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.
Update to handle complex inputs |
Adi Purwandana (view profile)
great! works fine to estimate noise level of the density profile measured using CTD probe (oceanography)
Ashkan (view profile)
dang thanh (view profile)
Please give me the paper/reference to explain this method! thanks
Antoni J. Canós (view profile)
Very nice. Thank you John!!
lara presa (view profile)
Thanks!
Mark Shore (view profile)
@ 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 presa (view profile)
Can we estimate the noise in an image with this function?? what happen with sampling times t?? is necessary for a matrix? Thanks!!
hbu (view profile)
thank you .
Jody Klymak (view profile)
^ Sorry, the first cumsum above should not be multiplied by 3!
Jody Klymak (view profile)
This 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 Eaton (view profile)
For 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.
I would like to see the extension to unevenly distributed data
An extremely useful piece of code. The planned extension to unevenly spaced data sets would be fantastic.
Yes well tested, verified, simple, mature and well documented code.
We should appreciate such job.
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.