3
$\begingroup$

Problem Definition

I need to calculate the convolution between a Gaussian function $g(x)$ and $d(x)$, the derivative of the inverse error function,

$$d(x) * g(x) = \int_{-\infty}^{\infty} d(\tau) g(x-\tau) \text{d}\tau = \int_{-\infty}^{\infty} d(x-\tau) g(\tau) \text{d}\tau $$

with

$$g(x) = \frac{1}{\sqrt{2 \pi}} \text{e}^{-\frac{1}{2} x^2},$$

$$ d(x) = \left\{ \begin{array}{ll} 0 & x < a \\ \frac{\text{d}}{\text{d}x} erf^{-1}(x) & -a\leq x\leq a \\ 0 & x > a \\ \end{array} \right. $$

with the constant $a<1$.

Since $d(x)$ is only non-zero on the interval $x \in [-a, a] $ I assume I can take these as limits for the convolution integral instead of $[-\infty, \infty]$ (please correct me here).

Attempted solution:

Since $d(x)$ is defined as the derivative of a function, I thought about using integration by parts:

$$\int_{-a}^{a} \frac{\text{d}}{\text{d}\tau} erf^{-1}(\tau) g(x-\tau) \text{d}\tau = \Big[ erf^{-1}(\tau) g(x-\tau) \Big]_{-a}^{a} - \int_{-a}^{a} erf^{-1}(\tau) \frac{\text{d}}{\text{d}\tau} g(x-\tau) \text{d}\tau $$

Then, using the convolution-derivative theorem (see below), we see that the last term is equal to the negative original term,

$$ \frac{\text{d}}{\text{d}x} erf^{-1}(x) * g(x) = erf^{-1}(x) * \frac{\text{d}}{\text{d}x} g(x)$$

Moreover, $erf^{-1}(-x) = - er^{-1}(x)$ (property of an odd function).

Rearranging:

$$ 2 \frac{\text{d}}{\text{d}\tau} erf^{-1}(x) * g(x) = 2 erf^{-1}(a)\Big[g(x-a) + g(x+a) \Big],$$

so, (surprisingly!) the result is a sum of two Gaussian functions,

$$d(x)*g(x) = erf^{-1}(a)\Big[g(x-a) + g(x+a) \Big]$$

I believe this can not be correct, as the numerical convolution doesn't look like a sum of Gaussians at all (see attached plots). There has to be a mistake in my math! Can anyone please point it out to me? Also, if you have suggestions on how to attack the problem, let me know!


Convolution-derivative theorem:

For any two functions $f_1(x)$ and $f_2(x)$, with ' denoting the derivative $\frac{\text{d}}{\text{d}x} $:

$$ ( f_1(x) * f_2(x) )' = f_1(x) * f_2'(x) = f_1'(x) * f_2(x) $$


If it helps, I found that the derivative of the inverse funciton is defined as follows:

$$\frac{\text{d}}{\text{d}u} erf^{-1}(u) = \frac{1}{\frac{\text{d}}{\text{d}u} erf\Big(erf^{-1}(u)\Big)}, $$

with

$$\frac{\text{d}}{\text{d}u} erf(u) = \frac{2}{\sqrt{\pi}} \text{e}^{-x^2}. $$

This would give

$$d(x) = \frac{\sqrt{\pi}}{2} \text{e}^{\Big(\text{erf}^{-1}(u) \Big)^2}, $$

and the following integral for the convolution,

$$ d(x) * g(x) = \frac{\sqrt{\pi}}{2} \frac{1}{\sqrt{2 \pi}} \int_{-\infty}^{\infty} \text{e}^{\Big(\text{erf}^{-1}(\tau) \Big)^2} \text{e}^{-\frac{1}{2} (x-\tau)^2} \text{d}\tau. $$

It looks difficult to attach directly, which is why I'm guessing playing with the derivatives might be the way to go.

$\endgroup$
0
$\begingroup$

I think the convolution-derivative theorem assumes that your functions vanish at the boundary, so in that sense where you are mistaken is in $$\int_{-a}^{a} \frac{\text{d}}{\text{d}\tau} erf^{-1}(\tau) g(x-\tau) \text{d}\tau = \Big[ erf^{-1}(\tau) g(x-\tau) \Big]_{-a}^{a} - \int_{-a}^{a} erf^{-1}(\tau) \frac{\text{d}}{\text{d}\tau} g(x-\tau) \text{d}\tau$$ which - after changing the $\tau$-derivative against the negative $x$-derivative - just becomes your convolution-derivative theorem when you neglect the boundary term. So I think you can not apply this theorem here in that form or you have to assume $\pm \infty$ as boundary in which case it vanishes. I think you have a sign error also, since you add these terms up to 2 times, but they actually cancel.

