0
$\begingroup$

I have the following equation to solve $$ z+e^{z^2}\operatorname{erfc}(z)=0 $$ being $$ \operatorname{erfc}(z)=1-\frac{2}{\sqrt{\pi}}\int_0^ze^{-s^2}ds. $$ I solved it numerically and appears to have an infinity of complex solutions. I am in need to find an approximate analytical approximation for the first root. I made several attempts, e.g. using either Taylor or asymptotic expansions, but to no avail. I was not able to recover the first root that seems to be $-.3378723369\ldots-.9499598705\ldots i$.

My question: is it possible to find a successful approximation scheme to get an analytical expression for the roots of the given equation?

Edit: There appear to be infinite roots in $\mathbb{C}$. This can be seen straightforwardly in Fig. 1 where both the real and imaginary part are given of $-z$ and $e^{z^2}\operatorname{erfc}(z)$ Fig. 1

Orange is $Re(-z)$, green is $Im(-z)$. There is a common line between this twos crossing the part with the erfc showing a possible infinite set of roots.

I was able to find several roots by choosing a different starting point for the search algorithm. E.g. I have got, listing just a few along with their complex conjugates implied,

$-1.90593 - 1.82833 i$

$-3.19368 - 3.06637 i$

$-4.07604 - 3.94595 i$

$-6.71931 - 6.60297 i$

$-8.00551 - 7.89692 i$

$-10.2496 - 10.1527 i$

The aforementioned first root appears a recurring one in the search.

$\endgroup$
  • $\begingroup$ Please, have a look to my edit. I do not think that the equations has many roots. Did you find other solutions ? If you did, please tell me. Cheers. $\endgroup$ – Claude Leibovici Feb 13 at 4:39
  • $\begingroup$ @ClaudeLeibovici From the plot I made with Mathematica, it seemed to me that the number of roots was large. E.g. I also get $-5.114532877548192 + 4.9884679626263475i$ and finding or not another root depends on the starting point in FindRoot.. I think you can easily check by yourself that the one I showed you is another zero. $\endgroup$ – Jon Feb 13 at 7:09
  • $\begingroup$ Have a look to my edit. Using the contour plot of $\Phi(x,y)=1$, I had to zoom crazy to "see" the solution you reported in your last comment (in fact, I did not see it - my wife did it for me !). I wonder if we could face some accuracy problems here with $\Phi(x,y)$, $\endgroup$ – Claude Leibovici Feb 13 at 8:01
  • $\begingroup$ Well, I plotted the imaginary and real part in a 3d plot and drawn the x and y plans. It is seen in this way that there seems to be a multitude of roots. Maybe, a contour plot could be not so easy to read. I used FindRoot starting from $-5-5i$ and the root come out. $\endgroup$ – Jon Feb 13 at 9:04
  • $\begingroup$ Good to know that ! It would be good you list in the post some of the roots. I am sure that you know that if $(a+ib)$ is a root $(a-ib)$ is another one. $\endgroup$ – Claude Leibovici Feb 13 at 9:42
1
$\begingroup$

You can obtain the first root using Newton method.

Using for example $z_0=1+i$, it would give the following iterates $$\left( \begin{array}{cc} n & z_n \\ 0 & 1+ i \\ 1 & -0.570752+0.455687 i \\ 2 & -0.217277+0.691623 i \\ 3 & -0.298698+1.012267 i \\ 4 & -0.336117+0.946645 i \\ 5 & -0.337868+0.949970 i \\ 6 & -0.337872+0.949960 i \end{array} \right)$$

Using $z_0=1-i$, it would give the following iterates $$\left( \begin{array}{cc} n & z_n \\ 0 & 1-i \\ 1 & -0.570752-0.455687 i \\ 2 & -0.217277-0.691623 i \\ 3 & -0.298698-1.012267 i \\ 4 & -0.336117-0.946645 i \\ 5 & -0.337867-0.949968 i \\ 6 & -0.337872-0.949960 i \end{array} \right)$$

