1
$\begingroup$

Let $\mathscr{C}:=\{\gamma : \mathbb{R}_+\rightarrow\mathbb{R}^n \mid \gamma \ \text{ continuous}\}$ be the set of all $\mathbb{R}^n$-valued paths over $[0,\infty)$. Endow $\mathscr{C}$ with the $\sigma$-algebra $\mathfrak{C}$ generated by all projections $\gamma \mapsto \gamma_t$ (for $t\geq 0$ fixed). Let further

$$\phi : \mathscr{C}\rightarrow\mathscr{C} \ \ \text{ be given by } \ \phi(\gamma):=(\gamma_{t+1})_{t\geq 0} \ \text{ (left-shift),}$$ and call $\phi$-invariant any event $S\in\mathfrak{C}$ with $\phi^{-1}(S)=S$.

Do you know of examples for $\mathscr{C}$-valued stochastic processes $Y$ that attain $\phi$-invariant events with trivial probability, i.e. are such that $\mathbb{P}_Y(S)\in\{0,1\}$ whenever $S\in\mathfrak{C}$ is $\phi$-invariant?

$\endgroup$

1 Answer 1

1
+50
$\begingroup$

Do you not also want that $\mathbb{P}_Y$ is $\phi$-invariant?

In any case, yes there are extremely many continuous-time continuous-path real-valued stochastic processes whose law is ergodic under the time-$1$-shift map. Of course, the most trivial example would be where $$ \mathbb{P}(Y_t=c \ \ \forall t \geq 0) = 1 $$ for some $c \in \mathbb{R}$. An almost-as-trivial case would be $$ Y_t = p(2\pi ft + \Phi) $$ where $p(\cdot)$ is a $1$-periodic continuous function, $f$ is an irrational number, and $\Phi$ is a random variable with $\Phi \sim \mathrm{Uniform}(0,2\pi)$. (I don't know if this would count as a "non-trivial example class" under your bounty statement!)

Perhaps the simplest "non-trivial" example would be the Ornstein-Uhlenbeck process:

(I hope there are no typos in the following!) Fix two parameters $\theta,\sigma>0$. Starting off with a two-sided Wiener process $(W_t)_{t \in \mathbb{R}}$ (meaning that $(W_t)_{t \geq 0}$ and $(W_t)_{t \leq 0}$ are independent Wiener processes), we can define a stochastic process $(Y_t)_{t \in \mathbb{R}}$ where for almost every sample path of the Wiener process, for all $t \in \mathbb{R}$, \begin{align*} Y_t &= \sigma \! \int_{-\infty}^t e^{\theta(s-t)} \, dW_s \quad\quad\quad\quad\quad\quad\ \, \text{(Riemann-Stieltjes integral)} \\ &= \sigma \left( W_t - \theta \! \int_{-\infty}^t e^{\theta(s-t)} W_s \, ds \right) \quad \text{(classical integral, obtained by integration by parts)}. \end{align*} This is the unique two-sided-time strong solution of the stochastic differential equation $$ dY_t = -\theta Y_t \, dt + \sigma dW_t, $$ or equivalently (as a random differential equation) $$ \frac{d(Y_t - \sigma W_t)}{dt} = -\theta Y_t. $$ $\text{[}$The one-sided-time solution $(Y_t^{(y_0)})_{t \geq 0}$ of this equation starting at any $y_0 \in \mathbb{R}$ would be \begin{align*} Y_t^{(y_0)} &= e^{-\theta t} \left( y_0 + \sigma \! \int_0^t e^{\theta s} \, dW_s \right) \\ &= \sigma W_t + e^{-\theta t} \left( y_0 - \sigma\theta \! \int_0^t e^{\theta s} W_s \, ds \right). \text{]} \end{align*} A one-sided-time stochastic process with the law of $(Y_t)_{t \geq 0}$ (or a two-sided-time stochastic process with the law of $(Y_t)_{t \in \mathbb{R}}$) is called an Ornstein-Uhlenbeck process.

Sketch-proof that OU process is shift-ergodic. Firstly, working in two-sided-time,

  • one can check that replacing $(W_t)_{t \in \mathbb{R}}$ with $(W_{t+1}-W_1)_{t \in \mathbb{R}}$ transforms $(Y_t)_{t \in \mathbb{R}}$ to $(Y_{t+1})_{t \in \mathbb{R}}$;
  • since the Wiener process has memoryless stationary increments, one can show that the law of $(W_t)_{t \in \mathbb{R}}$ is mixing and hence ergodic with respect to the map $(\gamma_t) \mapsto (\gamma_{t+1}-\gamma_1)$;

and so the law of $(Y_t)_{t \in \mathbb{R}}$ is ergodic with respect to the map $(\gamma_t) \mapsto (\gamma_{t+1})$. Finally, having ergodicity in two-sided time guarantees ergodicity in the one-sided-time restriction.


In general, given any $\phi$-invariant probability measure $\mathbb{P}$ on $\mathscr{C}$, one can express $\mathbb{P}$ as a weighted average of ergodic $\phi$-invariant probability measures on $\mathscr{C}$. This is essentially the ergodic decomposition theorem, which holds for any Borel-measurable dynamical system on a Polish space; in this case, the $\sigma$-algebra $\mathfrak{C}$ is precisely the Borel $\sigma$-algebra of the topology of uniform convergence on compact sets, which is a Polish topology.

$\endgroup$

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.