How to solve this integral? (Hilbert transform of a Gaussian)

cauchy-principal-valuecontour-integrationdefinite integralsgaussian-integralintegration

As a follow-up to this question about Mathematica's (in)ability to solve the following integral, I am curious how to solve it analytically:

$$\operatorname{PV}\int_{-\infty}^{+\infty}\frac{e^{-t^2}}{t-x}dt=-2\sqrt\pi F(x),$$

where $x\in\mathbb R,$ $F$ is the Dawson function defined by $F(x)=e^{x^2}\int_0^xe^{-u^2}du$, $\operatorname{PV}$ denotes the Cauchy principal value at $x$, namely, the limit as $\epsilon\to0+$ of the integral over $\mathbb R\setminus(x-\epsilon,x+\epsilon).$

I am comfortable with complex analysis, but not Schwartz distributions. There is a proof of this identity in the "Dawson function" Wikipedia link above, but it assumes knowledge of distributions and uses the identity $\frac{1}{u}=\int_0^\infty \sin(ku)dk$, which I am not going to accept casually. I suspect a contour integral + residue theory approach might be possible.

Best Answer

I found my own method that doesn't involve complex analysis or Fourier transforms. Just a bit of trickery with commuting derivatives and limits with integrals, which could actually be justified with convergence theorems from measure theory that I will neglect:


First, convert the integral to a different form:

$\begin{align} I(x)&=\operatorname{PV}\int_{-\infty}^{+\infty}\frac{e^{-t^2}}{t-x}dt\\\\ &=\lim_{\epsilon\to 0}\left(\int_{-\infty}^{x-\epsilon}\frac{e^{-t^2}}{t-x}dt +\int_{x+\epsilon}^{\infty}\frac{e^{-t^2}}{t-x}dt\right)\\\\ &=\lim_{\epsilon\to 0}\left(-\int_{\epsilon}^{\infty}\frac{e^{-(t-x)^2}}{t}dt +\int_{\epsilon}^{\infty}\frac{e^{-(t+x)^2}}{t}dt\right)\\\\ &=\int_0^\infty\frac{-e^{-(t-x)^2}+e^{-(t+x)^2}}{t}dt\\\\ &=-2e^{-x^2}\int_0^\infty\frac{\sinh(2xt)}{t}e^{-t^2}dt\\\\ \end{align}$

Now the remaining integral can be solved by differentiating under the integral sign:

$\begin{align} J(x)&=\int_0^\infty\frac{\sinh(2xt)}{t}e^{-t^2}dt\\\\ J'(x)&=2\int_0^\infty\cosh(2xt)e^{-t^2}dt\\\\ &=\sqrt\pi\,e^{x^2} \end{align}$

where the previous step comes from expanding the $\cosh$ as exponentials and completing the squares to obtain simple Gaussians. Then we have:

$\begin{align} J(x)&=J(0)+\int_0^x J'(t)dt\\\\ &=0+\sqrt\pi\int_0^xe^{t^2}dt\\\\ &=\frac{\pi}{2}\operatorname{erfi}(x) \end{align}$

Finally we have

$\begin{align} I(x)&=-\pi e^{-x^2}\operatorname{erfi}(x)\\\\ &=-2\sqrt\pi F(x) \end{align}$

Related Question