4
$\begingroup$

In stochastic analysis, for an Ito diffusion $X_t$ such that $dX_t=\mu(X_t)dt+\sigma(X_t)dB_t$, we can exlpicitly compute a "natural scale function" $$S(x)=\int^x\exp\left(-\int^y\frac{2\mu(z)}{\sigma^2(z)}dz\right)dy$$ with suitable conditions on $\mu$ and $\sigma$, making $S(X_t)$ a martingale(or at least a local martingale).

My question is: can we do something similar in more general processes, such as discrete time Markov processes?

One example is the asymmetric simple random walk on $\mathbb{Z}$. Let $\xi_i$ be i.i.d. taking value $1$ with probability $p\not=\frac 12$, and value $(-1)$ with probability $q=1-p$. The random walk is then $X_n=\sum_{i=1}^n\xi_i$. Then we know the function $S(x)=(q/p)^x$ is a natural scale function since $S(X_n)$ is a martingale.

Is there a systematic way to compute this function out, like in the Ito diffusion case?

This question is also posted on MSE here.

$\endgroup$
3
  • $\begingroup$ If you need additional conditions on "Markov processes", just add them--if they don't rule out the asymmetric SRW case above. $\endgroup$
    – MikeG
    Apr 25, 2021 at 12:01
  • 1
    $\begingroup$ Let $L$ be the generator of the Itô diffusion, and note that $LS = 0$. This means that $S$ is a harmonic function associated with the process. In general, for a Markov process $X_t$ with generator $L$, a non-negative function is a harmonic function if and only if $f(X_t)$ is a martingale. See, e.g., Section 1.4 of wt.iam.uni-bonn.de/fileadmin/WT/Inhalt/people/Anton_Bovier/… $\endgroup$ Apr 25, 2021 at 12:02
  • 1
    $\begingroup$ @NawafBou-Rabee: Oh that makes a lot of sense! Then the problem of finding the function f is reduced to solving the differential equation Lf=0, right?(So, the first step is to calculate the corresponding generator of the markov process... very reasonable!) $\endgroup$
    – MikeG
    Apr 25, 2021 at 13:53

1 Answer 1

3
$\begingroup$

Just figured this out with the help of @Nawaf Bou-Rabee. Thank you!

The idea is to calculate the infinitesimal generator $L$ for the markov process $X_n$ or $X_t$, and a function $f$ making $f(X_n)$ or $f(X_t)$ a martingale(or local martingale for continuous time) if and only if $Lf=0$. See reference here.

This is the key step, and this actually gives the explicit expression of S(x) in the case of Ito diffusion:

now the generator is $L=\mu\frac{d}{dx}+\frac{\sigma^2}{2}\frac{d^2}{dx^2}$, and the natural scale function mentioned above is just the solution of the ODE: $$\mu S'+\frac{\sigma^2}{2}S''=0.$$

For the asymmetric simple random walk, the generator by definition is $$Lf(x)=\mathbb{E}^x[f(X_1)]-f(x)=pf(x+1)+qf(x-1)-f(x).$$

And solving this difference equation we get that function $S(x)=(q/p)^x$ above as well.

$\endgroup$
2
  • 2
    $\begingroup$ A minor correction: in continuous time, $f(X_t)$ should be a local martingale rather than martingale. $\endgroup$ Apr 25, 2021 at 16:57
  • $\begingroup$ @MateuszKwaśnicki: Thanks for the correction! $\endgroup$
    – MikeG
    May 1, 2021 at 16:01

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.