6
$\begingroup$

I have the following function

$$ \int_{-\infty}^{\infty} \left\{{2^{1/\beta-1/2} \over v}\right\}^{it} { \Gamma\{(it+1)/\beta\}\over \Gamma\{(it+1)/2\} }dt $$

where $1<\beta<2$, $v>0$. Need to show it is positive.

The inverse Mellin transform of

$$ \left\{2^{1/\beta-1/2} \right\}^{it} { \Gamma\{(it+1)/\beta\}\over \Gamma\{(it+1)/2\} } $$ is

$$ {C \over v}\int_{-\infty}^{\infty} \left\{{2^{1/\beta-1/2} \over v}\right\}^{it} { \Gamma\{(it+1)/\beta\}\over \Gamma\{(it+1)/2\} }dt $$

$\endgroup$
9
  • $\begingroup$ It seems $v$ should be positive $\endgroup$ Commented Jul 18, 2020 at 17:53
  • 2
    $\begingroup$ @CarloBeenakker, it is not complex. Do you want me to draft the proof? $\endgroup$
    – Vova
    Commented Jul 18, 2020 at 18:07
  • 2
    $\begingroup$ I just evaluated it numerically for $\beta=3/2$ and $v=1$ and find $1.50029 + 1.65104\, i$ --- see wolframalpha.com/input/… $\endgroup$ Commented Jul 18, 2020 at 18:33
  • 2
    $\begingroup$ @CarloBeenakker, I think the integral should be from $-\infty$ to $+\infty$ $\endgroup$ Commented Jul 18, 2020 at 18:37
  • 1
    $\begingroup$ inverse Mellin transform of ${ \Gamma\{(it+1)/\beta\}\over \Gamma\{(it+1)/2\} $ $\endgroup$
    – Vova
    Commented Jul 19, 2020 at 3:01

1 Answer 1

11
$\begingroup$

$\newcommand\Ga\Gamma \newcommand{\R}{\mathbb{R}} \newcommand{\de}{\delta} \newcommand{\ga}{\gamma} \newcommand{\Si}{\Sigma}$ We have to show that for $a:=-\ln(2^{1/b-1/2}/v)\in\R$ and $b:=\beta\in(1,2)$, \begin{equation*} I(a):=\int_{-\infty}^{\infty} e^{-iat}R(t)\,dt>0, \tag{1} \end{equation*} where \begin{equation*} R(t):=\frac{\Ga\big((1+it)/b\big)}{\Ga\big((1+it)/2\big)}. \tag{2} \end{equation*}

The key is Euler's product formula
\begin{equation*} \Ga(z)=\frac1z\,\prod_{j=1}^\infty\frac{(1+1/j)^z}{1+z/j} \end{equation*} for $z\in\mathbb C\setminus\{0,-1,-2,\dots\}$, which yields \begin{equation*} \frac{\Ga(s+it)}{\Ga(s)}=\prod_{j=1}^\infty(1+1/j)^{it} \Big/\prod_{j=0}^\infty\Big(1+\frac{it}{j+s}\Big); \tag{3} \end{equation*} here and in what follows, $s$ is any positive real number and $t$ is any real number.

Based on (3), it is easy to obtain

Lemma 1: $\ln|\Ga(s+it)|\sim-\pi|t|/2$ as $|t|\to\infty$.

The proof of Lemma 1 will be given at the end of this answer.

