1
$\begingroup$

Is there a way to compute transition operators for Markov processes? To ask something much more tractable, suppose I have an Ito diffusion $$dX_t \ = \ \sigma(X_t) dB_t \ + \ b(X_t) dt$$

(or given by its generator). Is there any way to write down the transition kernels for the process $X_t$ from the above SDE? Also, maybe with assumptions on the coefficients $\sigma, b$, is there any way to establish any niceness in the sense of $L^p(\mathbb{R})$ for the transition kernels given just the diffusion? I'm guessing the drift component gives Dirac masses along some curve defined by $b(x)$ and the diffusion component gives a Gaussian process, but I'm not seeing quite how to write down the kernels themselves...

Thanks!!

Quick edit: I know one can solve the corresponding Fokker-Planck equation (with a given initial condition) with Green's functions, Fourier series, etc., but is there another way to detect any properties like $L^p$-regularity without PDE methods?

$\endgroup$
2
$\begingroup$

1) Explicit expressions for transition densities

In the case of linear systems with additive noise of the form

$d X(t) = (A X_t + b) \, d t + \sigma \, d B_t$

it is possible to obtain an explicit solution, see Ioannis Karatzas and Steven E. Shreve, MR 1241411 Brownian motion and stochastic calculus, Section 5.6.

There are a few other explicitly solvable cases (in terms of transition density), for example the SDE for the Bessel process

$d X_t = (d-1)/(2X_t) \, d t + B_t$,

and the exponential Brownian motion,

$d X_t = a X_t \, d t + \sigma X_t \, d B_t$,

and quite a few other examples. In general finding the transition density of an SDEs means that you have solved the associated Fokker-Planck equation (equivalently known as forward Kolmogorov equation) which is a parabolic PDE. So finding transition densities is in general of the same difficulty as solving a linear PDE of a particular type, and solving such PDEs explicitly is typically not possible.

Also noteworthy is the Lamperti transform to change from general one-dimensional diffusion coefficients to a diffusion $Z$ with unit diffusion coefficient, by defining $Z_t = \varphi(X_t)$ where $\varphi(x) = \int^x \frac{1}{\sigma(\xi)} \ d \xi$.

I would recommend to consult a basic textbook focussing on explicit computations such as Øksendal's and look at the exercises to get a feel for what is possible in this direction.

2) Regularity of solutions

Of course, if $\sigma = 0$, your solution will be a deterministic trajectory, with corresponding Dirac measure as transition kernel, i.e. not belonging to any $L^p$. A simple and well known result is that if $\sigma(x)$ is globally bounded away from 0 (known as ellipticity), you will have smooth transition densities. The important generalization of this result for multi-dimensional SDEs where some of the components of $\sigma$ may be zero is given by Hörmander's theorem. I recommend Martin Hairer's lecture notes, Convergence of Markov processes, for a gentle introduction.

Once you establish the probability density is smooth (using ellipticity/Hörmander) and vanishing at infinity (corresponding to boundedness in probability), it will be bounded and therefore in $L^p(\mathbb R)$ for any $p \geq 1$. Boundedness in probability can be established using e.g. bounds on $\mathbb E|X(t)|$, see the book by Xuerong Mao, or various other sources.

$\endgroup$

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.