Multiscale dynamics and multiscale
sampling through statistical projections
Alexandre
Chorin, University of California, Berkeley
Abstract:
I will present a mathematical derivation of equations of motion for
the
conditional averages of dynamical variables in stochastic multiscale
problems. The surprising feature of the resulting formalism is that it
yields
exact reduction schemes for the dynamics of multiscale problems, and
also
reduces the problem of evaluating the marginals in large sampling
problems to the much easier problem of evaluating conditional
expectations. The
resulting formulas are of course too complex to be fully evaluated,
but having an exact result to start with makes it easier to find good
approximations.
I will focus on two applications: the derivation of simplified dynamics
for systems with long memory (thus, no separation of scales), with an
application to hydrodynamics, and the development of effective
chain-free
sampling schemes, with an application to glassy systems.