Let $\alpha\in\mathbb R^d$, with $\alpha\neq 0$. Take a sequence of iid random variables $X^{1},\dots,X^{n}$ with values in $\mathbb R^d$ and denote $R$ the cdf of $\alpha X^1$ (where the product is a scalar product). Now consider $\alpha^n=g(X^{1},\dots,X^{n})$ for some function $g$, and suppose we know that $\alpha^n\to\alpha$ almost surely as $n\to\infty$. Define
$\hat R_n(t)=\frac{1}{n}\displaystyle\sum_{i=1}^n \mathbf 1_{\left\{\alpha^nX^i\leq t\right\}},$
which is the empirical distribution of $\alpha^nX^{1},\dots, \alpha^nX^{n}$.
Can we say that $\hat R_n(t)\to R(t)$ as $n\to\infty$?

  • $\begingroup$ What kind of convergence are you looking for? Point-wise or uniform in $t$? $\endgroup$ – Santiago Jan 2 '14 at 16:58
  • $\begingroup$ I'm interested in anything that can be said about this convergence, of course the stronger result would be better! $\endgroup$ – splinter123 Jan 3 '14 at 11:54

The answer is yes for $d=1$. By denoting $F$ the cdf of $X^1$, you have that $\hat{F}_n(t)$ converges uniformly to $F(t)$ by the Glivenko–Cantelli theorem. You want to show that $\hat{F}_n(t/\alpha_n)$ converges uniformly to $F(t/\alpha)$. You can check this on open sets $(-\infty,t)$. You get $$ \left|\int_{-\infty}^{t/\alpha_n} d\hat{P}_n-\int_{-\infty}^{t/\alpha} dP\right|\le \left|\int_{-\infty}^{t/\alpha} d\hat{P}_n-dP\right|+\left|\int_{-\infty}^{t/\alpha_n-t/\alpha} d\hat{P}_n-dP\right|+\left|\int_{-\infty}^{t/\alpha_n-t/\alpha} dP\right|$$

Now use uniform convergence of $\hat{F}_n$ to $F$ to make the first two terms as small as you want. Then you can use uniform convergence of $t/\alpha_n\rightarrow t/\alpha$ to make the third term small (assuming that $\alpha\neq 0)$.

  • $\begingroup$ thanks a lot, this works in a univariate setting. But I was actually mainly concerned with the multivariate version of the result, where this kind of technique doesn't seem to be applicable... I restated the question in its more general form. $\endgroup$ – splinter123 Jan 3 '14 at 13:31
  • $\begingroup$ for d>1, same proof goes through once you have a Glivenko-Cantelli type result. This is the case for instance if your random variables are compactly supported. I do not know what is the most general form of result you can get. I think with suitable choice of estimator (e.g. good kernel estimators instead of empirical mean) or weaker modes of convergence (e.g. in probability) you can make things work. this is perhaps a good reference: stat.washington.edu/jaw/RESEARCH/TALKS/talk2.pdf $\endgroup$ – adrido Jan 3 '14 at 18:13

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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