# Moments of diffusion processes

Suppose I have a diffusion $dX_t = a(X_t)dt + b(X_t)dW_t$. Is there a straightforward way of estimating the variance of $X_T$ for some time $T$, assuming that $T$ is large enough so that a simple Euler approximation isn't accurate?

Clearly, Monte-Carlo methods could be used here, but I'd like something more analytical.

Any ideas?

Many thanks.

-
What conditions over a and b ? are they deterministic, are they adapted ? how smooth are they ? –  TheBridge Aug 11 '11 at 22:39

One of the methods is the forward Kolmogorov's equation: $$\begin{cases} \frac{\partial m}{\partial t} &= a(x)\frac{\partial m}{\partial x} + b(x)\frac{\partial^2 m}{\partial x^2}, \\ m(0,x) &= f(x) \end{cases}$$ where $m(t,x) = \mathsf E_xf(X_t) = \mathsf E[f(X_t)|X_0=x]$.
In your case you should make calculations for $f_1 = x$ and $f_2 = x^2$. Then the variance will be given by $$V[X_T] = m_2(T,x) - (m_1(T,x))^2$$
Does $\mathsf E_xf(X_t)$ stand for expectation with respect to distribution of $X_t$ with initial condition being $x$ ? –  Sasha Aug 11 '11 at 15:42
@Simon: as far as I know, there are MC methods and FKE. I also would be happy to learn if there are others. The usual point is tha MC are working better in large dimensions. If you're looking for the steady behavior ($T\approx \infty$) or $T$ is just a large number (sorry for the informal statement)? –  Ilya Aug 11 '11 at 16:20
I'm just assuming $T$ is a large-ish number. A steady state does not necessarily exist. I was wondering if I could get some useful information from stochastic Taylor expansions, maybe. –  Simon Aug 11 '11 at 16:37