Let $W_t$ be a standard Brownian motion, i.e., $W_t \sim N(0,t)$. Define the random variable


A similiar process, $Y_t=\cos(\omega t+\sigma W_t+\theta)$, with the uniform random variable $\theta\sim U(-\pi,\pi)$ and $\omega\in\Bbb R$, was analyzed by Cai and Wu [1], where the probability density was defined as


Seeing that the process $Y_t$ has a density, how can I obtain the density of the simplified process $\cos(W_t)$ ? I expanded the integral as a Riemann sum,

$$X\approx\frac{1}{n}\sum_{i=1}^n{\cos\left( W_\frac{i}{n}\right)}=\frac{1}{n}\left(n\cos\left(W_\frac{1}{n}\right)+ (n-1)\left(\cos\left(W_\frac{2}{n}\right)-\cos\left(W_\frac{1}{n}\right)\right)+ (n-2)\left(\cos\left(W_\frac{3}{n}\right)-\cos\left(W_\frac{2}{n}\right)\right)+ \cdots + 2\left(\cos\left(W_\frac{n-1}{n}\right)-\cos\left(W_\frac{n-2}{n}\right)\right)+ \cos\left(W_1\right)-\cos\left(W_\frac{n-1}{n}\right) \right)$$

and after rearranging the terms I got the sum $$\frac{1}{n}\sum_{i=1}^n{\cos\left( W_\frac{i}{n}\right)}= \frac{1}{n}\sum_{i=1}^n{-2(n-i+1) \sin\left(\frac{W_\frac{i}{n}+W_\frac{i-1}{n}}{2}\right) \sin\left(\frac{W_\frac{i}{n}-W_\frac{i-1}{n}}{2}\right)}$$

How can I progress from here?

[1] CAI, G. Q.; WU, C. Modeling of bounded stochastic processes. Probabilistic Engineering Mechanics, v. 19, n. 3, p. 197–203, 2004.

  • $\begingroup$ I find your question somewhat misleading. The process considered in the article is of the form $Y_t = \cos(\sigma W_t+\theta)$ for a uniformly distributed random variable $\theta$... your process $X_t$ is not exactly "similar" to $Y_t$ . In particular, it is not to be expected that $p$ is the density of $X_t$. Please feel free to correct me if I missed something about the article. $\endgroup$ – saz Jan 28 at 15:24
  • $\begingroup$ sure, but the random variable $\theta$ is constant in time. The cosine can them be partitioned in $\cos(\sigma W_t)\cos(\theta)-\sin(\sigma W_t)\sin(\theta)$. In order to understand the final form, I thought that I needed to break the integral into fundamental blocks, which is why I'm specifically asking about this one. But you are right. The process don't need to have the same distributions. I will edit it. $\endgroup$ – Kaio César Borges Benedetti Jan 28 at 15:30
  • $\begingroup$ I'm just saying that $p$ is certainly not going to be the density of $X$. Using that $W_t \sim N(0,t)$ you can verify that $\mathbb{E}(X)>0$. If $Y$ is a random variable with density $p(y)=\frac{1}{\pi \sqrt{1-y^2}} 1_{(-1,1)}(y)$ then it has expectation zero... this shows that something doesn't fit here (or I'm misunderstanding your question). I think it's highly non-trivial to determine the distribution of $X$. $\endgroup$ – saz Jan 28 at 16:03
  • 1
    $\begingroup$ If $\theta$ is uniform on $(0,2\pi)$, then, for every real number $z$, $\theta+z$ mod $2\pi$ is also uniform on $(0,2\pi)$, thus, for every $Z$ independent of $\theta$, $\cos(\theta+Z)\stackrel d=\cos\theta$. Now, surely you can compute the PDF of $\cos\theta$... $\endgroup$ – Did Jan 28 at 17:10
  • 1
    $\begingroup$ Related: this and this. $\endgroup$ – Rodrigo de Azevedo Jan 29 at 18:08

I edited my response many times to correct some mistakes. Sorry.

We can compute the density of $U_t=\cos(W_t)$ by appropriately separating the integral and a change of variable. For any measurable $f$

\begin{align} \mathbb{E}(f(U_t)) &= \mathbb{E}(f(\cos(W_t))) = \int_\mathbb{R} f(\cos(w)) \mathbb{P}_{W_t}(w) \mathrm{d} w = \sum_{k \in \mathbb{Z}} \int_{k\pi}^{(k+1)\pi} f(\cos(w)) \mathbb{P}_{W_t}(w) \mathrm{d} w \\ & = \sum_{k \in \mathbb{Z}} \int_{-1}^{1} f(u) \left( \mathbb{P}_{W_t}(\arccos(u)+2k\pi)+\mathbb{P}_{W_t}(\arccos(-u)+(2k+1)\pi) \right)\frac{ \mathrm{d} u}{\sqrt{1-u^2}} \\ &= \int_{-1}^1 f(u) \frac{\sum_{k\in \mathbb{Z}} \mathbb{P}_{W_t}(\arccos(u)+2k\pi)+\mathbb{P}_{W_t}(\arccos(-u)+(2k+1)\pi)}{\sqrt{1-u^2}} \mathrm{d}u \end{align}

From the first to the second line we need to separate odd and even integers but we can group the sums together after that.

So the density of $U_t=\cos(W_t)$ is \begin{equation} \mathbb{P}_{U_t}(u) = \frac{\sum_{k\in \mathbb{Z}} \mathbb{P}_{W_t}(\arccos(u)+2k\pi)+\mathbb{P}_{W_t}(\arccos(-u)+(2k+1)\pi)}{\sqrt{1-u^2}} \end{equation}

Numerical simulations for $\mathbb{P}_{U_t}$:

Histogram and density of U_t

The numerator is an infinite sum as it appears in the wrapped Gaussian. But there is this $\arccos$ and the denominator. I don't know if it can help to evaluate the density of the integral.

The problem is that $U_t$ is neither Gaussian nor independent at different time step which makes difficult to evaluate the density of the Riemann sum.

Numerical simulations give the following histogram for X:

Histogram of X

I don't have any idea of what function it could be.

Interestingly, using a sinus ie $U_t = \sin(W_t)$ seems tractable. It's probably because $U_t$ is centered. To get the density, just replace $\arccos$ by $\arcsin$ above. See the figures. For $\mathbb{P}_{U_t}$

Histogram and density of U_t

And for the density of $X$, I'm able to fit a density $ \propto \exp\left( -\frac{s}{1-x^2} \right) $:

Histogram of X

I hope it could be helpful.

  • $\begingroup$ I edited my answer many times because of calculus errors and wrong simulations. Now it should be correct. $\endgroup$ – Jonathan Vacher Jan 30 at 4:45

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.