3
$\begingroup$

For a diffusion $X$ define the cumulative distribution function for $X_T$ started with $X_t=x$:

$$u(t,x):=E^{t,x}(1_{X_T\ge y})$$ Under what conditions does $u$ solve $X$'s Kolmogorov backward equation? Where can I read up on this question?

The Wiki article on Kolmogorov backward equations simply assumes this, but where can I find a proof that it works?

The books listed below give conditions (via the Feynman-Kac Theorem) under which $v(t,x):=E^{t,x}(f(X_T))$ solves the Kolmogorov equation, but only for continuous $f$. So $u$ as defined above does not qualify.

I could find many references for the existence of a transition density that solves the Kolmogorov backward equation, but is this also enough to ensure that $u$ solves it?

References

Oksendal, Stochastic Differential Equations

Karatzas and Shreve, Brownian Motion and Stochastic Calculus

Friedman, Stochastic differential equations and applications

$\endgroup$

1 Answer 1

6
$\begingroup$

If the function $f : \mathbb{R} \to \mathbb{R}$ is discontinuous, then $u(t,x)= \mathbb{E}_xf(X(t))$ may not satisfy the initial condition, in the sense that the limit statement: $$ \lim_{(t,s) \to (0^+,x)} u(t,s) = f(x) \quad \forall x \in \mathbb{R} \tag{$\star$} $$ may not hold. To be concrete, consider $$ d X(t) = d B(t) \;, \quad X(0) = x $$ where $B(t)$ is a standard Brownian motion on $\mathbb{R}$ and let $$ f(x) = \begin{cases} 1 & \text{if $x\ge 0$} \\ 0 & \text{otherwise} \end{cases} $$ In this case, for any $t>0$: $$ u(t,x) = \mathbb{E} f(x+B(t)) = \frac{1}{2} ( 1 + \operatorname{erf}( \frac{x}{\sqrt{2t}} ) ) \;. $$ However, $ \lim_{(t,s) \to (0^+,0)} u(t,s) $ does not exist, since for any $\alpha>0$ we have that: $$ \lim_{s \to 0^+} u(s^2,s^{\alpha}) = \lim_{s \to 0^+} \frac{1}{2} ( 1 + \operatorname{erf}( \frac{s^{\alpha-1}}{\sqrt{2}} ) ) = \begin{cases} 1/2 & \alpha>1 \\ 1/2 (1 + \operatorname{erf}(1/\sqrt{2}) ) & \alpha=1 \\ 1 & \alpha < 1 \end{cases} \tag{$\star \star$} $$ which depends on $\alpha$ (or in words, the limit depends on the path taken towards the origin). Nevertheless, $u(t,x) \in C^{\infty}((0, \infty)\times \mathbb{R})$, and by dominated convergence, one can show that $u(t,x)$ satisfies the heat equation for any $t>0$. Here is a graphical illustration of the function $u(t,x)$ and three paths from ($\star \star$) corresponding to the choices $\alpha=2$ (red), $\alpha=1$ (black), and $\alpha=1/2$ (blue). Note that each path towards the origin has a different terminus.

enter image description here

To summarize, continuity of $f$ is typically applied to show that the limit statement $(\star)$ holds, which (loosely speaking) is why theorems relating SDEs to parabolic equations often assume that the initial condition $f$ is at least continuous.

$\endgroup$
4
  • $\begingroup$ Dominated convergence allows you to interchange differentiation and integration, in order to conclude that: $\partial_t u = 1/2 \Delta u$. $\endgroup$ Mar 16, 2015 at 1:03
  • $\begingroup$ Thank you for this comprehensive answer! I still fail to grasp how to use dominated convergence to show that $u$ satisfies the heat equation and would really appreciate more details or a reference. Thanks! $\endgroup$
    – JSG
    Mar 16, 2015 at 11:32
  • $\begingroup$ Like this: Using the fact that $u(t,x)=\int f(y) p(y,t,x)dy$, for bounded $f$ and due to $p$ having bounded derivatives, I can use dominated convergence to interchange differentiation and integration. As $p$ solves the heat equation I get a zero integral and thus $\partial_t u-1/2\Delta u=0$? $\endgroup$
    – JSG
    Mar 16, 2015 at 17:00
  • $\begingroup$ Yes, like that. $\endgroup$ Mar 16, 2015 at 18:39

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.