bu-cms / rebsmear

Rebalance and smear

Flexible implementation of the rebalance and smear method in python + RooFit.

Example

from rebalance import Jet, RebalanceWSFactory

# Event with two arbitrary reco-level jets
jets = [
    Jet(pt=100, eta=1.3, phi=0.),
    Jet(pt=100, eta=1.3, phi=2.8),
]

# Factory creates RooWorkspace for us
# default workspace has dummy assumptions for priors, resolution functions
# -> Simple extension possible via inheritance!
rbwsfac = RebalanceWSFactory(jets)
rbwsfac.build()
ws = rbwsfac.get_ws()

# Ready to minize the negative log likelihood!
m = r.RooMinimizer(ws.function("nll"))
m.migrad()

rebsmear's Issues

Possibly wrong HTmiss calculation

Hey @AndreasAlbert , while deriving HTmiss distribution from data using a different coffea-based processor, I realized the HTmiss values calculated by that processor are not consistent with the ones that are calculated by rebsmear code. I believe the issue is that the "Jet phi" values read by rebsmear are different. I'll explain what I saw below.

I ran over the first event of our input for rebsmear, using both frameworks. From the input root file, I get the following jet pt and phi:

First running with coffea-based processor, I print out the pt, phi and resulting HTmiss values:

They are all consistent with the input file, the resulting HTmiss is wildly different from what rebsmear calculates though, its around 30 GeV, while rebsmear calculates 762 GeV for the same event.

To check rebsmear code, I looked at the output workspace, and you can see the pt of the leading jet is consistent, but phi is completely different somehow:

This is true for all jets in the event, and hence HTmiss ends up being very different. I don't yet know how exactly this happens but I'll keep investigating.

RooHistPdf check

Hey @AndreasAlbert , starting from the code snippet you posted earlier, I developed a small test script here:

https://github.com/alpakpinar/rebsmear/blob/2021-01-27_rootest/rootest.py

Here, I follow the workflow of the main code: Define a derived variable, re-evaluate it as a RooRealVar and feed it into RooDataHist and RooHistPdf. You can see what the code outputs at the bottom of the script.

I'm not sure if that's how the minimization software acquires the values for each iteration, so the comparison might not be entirely valid. But I see that the values returned from the pdf seem consistent. I'm also attaching a plot of the PDF for reference. I believe getValV() returns the un-normalized (but bin-width divided) values, seems consistent with the docs in 1.

Let me know what you think, thanks!