Jets faking hadronic taus get a new background estimate: multidimensional reweighting cuts the systematic uncertainty by up to 70%

When a jet is misidentified as a hadronic tau decay, the background estimate no longer relies on fixed bins in a few kinematic variables but comes from per-object multidimensional weights. The previous fake-factor method parametrised transfer factors with fixed binning in a limited number of variables and assumed the misidentification probability factorised between those variables and a separate set used to define analysis categories. On 62.4 inverse femtobarns of proton-proton collision data recorded by the CMS detector in 2022–2023, modelling of key kinematic observables improves relative to low-dimensional parametrisation methods, particularly in regions of high transverse momentum.

The approach trains a boosted decision tree in background-enriched determination regions to separate candidates failing and passing hadronic tau identification, then converts the calibrated outputs into continuous weights. A boosted decision tree is an ensemble classifier that adds weak learners sequentially; after temperature scaling its outputs are interpreted as class probabilities, and the ratio of two such probabilities acts as a continuous multidimensional transfer factor. Training is performed separately per analysis channel and background process, with inputs including the tau candidate decay mode, transverse momentum, pseudorapidity, the jet-to-tau transverse momentum ratio and event-level jet counts, while validation regions stay orthogonal to determination and signal regions.

In three representative analysis setups, the bin-by-bin non-closure is taken directly as the systematic uncertainty on the background estimate, and the multidimensional reweighting gives smaller or comparable deviations in most categories. The previous fake-factor method shows residual mismodelling in observables it cannot efficiently parametrise, often requiring additional closure corrections and associated uncertainties. In a QCD multijet-dominated validation region, the fitted tail residual corresponds to a tail systematic uncertainty falling from about 20% with the fake-factor method to about 6%, a reduction of about 70%.

The method is developed and validated on 62.4 inverse femtobarns of proton-proton collision data collected by the CMS detector in 2022–2023, and demonstrated in the fully hadronic and semileptonic channels. The intended readers are analysts working on tau final states and engineers needing data-driven background estimation. The approach does not depend on the specific hadronic tau identification algorithm and can be adapted as identification methods evolve.

The extrapolation uncertainty is the single largest source of systematic uncertainty in both channels, while all other components remain below the 2% level. A residual shape dependence remains in the semileptonic channel distributions, more pronounced than in the fully hadronic channel and qualitatively similar to that seen with the fake-factor method. In statistically limited high-score bins and some categories, the non-closures of the two methods are comparable within current statistical precision.