Jethro's Braindump

Laplace's Method

Suppose we have an unnormalized probability density P(x), whose normalizing constant:

ZPP(x)dx

is of interest, and has a peak at point x0.

We perform a Taylor expansion of lnP(x) at this peak:

lnP(x)lnP(x0)c2(xx0)2+

where c=2x2lnP(x) where x=x0.

P(x) can be approximated by an unnormalized Gaussian:

Q(x)P(x0)exp[c2(xx0)2]

and the normalizing constant is approximated with:

ZQP(x0)2πc

This is easily generalizable to a K-dimensional space x.