From Erik Brubaker

*: We should do/study it.
+: We should think about it and seriously consider doing/studying it.
-: Probably not worth spending extra time on.

 +Optimize cut on 4th jet et (i.e. half jet for 3.5 jet events).

 -SLT jet corrections (in progress, probably not feasible).

 +Optimize W mass constraint. Is PDG value the right thing here, or can we
 scan over some range and choose something more experimentally based? Mean
 of dijet W mass is not 80.41.

 *Use Breit-Wigner instead of Gaussian shape for W and top lineshapes. How
 to do this in a chi2?

 *Use top width of 1.5, not 2.5. Easier to explain even if Gaussian
 approximation is more important than actual width used. Check what the
 generators use; be consistent.

 *Reoptimize chi2 cuts (in progress).

 -Probabilistic treatment of tags<->j-p assignment (takes lots of study,
 not feasible).

 -Any better signal template parameterization?

 +Fit background components separately in likelihood?

 +Better fits to background templates. In particular, the mistag template
 is a collection of weighted events and the Poisson statistics assumed in a
 binned likelihood fit don't apply. Similarly for the combined template,
 used in our current approach. Because of this, it's not clear what
 normalization to give the weighted templates. The chi2/ndof is affected
 also (it's currently quite low).

 +For PDF uncertainty, weight background template as well? (Lots of work,
 probably not much reward.)

 *Re-evaluate the b-tagging systematic.

 +Better approach/prescription for getting 68% coverage than scaling
 uncertainties?

From Jean-Francois

 + Better estimate of the MC statistics uncertainty (re-extract machinery
 from fluctuated templates)

 * Try combining subsamples based on total uncertainty

 * For 2-tag mistag background, a more realistic shape would be 1 real
  tag+1 mistag (instead of 2-mistag)

 * Take correlation in the background into account

Pekka Sinervo

 Check out whether whether we would gain significantly from use 
 of an enhanced SVX tagging algorithm.  
 A study of a combined SecVTX+JetProb tagger by Stan Lai (CDF 6874)  
 would suggest a 15-20% increase in b-tags with only a modest increase in 
 fake tags.  This would result in something like a 7-10% improvement in 
 statistical precision, perhaps more if we significantly increase our 
 double-tag yield.  Of course, we would need to make sure that the 
 studies done in 4.11 are repeated with 5.3.  This is a different 
 strategy than using a probabilistic approach to b tagging in the fitting 
 (which I agree is not feasible).