1
$\begingroup$

I want to calculate the convolution $F * G$ of two Gaussian functions without resorting to Fouritertransforms:

$F(t) := \exp(-at^2), G(t) := \exp(-bt^2) \qquad a,b>0$

But intuitively I expected the convolution to result again in a non constant function. Can anyone find my mistake / confirm that this calculation is correct?


Let $\Omega = \mathbb R$, then

$\begin{align*} (F*G)(x) &= \int_\Omega F(t)G(x-t)dt &\\ & = \int_\Omega e^{-at^2-b(x-t)^2} dt \qquad\qquad \quad \text{substitute }u = t+\frac{1}{2} x \implies "du=dt" \\ &=\int_\Omega e^{-a(u-\frac{1}{2}x)^2-b(\frac{1}{2}x-u)^2}dt \qquad \text{substitute }v = u-\frac{1}{2} x \implies "du=dv"\\ &=\int_\Omega e^{-(a+b)v^2} dv \qquad\qquad \qquad\,\, \text{substitute } w = \sqrt{a+b}v \implies"dw = \sqrt{a+b}dv" \\ &=\frac{1}{\sqrt{a+b}}\int_\Omega e^{-w^2}dw \\ &=\frac{\sqrt{\pi}}{\sqrt{a+b}} \end{align*}$

$\endgroup$
  • 1
    $\begingroup$ Hi @flawr. There is a mistake after the first substitution : indeed, if $t=u-x/2$, then $x-t=x-(u-x/2)=3x/2-u$. You can also see that there is something that goes wrong by seeing that $v=u-x/2=(t+x/2)-x/2=t$. You should end up with a new gaussian : take the Fourier tranform of the convolution to get the product of two new gaussians (as the Fourier transform of a gaussian is still a gaussian), then take the inverse Fourier transform to get another gaussian. $\endgroup$ – Nicolas Apr 16 '16 at 15:54
  • $\begingroup$ Oh I see now, thank you. $\endgroup$ – flawr Apr 16 '16 at 16:37
1
$\begingroup$

You seem to have lost the constant term modified by the completion of the square: $$ \begin{align} \int_{-\infty}^\infty e^{-a(x-t)^2}e^{-bt^2}\,\mathrm{d}t &=\int_{-\infty}^\infty e^{-ax^2+2axt-at^2-bt^2}\,\mathrm{d}t\\ &=e^{-ax^2+\frac{a^2x^2}{a+b}}\int_{-\infty}^\infty e^{-\frac{a^2x^2}{a+b}+2axt-(a+b)t^2}\,\mathrm{d}t\\ &=e^{\frac{-abx^2}{a+b}}\int_{-\infty}^\infty e^{-(a+b)\left(t-\frac{ax}{a+b}\right)^2}\,\mathrm{d}t\\ &=\sqrt{\tfrac\pi{a+b}}\,e^{\frac{-abx^2}{a+b}} \end{align} $$

$\endgroup$
  • $\begingroup$ Which, BTW, should match the result from the Central Limit Theorem that $N(\mu_1,\sigma_1) * N(\mu_2,\sigma_2) = N(\mu_1+\mu_2,\sqrt{\sigma_1^2+\sigma_2^2})$ $\endgroup$ – Acccumulation Oct 8 at 14:50
  • $\begingroup$ Which, I believe, it does. That might indicate that this question might merit the probability-distributions tag. $\endgroup$ – robjohn Oct 8 at 15:04

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.