5
$\begingroup$

A sequence $z_n$ with $\lim_n z_n = z$ is said to have $Q$-linear convergence if a constant $r\in (0,1)$ exists such that

$\displaystyle |z_{n+1} - z| \leq r \, |z_n - z|$,

where $r$ is called the rate of convergence.

My question is if this notion of rate of convergence can be applied to a sequence of random variables (RV) $X_n$ converging in probability to a RV $X$:

$\lim_n P(|X_n - X| \geq \varepsilon) \to 0, \quad \forall \varepsilon > 0.$

Assume that we know that

$\displaystyle P(|X_{n+1} - X| \geq \varepsilon) \leq r_\varepsilon \, P(|X_n - X| \geq \varepsilon),\quad \forall \varepsilon > 0.$

Here $r_\varepsilon \in (0,1)$, is a non-decreasing function of $\varepsilon$, with $\lim_{\varepsilon \to 0^+} r_\varepsilon = 1.$

Is it OK to say that $r_\varepsilon$ is a ($\varepsilon$-dependent) rate of convergence in probability for $X_n$?

More generally, I would like to know if there a widely accepted definition of rate of convergence for convergence in probability, and if so, what would be the definition be?

$\endgroup$
  • $\begingroup$ I don't believe it makes sense to call $r_\epsilon$ a rate of convergence because it doesn't involve $n$ in any way. The point seems to be that $n$ is in the exponent in your condition so the rate of convergence is at least geometric, and I'm not sure what else there is to say. $\endgroup$ – dsaxton Jul 6 '16 at 14:41
  • $\begingroup$ Some definitions of convergence rate do not involve $n$. I've edited the question to make clear what I mean by rate of convergence. $\endgroup$ – pariasm Jul 7 '16 at 8:05
6
$\begingroup$

I would argue that the most widely accepted definition of a convergence rate uses the "big-Oh" and "small-oh" notation.

That is, convergence in probability is written as $z_n-z=o_p(1)$, while a rate of convergence could be indicated via a statement like $z_n-z=O_p(n^{-\alpha})$, which says that $z_n-z$ remains stochastically bounded even when divided by $n^{-\alpha}$, or multiplied by $n^{\alpha}$. A leading case is $\alpha=1/2$. Hence, $z_n-z$ must vanish (converge) at rate $n^{\alpha}$.

Of course, you are right that not all convergence rates need to directly relate to $n$. In nonparametric estimation, it is for example of interest to see what happens to the bias of a density estimator as the underlying bandwidth $h\to0$.

$\endgroup$

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.