SindhuChat 15 July 2024

Discussed that there has to be an excess of data events in the bins in which the data probability is higher than 0.5, by construction.

\[p_D = \frac{D}{D+B}\] \[ p_D = \frac{D/B}{1+D/B}\]

\[ p_D = D/B - p_D D/B\] \[ D/B = \frac{p_D}{1-p_D} \]

eg: \(p_D = 0.7\) => D/B = 0.7 / 0.3 ~ 2.3

Sindhu not convinced, will try out with a toy with unweighed background events then will add background events.

Plot form the optimal transport paper that shows this effect: https://arxiv.org/abs/2208.02807

The orange (before the FvT kinematic reweight) has data vs Bkg mis-modelling that is in a one-to-one correspondence with the FvT classifier output value. (here the FvT is defined as p4/p3 not p4 / (p4 + p3)).