Consider the following probability density function parameterized by $\sigma>0$: $$g_\sigma(x)=2\sum_\limits{n=0}^\infty\phi_\sigma(x+n), \quad 0<x<1$$ where $$\phi_\sigma(t)={1\over \sigma\sqrt{2\pi}}\exp\left({-{1\over 2}\left({t\over\sigma}\right)^2}\right),\quad -\infty<t<\infty.$$

Context: $g_\sigma$ and $\phi_\sigma$ are the probability density functions of $\{|\sigma Z|\}$ and $Z$ , respectively, where $Z$ is a standard normal random variable, $\sigma>0$, and $\{x\}:=x-\lfloor x\rfloor$ is the fractional part of $x$. These arose in my answer to a related question.

Conjecture 1: For any $x\in(0,1)$, $\lim_{\sigma\to\infty}g_\sigma(x)=1$; i.e., $\{|\sigma Z|\}\overset{d}{\to}\text{Uniform}(0,1).$

This seems quite clear from plots of $g_\sigma$ for various values of $\sigma$: enter image description here

In fact, the asymptotic behavior is so regular, that the following seems to hold:

Conjecture 2. There is a constant $c\approx 0.4$ such that for all $\sigma>1$, $$g_\sigma(x) = 1+{c\over \sigma}(1-2x) + o(\sigma^{-1}),\quad 0<x<1.$$ (But what is the exact value of $c$?)

Here's a picture of the corresponding linear approximations using $c=0.4$, superimposed onto the previous plot:

enter image description here

Question: How to prove these conjectures?

Here's what I've tried ...

Fourier transforms can be used for the probability density (call it $f_\sigma$) of the fractional part without the absolute value, i.e. the density of $\{\sigma Z\}$. Using this method for $f_\sigma$ turns out to give $$\begin{align}f_\sigma(x) &= \sum_{n=-\infty}^\infty\phi_\sigma(x+n)\quad (0<x<1)\\[2ex] &=\sum_{k=-\infty}^\infty\hat\phi_\sigma(k)e^{i2\pi k x},\quad \hat\phi_\sigma(k):=\int_{-\infty}^\infty\phi_\sigma(x)e^{-i 2\pi k x}dx=2\int_{0}^\infty\phi_\sigma(x)\cos(2\pi k x)\,dx\\[2ex] &=1 + 2\,\sum_{k=1}^\infty\,\exp\left(-2(\pi k\sigma)^2 \right)\ \cos(2\pi k x)\quad (0<x<1)\\[2ex] &=1 + 2\exp(-2\pi^2\sigma^2)\cos(2\pi x) + o(e^{-\sigma^2})\quad (0<x<1) \end{align}$$ which has a very much more rapid convergence to uniformity than does $g_\sigma$.

Proceeding similarly with our case of $g_\sigma$, let $\psi_\sigma(x) = \phi_\sigma(x)\,1_{(0,\infty)}(x)$, so we have $$g_\sigma(x)=2\sum_{n=-\infty}^\infty\psi_\sigma(x+n)\quad (0<x<1)$$ and it seems that similar Fourier transform methods ought to apply here as well.

However, when I write (using Poisson's summation formula) $$g_\sigma(x)=2\sum_{k=-\infty}^\infty\hat\psi_\sigma(k)e^{i2\pi k x} $$ with $$\begin{align}\hat\psi_\sigma(k)&=\int_{-\infty}^\infty\psi_\sigma(x)e^{-i 2\pi k x}dx\\[2ex] &=\int_{0}^\infty\phi_\sigma(x)e^{-i2\pi k x}dx \end{align}$$ I get, using Sage, that

$$\hat\psi_\sigma(k)={1\over 2}e^{-2(\pi k\sigma)^2}\left( 1 - \text{erf}\left( i \sqrt{2}\pi k\sigma\right)\right). $$ The trouble is that inserting this into the summation gives $$g_\sigma(x) = \sum_{k=-\infty}^\infty e^{-2(\pi k\sigma)^2}\left( 1 - \text{erf}\left( i \sqrt{2}\pi k\sigma\right)\right)e^{i 2\pi k x} $$ and I don't know how to proceed further. (I'm rusty with Fourier methods, so I would appreciate any advice as well as corrections.)

  • $\begingroup$ At one point, due to some simulation discrepancies, I doubted whether $g_\sigma$ (as defined in my opening sentence) is truly the pr. dens. of $\{|\sigma Z|\}$, so I edited to treat it as just a given pr. density function; however, I then discovered that the trouble was in the simulation, and my doubt was unfounded -- so I rolled back the edit to the original version and added a link to a derivation of $g_\sigma$. $\endgroup$ – r.e.s. Apr 9 '18 at 3:18

Here's a proof of the conjectures ...

