This question is stimulate by the previous two question here and here. We are interested in studying the following special case of Fox-Write function \begin{align} \Psi_{1,1} \left[ \begin{array}{l} (1/k,2/k) \\ (1/2,1)\end{array} ; -x^2\right], x\in \mathbb{R}, k\in (1,\infty. \end{align}

Where Fox-Write function $\Psi_{1,1} \left[ \begin{array}{l} (a,A) \\ (b,B)\end{array} ; z\right]$ is defined as

\begin{align} \Psi_{1,1} \left[ \begin{array}{l} (a,A) \\ (b,B)\end{array} ; z\right] =\sum_{n=0}^\infty \frac{\Gamma(a+An)}{\Gamma(b+Bn)} \frac{z^n}{n!}. \end{align}

The reason this is interesting is because of the following Fourier cosine: \begin{align} \Psi_{1,1} \left[ \begin{array}{l} (1/k,2/k) \\ (1/2,1)\end{array} ; -x^2\right]= \int_0^\infty \cos(xt) e^{-t^k} dt \end{align} see this question for more details.

Goal We are interest in finding conditions on $k$ such that $ \Psi_{1,1} \left[ \begin{array}{l} (1/k,2/k) \\ (1/2,1)\end{array} ; -x^2\right]$ has no zeros.

The current conjecture is that there are no zeros for $k \le 2$ and at least one zero for $k>2$. Thanks for any ideas you might have.

Edit: Perhaps there is a connection to the notion of Stable Distribution in probability. See comments below.

  • 2
    $\begingroup$ Given the hypergeometric behaviour of the involved series, maybe it is possible to apply en.wikipedia.org/wiki/Sturm%27s_theorem. $\endgroup$ – Jack D'Aurizio Aug 4 '16 at 17:22
  • $\begingroup$ Thanks. You always great ideas. Cheers. Let you know if I have something. $\endgroup$ – Boby Aug 4 '16 at 17:47
  • $\begingroup$ I just read through the theorem. But I don't see the connection. Could you give more on the idea that you had. $\endgroup$ – Boby Aug 4 '16 at 17:59
  • $\begingroup$ my idea is to take a suitable truncation of the power series and compute the discriminant of such a polynomial through a resultant or other techniques. With a bit of luck, the positivity/negativity of the discriminant just depends on $k$ and should tell us if there are real zeroes or not. $\endgroup$ – Jack D'Aurizio Aug 4 '16 at 18:06
  • 2
    $\begingroup$ Notice that $$ \int_0^\infty \cos(xt) e^{-t^k} \, \rm{d} t = \frac12 \int_{-\infty}^\infty e^{itx} e^{-|t|^k} \, \rm{d} t. $$ Your conjecture corresponds to a theorem of Paul Lévy about the characteristic functions of stable distributions (that only exist for $k \le 2$). $\endgroup$ – sometempname Aug 10 '16 at 15:24

I'll sketch an answer.

I use the notation: $$ \psi_k(z) = \sum_{n=0}^\infty \frac{\Gamma(\frac{1}{k} + \frac{2n}{k})}{\Gamma(\frac12 + n)} \cdot \frac{z^n}{n!}, $$ and $$ \phi_k(z) = \int_{0}^\infty \cos(z t)e^{-t^k} \, \rm{d} t. $$

First of all, the correct identity seems to be: $$ \phi_k(x) = \frac{\sqrt{\pi}}{k} \psi_k\left(-\left(\frac{x}{2}\right)^2\right).$$

Elementary cases ($\phi_k > 0$): $$k=1\rm{:}\,\, \phi_1(x) = \frac{1}{1+x^2}, \quad k=2\rm{:}\,\, \phi_2(x) = \frac{\sqrt{\pi}}{2} e^{-\frac{x^2}{4}}. $$

The case $k>2$:

The function $e^{-t^k}$ is differentiable at $t=0$ (from the right, at least twice), so integrating by parts twice, we have $\phi_k(x) = O(x^{-2})$. Thus, $\phi_k(x)$ is integrable (it is clearly bounded). Since $\cos(z) = \frac12 (e^{iz} + e^{-iz})$, we can write $$ \phi_k(z) = \frac12 \int_{-\infty}^\infty e^{izt} e^{-|t|^k} \, \rm{d} t.$$ We might think of $\frac{1}{\pi} \phi_k(x)$ as the "density" of a random variable $X$, whose characteristic function is $\varphi_k(t) = e^{-|t|^k}$. If $\varphi_k(t)$ was a characteristic function, then since it is at least twice differentiable at $t=0$, we would have $\mathbb{E} [X^2] = -\varphi_k^{(2)} (0) = 0$. Thus $X$ is a trivial random variable, and the contradiction implies that $\frac{1}{\pi} \phi_k$ is not a density.

The case $k<2$:

I'm not sure if any easy proof is available. An indirect way to prove $\phi_k > 0$ is to show $$ \varphi_k(t) = \int_0^\infty e^{-t^2 y} \, \rm{d} \mu_k(y) ,\quad t \in \mathbb{R}, \tag{1} \label{eq:sample} $$ for some non-negative Radon measure $\mu_k$. Then we have, $$ \phi_k(x) = \frac12 \int_{-\infty}^\infty e^{i x t} \varphi_k(t) \, \rm{d} t = \frac12 \int_0^\infty \left[ \int_{-\infty}^\infty e^{i x t - t^2 y} \, \rm{d} t \right] \, \rm{d} \mu_k(y) = \frac{\sqrt{\pi}}{2} \int_0^\infty y^{-\frac12} e^{-\frac{x^2}{4y}} \, \rm{d} \mu_k(y). $$ However, it is not straightforward to find the representation \eqref{eq:sample}. The details can be found in the paper "Stable laws of probability and completely monotone functions" by S. Bochner.

  • $\begingroup$ Why can you think of $\frac{1}{\pi} \phi_k(x)$ as a density? It can be negative. $\endgroup$ – Boby Aug 19 '16 at 15:02
  • 1
    $\begingroup$ By the properties of the Fourier transform you have that the total integral of $\frac{1}{\pi} \phi_k(x)$ is $1$. In the case $k > 2$, you assume that it is a density and reach a contradiction (so it is negative). $\endgroup$ – sometempname Aug 19 '16 at 17:18
  • $\begingroup$ Question: if the function integrates to one and can not be a pdf, the this implies that this function is negative? Never saw this argument before. I hope you don't mind if I ask you a few more question? I can be slow at times? $\endgroup$ – Boby Aug 19 '16 at 18:53
  • $\begingroup$ Also for the part $k>2$ at what point do we use the assumption $k>2$. $\endgroup$ – Boby Aug 19 '16 at 19:00
  • $\begingroup$ I think I got that par, for $k \le2$ the second derivative at zero is not zero. $\endgroup$ – Boby Aug 19 '16 at 19:04

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.