3
$\begingroup$

Consider a sequence $X_n$ of Gaussian random variables with mean $\mu_n$ and variance $\sigma_n^2$, which converges in distribution (to some limiting distribution). Can I then conclude that $\mu_n$ converges to some $\mu$ and $\sigma_n^2$ converges to some $\sigma^2$, with the limiting distribution being $N(\mu, \sigma^2)$? Here, I think I'd allow the degenerate Gaussian distribution with zero variance.

I've tried looking this up elsewhere on this site, but all those I've found assume almost sure/L2 convergence. I was thinking this should be true, and was trying to do this by characteristic functions but without assuming the convergence of the two parameters, I can't seem to conclude anything.

$\endgroup$
2
$\begingroup$

The following holds: Let $(\xi_n)$ and $(\sigma_n)$ be sequences in $\mathbb{R}$, where $\sigma_n>0$. Let $\mu_n$ denote the Gaussian distribution with mean $\xi_n$ and variance $\sigma_n^2$. Then $(\mu_n)$ converges weakly if and only if $(\xi_n)$ and $(\sigma_n^2)$ both converge, and in the affirmative case, the limiting distribution is a Gaussian distribution with mean $\xi = \lim_n \xi_n$ and variance $\sigma^2 = \lim_n \sigma_n^2$. Here, the case $\sigma=0$ is understood to indicate the Dirac measure in $\xi$.

You are correct in observing that when proving this, the difficult part is to show that $(\xi_n)$ and $(\sigma_n^2)$ when $(\mu_n)$ converges weakly. Here are some hints. Assume that $(\mu_n)$ converges weakly. It then holds that

$$ \lim_{M\to\infty} \sup_{n\ge1} \mu_n([-M,M]^c) = 0, $$ essentially meaning that $(\mu_n)$ is a tight family of measures. Use this and the properties of Gaussian distributions to show that both $(\xi_n)$ and $(\sigma_n^2)$ are bounded sequences. Assume, expecting a contradiction, that the sequences are not convergent. As the sequences are bounded, it must in particular hold that $(\xi_n)$ has two different limit points. Use this to obtain a contradiction. Thus, $(\xi_n)$ is convergent. Use a similar technique to obtain that $(\sigma_n^2)$ is convergent.

$\endgroup$
  • $\begingroup$ Thanks! I've tried to prove the first part about tightness for a while now, but I'm not getting anywhere... what properties of convergence in distribution am I supposed to use? The definition I have got is that there is pointwise convergence of the cdf's wherever the limiting distribution is continuous, but that's all I know. $\endgroup$ – Richard Feb 14 '15 at 9:46
  • 1
    $\begingroup$ It's a standard result about weak convergence, in fact. See e.g. Lemma 3.1.6 of these notes for a proof: math.ku.dk/noter/filer/vidsand12.pdf Essentially, the proof works by approximating and indicator function with a continuous (here, even uniformly continuous) and bounded mapping. $\endgroup$ – Alexander Sokol Feb 15 '15 at 11:30
  • $\begingroup$ I've given this another go, but I believe I'm not getting anywhere -- I can't show that the sequences are bounded, nor arrive at any sort of contradiction assuming boundedness and that there were two limits. It might be just that I just haven't got sufficient mathematical machinery to deal with such a problem yet, but would it be possible for you to type out a fuller solution to this? Though if you're busy and can't do it, it's okay as well. Thanks :) $\endgroup$ – Richard Feb 15 '15 at 15:47
  • $\begingroup$ show $\lim_{n \to \infty} exp(i \mu _n t)= \phi (t) exp(\sigma^2 t^2 /2)$ and then show above equality with two different limit points of $\mu_n$ (or just consider limsup, liminf) is a contradiction. $\endgroup$ – izimath Jun 6 '18 at 8:20
0
$\begingroup$

A sequence of complex number $z_n$ converges to a limit $z$ iff $|z_i| \to |z|$ and $\arg(z_i) \to \arg(z)$. Now we have the following from Levy continuity theorem, $$\lim_{n \to \infty}e^{-\sigma_n^2t^2/2}e^{i\mu_nt} \to r(t)e^{i\theta(t)}, \forall t \in \mathbb{R}$$ $$\iff \lim_{n \to \infty}e^{-\sigma_n^2t^2/2} \to r(t), \lim_{n \to \infty}\mu_nt \to \theta(t), \forall t \in \mathbb{R}$$ Now you can see this implies $\mu_n,\sigma_n$ does converge to constant. For example, $$\lim_n \mu_n=\theta(t)/t=const$$ as LHS does not contain $t$. Some details here to verify to be completely rigorous.

$\endgroup$

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.