Strong Nonlinearity and Large Dimensionality: Should We Go to Work or to Sleep?

Wednesday, April 24, 2002 - 1:30pm - 2:30pm
Keller 3-180
Albert Tarantola (Institut de Physique du Globe de Paris)
This speaker believes that any serious formulation of an inverse problem leads to the definition of a probability distribution in the parameter space. and that solving the inverse problem means sampling this probability distribution. When using realistic parameterizations, many inverse problems lead to terribly multimodal probability distributions, that are trivial to sample in low dimension but difficult (or impossible?) to sample in the typical situation where the parameter space has thousands or millions of dimensions. As an introduction to this serious issue, the speaker will explain how an inverse problem should be formulated, pointing his finger to some of the most common mistakes.