Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Define, on $\mathbb{C}$: $$G(z)=\sum_{n=1}^{\infty} \frac{e^{-n^{2}}}{n^{z}}$$

A domain colored portrait of $G(z)$ (boxes are supposed to be negative signs):

plot of G(z) in the complex plane

suggests that the roots of $G(z)$ are equally spaced along lines of fixed real component. So:

  • Are the roots of $G(z)$ regularly spaced for some fixed $\Re(z)$? and if so:
  • Are there analytic expressions for the roots?
  • Can $G(z)$ expressed as a Weierstrass product?
share|improve this question
add comment

2 Answers

up vote 7 down vote accepted

Approximate $G(z)$ by the first two terms, $g(z) := e^{-1} + e^{-4}/2^z$. The zeroes of $g$ are $$- \frac{3}{\log 2} + (2k+1) \frac{\pi}{\log 2} i \approx - 4.328 + (2k+1) 4.532 i \ \mathrm{for} \ k \in \mathbb{Z}.$$ In particular, the zeroes of $g$ are perfectly regularly spaced.

Your zeroes are pretty close to these points, so I suspect that the regularity you are seeing just means that the zeroes of $G$ are pretty close to those of $g$.

share|improve this answer
    
+1, I was about to post pretty much the same answer. The coefficients decrease so quickly, only the first few contribute if $\text{Re}(z)$ is not too small. –  Eric Naslund Jan 6 '12 at 19:19
add comment

Numerically I got these roots with growing imaginary parts :

$-4.3409477472 + i\ 4.4770216332$
$-4.2880579538 + i\ 13.633957460$
$-4.3851693625 + i\ 22.647453867$
$-4.2763847378 + i\ 31.711259005$
$-4.3658043306 + i\ 40.8352457753$
$-4.3147004415 + i\ 49.8019753881$
$-4.3083714889 + i\ 58.9723863592$
...

Regularity and analytic expressions for the exact roots seem both compromised.

EDIT2: Since this is a Dirichlet series (with coefficients $c_n=e^{-n^2}$) and since there is absolute convergence everywhere (by the Weierstrass M-test as noticed by Eric Naslund) we may use the analytic properties of Dirichlet series to conclude that the abscissa of convergence of the Dirichlet series is $-\infty$ so that $G$ is analytic on the whole plane. Applying the Weierstrass factorization theorem we deduce that $G(z)$ may indeed be written as a product involving its roots.


The approximate values were nicely presented by David Speyer so let's compare them to the actual values. I tabulated the distances between the actual and approximative roots to get : $[0.056814, 0.054425, 0.058860, 0.053906, 0.057958, 0.055621, 0.055335]$

Concerning the arguments (from the actual roots) the difference between consecutive arguments $\mod (2 \pi)$ are : $[2.5436, 2.6434, 2.6083, 2.5665, 2.6760, 2.5343]$

Both distance and rotation look regular. This could indicate that the exact roots 'turn regularly around' the approximative roots as $k$ grows. Indeed a 'second order approximation' for the root of index k is :

$$\frac{-3}{\ln(2)} + \frac{(2k+1)\pi}{\ln(2)} i - \alpha e^{i(2k+1)\beta} \approx - 4.328 + (2k+1) 4.532 i - 0.05621 e^{-4.9793i(2k+1)}$$ $$\mathrm{with}\ \alpha=\frac{e^{3\frac{ln(3)}{\ln(2)}-8}}{\ln(2)},\ \beta=-\pi \frac{\ln(3)}{\ln(2)}\ \mathrm{and}\ k \in \mathbb{Z}.$$

To get this I added one term to David's approximation : $g_3(z) := e^{-1} + e^{-4}/2^z + e^{-9}/3^z$
multiplied by $e$, considered the $e^{1-9}/3^z$ term a 'small perturbation' of $1$ to get : $-e^{e^{-8}/3^z}\approx e^{-3-z\ln(2)}$. Taking the logarithms on both sides and replacing the z at the left with the 'first order' solution returns $z\approx \left(-3+(2k+1)i\pi-e^{-8-\ln(3)(-3+(2k+1)i\pi)/\ln(2)}\right)/\ln(2).$

With this new approximation the median distance is now $\approx 0.0023$ but I'll refrain the search of epicycles in this new astronomy that still requires a Kepler...


