Next: The Main Idea
Up: Infinite Dimensional Preconditioners for
Previous: Infinite Dimensional Preconditioners for
In this lecture we discuss the use of elementary theory of pseudo-differential
operators to analyze Hessians for optimization problems governed by
partial differential equations. The role of this analysis is to come up with
Hessian approximation that will accelerate gradient based methods.
Quasi-Newton methods such as BFGS
cannot cope efficiently with a very large dimension of the design space.
Their efficiency is very good for a small dimensional design space,
and it deteriorate as the dimension of the design space increases.
The approach presented here is a complementing approach for the classical
quasi-Newton methods. It uses the asymptotic behavior
of the symbol of the Hessian to construct an accurate Hessian approximation
for the high frequency range, in the representation of the design space.
The resulting approximate Hessians are differential or pseudo-differential