For any $\sigma>0$, $$\begin{align}\Pr\left( \{|\sigma Z|\} \le x \right) &=\sum\limits_{k=0}^\infty\Pr\left(-k-x< \sigma Z \le -k \right) +\sum\limits_{k=0}^\infty\Pr\left(k< \sigma Z \le k+x \right)\\[2ex] &=2\,\sum\limits_{k=0}^\infty\Pr\left(k< \sigma Z \le k+x \right)\\[2ex] &=2\,\sum\limits_{k=0}^\infty{\large(}\Pr\left(\sigma Z \le k+x \right)-\Pr\left(\sigma Z \le k\right){\large)}\\ \end{align}$$ and differentiating to obtain the probability density ... $$\begin{align}g_\sigma(x) &=2\,\sum\limits_{k=0}^\infty \phi_\sigma(k+x)\quad (0<x<1). \end{align}$$ Now let $\psi_\sigma(x) = \phi_\sigma(x)\,1_{(0,\infty)}(x)$, so we have $$g_\sigma(x)=2\sum_{n=-\infty}^\infty\psi_\sigma(x+n)\quad (0<x<1)$$ and Poisson's summation formula can be applied to give $$g_\sigma(x)=2\sum_{k=-\infty}^\infty\hat\psi_\sigma(k)e^{i2\pi k x} $$ where $$\begin{align}\hat\psi_\sigma(k)&=\int_{-\infty}^\infty\psi_\sigma(x)e^{-i 2\pi k x}dx\\[2ex] &=\int_{0}^\infty\phi_\sigma(x)e^{-i2\pi k x}dx\\[2ex] &={1\over 2}e^{-2(\pi k\sigma)^2}\left( 1 - \text{erf}\left( i \sqrt{2}\pi k\sigma\right)\right). \end{align}$$ Inserting this into the summation gives

$$\begin{align}g_\sigma(x) &= \sum_{k=-\infty}^\infty e^{-2(\pi k\sigma)^2}\left( 1 - \text{erf}\left( i \sqrt{2}\pi k\sigma\right)\right)e^{i 2\pi k x}\\[2ex] &=\sum_{k=-\infty}^\infty e^{-2(\pi k\sigma)^2}e^{i 2\pi k x} - \sum_{k=-\infty}^\infty e^{-2(\pi k\sigma)^2}\text{erf}\left( i \sqrt{2}\pi k\sigma\right)e^{i 2\pi k x}.\\[2ex] \end{align}$$

Note that $e^{-2(\pi k\sigma)^2}$ and $\cos(2\pi k x)$ are even functions of $k$, while $\text{erf}\left( i \sqrt{2}\pi k\sigma\right)$ and $\sin(2\pi k x)$ are odd functions of $k$, so the summations simplify to give

$$\begin{align}g_\sigma(x) &= 1+2\sum_{k=1}^\infty e^{-2(\pi k\sigma)^2}\cos( 2\pi k x) - i\,2\sum_{k=1}^\infty e^{-2(\pi k\sigma)^2}\text{erf}\left( i \sqrt{2}\pi k\sigma\right)\sin(2\pi k x)\\[2ex] &=\underbrace{1+2\sum_{k=1}^\infty e^{-2(\pi k\sigma)^2}\cos( 2\pi k x)}_{f_{\sigma}(x)} + \,2\sum_{k=1}^\infty e^{-2(\pi k\sigma)^2}\text{erfi}\left(\sqrt{2}\pi k\sigma\right)\sin(2\pi k x)\\[2ex] \end{align}$$

where $f_\sigma(x)=1+o(e^{-\sigma})$ is the probability density of $\{\sigma Z\}$ as derived earlier, and $\text{erfi}(x):=i\,\text{erf}(i\,x)$ for real $x$. Now, it can be shown that asymptotically for $x\to\infty$, $$\text{erfi}(x) = {1\over\sqrt{\pi}}x^{-1}e^{x^2}(1+O(x^{-2})) $$ so we have $$\begin{align}g_\sigma(x) &= f_\sigma(x)+{1\over\sigma}\sqrt{2\over\pi^3}\sum_{k=1}^\infty {\sin(2\pi k x)\over k}+O(\sigma^{-2}).\\[2ex] \end{align}$$

The conjectures are then established by noting that $$\sum_{k=1}^\infty {\sin(2\pi k x)\over k}={\pi\over 2}(1-2x) \quad(0<x<1) $$ because the LHS is just the Fourier series representation of the odd periodic extension of the RHS, the $k$th Fourier coefficient being given by $$2\int_0^1 {\pi\over 2}(1-2x)\sin(2\pi k x)\,dx={1\over k}. $$ Consequently, both conjectures are true, and in Conjecture 2 the constant is $c={1\over\sqrt{2\pi}}=0.39894...\,:$ $$g_\sigma(x) = 1+{1\over\sigma\sqrt{2\pi}}(1-2x) + O(\sigma^{-2}),\quad 0<x<1. $$


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.