I have been working on a calculation that involves the following type of integral:

$$ f(z)={\frac{1}{i\sqrt{\pi}}}\int_{-\infty}^{\infty}{\frac{e^{-t^2}}{t-iz} dt} \hspace{1.5cm} z \in \Bbb{C} \hspace{3cm} \rm{[1]} $$

typically this integral is represented as

$$f(z)=e^{z^2}(1-\mathrm{erf}[z]) \hspace{5cm} \rm{[*]}$$

But is this representation valid for any $z$?

Here are the results I have obtained thus far.

One can find that the integral (1) satisfies this ODE:

$$ \frac{\partial f}{\partial z}=-2+2zf \hspace{4cm} \rm{[2]} $$

So, if this ODE has a unique solution, an expression for $f(z)$ will be found by integrating it and applying the appropriate initial condition.

[2] has a solution

$$ f(z)=e^{z^2} \left( e^{-z_0 ^2} f(z_0) -2\int_{z_0}^{z}{e^{-t^2} dt} \right) \hspace{3cm} \rm{[3]} $$

Now, what I am less confident about is how to calculate the initial conditions from 1. But here are my attempts:

(a) take $z_0=0$ ; in that case I get $f(0)=2\sqrt{\pi}$ leading to

$$ f(z)=\sqrt{\pi}e^{z^2}(2-\mathrm{erf}[z]) \hspace{4cm} \rm{[4a]} $$

which is not the same as [$*$]

(b) If letting $|z_0| \rightarrow -\infty$ implies $e^{-z_0^2}f(z_0)\rightarrow 0$ then [3] becomes

$$ f(z)=\sqrt{\pi}e^{z^2}(1-\mathrm{erf}[z]) \hspace{4cm} \rm{[4b]} $$

which is more similar to [$*$] but still different to a factor of $\sqrt{\pi}$

So what is the problem? is [$*$] in fact incorrect? Which one (of [4a] and [4b]) is correct? And is any of them valid for all $z$?


Okay, I think I've got it. The solution to your integral is likely $$f(z)=\sqrt{\pi}e^{z^2}(\operatorname{sgn}(\operatorname{Re}(z)) - \operatorname{erf}(z) ).$$ To see this, let first $\operatorname{Im}(z)=:v> 0$ and denote $z=u+iv$.

First, note that $$\int_{-\infty}^\infty \frac{e^{-t^2}dt}{t+z}= e^{-z^2} \int_{-\infty+iv}^{\infty+iv} \frac{e^{-t^2+2t(u+iv)}dt}{t}=:e^{-z^2} I(u,v)$$ via substitution. Differentiating $I(u,v)$ with respect to $u$ yields $$\frac{\partial I(u,v)}{\partial u} = 2\int_{-\infty+iv}^{\infty+iv} e^{-t^2+2t(u+iv)}dt=2e^{-v^2+2iuv}\int_{-\infty}^{\infty} e^{-t^2+2ut}dt$$ via substition. Using the fact that $\int_{-\infty}^\infty e^{-t^2+2at}dt = \sqrt{\pi} e^{a^2}$, it follows that $$\frac{\partial I(u,v)}{\partial u}=2\sqrt{\pi}e^{(u+iv)^2}.$$ Therefore, $$I(u,v)=2\sqrt{\pi}\int_0^u e^{(t+iv)^2}dt + g(v)=2\sqrt{\pi}\int_{iv}^{u+iv} e^{t^2}dt + g(v).$$ (Note that $g$ may depend on $v$.)

