...............................................
Return of the MINISEMINAR, g wahba prop. 

Who: Doug Nychka, Project Leader, 
     Geophysical Statistics Project
     National Center for Atmospheric Research

When: 11:00 am, Tuesday, April 20, 1999
Where:  Room 4310 CS-Stat 
      (enter 1210 W. Dayton, go straight
      back, take elevators on your left to 
      4th Floor, turn right and go down 
      1/2 flight of stairs, turn right and
      4310 will be first room on your right. 

What: Approximating posterior distributions in ensemble weather forecasting 
            
	    Abstract: 


A basic problem in forecasting weather is updating current knowledge of
the atmosphere (the state vector) with new observational data. In
statistical terms, the current knowledge would be expressed as a prior and
the updating is computing a posterior distribution using Bayes theorem.  
Also, statisticians will recognize the updating step as solving a
variational problem.  While conceptually simple, there are significant
difficulties in implementing this framework. Most notably, the prior
distribution may be complicated and the updating step for practical
problems involves very large data sets and so is computationally
intensive.  As a scheme for representing prior uncertainty the geophysical
researchers have used ensembles (i.e. groups) of state vectors to
represent prior uncertainty. The contribution of this talk is to recognize
that ensemble methods at their best are using multivariate normal mixtures
as priors and posteriors. This insight allows one to suggest improvements
to the methods.