File Exchange

## Noise variance estimation

version 1.5 (2.68 KB) by

EVAR estimates the noise variance from 1-D to N-D data

Updated

Suppose that you have a signal Y (Y can be a time series, a parametric surface or a volumetric data series) corrupted by a Gaussian noise with unknown variance. It is often of interest to know more about this variance. EVAR(Y) thus returns an estimated variance of the additive noise.

EVAR provides better results if the original function (i.e. the function without noise) is relatively smooth i.e. has continuous derivatives up to some order. Several tests, however, showed that EVAR works very well even with multiple discontinuities.

Note: EVAR only works with evenly-gridded data in one and higher dimensions.

Here are two examples:

%-- Let us estimate the noise variance from a corrupt signal --
% First create a time signal
t = linspace(0,100,1e6);
y = cos(t/10)+(t/50);
% Make this signal corrupted by a Gaussian noise of variance 0.02
var0 = 0.02; % noise variance
yn = y + sqrt(var0)*randn(size(y));
% Now estimate the variance with EVAR and compare with the "true" value
evar(yn)

%-- Now, let us estimate the noise variance from volumetric data --
% Create a volume array
[x,y,z] = meshgrid(-2:.2:2,-2:.2:2,-2:.2:2);
f = x.*exp(-x.^2-y.^2-z.^2);
% Make these data corrupted by a Gaussian noise of variance 0.5
var0 = 0.5; % noise variance
fn = f + sqrt(var0)*randn(size(f));
% Estimate the variance with EVAR and compare with the "true" value
evar(fn)

------
Several tests are also given in:
http://www.biomecardio.com/matlab/evar.html
-----

Vincent

Vincent

### Vincent (view profile)

Hi,
I just downloaded the file and port it to work with Octave.
But noise variance is half the true variance.
Any idea about modifications (not multiply by 2 ;) ) to be closer the true noise variance ?
Thanks

Vincent

John D'Errico