Proof that $L^2 (Ω, F, P)$ is a Hilbert space

126 Views Asked by At

I'm working on the probabilistic space of square random variables ($L^2 (Ω, F, P)$). I need to prove that it is a Hibert Space (the exercise doesn't say with which norm). The usual inner product is $<X,Y>=\mathbb{E}[XY]$, it verifies the necessary properties. So I decide to consider the norm $\lVert X\rVert_2=\sqrt{\mathbb{E}[X^2]}$.

So I know that I need to prove that this space is complete with relation to that norm, and that's what I can't seem to do. I need to consider a Cauchy sequence $\{X_n\}$ in $L^2 (Ω, F, P)$. So for every $\epsilon>0$, there is a N such that for all $n,m>N$ we have : $\lVert X_n-X_m\rVert_2=\sqrt {\mathbb{E}[(X_n-X_m)^2]}\leq \varepsilon$.

I need to prove that there is a X in $L^2 (Ω, F, P)$ such that $\lim_{n\to\infty}\lVert X_n-X\rVert_2=0$

And now from there I don't really know where to go, I saw things about the dominated convergence theorem but I'm unsure about how to use it in this siuation. I hope my question is not too dumb.

1

There are 1 best solutions below

0
On

First of all, the correct inner product depends on whether you're considering real or complex random variables. If you're looking at complex RV's, then the inner product you want is $\left< X, Y \right> : = \mathbf{E} \left[ X \overline{Y} \right]$, but if you're looking at real RV's, then $\mathbf{E}[XY]$ is fine. The norm is then $\left\| X \right\| : = \sqrt{\left< X, X \right>} = \mathbf{E} \left[ |X|^2 \right]^{1/2}$. Either way, the distinction won't make a difference in what follows. I should also note that what I'm about to write out can be found in basically any measure theory text, since any introductory measure theory text will prove that $L^p$ spaces are complete.

Let $(X_n)_{n = 1}^\infty$ be a Cauchy sequence in $L^2$. I'm going to prove that an appropriately chosen subsequence of $(X_n)_{n = 1}^\infty$ converges in $L^2$. This will then imply that $(X_n)_{n = 1}^\infty$ is convergent. I'll prove this at the end, but you can also take it as a worthwhile exercise.

Because $(X_n)_{n = 1}^\infty$ is Cauchy, for each $k \in \mathbb{N}$, there exists $n_k \in \mathbb{N}$ such that if $n \geq n_k$, then $\|X_n - X_{n_k}\| \leq 4^{-k}$. We can assume without loss of generality that $n_1 < n_2 < n_3 < \cdots$. We then have in particular that $\|X_{n_{k + 1}} - X_{n_k}\| \leq 4^{-k}$. I'm going to prove that $(X_{n_k})_{k = 1}^\infty$ converges in $L^2$. In fact, I'm going to do this by proving a slightly different claim: that $(X_{n_k})_{k = 1}^\infty$ converges pointwise, meaning I can use Fatou's Lemma. For convenience, let's write $Y_k = X_{n_{k + 1}} - X_{n_k}$.

For each $k \in \mathbb{N}$, set $$E_k = \left\{ \omega : |Y_k(\omega)| \geq 2^{-k} \right\} = \left\{ \omega : |Y_k(\omega)|^2 \geq 4^{-k} \right\} .$$ By the Chebyshev Inequality, we have that $P(E_k) \leq 4^k \mathbf{E} \left[|Y_k|^2 \right] = 4^k \|Y_k\|^2 \leq 4^k / 16^k = 4^{-k}$. Therefore $\sum_{k = 1}^\infty P(E_k) < \infty$. By the Borel-Cantelli Lemma, it follows that $$P \left( \left\{ \omega \in \Omega : |Y_k(\omega)| \geq 2^{-k} \textrm{ for infinitely many $k \in \mathbb{N}$} \right\} \right) = 0 .$$ This means that for almost all $\omega \in \Omega$, we have that $|Y_k(\omega)| < 2^{-k}$ for sufficiently large $k \in \mathbb{N}$. Write $$F = \left\{ \omega \in \Omega : \exists K \in \mathbb{N} \; \forall k \geq K \; \left( \left| Y_k(\omega) \right| < 2^{-k} \right) \right\} .$$ If $\omega \in F$ (which is almost all $\omega \in \Omega$), then \begin{align*} X_{n_k}(\omega) = X_{n_1}(\omega) + \sum_{j = 1}^{k - 1} Y_j(\omega) , \end{align*} and by the comparison test, this series converges as $k \to \infty$. Therefore $(X_{n_k}(\omega))_{k = 1}^\infty$ converges for all $\omega \in F$, i.e. almost everywhere.

Now write $Z(\omega) = \lim_{k \to \infty} X_{n_k}(\omega)$ for all $\omega \in F$. Since $\Omega \setminus F$ has probability $0$, we can set $Z$ to whatever we want there. Fatou's Lemma then tells us that \begin{align*} \mathbb{E} \left[ |Z|^2 \right]^{1/2} & \leq \liminf_{k \to \infty} \mathbf{E} \left[ |X_{n_k}|^2 \right]^{1/2} \\ & = \liminf_{k \to \infty} \left( \mathbf{E} \left[ \left| X_{n_1} + \sum_{j = 1}^{k - 1} Y_j \right|^2 \right]^{1/2} \right) \\ & = \liminf_{k \to \infty}\left( \left\| X_{n_1} + \sum_{j = 1}^{k - 1} Y_j \right\| \right)\\ & \leq \liminf_{k \to \infty} \left( \|X_{n_1}\| + \sum_{j = 1}^{k - 1} \| Y_j \| \right) \\ & \leq \liminf_{k \to \infty} \left( \|X_{n_1}\| + \sum_{j = 1}^{k - 1} 4^{-j} \right) \\ & = \|X_{n_1}\| + \sum_{j = 1}^{\infty} 4^{-j} \\ & < \infty . \end{align*} Therefore $Z \in L^2$, meaning that $(X_{n_k})_{k = 1}^\infty$ converges to a limit $Z$ in $L^2$. Finally, to prove that this is the limit of $(X_n)_{n = 1}^\infty$, let $\epsilon > 0$. Choose $N \in \mathbb{N}$ such that if $n \geq N$, then $\|X_n - X_N\| \leq \epsilon / 3$. Since $X_{n_k} \to Z$, there exists $K \in \mathbb{N}$ such that if $k \geq K$, then $\|X_{n_k} - Z\| \leq \epsilon / 3$. Assume without loss of generality that $n_K \geq N$. Then for all $n \geq n_K$, we have that $\|Z - X_n\| \leq \|Z - X_{n_K} \| + \|X_{n_K} - X_N\| + \|X_N - x_n\| \leq \epsilon$, where the estimates on $\|X_{n_K} - X_N\|, |X_N - x_n\|$ follow from the fact that $n_K,n \geq N$.