RooDecay getVal() normalization problem

xu_li · April 3, 2022, 3:52pm

Dear experts,

I found a problem about RooDecay getVal().
I have two observables tz and tzErr. The resolution model is the sum of two Gauss models and their widths are proportional to the tzErr.
I thought getVal(ROOT.RooArgSet(tz, tzErr)) and getVal(ROOT.RooArgSet(tz)) should return different values, but it doesn’t work.
Here is the code:

import ROOT
tz = ROOT.RooRealVar("tz", "tz", -5, 5)
tzErr = ROOT.RooRealVar("tzErr", "tzErr", 0.5, 1.0)
zero = ROOT.RooRealVar("zero", "zero", 0)
one = ROOT.RooRealVar("one", "one", 1)
sigma1 = ROOT.RooRealVar("sigma1", "sigma1", 0.5)
sigma2 = ROOT.RooRealVar("sigma2", "sigma2", 1.0)
beta = ROOT.RooRealVar("beta", "beta", 0.6)
res1 = ROOT.RooGaussModel("res1", "res1", tz, zero, sigma1, one, tzErr)
res2 = ROOT.RooGaussModel("res2", "res2", tz, zero, sigma2, one, tzErr)
res = ROOT.RooAddModel("res", "res", ROOT.RooArgList(res1, res2), ROOT.RooArgList(beta))
tz.setVal(0.0)
tzErr.setVal(0.8)
decay = ROOT.RooDecay("decay", "decay", tz, zero, res, ROOT.RooDecay.DoubleSided)
print("Norm[tz]", decay.getVal(ROOT.RooArgSet(tz)))
print("Norm[tz, tzErr]", decay.getVal(ROOT.RooArgSet(tz, tzErr)))
decay2 = ROOT.RooDecay("decay2", "decay2", tz, zero, res, ROOT.RooDecay.DoubleSided)
print("Norm[tz, tzErr]", decay2.getVal(ROOT.RooArgSet(tz, tzErr)))
print("Norm[tz]", decay2.getVal(ROOT.RooArgSet(tz)))

The printed results are:

(‘Norm[tz]’, 0.7978845609338628)
(‘Norm[tz, tzErr]’, 0.7978845609338628)

(‘Norm[tz, tzErr]’, 1.5957691471135516)
(‘Norm[tz]’, 1.5957691471135516)

The second call of getVal() function returns the same value as that in the first call.
Is this a bug or some feature that I did not understand?

The ROOT version is 6.24/00.

Best regards,
Li

jonas · April 6, 2022, 12:03am

Hi xu_li!

This is a bug, thanks for reporting. It will be fixed with this PR:

github.com/root-project/root

[RF] Use `RooArgSet::uniqueId` instead of pointers in RooNormSetCache

root-project:master ← guitargeek:RooNormSetCache_1

opened 11:57PM - 05 Apr 22 UTC

guitargeek

+31 -26

This change was considered for some time now, and a bug recently reported on th…e forum was the final motivation to open this PR: https://root-forum.cern.ch/t/roodecay-getval-normalization-problem/49457 There are many problems with the pointer comparisons: 1. If normalization sets are created on the stack it can often happen that the same address gets reused (as happened in that forum post, more precisely in the implementation of `RooAddModel`) 2. Memeory might also be reused on the heap, which lead to the developement of a memory arena where memory is not reused. This memory arena often causes other problems related to heap fragmentation or dictionaries. 3. Pointer comparisons still work if the RooArgSets for the cache entry are already out of scope, which can also cause problems The `UniqueId` is unique for each `RooArgSet` instance created in a RooFit process, no matter if on the heap or on the stack. The IDs are also never reused. This makes the `UniqueId` much more practical for cache keys than RooArgSet pointers.

It will land in the upcoming ROOT version 6.28.

I’m afraid the fix is too fundamental to backport to ROOT 6.24 which you are using. Would it be fine if this fix goes also int 6.26.02, the next upcoming patch release, and then you migrate to that version?

Cheers,
Jonas

PS: you should always use the most recent patch release of a given ROOT release branch. For 6.26, this is 6.24/06. Better not to use 6.24/00 anymore.

xu_li · April 6, 2022, 6:40am

Hi Jonas,

Thanks a lot for your reply and helpful suggestion. It is fine for me to migrate to another version.

Best regards,
Li

system · April 20, 2022, 6:40am

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.