For finite-dimensional Brownian motion $W_t$, it is well known that \begin{equation} \lim_{t\to \infty}\frac{W_t}{t}=0,\text{ a.s. }\ \ \ \ \hspace{1cm} \langle 1\rangle \end{equation}

Now suppose we are given an $L^2(\mathcal{D})$-valued Brownian motion $W_t$ defined by $$W_t:=\sum_{k=1}^{\infty}\sqrt{\sigma_k}W_t^k\phi_k(x),$$ where $\mathcal{D}$ is bounded domain in $\mathbb{R}^d$, $\{\phi_k(x)\}$ forms the complete orthogonal basis of $L^2(\mathcal{D})$, $\{W_t^k\}_{k\in \mathbb{N}^{+}}$ are mutually independent one-dimensional standard BM, and $\sigma_k$ satisfies $$\sum_{k=1}^{\infty}\sigma_k<\infty.$$

I wonder does $\langle1\rangle$ still holds for the infinite-dimensional BM introduced above?


2 Answers 2



One way to prove it is using Fernique's theorem:

Let $\mu$ be a centered Gaussian measure on a separable Banach space $(X, \|\cdot\|)$. Then there exists $\alpha > 0$ such that $$\int_X e^{\alpha \|x\|^2} \,\mu(dx) < \infty.$$

Using Chebyshev's inequality, it follows that for all $r$,

$$\mu(\{x : \|x\| > r\}) \le C e^{-\alpha r^2}$$

where $C := \int_X e^{\alpha \|x\|^2} \,\mu(dx) < \infty$.

Let's replace $L^2(\mathcal{D})$ with an arbitrary separable Banach space $Y$. Now, the law of your Brownian motion $\{W_t : 0 \le t \le 1\}$ is a centered Gaussian measure on the separable Banach space $X = C([0,1]; Y)$ equipped with the sup norm. Fernique's theorem then tells us that

$$\mathbb{P}\left(\sup_{0 \le s \le 1} \|W_s\|_{Y} > r\right) \le C e^{-\alpha r^2}$$ for appropriate constants $C,\alpha$. Using the Brownian scaling, we thus have for any integer $n$, $$\mathbb{P} \left( \frac{1}{n} \sup_{0 \le s \le n} \|W_s\|_Y > \epsilon \right) = \mathbb{P} \left( \sup_{0 \le s \le 1} \|W_s\|_Y > \epsilon \sqrt{n} \right) \le C e^{-\alpha \epsilon^2 n}.$$

By the Borel-Cantelli lemma, it follows that $$\mathbb{P} \left( \frac{1}{n} \sup_{0 \le s \le n} \|W_s\|_Y > \epsilon \quad \text{i.o. $n$}\right) = 0$$ and letting $\epsilon \downarrow 0$ along a sequence, $$\frac{1}{n} \sup_{0 \le s \le n} \|W_s\|_Y \to 0, \quad \text{a.s.}$$

Noting that when $n-1 \le t \le n$, we have $\frac{\|W_t\|_Y}{t} \le \frac{1}{n-1} \sup_{0 \le t \le n} \|W_t\|_Y$, it follows that $\frac{\|W_t\|_Y}{t} \to 0$ almost surely. That is $\frac{W_t}{t} \to 0$ in $Y$-norm almost surely.


One can see it much easier by using the subadditive ergodic theorem, which is a standard tool for dealing with the law of large numbers.

Since the second moment of $W_t$ is obviously finite (this is precisely summability of $\sigma_k$), its first moment is also finite, so that the subadditive ergodic theorem implies that there exists a constant $C$ such that $\|W_t\|/t \to C$ a.e. and in $L^1$. In particular, $\|W_t\|^2/t^2\to C^2$ in probability. On the other hand, $\mathbb E\|W_t\|^2=t\,\mathbb E\|W_1\|^2$ grows linearly, so that $C=0$.

  • $\begingroup$ Really neat, but could we develop the problem more detailed, by which I mean does the subadditive ergodic theorem help with the Law of the iterated logarithm for infinite dimensional BM? $\endgroup$
    – Yue
    Commented Sep 27, 2013 at 11:50
  • $\begingroup$ Have never seen any applications of the subadditive theorem to the LIL - it's too coarse for such things (which is why when it works it works in a very simple fashion) $\endgroup$
    – R W
    Commented Sep 27, 2013 at 14:11

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.