1
$\begingroup$

For a time discrete signal, the rect-filter (for 1D signal [1,1]) is a perfect low pass filter. I was wondering, what is the best gaussian approximation of this filter?

For a current problem, I cannot use binomial filters but have to use Gaussian filters for smoothing. Ths question is thus, what is the standard deviation of a Gaussian filter best approximating the [1,1] rect-filter. 'Best' could e.g. mean is has similar probperties in the frequency domain.

$\endgroup$
1
$\begingroup$

Suppose you want a low pass filter supported on $f \in [-1,1]$. The Fourier transform of a Gaussian is a Gaussian. So let's assume our Gaussian filter has a frequency response of the form $G_\sigma(f) = \frac{1}{\sigma \sqrt{2 \pi}} e^{-f^2/\sigma^2}$. For the best Gaussian fit in least squares sense, let us minimize the squared error ($\mathsf{SE}$): $$ \mathsf{SE}(\sigma) = \int_{-1}^1 |G_\sigma(f)-1|^2 \;\;\; d\,f $$ by choice of $\sigma$. This calls for some numerical evaluation.

I used Mathematica:

Integrate[ (1/s/Sqrt[2 Pi] Exp[-f^2/s^2] - 1)^2, {f, -1, 1}]

which gave me

2 - Sqrt[2] Erf[1/s] + 1/(2 Sqrt[2 Pi] s) Erf[ Sqrt[2]/s ]

Then plot the result as a function of $\sigma$ (i.e. s).

I see a minimum around $\sigma =0.7$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.