0
$\begingroup$

I'm trying to figure out the connections between two contructions of Gaussian measure.

Let $(U, \langle\cdot,\cdot\rangle_U)$ be a seprable Hilbert space, and $\mathcal{B}(U)$ be the Borel sigma-algebra.


Definition 2.1, page 10 gave the following definition of Gaussian measure:

A measure $\mu$ on $(U, \mathcal{B}(u), \mu)$ is Gaussian if for all $v\in U$, there are $m_v\in \mathbb{R}$, $\sigma_v\in \mathbb{R}^+$ such that if $\sigma_v>0$,

$$ \mu(u \in U\colon \langle v,u\rangle_U \in A) = \frac{1}{\sqrt{2\pi\sigma_v}}\int_A\exp(-\frac{(s-m_v)^2}{2\sigma_v^2})ds, $$

for all $A\in\mathcal{B}(\mathbb{R})$.


However, I recall from when we construct the Brownian motion. See, page 23 of https://www.springer.com/gp/book/9780387287201

Let $U$ be the space of continuous funcitons $u(t)$ on $[0,1]$, and $D$ be a cylindrical set

$$ D = \{u\in U: (u(t_1), u(t_2), \ldots, u(t_N))\in E\in \mathcal{B}(\mathbb{R}^N)\} $$

Then the measure is defined by

$$ \mu(D) = \int_E \prod_{i=1}^N (\frac{1}{\sqrt{2\pi(t_i-t_{i-1})}} \exp(-\frac{(h_i-h_{i-1})}{2(t_i-t_{i-1})}))dh_1dh_2\cdots dh_N $$

on $0<t_1<t_2<\cdots<t_N<1$.


My questions is: Is it possible to define the infintite Gaussian measure on finite-dimensional distributions similar to we did for Brownian motion? I don't know how to find the connection bewtween the inner product $\langle v,u\rangle_U$ and a finite-dimensional distribution.

$\endgroup$
9
  • $\begingroup$ I'm not 100% sure what you're looking for, but you might recall the fact that a set of random variables is jointly Gaussian iff every linear combination is Gaussian. Each $v \in U$ can be viewed as a random variable on the probability space $(U, \mu)$, identified with $u \mapsto \langle u, v \rangle$. With this in hand, you can determine the covariance between the random variables corresponding to $v_1, \dots, v_n$, and thus write down their joint distribution. $\endgroup$ – Nate Eldredge Mar 1 at 19:51
  • $\begingroup$ So you end up with a formula for $\mu(\{u \in U : (\langle u, v_1 \rangle, \dots \langle u, v_n \rangle) \in E\})$ for any Borel $E \subset \mathbb{R}^n$ which will look similar to what you have for Brownian motion. $\endgroup$ – Nate Eldredge Mar 1 at 19:52
  • $\begingroup$ I think the problem is that I do't understand why $\langle u,v \rangle$ having Gaussian law induces the Gaussian measure. I can't find the intuition/motivation. To be the finite-dimensional construction is quite straightforward. $\endgroup$ – Nathan Explosion Mar 1 at 20:01
  • 2
    $\begingroup$ At the risk of self-promotion, I have some notes that you may find helpful. $\endgroup$ – Nate Eldredge Mar 1 at 20:26
  • 1
    $\begingroup$ You say construction though, and neither of these statements are really constructions. The Wiener measure paragraph states a certain property for a measure to satisfy; it's still difficult work to prove that there exists a (unique) measure on $C([0,1])$ satisfying this property. Likewise, the Hilbert space paragraph just states a property, and in this case one needs some fairly strong conditions on the $m_v, \sigma_v$ in order to be able to prove there exists a measure with that property. $\endgroup$ – Nate Eldredge Mar 1 at 20:34
2
$\begingroup$

If $U$ is a Banach space, then the natural thing to do is replacing the $\langle v,\cdot\rangle_U$ by any continuous form $\phi\in U^*$. Then the definition of a Gaussian measure is a measure $\mu$ such that the pushforward $\phi^*\mu$ is Gaussian for all $\phi\in U^*$ (note that this works in finite dimension too). In general, there are subtleties about the mean being in $U$ or $U^{**}$, but for centred examples everything works well.

Your definition of Brownian motion implies that the every form $\phi:u\mapsto\alpha_1u(t_1)+\cdots+\alpha_n\phi(t_n)$ is Gaussian (it is in fact equivalent, using the above remark). Now one can show that the subspace of those forms, $\mathrm{Vect}(\mathrm{ev}_t,t\in[0,1])$, is actually dense in $U^*$ in the weak-$*$ topology. Let us show that this proves that $\phi^*\mu$ is Gaussian for all $\phi\in U^*$.

Let $\phi\in U^*$. Let $\phi_n\in U^*$ be a sequence of functions in $\mathrm{Vect}(\mathrm{ev}_t,t\in[0,1])$ that converges to $\phi\in U^*$ in the weak-$*$ topology (here we use the fact that the topology is metrisable, which is true since $U$ is second countable). We show that $\phi^*\mu$ is the weak limit of $\phi_n^*\mu$ (in the sense of probability measures on $\mathbb R$), and that the limit is Gaussian.

The first point is a matter of unfolding the definitions: for every bounded continuous $f:\mathbb R\to\mathbb R$, $f\circ\phi_n(u)\to f\circ\phi(u)$ for all $u$, and since $f$ is bounded we apply Lebesgue's dominated convergence theorem to get $$ \int f(x)\phi_n^*\mu(x) = \int f\circ\phi_n(u)\mu(du) \to \int f\circ\phi(u)\mu(du) = \int f(x)\phi^*\mu(dx). $$ It remains to show that a weak limit of Gaussian measures stays Gaussian. This can be seen by showing first that the means $m_n$ and variances $\sigma^2_n$ have to be bounded, then up to extraction we can assume they converge to $m$ and $\sigma$, and at this point one can see that the characteristic function of the limit is the limit of the characteristic functions $t\mapsto\exp(im_nt-\sigma^2_nt^2/2)$, which is of course the characteristic function of $\mathcal N(m,\sigma^2)$

$\endgroup$
2
  • $\begingroup$ I guess this might not be what you are interested in. In hindsight, I show that Brownian motion is an example of a Gaussian measure using your definition, and you're asking for a construction à la Brownian motion for all Gaussian measure. Is that it? $\endgroup$ – Pierre PC Mar 1 at 20:39
  • $\begingroup$ My answer is more or less the exercise that Nate Eldredge suggests in his comments. $\endgroup$ – Pierre PC Mar 1 at 20:40

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.