Speaker
Erik Aurell
(KTH Stockholm)
Description
The cavity method or, in computer science, Belief
Propagation, is an efficient method to approximately
compute marginals of equilibrium probability
distributions e.g. magnetizations in spin glasses. The
method is exact if the underlying graph of interactions is
a tree, and generally expected to be accurate if that
graph is locally tree-like.
We have investigated a similar approximation scheme
for the diluted asymmmetric Ising spin glass with
synchronous or sequential update rules. The cavity
method can be formally set up in this context, but
requires an additional assumption of stationarity to be
computationally feasible: the approach is hence limited
to steady (but non-equilibrium) states. I will present
the dynamic cavity method, and numerical results for a
few examples.
This is joint work with Hamed Mahmoudi (Helsinki),
other recent relevant contributions are Neri & Bolle
(2009) and Montanari (2009).