Consider the following stochastic differential equation (SDE)

$$d X_s= \mu (X_s + b)ds + \sigma X_s d w_s $$

where constants $\mu, \sigma, b > 0$ and initial position $X_0$ are given.

If $b=0$, then the above equation is a geometric Brownian motion (GBM) and the distribution of $X_t$ at time $t$ is lognormally distributed. If $b>0$, can I say anything about the distribution of $X_t$ at a later time $t$? Is it possible to find the probability that $X_t \in (B, B+1)$?


2 Answers 2


Simplifying the Drift

As will all linear SDEs, let $Y_t=e^{-\mu t}X_t$. Then, \begin{align*} \text{d}Y_t &=-\mu e^{-\mu t}X_t\text{d}t+e^{-\mu t}\text{d}X_t \\ &=\mu b e^{-\mu t}\text{d}t+\sigma Y_t\text{d}W_t. \end{align*}

Product Rule

Consider the geometric Brownian motion $Z_t$ with $\text{d}Z_t=\sigma^2Z_t\text{d}t-\sigma Z_t\text{d} W_t$ and $Z_0=1$ such that $Z_t=\exp\left(\frac{1}{2}\sigma^2 t-\sigma W_t\right)$.

Then, \begin{align*} \text{d}Y_tZ_t&= Y_t\text{d}Z_t+Z_t\text{d}Y_t+\text{d}Y_t\text{d}Z_t \\ &=\sigma^2Y_tZ_t\text{d}t-\sigma Y_tZ_t\text{d} W_t+\mu b e^{-\mu t}Z_t\text{d}t+\sigma Y_tZ_t\text{d}W_t-\sigma^2Y_tZ_t\text{d}t \\ &=\mu b e^{-\mu t}Z_t\text{d}t. \end{align*} Thus, \begin{align*} Y_tZ_t-Y_0Z_0=\mu b\int_0^te^{-\mu s}Z_s\text{d}s. \end{align*} Finally, \begin{align*} X_t&=X_0e^{\mu t}Z_t^{-1}+\mu be^{\mu t}Z_t^{-1}\int_0^te^{-\mu s}Z_s\text{d}s\\ &=e^{\mu t}Z_t^{-1}\left(X_0+\mu b\int_0^te^{-\mu s}Z_s\text{d}s\right). \end{align*} However, I do not think the distribution of $X_t$, which includes an integrated geometric Brownian motion, is known? This is the entire struggle of pricing Asian options.

Special Cases

We can recover two special cases:

  • If $\mu=0$, we get $X_t=X_0\exp\left(-\frac{1}{2}\sigma^2 t+\sigma W_t\right)$.
  • If $b=0$, we get $X_t=X_0\exp\left(\left(\mu-\frac{1}{2}\sigma^2\right) t+\sigma W_t\right)$.
  • $\begingroup$ Thanks. A follow up question. In this case, for any finite B>0 and X_o > 0, can I say that P(X_t ∈ (B,B+1) ) > 0 for any t>0? I think in case of GBM we can say that because CCDF of lognormal distribution has a error function which goes to 1 only if B tends to infinity. $\endgroup$
    – S_R
    Feb 25, 2021 at 11:40
  • 1
    $\begingroup$ Yes, I think you can say this. $e^{\mu t}Z_t^{-1}$ is a geometric Brownian motion. Thus, $X_t$ is the product of a GBM with another factor which is positive (assuming $X_0,b,\mu>0$). Because the support of a log-normal distribution is $(0,\infty)$, I'd guess the same applies $X_t$ and the process should be able to attain values in any given interval $(B,B+1)$, i.e. $\mathbb{P}[\{X_t\in(B,B+1)\}]>0$ for all time points $t>0$ and constants $B>0$. $\endgroup$
    – Kevin
    Feb 25, 2021 at 12:11

If b>0, can I say anything about the distribution of 𝑋𝑡 at a later time t?

Yes - The solution is in Kloeden and Platen

You want to refer to section 4.4 of Numerical solutions of stochastic differential equations by Kloeden and Platen (which is my go-to book for SDEs). Under the section on linear SDEs with multiplicative noise on page 119, an SDE of this form is explicitly listed in (4.9), which I repeat here: $$ \mathrm{d}X_t = (aX_t + c) \,\mathrm{d}t + (bX_t + d)\, \,\mathrm{d}W_t $$ has the solution $$ X_t = \Phi_t \left(X_0 + (c-bd)\int_0^t \Phi^{-1}_s\,\mathrm{d}s + d\int_0^t \Phi^{-1}_s\,\mathrm{d}W_s\right) $$ with the fundamental solution $$ \Phi_t = \exp\left( \left(a-\tfrac{1}{2}b^2\right)t + bW_t\right). $$

A few equations later they give the solution for when the coefficients are all functions of time too, but it's a bit too much for me to type up here.


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.