3
$\begingroup$

Let $\varepsilon$ be a number in $(0, 1)$, consider the following random walk on the real line $X^{(0)}, X^{(1)}, \dots$, where

  • $X^{(0)}=0$
  • If $X^{(t)} > 0$, then with probability $.5$, $X^{(t+1)} = X^{(t)} + (1-\varepsilon)$, and with probability $.5$, $X^{(t+1)} = X^{(t)} - (1+\varepsilon)$
  • If $X^{(t)} < 0$, then with probability $.5$, $X^{(t+1)} = X^{(t)} - (1-\varepsilon)$, and with probability $.5$, $X^{(t+1)} = X^{(t)} + (1+\varepsilon)$
  • If $X^{(t)} = 0$, then we do the same thing as the original random walk: with probability $.5$, $X^{(t+1)} = X^{(t)} + 1$, and with probability $.5$, $X^{(t+1)} = X^{(t)} - 1$.

In other words, if $X^{(t)}$ is biased from $0$, we want to "drag it towards $0$," and the strength is controlled by $\varepsilon$.

Is it the case that for nontrivial setting of $\varepsilon$ (for example, $\varepsilon = .3$), that $X^{(t)}$ is well concentrated around $0$ with smaller variance, compared to variance $t$ (thus a deviation of $\sqrt{t}$) in the case of $\varepsilon=0$?

$\endgroup$
2
  • $\begingroup$ Isn't it positively recurrent for any $\epsilon>0$? $\endgroup$
    – R W
    Commented Oct 29, 2018 at 19:07
  • $\begingroup$ I think so, but I am not sure how that would help the estimate of the variance. Maybe there is something in your mind, if so feel free to teach me about that. $\endgroup$
    – Xi Wu
    Commented Oct 30, 2018 at 2:16

1 Answer 1

4
$\begingroup$

Let me try an answer. [Edit: simplified and (hopefully) corrected.]

Let $\alpha$ be the only positive solution of $\mathrm{ch}\alpha = \exp(\varepsilon\alpha)$, so that $$ \exp(\alpha x) = \frac12\left(\exp(\alpha(x+1-\varepsilon))+\exp(\alpha(x - 1 - \varepsilon))\right)\text. $$ The purpose of the definition lies in the fact that $\mathbb E[\mathbf e^{\alpha|X^{(n+1)}|}|\mathcal F_n]=\mathbf e^{\alpha|X^{(n)}|}$ whenever $|X^{(n)}|\geq 1+\varepsilon$. I claim that:

For $C>0$ large enough, $$\mathbf e^{\alpha|X^{(n)}|}-Cn$$ is a supermartingale.

Indeed, because of the remark above, it is enough to show that for some constant $C>0$, $$\mathbb E[\mathbf e^{\alpha|X^{(n+1)}|}|\mathcal F_n]\leq \mathbf e^{\alpha|X^{(n)}|} + C$$ whenever $|X^{(n)}|< 1+\varepsilon$. This is true for $C=\max\{\mathbf e^{\alpha x},0\leq x\leq 2\}=\mathbf e^{2\alpha}$, for example.

This means that $\mathbb E[\mathbf e^{\alpha |X^{(n)}|}]\leq 1+Cn$, so $X^{(n)}$ cannot explode that much. For example, about the variance, for any positive exponent $p\in\mathbb N$, $$ \mathbb E[|X^{(n)}|^2] \leq \mathbb E[|X^{(n)}|^{2p}]^{1/p} \leq K\mathbb E[\mathbf e^{\alpha|X^{(n)}|}]^{1/p} \leq K'(1+n)^{1/p}\text,$$ so the variance increases slower that any root.


As Martin pointed out below (thank you!), $\mathbb E[\exp(a|X^{(n)}|)]\leq C'$ whenever $a<\alpha$. It can be seen as a consequence of the modified claim:

For $C\geq1$ large enough and $\lambda=\mathrm{ch}(a)\exp(-\varepsilon a)$, $$\mathbb E[\mathbf e^{a|X^{(n+1)}|}|\mathcal F_n]\leq \lambda\mathbf e^{a|X^{(n)}|} + C\text.$$

When $X^{(n)}\geq 1+\varepsilon$, this is a consequence of \begin{align*} \mathbb E[\mathbf e^{a|X^{(n+1)}|}|\mathcal F_n] = & \frac12\left(\exp(a(X^{(n)}+1-\varepsilon))+\exp(a(X^{(n)} - 1 - \varepsilon))\right)\\ = & \mathbf e^{a|X^{(n)}|}\mathrm{ch}(a)\exp(-\varepsilon a)\text, \end{align*} and similarly when $X^{(n)}\leq -1-\varepsilon$; when $|X^{(n)}|<1+\varepsilon$, $$\mathbb E[\mathbf e^{a|X^{(n+1)}|}|\mathcal F_n]\leq\mathbf e^{2a}\text,$$ and the claim follows.

From the claim, and noting that $\lambda<1$, we see that $$ \mathbb E[\mathbf e^{a|X^{(n)}|}|] \leq C(1+\lambda+\cdots+\lambda^n) \leq \frac C{1-\lambda}$$ so $\mathbb E[\exp(a|X^{(n)}|)]$ (hence the variance) is uniformly bounded.

$\endgroup$
4
  • 1
    $\begingroup$ By essentially the same argument you'll find that $\mathbb{E} \exp(a|X^{(n)}|)$ remains bounded uniformly in $n$ whenever $a < \alpha$. $\endgroup$ Commented Oct 29, 2018 at 21:19
  • $\begingroup$ Thanks and appreciated. I get the solution. Since I am far from being expert I would like to ask some questions for clarifications: (1) How do you come up with the choice of exp(\alpha|X^{(n)}|)? For a while I was considering supermartingale by transforming |X^{(n)}|, but clearly I failed. Is this related to some consideration of moment generating function? (2) (Just to confirm) Regarding constants here, so \alpha, C only depend on \varepsilon, and K only depends on \alpha and p. $\endgroup$
    – Xi Wu
    Commented Oct 30, 2018 at 2:14
  • $\begingroup$ I edited my answer to include the comment of @martin-hairer. Concerning yours: (1) I was looking for a function $f$ such that $\mathbb [f_{n+1}|\mathcal F_n]=f_n$ ($f_n=f(X^{(n)}$), at least away from the origin. I had the intuition that such a function would grow exponentially, and then the calculations (see above) made the appearance of $\alpha$ clear. $\endgroup$
    – Pierre PC
    Commented Oct 30, 2018 at 8:31
  • $\begingroup$ (2) Yes, $\alpha$ and $C$ depend on $\varepsilon$ only in the first approach, whereas $K$ and $K'$ depend on $\alpha$ (hence on $\varepsilon$) and $p$. In the second, $\lambda$ depend on $\varepsilon$ and $a$, and $C$ depends essentially on $\alpha$ (because $a\leq\alpha$). $\endgroup$
    – Pierre PC
    Commented Oct 30, 2018 at 8:41

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .