🚀 go-pugleaf

We are working on an algorithm  that  incorporates  a  Metropolis
sampler  (*).   Therefore  we  are  interested  in references and
information concerning:

1. The convergence properties of a Metropolis sampler:
   - when  is  the  starting  state  of  a   Metropolis   sampler
     'forgotten'?
   - how fast and how  good  does  the  sampler  approximate  the
     Boltzmann distribution?
2. Variants of the Metropolis algorithm and recent developments.
3. Possibilities to accelerate a Metropolis sampler.

Concerning the 3rd question, we know of approaches by  Aarts  and
Van   Laarhoven  to  parallelize  simulated  annealing  known  as
systolic and clustered algorithms.  These might be applied to the
Metropolis  sampler  running  on  one  fixed  temperature.  Other
approaches distribute the generation of new states of one sampler
among  processors  but  have  a  speed-up  that  depends  on  the
temperature.  We prefer to avoid this.

Thankyou in advance for any references and hints;  please  e-mail
your reply to me: maurits@cs.kuleuven.ac.be.

(*) A Metropolis sampler generates a sequence of states such that
the  frequency  of  the  states  in this chain is proportional to
their probability in a given Boltzmann distribution.   Metropolis
samplers  are often used to simulate the equilibrium distribution
in the inner loop of simulated annealing.  We are  interested  in
the  distribution  at  just  one  fixed  temperature,  whereas in
simulated annealing one decreases the temperature  in  the  outer
loop.

Maurits Malfait
--
+-----------------------------------+--------------------------------+
| Email: na.malfait@na-net.ornl.gov | Maurits Malfait                |
|                    or, preferably | Katholieke Universiteit Leuven |
| Maurits.Malfait@cs.kuleuven.ac.be | Department of Computer Science |
| Phone:       + 32 16 201015 x3080 | Celestijnenlaan 200A           |
| Fax:               + 32 16 205308 | B-3001 Leuven (Belgium-Europe) |
+-----------------------------------+--------------------------------+