Lecture 1 From random walks to fractional diffusion equations

Dr Qianqian Yang
AMSI Winter School 2019

Fractional calculus is a rapidly growing field of research, at the interface between probability, differential equations, and
mathematical physics. Fractional calculus is used to model anomalous diffusion, in which a cloud of particles spreads
in a different manner than traditional diffusion. In our first lecture, we will derive the space fractional diffusion equations
from the point of view of probability, using random walks.

Case study 1: Random walks with constant jump sizes

can be derived in many ways. We will be interested in its derivation using random walks, since that approach
generalises nicely to the fractional diffusion case. Notice that we have written the equation using the
operator (which can also be written as ). The reason for the minus sign is that is then a positive
definite operator, so its eigenvalues are positive. In one dimension this operator is just . We will derive the

equation in one dimension only, but the idea generalises to any number of space dimensions.

Consider a large number of particles all undergoing independent random walks. For simplicity, we suppose that they all
begin at the origin at time . Then at each discrete time step , they randomly jump one unit left, or
one unit right, .

Let denote the density of particles at position x at some time t: the number of particles per unit distance.

At the next time , all of these particles at position x will jump away. And since we're dealing with a large number
of particles, we can say that half of the particles from the left and half of the particles from the right will each jump here.
So the new number of particles will be

Clearly we can cancel the factors of and just talk about particle density:

Taking Taylor expansions we find that

1
 .

Now we take the limit as the stepsize and time between jumps go to zero, keeping the factor fixed.

We derive the diffusion equation

where D is called the diffusivity, or diffusion coefficient. In nature, there is never really the limit as and go to
zero. Instead, the diffusion coefficient is whatever it is, for the small but finite values of and that govern the
diffusion (e.g. molecular scale).

The initial condition that all the particles begin at the origin can be written as (Dirac delta function),
where for convenience we can measure the density in units such as "billions of particles per millimetre" (or whatever),
so that that the initial number of particles is just

(billion particles)

With this initial condition, the diffusion equation has the solution

You might recognise this as the equation for the normal distribution, also called Gaussian distribution probability density
function. That's no coincidence! The particles' positions at time t are the sum of all the independent random jumps up
to that time. From statistics, we know that the sum of many independent random variables approaches the Gaussian
distribution. This is called the Central Limit Theorem.

Let's do a simulation! The code is provided in the file random_walk_simulation.m. This simulation uses particles
each taking steps of size . The final time t is arbitrarily set to , meaning the effective timestep
is and so the diffusivity for the PDE model is . Let's compare the histogram of
particle locations after steps with the prediction of the continuum PDE model.

random_walk_simulation

2
 D = 0.1250

The solutions are in good agreement. Better agreement would be obtained with more particles (thereby smoothing over
the random bumps) and more steps (meaning the random walk was closer to the continuum limit with and going
to zero).

Case study 2: Random walks with jump sizes drawn from a symmetric distribution
with finite variance
We might wonder how general this diffusion equation is. Instead of assuming the jumps are all constant size , could
we have drawn each jump size from a probability distribution? Let the random variable X denote the jump size now,
and suppose it comes from a probability distribution with probability density function . We suppose that this is
a symmetric distribution around zero, so the jump size can be positive or negative with equal probability. Thus, the
mean µ of the distribution is

and we suppose its variance is some finite value

Now, previously we had the equation

3
to update the density at time . But that assumed the only possible jump sizes were and . Now, in
theory any jump size is possible (though, very large jumps are less likely because of finite variance). So we have to
consider all the ways that particles could jump to location x at time . We generalise the two terms in the sum
(each with probability ) to a continuous range of possibilities, each with their own probability given by q. Note that to
jump from some arbitrary position to position x requires a jump size of . So we have,

How to proceed from here? We recognise the integral on the right hand side as the convolution of u and q. Any time
you see convolutions, there's a good chance that Fourier transforms can help, because the Fourier transform of a
convolution is the product of the individual Fourier transforms.

So take Fourier transforms to obtain

and we will use the convention from statistics in defining the Fourier transform to be

(In statistics, if f is a probability density function, then is known as the characteristic function.)

We can now expand q in terms of its moments:

Expanding in a Taylor series on the left as before, we find that

Remember the big O notation means that term goes to zero at least as fast as , while the 'little o
notation' means that term goes to zero faster than .

So, letting and taking limits as and go to zero, keeping D fixed, we have
σ

With initial condition (corresponding to ) we obtain

and by Fourier inversion

4
as before. But now we see that the result is true much more generally than for constant-size jumps, which was just a
special case with . Instead, we can choose any symmetric distribution of jump sizes with variance . This
includes distributions for which there is considerable probability for taking very long jumps.