EXACT ROOTS - Using the same method (with the exact terms at the left instead of $-e^{e^{-8}/3^z}$) we may find the exact roots by iterations. Let's define : $$f_n(z)=1+e\sum_{m=3}^{n+1} \frac{e^{-m^2}}{m^z}\ \mathrm{and} \ r_1(k)=\frac{-3+(2 k+1) i \pi}{\ln(2)}$$

then the root corresponding to $k$ is obtained by iterations of : $$\ r_n(k)=r_1(k)-\frac{\ln(f_n(r_{n-1}(k))}{\ln(2)}$$

Example for $k=6$ :

$r_1(6)= -4.328085123 + i\ 58.92068184$
$r_2(6)= -4.310848945 + i\ 58.97454441$
$r_3(6)= -4.308195489 + i\ 58.97248418$
$r_4(6)= -4.308369505 + i\ 58.97237420$
$r_5(6)= -4.308372230 + i\ 58.97238650$
$r_6(6)= -4.308371466 + i\ 58.97238640$
$r_7(6)= -4.308371487 + i\ 58.97238636$
$r_8(6)= -4.308371489 + i\ 58.97238636$

(the distance from the actual value is more than $16$ times smaller at each iteration, the same is true for $k=0$).
Nice problem!


OTHER LAYERS - Let's investigate the 'second layer' detected by deoxygerbe with first roots at :

$-12.3840905498065732 + i\ 7.69101465838042938$
$-12.2523952898036964 + i\ 23.2196214862555024$
$-12.3153823799387642 + i\ 38.8071497678513693$

The 'first layer' was generated by the solutions of $0=e^{-1^2}/1^z+e^{-2^2}/2^z$ (first+second term of the series).

The 'second layer' seems to have been generated by the solutions of $0=e^{-2^2}/2^z+e^{-3^2}/3^z$ (second+third term of the series). Solutions may be obtained from the previous ones with the replacement of $-3=1^2-2^2$ by $-5=2^2-3^2$ and of $\ln(2)=\ln(2/1)$ by $\ln(3/2)$ to get : $$r^{(2)}(k)=\frac{-5+(2k+1)i\pi}{\ln(\frac32)}$$

I used the notation $r^{(2)}_k$ with the thought (comforted by your clear picture!) that other layers should exist near : $$r^{(L)}(k)=\frac{-(2L+1)+(2k+1)i\pi}{\ln(\frac{L+1}{L})}$$

Indeed a solution near $r^{(3)}(0)$ (the 'third layer' at $x\approx -24.33$) is : $-24.400057077391736+ i\ 10.878966908988947$.

Further the abscissa of the 'fourth layer' near $-40.33$ with ordinates near $14.0$ and $42.2$ seem to correspond rather well to your picture!

Anyway let's try to handle all the layers at once and define : $$f^{(L)}_n(z)=e^{L^2}L^z\left(\sum_{m=1}^{n+1} \frac{e^{-m^2}} {m^z}- \frac{e^{-(L+1)^2}}{(L+1)^z}\right)$$ $$r^{(L)}_1(k)=\frac{-(2 L+1)+(2 k+1) i \pi}{\ln(\frac{L+1}{L})}$$

then the root corresponding to $k$ on the layer $L$ should be obtained by iterations of : $$\ r^{(L)}_n(k)=r^{(L)}_1(k)-\frac{\ln(f^{(L)}_n(r^{(L)}_{n-1}(k))}{\ln(\frac{L+1}{L})}$$

This seems to provide all the 'visible' roots but I won't exclude the existence of other families (involving for example the first and third term) after all I missed an infinity of layers... :-)

share|improve this answer
    
We don't need to use analytic continuation, or properties about Dirichlet series, since the coefficients decrease so quickly. The series converges absolutely on the entire plane, so the $M$-test tells us it is analytic everywhere. –  Eric Naslund Jan 6 '12 at 19:17
    
The M-test proves convergence of course but concerning analyticity see this (en.wikipedia.org/wiki/…). Anyway you are right : since we have (absolute) convergence in any half-plane for this Dirichlet series we have convergence in the whole plane. –  Raymond Manzoni Jan 6 '12 at 19:59
    
@RaymondManzoni: thank you! I am going to probably make a Newton's method fractal of $G(z)$ shortly. (the fact that I can't immediately generate an Euler product of this function because $e^{-n^{2}}$ isn't completely multiplicative is currently bugging me) –  deoxygerbe Jan 9 '12 at 1:43
    
Also, what about the other layers of roots? I used mpmath's findroot with the muller method and initial value of $-20-3i$ and got mpc(real='-12.3840905498065732526323865513203', imag='7.69101465838042938331130479138366') –  deoxygerbe Jan 9 '12 at 2:01
    
Also, since I don't have edit privelidges: could you change the index variable in your summation to something not $i$: the lhs and rhs of that equation are jarring. –  deoxygerbe Jan 9 '12 at 4:26
show 2 more comments

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.