= Fokker-Planck equation
{c}
{wiki=Fokker–Planck_equation}
A <Fokker-Planck equation> describes how a <probability density function>[probability density] evolves under <drift coefficient>[drift] and <diffusion process>[diffusion]. For an <Itô process>[Itô diffusion] $dX_t=b(X_t,t)\,dt+\sigma(X_t,t)\,dW_t$, put $a=\sigma\sigma^{\mathsf T}$. When a density exists and the required regularity holds, its equation is
$$
\partial_t p=-\sum_i\partial_i(b_i p)+\frac12\sum_{i,j}\partial_i\partial_j(a_{ij}p).
$$
The first term transports <probability> with the <drift coefficient>[drift]; the second spreads it through the local increment <covariance>. It is the forward equation adjoint to the backward <infinitesimal generator (stochastic processes)>[infinitesimal generator] $Lf=b\cdot\nabla f+\tfrac12\sum_{i,j}a_{ij}\partial_i\partial_jf$.
Writing $J_i=b_ip-\tfrac12\sum_j\partial_j(a_{ij}p)$ gives the <continuity equation> $\partial_t p+\nabla\cdot J=0$. Boundary conditions specify whether <probability> is reflected, absorbed or exchanged with the exterior. An invariant density must satisfy both the stationary equation and the boundary conditions. A zero divergence of current need not <mean> zero current.
For constant <drift coefficient>[drift] zero and $a=2DI$, the equation reduces to the <heat equation> $\partial_t p=D\Delta p$; a point source spreads into a <Normal distribution> with <covariance> $2DtI$. For one-dimensional <drift coefficient>[drift] $b(x)=-\gamma x$ with $\gamma>0$ and constant $a=2D>0$, the stationary density is normal with <variance> $D/\gamma$. Some texts call $a$ the <diffusion process>[diffusion] matrix, while others use $a/2$; stating the displayed convention avoids a factor-of-two ambiguity.
Back to article page