3
$\begingroup$

I am a physicist caught in the following situation: I have two probability measures $\mathbb{P}_1$ and $\mathbb{P}_2$ and have to deal with the following integral where $X_i$ are random iid:

$$\int_{B} \mathbb{P}_1\left(\frac{1}{N} \sum_{i=1}^N X_i \ge x\right) d\mathbb{P}_2(x).$$

I was able to obtain a large deviation principle for $X$, i.e. I was able to show that

$$\frac{1}{N} \log\left( \mathbb{P}_1\left(\frac{1}{N} \sum_{i=1}^N X_i \ge x\right)\right) = -I(x).$$

Now it is tempting to replace $\mathbb{P}_1\left(\frac{1}{N} \sum_{i=1}^N X_i \ge x\right) $ in the above integral by $e^{-N I(x)}$ and compute $$\int_{B} e^{-NI(x)} d\mathbb{P}_2(x).$$

However, I do not know in which sense this now close to the expression I am looking for?

Are there any standard bounds to estimate the error between my approximation and the object I am interested in?

$\endgroup$
8
  • $\begingroup$ The function I has to be monotone and the rate on the integral ought to be the smallest x in the support of $P_2$. $\endgroup$
    – user83457
    Dec 18, 2017 at 11:39
  • $\begingroup$ @michael The function $I$ is indeed monotone. What precisely does the second part of your comment mean? $\endgroup$ Dec 18, 2017 at 11:57
  • $\begingroup$ That if V is the value of your integral $\frac {log(V)} N \rightarrow -\inf \lbrace I(x) \rbrace$ where the inf is over x in the support of $P_2$. If $P_2$ is really nice, like uniform on an interval, I think it is easy to write down, but if it has 0 or no density at the lower limit it is probably tedious. $\endgroup$
    – user83457
    Dec 18, 2017 at 13:34
  • $\begingroup$ On second thought, If $P_2$ is really nasty I don't know how bad it can get. What do you have in mind for $P_2$? $\endgroup$
    – user83457
    Dec 18, 2017 at 13:42
  • $\begingroup$ I am not quite sure you understood my question. I am interested in comparing the value of $$\int_{B} \mathbb{P}_1\left(\frac{1}{N} \sum_{i=1}^N X_i \ge x\right) d\mathbb{P}_2(x)$$ to the value of $$\int_{B} e^{-NI(x)} d\mathbb{P}_2(x).$$ $\endgroup$ Dec 18, 2017 at 14:02

1 Answer 1

2
$\begingroup$

It all depends on the shape of $P_2$ and on the assumptions you put on $X_i$. In what follows I'll assume that $\Lambda(\lambda)=\log E_1 e^{\lambda X_1}$ is finite for all $\lambda$. I will also assume that $E_1X_i=0$. Further I will assume that $P_2$ is supported on $R$ with density $f$.

Case 1: $B\cap (-\infty,0]>0$. In that case your formula may be false, and the answer is essentially $P_2(B\cap (-\infty,0))$.

Case 2: (which is what I guess you had in mind): $B\cap (-\infty,a)=\emptyset$ for some $a>0$. In that case, use Bahadur-Rao to approximate, for $x\in B$, $P_1(N^{-1}\sum_{i=1}^N X_i>x)\sim C(x)e^{-NI(x)}/\sqrt{N}$, with explicit $C=C(x)$. So the expression you write is correct at the exponential scale but if you meant up to precise asymptotics, you are missing a constant multiple and a factor $\sqrt{N}$. The constant $C(x)$ depends on whether the law of $X_1$ is lattice or not. See the original paper of Bahadur-Rao or Theorem 3.7.4 in Dembo-Zeitouni's large deviations book.

One can also deal with other cases ($B$ touching $0$ with density of $P_2$ vanishing there, etc.) but I'll stop here.

$\endgroup$
2
  • $\begingroup$ thank you. I wonder where the the first case comes from though. Let us say if $\mathbb{P}_2$ is Gaussian. Then, we are in Case $1$, right? So how does this one come about then? $\endgroup$ Dec 20, 2017 at 6:22
  • $\begingroup$ Depends what B is. if B charges the negative axis, the probability involving $P_1$ is essentially 1! $\endgroup$ Dec 20, 2017 at 7:24

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.