Changes between Version 1 and Version 2 of WeightTreatment

Timestamp:

Feb 27, 2014, 1:07:08 PM (10 years ago)

Author:

buckley

Comment:

Fix misleading statement re. negative weights treatment

WeightTreatment

@@ -5 +5 @@
  * Weight treatment must allow arbitrary partitioning of data: fill1 + fill2 + fill3 == fill1 + (fill2 + fill3) == (fill1 + fill2) + fill3, etc.
  * The whole point of weights is that they allow more rapid convergence of distributions, but the errors must be equivalent to filling with a larger number of samples with smaller weights which add up to the same.
- * Negative weights must be treated carefully: filling a bin with +w and later with -w should behave as if neither fill ever occurred, ''including the error''. Hence, sum(w^2^) should be filled as sumw2 += sign(w) * w^2^.
+ * Negative weights must be treated carefully: filling a bin with +w and later with -w should contribute a value of zero, but the errors necessarily scale as the differing negative/positive contributions indicate significant uncertainty/fluctuation in the bin content.
 
-How to treat questions about bin/histo statistics where there is no weight (i.e. no information)? Throw exception? Return mean = 0, error = inf?
+Where bin/histo statistics questions are asked and there have been insufficient fills, a LowStatsError exception is thrown rather than return a magic number. These conditions are minimised as far as possible. The decision re. whether there is sufficient information in the bin to calculate e.g. the bin mean or its standard error is typically taken based on the effective number of fills/entries, sumW**2/sumW2.