Starting with $z_0=i$ seems to converge faster $$\left( \begin{array}{cc} n & z_n \\ 0 & i \\ 1 & -0.400169+0.909374 i \\ 2 & -0.339614+0.946181 i \\ 3 & -0.337864+0.949950 i \\ 4 & -0.337872+0.949960 i \end{array} \right)$$

The same starting with $z_0=-i$ $$\left( \begin{array}{cc} n & z_n \\ 0 & - i \\ 1 & -0.400169-0.909374 i \\ 2 & -0.339614-0.946181 i \\ 3 & -0.337864-0.949950 i \\ 4 & -0.337872-0.949960 i \end{array} \right)$$

Concerning approximations, I am stuck. The only thing I really tried is a series expansion around $z=i$ and got $$\frac{-i \text{erfi}(1)+1+i e}{e}+\frac{\left(2 \text{erfi}(1)+2 i+e-\frac{2 e}{\sqrt{\pi }}\right) (z-i)}{e}+\frac{\left(i \text{erfi}(1)-1-\frac{2 i e}{\sqrt{\pi }}\right) (z-i)^2}{e}+O\left((z-i)^3\right)$$ Ignoring the higher order terms, solving the quadratic in $(z-i)$ and evaluating leads to $$z\approx -0.351584 + 0.947053 i$$ Doing the same around $z=-i$ gives the other root.

Edit

In terms of approximation of the solution, we can do better using the simplest $[1,1]$ Padé approximant built at $z_\pm=\pm i$. For example, $$z_+=-\frac{i \left(-4 e \left(\pi -2 \sqrt{\pi }\right) (\text{erfi}(1)+i)-5 \pi (\text{erfi}(1)+i)^2+4 e^2 \left(\sqrt{\pi }-1\right)\right)}{e \left(5 \pi -6 \sqrt{\pi }\right) (\text{erfi}(1)+i)+3 \pi (\text{erfi}(1)+i)^2+e^2 \left(4-6 \sqrt{\pi }+\pi \right)}$$ which is $-0.347274 + 0.955038 i$ much better than the previous one.

We could continue with $[1,n]$ Padé approximant to get explicit approximations (the expressions are really messy). The values are reporteds below $$\left( \begin{array}{cc} n & z^{(n)}_+\\ 0 & -0.400169+0.909374 i \\ 1 & -0.347274+0.955038 i \\ 2 & -0.337395+0.951302 i \\ 3 & -0.337591+0.949901 i \\ 4 & -0.337859+0.949907 i \end{array} \right)$$

Update

Looking at the table of the roots given in Jon' edit, they really look like $z_k \sim -k \pm ki$. Using the $[1,1]$ Padé approximant built at $z=-k-ik$ (the formula is too messy to be reported here) an approximation of the root can be obtained.

Below are listed some values (to be compared to Jon's) $$\left( \begin{array}{cc} k & \text{approximation} \\ 1 & -0.34995-0.95803 i \\ 2 & -1.90291-1.81163 i \\ 3 & -3.23300-3.00433 i \\ 4 & -4.06884-3.94940 i \\ 5 & -5.10001-4.96837 i \\ 6 & -6.05119-5.85692 i \\ 7 & -7.09451-6.88121 i \\ 8 & -8.03695-7.91822 i \\ 9 & -9.07841-8.95282 i \\ 10 & -10.0648-9.96230 i \\ 11 & -11.0449-10.9348 i \\ 12 & -12.0518-11.9671 i \\ 13 & -13.0490-12.9660 i \\ 14 & -14.0285-13.9565 i \\ 15 & -15.0394-14.9589 i \end{array} \right)$$

$\endgroup$
  • $\begingroup$ Thanks a lot. Very nice. $\endgroup$ – Jon Feb 12 at 13:14
  • $\begingroup$ Very neat answer. This is what I needed. Thank you very much. $\endgroup$ – Jon Feb 13 at 7:03

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.