2
$\begingroup$

I have been working on bounding $\rho_{\nu_1^{e_1} \nu_2^{e_2}}(u)$ for $\nu_i \sim \mathcal{N}(0,1)$ i.i.d. for large $u$, i.e. the density of a product of (positive integer) powers of independent Gaussians, as

$$ \rho_{\nu_1^{e_1} \nu_2^{e_2}}(u) = \Omega\big(\exp(-u^c /2)\big) \tag{1} \label{eq1}$$ for some $c$. It is not too difficult to show that $$\rho_{\nu_i^{e_i}}(u) = \Omega\big(\exp(-u^{2/e_i}/2)\big).$$ I suspect that the tails of the product should be at most as large as the tails of the bigger of the two, i.e. of $\nu_1^{e_1}$, assuming $e_1 \geq e_2$. I have however only been able to show a statement in that direction for the probability, namely $$P(\lvert \nu_1^{e_1} \nu_2^{e_2} \rvert \geq u) \leq C \exp(-x^{1/e_1}/2),$$ using the Markov inequality. My question is, is there a $c$ such that equation \eqref{eq1} holds, and if yes, how large can $c$ be?

Thank you for any suggestions.

$\endgroup$

1 Answer 1

1
$\begingroup$

Let $Z_i:=\nu_i$, $m:=e_1$, and $n:=e_2$. We want to find a bound on the tails of the pdf $f_X$ of the random variable (r.v.) \begin{equation*} X:=UV, \end{equation*} where \begin{equation*} U:=|Z_1|^m,\quad V:=|Z_2|^n. \end{equation*} Indeed, if $m$ and $n$ are both even, then $X=Y:=Z_1^m Z_2^n$, and otherwise $X=|Y|$ and the r.v. $Y$ is symmetric, so that $f_Y(y)=f_X(y)1(y>0)/2+f_X(-y)1(y\le0)/2$ for all real $y$, where $f_Y$ is the pdf of $Y$.

The pdf $f_X$ is the multiplicative convolution of the pdf $f_U$ and $f_V$ of the positive r.v.'s $U$ and $V$: for $x>0$, \begin{equation*} f_X(x)=\int_0^\infty\frac{dy}y\, f_U(x/y)f_V(y). \end{equation*} Also, \begin{equation*} f_U(u)=\frac2m u^{1/m-1}f(u^{1/m}),\quad f_V(v)=\frac2n v^{1/n-1}f(v^{1/n}) \tag{0} \end{equation*} for positive $u,v$, where $f$ is the pdf of $N(0,1)$. So, for $x>0$, \begin{equation*} f_X(x)=\frac2{\pi mn}\int_0^\infty dy\,x^{1/m-1}y^{1/n-1/m-1}\exp\{-[(x/y)^{2/m}+y^{2/n}]/2\}. \end{equation*} Making here the substitution $y=x^{an/2}z$, where \begin{equation*} a:=\frac2{m+n}, \tag{1} \end{equation*} we have \begin{equation*} f_X(x)=\frac2{\pi mn}\,x^b\int_0^\infty dz\,z^{1/n-1/m-1}\exp\{-x^a g(z)/2\}, \end{equation*} where $b:=1/m-1+(1/n-1/m)an/2=a-1$ and \begin{equation*} g(z):=z^{-2/m}+z^{2/n}. \end{equation*} So, by standard asymptotic analysis, \begin{equation*} f_X(x)\sim A\,x^{a/2-1}\exp\{-Gx^a/2\} \tag{2} \end{equation*} as $x\to\infty$, where \begin{equation*} G:=\min_{z>0}g(z)=g(z_*)=m^{-\frac{m}{m+n}} n^{-\frac{n}{m+n}} (m+n), \end{equation*} \begin{equation*} z_*:=\left(\frac{n}{m}\right)^{\frac{m n}{2 (m+n)}}, \end{equation*} and
\begin{equation*} A:=\frac2{\pi mn}\,\sqrt{2\pi}\, \frac{z_*^{1/n - 1/m - 1}}{\sqrt{g''(z_*)/2}} =\frac2{\sqrt{\pi(m+n)}} \, m^{-\frac{m}{2 (m+n)}} n^{-\frac{n}{2 (m+n)}}. \end{equation*} Thus, the exact exponent $a$ of $x$ in the term $\exp\{-Gx^a/2\}$ in (2) is given by (1). (This exponent is better than the corresponding exponent you had in the upper bound on the tail probability.)


Note that (2) will hold for all real $m,n>0$. That the exact exponent $a$ depends only on the total $m+n$ of the exponents of $Z_1$ and $Z_2$ in $Z_1^m Z_2^n$ seems quite reasonable. Moreover, the roles of $m$ and $n$ on the right-hand side of (2) are interchangeable, as they should be. Furthermore, in the limit cases when $m\downarrow0$ or $n\downarrow0$, (2) is in agreement with (0).


Here is the graph of the ratio of $f_X(x)$ to the right-hand side of (2) for $m=3$, $n=4$, and $x\in[1,200]$:

enter image description here

$\endgroup$
2
  • $\begingroup$ Thank you very much for the detailed explanation, it took me a while to work through it. In particular the link to Laplace's Method was very helpful, as I have no background in asymptotic analysis. Do you know if there is theory on whether the asymptotic expression is actually an upper bound for the density for some $u \geq \phi(m,n)$? It would seem so from the graph you plotted. $\endgroup$ Commented Nov 23, 2021 at 15:01
  • $\begingroup$ @MathsIsFun : I don't think that there is a theory on whether the asymptotic expression is actually an upper bound. Such an upper bound would probably be much harder to get than the asymptotic expression. $\endgroup$ Commented Nov 23, 2021 at 15:33

You must log in to answer this question.

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