It also follows from (3) that \begin{equation*} R(t)=c\prod_{j=1}^\infty(1+1/j)^{iht}f_j(t), \tag{4} \end{equation*} where $c:=\Ga(1/b)/\Ga(1/2)>0$, \begin{equation} h:=\frac1b-\frac12=\frac{2-b}{2b}, \end{equation} and \begin{equation} f_j(t):=\frac{1+it/(1+2j)}{1+it/(1+bj)}, \end{equation} so that $f_j$ is the characteristic function (c.f.) of a random variable (r.v.) $X_j\sim p_j\de_0+(1-p_j)Exp(-1/(1+bj))$, where in turn $p_j:=(1+bj)/(1+2j)\in(0,1)$, $\de_0$ is the Dirac distribution supported on the set $\{0\}$, and $Exp(-1/(1+bj))$ is the exponential distribution with mean $-1/(1+bj)$, supported on the interval $(-\infty,0]$. Here and in what follows, $j$ is any natural number. Note that $EX_j=-\frac{hj}{(j+1/b)(j+1/2)}$ and $Var\,X_j\le1/(bj)^2\le1/j^2$. So, the series \begin{equation} \sum_{j=1}^\infty(X_j-EX_j)=:S \end{equation} converges almost surely. Hence, by (4) \begin{equation*} R(t)=ce^{ihc_1t}f_S(t), \end{equation*} where $f_S$ is the c.f. of the r.v. $S$ and \begin{equation} c_1:=\sum_{j=1}^\infty\Big(\ln(1+1/j)+EX_j\Big) \\ =\sum_{j=1}^\infty\Big(\ln(1+1/j)-\frac{j}{(j+1/b)(j+1/2)}\Big)\in\R \end{equation} (in fact, $c_1=(\ga b-2 b+b \ln4+2 \psi\left(1+1/b\right))/(2-b)$, where $\ga=0.577\dots$ is the Euler constant and $\psi:=\Ga'/\Ga$; however, the actual value of $c_1$ does not matter here).

So, $R$ is the c.f. of the r.v. $T:=hc_1+S$. Also, by Lemma 1, $R$ is in $L^1$. It now follows that the function $I$, defined by (1), is $2\pi$ times the density of the r.v. $T$. Thus, $I(a)\ge0$ for all real $a$, as desired.

It remains to provide

Proof of Lemma 1: By (3), \begin{equation*} \frac{|\Ga(s+it)|}{\Ga(s)}=\prod_{j=0}^\infty\frac{j+s}{|j+s+it|} =\exp\{-\Si_{s,t}/2\}, \end{equation*} where \begin{equation} \Si_{s,t}:=\sum_{j=0}^\infty\ln\Big(1+\frac{t^2}{(j+s)^2}\Big). \end{equation} Since $\ln\big(1+\frac{t^2}{(j+s)^2}\big)$ is nonincreasing in $j$, we have \begin{equation} J_{s,t}\le\Si_{s,t}\le J_{s,t}+\ln\big(1+\frac{t^2}{s^2}\big), \end{equation} where \begin{equation} J_{s,t}:=\int_0^\infty\ln\big(1+\frac{t^2}{(x+s)^2}\big)\,dx\sim\pi|t| \end{equation} as $|t|\to\infty$, which completes the proof of Lemma 1 and the entire answer. (In fact, integrating by parts, for $t\ne0$ we find \begin{equation} J_{s,t}=\pi|t|-s \ln \left(s^2+t^2\right)-2 t \arctan(s/t)+2 s \ln s\sim\pi|t|.) \end{equation} The proof of Lemma 1 and the entire answer are now complete.

$\endgroup$
7
  • $\begingroup$ looks great, I do not quite see Fourier in the original problem, I have $\left\{{2^{1/\beta-1/2} \over v}\right\}^{it}$ though, where did it go? $\endgroup$
    – Vova
    Commented Jul 19, 2020 at 3:52
  • $\begingroup$ @Vova : this power equals $e^{iat} $ for some real $a$. $\endgroup$ Commented Jul 19, 2020 at 4:40
  • $\begingroup$ I still do not see how you got $(1+ui)/(1+cui)$ from ${1+(it)/(1+2j) \over 1+(it)/(1+\beta j) }$, if you do the change of variables, your $u$ depends on $j$ $\endgroup$
    – Vova
    Commented Jul 19, 2020 at 20:17
  • $\begingroup$ @Vova : I have rewritten the answer in a probabilistic and more detailed way. $\endgroup$ Commented Jul 19, 2020 at 21:36
  • $\begingroup$ Great one, I really like the original line of proof as well. I have a follow up question. My $b$ is actually $b_{kl}$ and by knowing how $b$ survives integration, in which functional form it appears on the other end - I can find such $A_{kl}$ such that $A_{kl}*f(b_{kl})_{v}$ is positive definite for any $v$ and any $n$, $k,l=1,\cdots n$ $\endgroup$
    – Vova
    Commented Jul 19, 2020 at 22:23

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .