# Is this a Brownian motion?

I am building a 2D stochastic process as follows. I start with a point $$P_0=(0,0)$$. Then $$P_k=(X_k,Y_k)$$ is defined as follows, for $$k>0$$: \begin{align} X_k & =X_{k-1}+R_k \cos(2\pi\theta_k) \\ Y_k & =Y_{k-1}+R_k \sin(2\pi\theta_k) \end{align} where the $$\theta_k$$'s are uniformly and independently distributed (iid) on $$[0,1]$$, the $$R_k$$'s are also iid, and independent from the $$\theta_k$$'s. The $$R_k$$'s are generated as follows:

$$R_k=\frac{1}{\lambda}\Big[-\log(1-U_k)\Big]^c,$$

where $$\lambda>0$$ and $$U_k$$ is uniform on $$[0,1]$$. If $$c=1$$, $$R_k$$ has an exponential distribution of parameter $$\lambda$$. If $$c>0$$, $$R_k$$ has a Weibull distribution; if $$c<0$$, $$R_k$$ has a Fréchet distribution. In all cases, $$E[R_k]=\Gamma(1+c)/\lambda$$ if $$c>-1$$, otherwise the expectation is infinite. It is possible to rescale the process, in the same way a random walk is rescaled to become at the limit, a Brownian motion.

My question is this: will my simulation always result in a Brownian motion regardless of the parameter values? My hope is that the answer is sometimes yes, sometimes no. For standardization purposes, assume that $$\lambda=\Gamma(1+c)$$ if $$c>-1$$. I am interested in finding parameter values such that the resulting process is not Brownian. What if $$c=4$$? Below is a realization with $$10^4$$ points and $$c=4$$. It does not look Brownian, it consists of well separated clusters, typical for a large value of $$c$$. And my goal is to illustrate clustering techniques (in particular, identifying the number of clusters) for processes that are Brownian-related, but exhibiting a much stronger cluster structure with well separated clusters.

• Perhaps related: Levy flight ... en.wikipedia.org/wiki/Lévy_flight Jan 29 at 11:58
• Come on, this is just CLT (a.k.a. invariance principle). If $R_k$ has finite variance (which is the case for $c > -\tfrac12$, I think), rescaled random walks converge to the Brownian motion. If the tail of $R_k$ decays as $x^{-\alpha}$ (or, more generally, is regularly varying of order $-\alpha$) as $x \to \infty$ and $\alpha \in (0, 2]$, the limit is the symmetric $\alpha$-stable Lévy process (which is just the Brownian motion if $\alpha = 2$). If I am not mistaken, this happens for $c \leqslant -\tfrac12$, with $\alpha = -1/c$. Jan 29 at 14:33
• @Gerald: thank you, this was very useful. Jan 29 at 17:15

I vote for Mateusz Kwaśnicki. The condition for whether the random walk you generate this way scales towards Brownian motion under taking long times and rescaling is whether or not the variance is finite. You already know how to take the expectation, your formula is correct. $$\int_0^1 \left(-\ln(1-x)\right)^c\, dx\, =\, \Gamma(1+c)\, .$$ So if $$\lambda=1/\Gamma(1+c)$$, then $$\mathbf{E}[R_k]=1$$. Note that you still have $$\mathbf{E}[X_k]\, =\, \mathbf{E}[X_{k-1}]\, ,\ \mathbf{E}[Y_k]\, =\, \mathbf{E}[Y_{k-1}]\, ,$$ because of the $$\theta_k$$ randomization. Then by the same integral formula $$\mathbf{E}[R_k^2]\, =\, \frac{1}{\lambda^2}\, \int_0^1 \left(-\ln(1-x)\right)^{2c}\, dx\, =\, \frac{\Gamma(1+2c)}{\Gamma(1+c)^2}\, .$$ So, as long as $$-1/2 then this is finite. (It is infinite if $$c\leq -1/2$$ and the Gamma formula above should not be used.) So if you take $$\widetilde{B}^{(n)}_t\, =\, \frac{\Gamma(1+c) \sqrt{2}}{\sqrt{n\Gamma(1+2c)}}(X_{\lfloor nt\rfloor},Y_{\lfloor nt\rfloor})$$ then this should converge in distribution to a process $$B_t = (B_t^{(1)},B_t^{(2)})$$ where $$B_t^{(i)}$$ for $$i=1,2$$ are two independent Brownian motions normalized so that $$\mathbf{E}[(B_t^{(i)})^2]=t$$.