1
$\begingroup$

I'm interested to see a result where for large degree of freedom $m,$ the chi distribution $\chi_m$ is increasingly well approximated by a family of normal distributions with parameters depending on $m.$ The motivation comes from the fact that for large $m, \chi^2_m \sim \mathcal{N}(m,2m)$ approximately and the difference between the corresponding CDF's approach zero as the d.f. $m \to \infty$, as asked and proved in this question. The problem is: we can't apply some sort of continuous mapping theorem here to pass onto square root. I said "some sort of" because we'd be looking here for a case where $X_m - Y_m \to_d 0 \implies g(X_m) - g(Y_m)\to_{d}0$ with $g$ being a continuous, real-valued function (in our case $X_m:=\chi^2_m, Y_m:=\mathcal{N}(m,2m), g:= \sqrt{}$), and this is not the continuous mapping theorem.

Regardless, the wiki page of chi-squared distribution seems to refer to this theorem, proved by Ronald Fisher:

Approximately, $\sqrt{2\chi^2_m} \sim \mathcal{N}(\sqrt{2m-1},1) \implies \chi_m \sim \mathcal{N}(\sqrt{m-\frac{1}{2}},\frac{1}{\sqrt{2}}).$

Despite the reference to a book on that wiki page, I'm unable to access it and am also unable to otherwise get a reference to this theorem. Do you happen to know the proof or could you refer me to it please? Thank you!

$\endgroup$
4
  • 1
    $\begingroup$ Related: stats.stackexchange.com/questions/241504/… which (since $\chi^2_1$ has a mean of $1$ and variance of $2$) implies $\dfrac{\chi_m -\sqrt{m-1/2}}{1/2}$ converges in distribution to a standard normal distribution $\mathcal N(0,1)$ as $m$ increases, the same result as you attribute to Fisher $\endgroup$
    – Henry
    Commented Aug 20, 2020 at 14:40
  • 1
    $\begingroup$ @Henry Thanks for pointing this out - this is indeed of help here - appreciate it! $\endgroup$ Commented Aug 21, 2020 at 10:25
  • 1
    $\begingroup$ @Henry BTW: a little typo I think: your denominator would be $1/\sqrt{2}$, not $1/2.$ $\endgroup$ Commented Sep 10, 2020 at 14:50
  • 1
    $\begingroup$ Learning math: you are correct but I cannot edit comments. I should have written $\dfrac{\chi_m -\sqrt{m-1/2}}{\sqrt{1/2}}$ converges in distribution to a standard normal distribution $\mathcal N(0,1) $ as $m$ increases $\endgroup$
    – Henry
    Commented Sep 10, 2020 at 15:17

1 Answer 1

2
$\begingroup$

Such a result can be obtained by the so-called delta method, which yields, in particular, the following: if $X_m\sim\chi^2_m$, then the distribution of $\sqrt X_m$ is approximately $N(\sqrt m,1/\sqrt2)$ (for large $m$).

Details: By the central limit theorem, $\overline X_m:=X_m/m$ is approximately normal with mean $\mu:=1$ and standard deviation $\sqrt{2/m}$. Hence, by the delta method, for $g(x)\equiv\sqrt x$, $\sqrt{\overline X_m}$ is approximately normal with mean $g(\mu)=\sqrt1=1$ and standard deviation $|g'(\mu)|\sqrt{2/m}=1/\sqrt{2m}$. Thus, $\sqrt X_m=\sqrt m\,\sqrt{\overline X_m}$ is approximately normal with mean $\sqrt m$ and standard deviation $1/\sqrt2$, as claimed.


On the uniform and nonuniform bounds on the rate of convergence to normality in the (possibly multivariate) delta method, see this and references there.

The convergence of the distribution of $\sqrt X_m$ to normality is in a sense monotonic; cf. formula (2.6).

$\endgroup$
4
  • $\begingroup$ @Ioself: thank you, I'll check out the delta method! Quick question before I do though: will this method imply that the distribution of the two CDF's converge to zero? $\endgroup$ Commented Aug 19, 2020 at 21:23
  • 1
    $\begingroup$ @Learningmath : The uniform (Kolmogorov) distance between the two cdf's does converge to $0$, actually with rate $O(1/\sqrt m)$; stronger than this nonuniform bounds also hold. See the first linked paper of the two added ones. $\endgroup$ Commented Aug 19, 2020 at 21:51
  • $\begingroup$ thanks again! I do wonder if we can also say that the $\sqrt{X_m} = \chi_m$ converges in density to the density of $N_m \sim N(\sqrt{m-1/2}, 1/\sqrt{2}), i.e. lim_{m \to \infty} ||f_{\chi_m} - f_{N_m}||_{L^{\infty}(\mathbb{R})} \to 0, m \to \infty?$ Here $f_Z$ denotes the PDF of the random variable $Z.$ $\endgroup$ Commented Sep 4, 2020 at 14:56
  • 1
    $\begingroup$ @Learningmath : Yes, the pdf convergence also holds, as can be checked directly. $\endgroup$ Commented Sep 4, 2020 at 21:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.