4
$\begingroup$

How to do stochastic integration $dS = a S^b dt + c S dW$, where $a$, $b$ and $c$ are constant, $b > 0$, and $W$ is the Wiener process.

I know how to do integration for $dS = aS dt + cS dW$, or $dS / S = a dt + c dW$:

According to Ito's lemma, for $dx = \mu dt + \sigma dW$, and $f(t,x)$, $$df(t,x) = (\frac{\partial f}{\partial t} + \mu \frac{\partial f}{\partial x} + \frac{1}{2} \sigma^2 \frac{\partial^2f}{\partial x^2} ) dt + \sigma \frac{\partial f}{\partial x} dW $$

Submitting in $S\rightarrow x$, $aS \rightarrow \mu$, $cS \rightarrow \sigma$, and $ln(S)\rightarrow f(t,x)$, will get:

$$d[ln(S)] = (a - \frac{1}{2} c^2) dt + c dW$$

, then we'll have $$S(t) = exp[(a - \frac{1}{2} c^2) t + cW_t]$$

But if $S^b$ is introduced, how could I solve it?

$\endgroup$

1 Answer 1

9
$\begingroup$

There are several possibilites to solve this SDE:

Solution 1 You already solved the case $b=1$, so without loss of generality $b \neq 1$.

First, consider the (deterministic) ODE

$$ds = a \cdot s^b \, dt$$

The solution $s(t)$ satisfies

$$\frac{s(t)^{-b+1}}{1-b} - a \cdot t = \text{const}$$

Therefore, we try the following Ansatz:

$$Y_t := \frac{S_t^{1-b}}{1-b} - a \cdot t \tag{1}$$

The idea behind this choice is simply the following: In the deterministic case, $Y_t$ is constant. But since we add a random perturbation, we should allow $Y$ to depend on $\omega$. By applying Itô's formula,

$$dY_t = c \cdot (1-b) \cdot (Y_t+a \cdot t) \, dW_t - \frac{b}{2} \cdot c^2 \cdot (Y_t+at) \, dt$$

Now let $Z_t := Y_t + a \cdot t$,

$$dZ_t = dY_t + a \, dt = c \cdot (1-b) \cdot Z_t \, dW_t + \left( - \frac{b}{2} \cdot c^2 \cdot Z_t+a \right) \, dt$$

Consequently, $Z_t$ solves a linear SDE. Sinc it is known, how to solve this SDE, we find an expression for $Z_t$ and -obviously- for $Y_t$. From (1), we get a candidate for the solution $S_t$.

(Hopefully, my calculations are correct. Don't hesitate to ask!)


Solution 2 For SDE's of the form $$dS_t = b(S_t) \, dt + \sigma(S_t) \, dW_t \tag{2}$$ there exist equivalent conditions for a transformation into a linear SDE, for example the following:

The SDE (2) can be transformed into a linear SDE $$dZ_t = (\alpha+ \beta \cdot Z_t) \, dt + (\gamma+\delta \cdot Z_t) \, dW_t$$ if and only if $$\frac{d}{dx} \left( \frac{\frac{d}{dx}(\kappa'(x) \cdot \sigma(x))}{\kappa'(x)} \right) = 0 \tag{3}$$ where $\kappa(x) := \frac{b(x)}{\sigma(x)}- \frac{1}{2} \sigma'(x)$. The transformation $Z_t = f(X_t)$ is given by $$f(x) := \begin{cases} e^{\delta \cdot d(x)} & \delta \neq 0 \\ \gamma \cdot d(x) & \delta = 0 \end{cases}$$ where $$d(x) := \int_0^x \frac{1}{\sigma(y)} \, dy \qquad \qquad \delta = - \frac{\frac{d}{dx}(\kappa'(x) \cdot \sigma(x))}{\kappa'(x)}$$

(cf. René L. Schilling/Lothar Partzsch: Brownian motion - An Introduction to Stochastic Processes, p.278.)

You can easily check that for the given SDE, condition (3) is satisfied and therefore this theorem tells you how to transform the SDE into a linear SDE.

$\endgroup$
4
  • $\begingroup$ thanks a lot for your detailed answer. sorry for the late reply. appreciate your time and elaboration! $\endgroup$
    – athos
    Oct 1, 2013 at 6:08
  • 2
    $\begingroup$ @athos You are welcome. By the way, you can upvote & accept answers which satisfy you. $\endgroup$
    – saz
    Oct 1, 2013 at 16:11
  • $\begingroup$ when you write $\sigma^\prime (x)$ do you mean the derivative? The same for $\kappa^\prime (x)$? $\endgroup$
    – user48672
    Aug 23, 2017 at 10:33
  • $\begingroup$ @user48672 Yes, exactly. $\endgroup$
    – saz
    Aug 23, 2017 at 12:35

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.