Also note that the integral as you wrote it above actually diverges for $a=1$, so you have to either keep track of the boundary term or use the derivative with respect to $x$ on the Gaussian.

In any case for the integration of the original problem I doubt you will find closed form. But you can get some approximation for $a\rightarrow 1$ by first substituting $t={\rm erf}^{-1}(\tau)$ and then integrating by parts \begin{align} &\quad \, \, \int_{-a}^{a} {\rm erf}^{-1}(\tau) \, \frac{{\rm e}^{-\frac{(x-\tau)^2}{2}}}{\sqrt{2\pi}} \, {\rm d}\tau \tag{1} \\ &= \int_{-{\rm erf}^{-1}(a)}^{{\rm erf}^{-1}(a)} \frac{{\rm e}^{-\frac{(x-{\rm erf}(t))^2}{2}}}{\sqrt{2\pi}} \, \frac{2t}{\sqrt{\pi}} \, {\rm e}^{-t^2} \, {\rm d}t \\ &= -\frac{{\rm e}^{-\frac{(x-{\rm erf}(t))^2}{2}}}{\sqrt{2\pi}} \, \frac{ {\rm e}^{-t^2}}{\sqrt{\pi}} \Bigg|_{-{\rm erf}^{-1}(a)}^{{\rm erf}^{-1}(a)} - \frac{2}{{\pi}} \frac{{\rm d}}{{\rm d}x} \int_{-{\rm erf}^{-1}(a)}^{{\rm erf}^{-1}(a)} \frac{{\rm e}^{-\frac{(x-{\rm erf}(t))^2}{2}}}{\sqrt{2\pi}} \, {\rm e}^{-2t^2} \, {\rm d}t \, . \end{align}

Approximating ${\rm erf}(t) \approx {\alpha t}$ with $\alpha=\frac{2}{\sqrt{\pi}}$ the latter integral becomes \begin{align} &\quad \,\, \frac{2}{\pi} \int_{-{\rm erf}^{-1}(a)}^{{\rm erf}^{-1}(a)} \frac{{\rm e}^{-\frac{(x-{\rm erf}(t))^2}{2}}}{\sqrt{2\pi}} \, {\rm e}^{-2t^2} \, {\rm d}t \\ &\approx \frac{2}{\pi} \int_{-{\rm erf}^{-1}(a)}^{{\rm erf}^{-1}(a)} \frac{{\rm e}^{-\frac{\left(x-{\alpha t}\right)^2}{2}}}{\sqrt{2\pi}} \, {\rm e}^{-2t^2} \, {\rm d}t \\ &=\frac{{\rm e}^{-\frac{2 x^2}{4+\alpha^2}}}{\pi\sqrt{4+\alpha^2}} \left\{ {\rm erf}\left( \frac{x \alpha + T(4+\alpha^2)}{\sqrt{2(4+\alpha^2)}} \right) - {\rm erf}\left( \frac{x \alpha - T(4+\alpha^2)}{\sqrt{2(4+\alpha^2)}} \right) \right\} \end{align} where $T={\rm erf}^{-1}(a)$. For $a=1$ this yields $$ \sim \frac{2}{\pi} \frac{{\rm e}^{-\frac{2 x^2}{4+\alpha^2}}}{\sqrt{4+\alpha^2}} \tag{2} \, . $$ The result fits qualitatively, but underestimates a little bit for $\alpha=\frac{2}{\sqrt{\pi}}$. In fact by choosing $\alpha=\frac{1.685}{\sqrt{\pi}}$ you get an almost perfect match. I doubt you get much closer with any other approximation. In the below figure you don't see the difference between $-\frac{{\rm d}}{{\rm d}x} (2)$ and the exact result of (1).

errorfunction approximation

For the derivative of the inverse error function as in your question you finally need to derive (1) with respect to $x$ another time as in the convolution-derivative theorem.

$\endgroup$

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.