Performance of the low-pT tau identification algorithm

There is an increasing demand to identify hadronically decaying tau leptons (denoted as $\tau_\text{h}$) below a $p_\text{T}$ of 10 GeV, such as for B-physics analyses and compressed SUSY searches. However, the orthodox $\tau_\text{h}$ identification algorithm is not applicable in the low $p_\text{T}$ regime, because the daughter particles from the $\tau_\text{h}$ decay will be spread over a large $\eta$--$\phi$ region and will not be confined in a single jet anymore from which $\tau_\text{h}$ is reconstructed. This note presents a new algorithm for identifying low-$p_\text{T}$ $\tau_\text{h}$ that specifically targets the $\tau \to \pi \pi \pi \nu $ decay without relying on the jet as a seed. The algorithm solely relies on reconstructed charged pions and exploits state-of-the-art machine learning algorithm to filter out irrelevant charged pions before building the $\tau_\text{h}$ candidate. The performance is evaluated using the $B_c \to J/\psi \tau \nu$ Monte Carlo simulation and superior performance compared to the existing algorithm is demonstrated down to a few GeV.