Let's do another simulation! This time we'll use a two-sided exponential distribution for our step sizes, with mean
absolute stepsize . Then the variance for the two-sided distribution is . The other parameters
are as before, so the diffusivity for the PDE model is now . Although the mean absolute step
size is the same as previously, it has more variance (large steps are possible), so the overall diffusivity is twice as large.
Let's again compare the histogram of particle locations after steps with the prediction of the continuum PDE model.

random_walk_simulation('exponential')

D = 0.2500

The solutions are again in good agreement, with more spreading of the particles compared to previously.

Case study 3: Random walks with jump sizes drawn from a symmetric distribution
with infinite variance
You might wonder then, how would we ever obtain anomalous diffusion, where the particle density doesn't follow a
Gaussian distribution? Doesn't the Central Limit Theorem guarantee that the sum of many jumps from any distribution
will approach a Gaussian? The catch is in the fine print of the theorem. It only applies to jump size distributions
with finite variance .

5
If the particle jump sizes are drawn from a probability distribution with infinite variance, then the Central Limit Theorem
doesn't apply. The prototype example of such a distribution is the Pareto distribution which has probability density
function

(This is the two-sided form of the Pareto distribution, which allows for positive and negative jumps with equal
probaiblity.)

The parameter α is the key here. It determines the "heaviness" of the tails. We will be concerned with the
range . For this range, the mean of this distribution is equal to zero, but the variance is infinite.

Let's use MATLAB to check these properties of the Pareto distribution.

syms x alpha x_m

pareto =

figure;
fplot(subs(pareto, {alpha, x_m}, {1.5, 1}), [-5,5], 'LineWidth', 2)

6
 int(pareto, x, -inf, inf)

Warning: Unable to check whether the integrand exists everywhere on the integration interval.
ans =

int(x*pareto, x, -inf, inf)

Warning: Unable to check whether the integrand exists everywhere on the integration interval.
ans =

int(x^2*pareto, x, -inf, inf)

Warning: Unable to check whether the integrand exists everywhere on the integration interval.
ans =

The equation for updating the density is still the same as before:

and its Fourier transform is

7
But the variance is infinite, we can't apply our moment expansion like we did before. We need a new technique to
tackle this equation. It will be great if we know the formula for .

MATLAB is too stupid to calculate the Fourier transform

fourier(pareto)

ans =

but from Mathematica we find that

or in simpler terms,

where (which is positive), and the term goes to zero faster than in the limit

as goes to zero (remember that ).

Now, substitute into the Fourier transformed equation

to get

Taylor series on the left as usual, and we find

So, letting and taking limits as and go to zero, keeping D fixed, we have

and so

But how do we invert this Fourier transform? Equivalently, for what probability distribution is this the characteristic
function? Statisticians will recognise the characteristic function as that of the so-called α-stable distribution. . Its density
function u has no analytic expression, so this formula for is as far as we can write down. But its properties as a
probability distribution are very well known. It looks "Gaussian-like", but with heavier tails.

8
 pd = makedist('Stable', 2, 0, 1, 0); % Gaussian is a special case of alpha-stable (with alpha = 2)
figure, fplot(@pd.pdf, [-5,5], 'LineWidth', 2, 'DisplayName', 'Gaussian')
pd = makedist('Stable', 1.5, 0, 1, 0); % alpha = 1.5
hold on, fplot(@pd.pdf, [-5,5], 'LineWidth', 2, 'DisplayName', '\alpha-stable')
legend

Now the question is: what partial differential equation does this characteristic function solve? We know for
the Gaussian density it's the standard diffusion equation

which in Fourier space is

So what's the PDE that leads to the solution

we found above? Simple! It's the fractional diffusion equation

where the fractional Laplacian operator is defined by its action in Fourier space:

9
Since there is no analytical solution for the fractional diffusion equation on infinite domain, to compare the PDE solution
to the random walk simulation we require a numerical method to solve the PDE. That will be the topic for the next two
lessons. But for now, we'll trust that our code can do the trick.

random_walk_simulation('pareto',1.5)

D = 1.7055

We see the characteristic "heavy tails" associated with fractional diffusion.

Although we derived this fractional PDE using the Pareto distribution, the result is more general. We recognise that our
two-sided Pareto distribution follows a form of two-sided power-law probability distribution

, for large .

Here is an arbitrary α-dependent constant, and σ is playing the role of the variance in determining how "spread out"
the distribution is (remember though the the variance itself is infinite). Clearly our two-sided Pareto distribution meets
this criteria, with and , but any other distribution with meeting that asymptotic condition will
lead to the same form of fractional PDE.

References:

10
R. Metzler, J. Klafter. The random walk's guide to anomalous diffusion: a fractional dynamics approach. Physics
Reports 339 (2000) 1-77.

11