)';

x = sin(t*50) + randn(size(t))/10;

mv = estimatenoise(x)

mv =

0.0096887

Pure gaussian noise, with a nominal variance of 9. (Note that var would have been a better estimator for this particular case...)

mv = estimatenoise(3*randn(2,3,1000),3)

mv =

9.6584 8.2696 8.632

9.2404 8.5346 9.7725

A piecewise constant function with multiple discontinuities. The true noise variance should be 0.01.

t = linspace(0,1,1000);

X = round(cos(t*6*pi)) + randn(size(t))/10;

plot(t,X)

var(X) % var will be wildly in error

ans =

0.68256

estimatenoise(X)

ans =

0.010882

Test if estimatenoise is able to recover the variance of a normally distributed random sample with unit variance. (Yes, it will be much slower than var.)

mean(estimatenoise(randn(1000,1000)))

ans =

1.0002

Estimatenoise can now handle non-uniformly spaced series (by request.) In the next example,

the actual noise variance was 1.0 here. Perform the operation 1000 times, then look at the median variance estimate. How well did we do?

t = sort([randn(1,100) , randn(1,100)+5]);

X = repmat(sin(t*5)*100,1000,1) + ...

randn(1000,length(t));

Estimatenoise is clearly wrong when the spacing is ignored.

median(estimatenoise(X,2))

ans =

16.438

Supplying the sampling "times", we get quite a reasonable result.

median(estimatenoise(X,2,t))

ans =

1.1307

Estimatenoise also works on data with replicates. In this example, each point will be replicated up to 3 times. The actual noise variance was again 1.0 here. I'll also compare the increase in times required for estimatenoise when t is supplied.

t = sort([0:1:100, 1:2:100, 1:4:100]);

X = repmat(sin(t/10)*100,1000,1)+ ...

randn(1000,length(t));

Again, estimatenoise is clearly wrong when the non-uniform spacing is ignored.

tic,median(estimatenoise(X,2)),toc

ans =

4.2056

Elapsed time is 2.690135 seconds.

Supplying the sampling "times", we again get quite a reasonable result. The time penalty is not quite 2x.

tic,median(estimatenoise(X,2,t)),toc

ans =

1.0116

Elapsed time is 4.864486 seconds.

John D'Errico (2020). Estimatenoise (https://www.mathworks.com/matlabcentral/fileexchange/16683-estimatenoise), MATLAB Central File Exchange. Retrieved .

Created with
R2007a

Compatible with any release

**Inspired:**
Noise variance estimation, Reverberation Time Calculator, Sound Power Directivity Analysis, Nth Octave Test Signal, Nth_Oct_Hand_Arm_&_AC_Filter_Tool_Box, Continuous Sound and Vibration Analysis

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!Create scripts with code, output, and formatted text in a single executable document.

ahmed youssefIvan Pessanecan i have the paper that explain this method? thak you in advance

Adi Purwandanagreat! works fine to estimate noise level of the density profile measured using CTD probe (oceanography)

AshkanJackiePlease give me the paper/reference to explain this method! thanks

Antoni J. CanósVery nice. Thank you John!!

lara presaThanks!

Mark Shore@ lara, John D'Errico recently announced he would be significantly reducing his input to MATLAB forums. This is a 1D tool, so you might want to consider using Damien Garcia's N-D noise estimation tool at http://www.mathworks.com/matlabcentral/fileexchange/25645

lara presaCan we estimate the noise in an image with this function?? what happen with sampling times t?? is necessary for a matrix? Thanks!!

hbu thank you .

Jody Klymak^ Sorry, the first cumsum above should not be multiplied by 3!

Jody KlymakThis is very nice.

Potential users should be aware that it assumes a spectral gap in the data. If you generate a red signal with noise:

>> x = cumsum(randn(1000,1000))*3 + randn(1000,1000);

>> mean(estimatenoise(x))

ans =

1.3021

>> x = cumsum(randn(1000,1000))*3 + randn(1000,1000);

>> mean(estimatenoise(x))

ans =

3.3083

you get a high biased estimate because there is still significant signal at high frequencies. As with any noise estimate, you need to understand the signal first!

I use this example because a large class of geophysical signals are red.

Kenneth EatonJohn D'ErricoFor those users who asked for the non-uniform version, its up there now. Sorry it took a while, I had a couple of aborted attempts before I got it right. The new code will use the old version on equally spaced series, but the unequally spaced option is not terribly slower in my tests.

A BI would like to see the extension to unevenly distributed data

John de MelloAn extremely useful piece of code. The planned extension to unevenly spaced data sets would be fantastic.

pule shuleIhsan ulhaqYes well tested, verified, simple, mature and well documented code.

We should appreciate such job.

G. A.M.Very well documented. Very clear code. Nice examples. The algorithm has been well-tested (Although the author feels he could do more testing and verification and further improve this code, don't we all feel that way about code we write?) I think this code is more than mature enough for release here. I appreciate the author making it available.