Consider a diffusion given by, $d X_t = \mu(X_t) dt + \sigma(X_t) dB_t$

$X_0 = x$.

Suppose the functions $\mu$ and $\sigma$ are as follows -

$f(x) = \mu(x) = \sigma(x) = \begin{cases} 2 & \text{ if } x \ge 0 \\ 1 & \text{ if } x < 0 \end{cases} $

The purpose of $f(x)$ will be clear in a moment.

By Nakao(1972) we know that there exists a strong solution. Now, suppose I am interested in computing the following -

$v(x) = \mathbb{E}^x \int_0^\infty e^{-t} f(X_t) d t $

I know how to do it mechanically. We have the following two DEs: \begin{align} v(x) - 2 v'(x) -2v"(x) -2 = 0 & \text{ if } x >0 \\ v(x) - v'(x) - \frac{1}{2} v"(x) -1 =0 & \text{ if } x < 0 \end{align}

Now, I will solve these 2 simple DEs. Each solution will have 2 constants to be determined. I will use the fact that $v(\infty) = 2$ and $v(-\infty) = 1$ to kill one constant on either side. Then, I will use continuity and differentiability (smooth-pasting) at $0$ to obtain the other 2 constants.

In doing so, however, I have assumed that $v$ is differentiable at $0$. I can prove that $v$ is continuous at $0$. But I do not know how to make the argument for differentiability. This sort of a question comes up often for applied people working with stochastic control and the "standard" method is to assume that it is smooth and then use a "verification theorem". Assuming I want to avoid that, what could be a direct way to prove differentiability?

  • $\begingroup$ I think you should add an expectation $\mathbb E^x$ in the definition of $v(x)$, and maybe $X_0=x$ in that of $X_t$. $\endgroup$ – Jean Duchon Jul 10 '16 at 10:34
  • $\begingroup$ Absolutely! Damn, how did I miss this!! Many thanks for pointing it out. $\endgroup$ – avk255 Jul 10 '16 at 20:35

Here is a rough probabilistic argument for differentiability of $\nu(x)$. At least formally, we have: $$ \nu^{\prime}(x) = \mathbb{E} \int_0^{\infty} e^{-t} \delta(X_t(x)) X_t^{\prime}(x) dt \tag{$\star$} $$ since the derivative of a step function, in the distributional sense, is a Dirac delta function $\delta(\cdot)$. Note that the notation $X_t(x)$ expresses the dependence of the SDE solution on its initial data $x$. In terms of the symmetric local time $L_t$ accumulated by the process $X(t)$ at the origin, we can write ($\star$) as: $$ v^{\prime}(x) = \mathbb{E} \int_0^{\infty} e^{-t} X_t^{\prime}(x) dL_t \tag{$\star\star$} $$ This local time is continuous with respect to $x$, since roughly speaking, the drift of the SDE for $X(t)$ does not involve the local time of the process. Moreover, at least formally, the process $X_t^{\prime}$ satisfies the SDE: $$ d X_t^{\prime} = X_t^{\prime} \left( d L_t + \delta(X_t) dB_t \right) \quad X_0^{\prime}(x) = 1 $$ which depends on the initial data only through $X_t$. So if $X_t$ is a.s. pathwise differentiable, then it seems from ($\star\star$) like $\nu^{\prime}(x)$ is continuous. This motivates revisiting that paper you reference by Nakao, and understanding better the regularity of $X_t$ with respect to its initial data.


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.