One way to generate Brownian motion is as follows: Define a waiting time probability distribution $\psi(t)$ and step length probability distribution $\lambda(x)$. Require also that $\langle \psi \rangle = \tau$, $\langle \lambda \rangle = 0$, and $\langle \lambda^2 \rangle = \sigma^2$. That is, the waiting time distribution has finite mean and the step length is symmetric about zero with finite variance. Then, we can generate a sequence of time steps $\delta t_i \sim \psi$, a sequence of step lengths $\delta x_i \sim \lambda$. We can then define the trajectory $x(\sum^N \delta t_i) = \sum^N \delta x_i$.

If one is simulating the process, convenient choices include $\psi(t) = \delta(t - \tau)$, and $\lambda(x) = \frac{1}{2}( \delta(x - \sigma ) + \delta(x + \sigma ))$. That is, at regular time intervals you pick whether the particle goes left or right, with equal probability. So far so good.

Now for simplicity, let anomalous diffusion be a process such that $\langle \Delta x^2 (t) \rangle \propto t^\alpha$, for $0 < \alpha < 1$.

To generate anomalous diffusion, it must be that the mean waiting time diverges $\langle \psi \rangle \to \infty$ or that the step length variance diverges $\langle \lambda^2 \rangle \to \infty$ or, I guess, possibly both. I know one way to do this: Let $\langle \psi \rangle$ have a Pareto distribution $\psi(t) = \frac{\alpha \tau^\alpha}{t^{1+\alpha}}$ for $t > \tau$, so that it has a diverging mean. We can then pick any $\lambda$ with finite variance, like the one in the previous paragraph.

My question is whether it is possible to require that $\psi(t) = \delta(t - \tau)$ and still get anomalous diffusion. Obviously $\lambda$ would need to have diverging variance, but I have no idea beyond that. I suppose $\lambda(x_i)$ could even depend on past history. That is, I'm looking to pick steps at regular time intervals $\Delta t$ from a certain distribution $\lambda$ such that the resulting trajectories display anomalous diffusion.

  • $\begingroup$ Have you had a look at Fractional Brownian Motion and generalized Langevin equaiton? It is an approach alternative to random walk to describe diffusion and incoporates memory effects. $\endgroup$
    – user35952
    Aug 29, 2022 at 7:40
  • $\begingroup$ I think, when simulating the Brownian normal diffusion process, you should choose $\psi(t) = e^{-t/\tau}$ and not the delta function, similarly, the jump length distribution has to be chosen as $\lambda(x) \approx e^{-x^2/(2\sigma^2)}$. Further, please note that in the random walk problem that you have stated, the waiting time and jump lengths are independent and therefore are governed by independent probabiity distributions - implying $\lambda(x)$ can't depend on time. There are of course processes in which you can couple the jump length and waiting time. $\endgroup$
    – user35952
    Dec 6, 2022 at 8:50

1 Answer 1


Unless you want a levy flight you do not need $\lambda$ with diverging variance!

For anomalous diffusion you can try $\lambda $ from a distribution with $var(\Lambda) =~Dt^\alpha$ where $0<\alpha<1$ and $D$ is a constant.


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.