Setting $u=0$, it is clear that $$g(v)=I(0,v)=\int_{-\infty+iv}^{\infty+iv} \frac{e^{-t^2+2tiv}dt}{t}.$$ Using again differentation under the integral (while being cautious that the limit involves $v$, which is no big deal since it vanishes in the limit), we arrive at $$\frac{\partial I(0,v)}{\partial v}=2i\sqrt{\pi}e^{-v^2}.$$ Thus, $$g(v)= 2i\sqrt{\pi}\int_0^v e^{-t^2}dt + C,$$ and $$I(u,v)=2\sqrt{\pi}\int_{iv}^{u+iv} e^{t^2}dt +2i\sqrt{\pi}\int_0^v e^{-t^2}dt + C$$ where $C$ is some constant. Via substitution, $$I(u,v)=2\sqrt{\pi}i\left(\int_{v}^{v-iu} e^{-t^2}dt + \int_0^v e^{-t^2}dt\right) + C =2\sqrt{\pi}\int_0^{u+iv} e^{t^2}dt+C.$$

We now have to calculate the value of $C$. First, note that $$\lim_{v\to\infty} I(0,v)=\lim_{v\to\infty} \pi i \operatorname{erf}(v)+C=\pi i + C.$$ Second, $$\lim_{v\to\infty} I(0,v)= \lim_{v\to\infty} \int_{-\infty}^{\infty} \frac{e^{-(t^2+v^2)}dt}{t+iv}=0.$$ Thus, $C=-\pi i$ and, finally, $$I(u,v)=2\sqrt{\pi}\int_0^{u+iv} e^{t^2}dt - \pi i.$$

For $\operatorname{Im}(z)=v< 0$, the derivation is completely analogous except $$\lim_{v\to -\infty} I(0,v)=\lim_{v\to -\infty} \pi i \operatorname{erf}(v)+C=-\pi i + C.$$ In this case $C=-\pi i$.

For $\operatorname{Im}(z)=v= 0$, note that the Cauchy principal value of $g(0)=0$. Analogous reasoning as above yields $C=0$.

To sum up, it follows that $$I(u,v)=2\sqrt{\pi}\int_0^{u+iv} e^{t^2}dt - \operatorname{sgn}(v) \pi i,$$ and $$\int_{-\infty}^\infty \frac{e^{-t^2}dt}{t+z}=e^{-z^2}\left(2\sqrt{\pi}\int_0^{z} e^{t^2}dt - \operatorname{sgn}(\operatorname{Im}(z)) \pi i\right).$$

Finally, turning to your integral: \begin{align} \frac{1}{i\sqrt{\pi}} \int_{-\infty}^\infty \frac{e^{-t^2}dt}{t-iz}&= -i e^{z^2} \left(2\int_0^{-iz} e^{t^2}dt - \operatorname{sgn}(\operatorname{Im}(-iz)) \sqrt{\pi} i\right)\\ &= -e^{z^2} \left(\sqrt{\pi} \operatorname{erf}(z) - \operatorname{sgn}(\operatorname{Re}(z)) \sqrt{\pi}\right)\\ &= e^{z^2} \sqrt{\pi} \left( \operatorname{sgn}(\operatorname{Re}(z)) - \operatorname{erf}(z))\right) \end{align}

This is quite similar to your original conjecture that $f(z)=e^{z^2} \left(1 - \operatorname{erf}(z))\right)$. Obviously, the answer seems to be that your your formula (despite the missing $\pi$) holds for all $z$ with $\operatorname{Re}(z)>0$.

  • $\begingroup$ Interesting, I will have a closer look. I generally followed the method given in here but then I arrive at contradictory results, which implies that something must be wrong. $\endgroup$ – JMK Jul 21 '14 at 17:56
  • $\begingroup$ @JMK I totally revised my result and gave a new proof. This time, it seems to be correct, I guess. $\endgroup$ – Dumdi44 Jul 23 '14 at 17:36
  • $\begingroup$ Yes, I was going to say that it seemed conspicuous that in the original answer the substitution didn't shift the integration limits on the complex plane - it seems fine now though. I did btw take the original formula from Abramowitz & Stegun and also DMLF , so I didn't just pull it out of a hat - or rather I used some 'reputable' sources. But for this and other related functions, there seems to be differences in the formulae on different quadrants of the complex plane, and I wanted to gt to the bottom of them :) $\endgroup$ – JMK Jul 23 '14 at 19:39

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.