1
$\begingroup$

I am confused by this problem. Let $F$ be a distribution function with $F(0-)=0$ and $F(1)=1$ and let $\mu$ be the associated law.

Let $m_k=\int_{[0,1]}x^k dF(x)$.

Define \begin{array}{c c c} \Omega=[0,1]\times [0,1]^{\mathbb{N}} &\Sigma=\mathcal{B}\times\mathcal{B}^{\mathbb{N}} & P=\mu\times Leb^{\mathbb{N}} \\ \Theta(w)=w_0 &H_k(w)=I_{[0,w_0]}(w_k) & S_n=H_1+...+H_n .\\ \end{array} Here $\mathcal{B}^{\mathbb{N}}$ and $Leb^{\mathbb{N}}$ denote respectively the product Borel sigma algebra and product lebesgue measure on $[0,1]^{\mathbb{N}}$.

The problem says that by the strong law and Fubini's theorem it follows that $$S_n/n\longrightarrow \Theta \quad a.s.$$

I am confused by this claim. The way I interpret the problem is that $w=(w_0,w_1,...)$ and $\Theta$ is a function defined on this sequence acting as the projection on the first component. The strong law of large numbers says that the averages converge to the expected value which is a constant, so how can they converge to a non constant function?

Also if I compute this expected value I get

$$ E(H_1)=\int_0^1\int_0^1 I_{[0,\omega_0]}(\omega_1) dF(\omega_0) \: dLeb(\omega_1)=\int_0^1 \omega_0 dF(\omega_0)=m_1$$ because all the other integrals are $1$ since the function doesn't depend on the other variables... Where am I going wrong?

$\endgroup$
  • $\begingroup$ It took me a few minutes to figure out that $Leb^{\mathbb{N}}$ means "the Lebesgue product measure on $\mathbb{R}^{\mathbb{N}}$", so you might clarify that in your question. $\endgroup$ – Ian May 3 '15 at 17:39
1
$\begingroup$

For each fixed $w_0$, the random variables

$$[0,1]^{\mathbb{N}} \ni \tilde{w} \mapsto H_k(w_0,\tilde{w})$$

are iid. Therefore by, the strong law of large numbers,

$$\frac{S_n(w_0,\cdot)}{n} \to \mathbb{E}H_1(w_0) \qquad \text{$\text{Leb}^{\mathbb{N}}$-almost surely.}$$

This means that there exists a set $N(w_0)$, $\text{Leb}^{\mathbb{N}}(N(w_0))=0$, such that

$$\frac{S_n(w_0,\tilde{w})}{n} \to \mathbb{E}H_1(w_0) \qquad \text{for all $\tilde{w} \in [0,1]^{\mathbb{N}} \backslash N(w_0)$.}$$

Since

$$\mathbb{E}H_1(w_0,\cdot) = \int_0^{w_0} \, \text{Leb}(dw_k) = w_0$$

we get

$$\frac{S_n(w_0,\tilde{w})}{n} \to w_0 \qquad \text{for all $\tilde{w} \in [0,1]^{\mathbb{N}} \backslash N(w_0)$.}$$

By Tonelli's theorem,

$$\begin{align*}& P\left(\{(w_0,\tilde{w}): \frac{S_n(w_0,\tilde{\omega})}{n} \, \text{does not converge to $\omega_0$}\} \right) \\ &= P(\{(w_0,\tilde{w}); \tilde{w} \in N(w_0)\})\\ &= \int_{[0,1]} \underbrace{\int_{[0,1]^{\mathbb{N}}} 1_{N(w_0)}(\tilde{w}) \, d\text{Leb}^{\mathbb{N}}(\tilde{w})}_{0} \, d\mu(w_0)\\ &=0. \end{align*}$$

Hence,

$$\frac{S_n}{n} \to \theta \qquad \text{$P$-almost surely}.$$

$\endgroup$
  • $\begingroup$ Thank you for the detailed explanation $\endgroup$ – bibo May 3 '15 at 21:53
1
$\begingroup$

The $H_k$ are random variables. For each fixed $w_0 \in [0,1]$, they are iid: specifically they are iid Bernoulli($w_0$) variables (since $w_k,k=1,2,\dots$ are iid U(0,1) variables). So by the strong law, for each fixed $w_0$, $S_n/n$ converges a.s. to the expected value of an iid Bernoulli($w_0$) variable, which is $w_0=\Theta(w)$. Now you just need to check that this implies that $S_n/n$ converges a.s. (in your full probability space) to $\Theta(w)$. This is the Fubini/Tonelli step.

What's going on here is only slightly more complicated than a situation where $X_k(t)$ are iid Bernoulli(t) variables for each $t \in [0,1]$, in which case $S_n/n \to t$ by the strong law. The difference is that in this setting $t$ is deterministic, whereas in your setting, $w_0$ is also random, but it is chosen according to a distribution independent of the distribution of the $w_k$ for $k \geq 1$.

$\endgroup$

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.