Efficient tau-pair invariant mass reconstruction with simplified matrix element techniques

The quality of the invariant mass reconstruction of the di-τ system is crucial for searches and analyses of di-τ resonances. Due to the presence of neutrinos in the final state, the τ τ invariant mass cannot be calculated directly at hadron colliders, where the longitudinal momentum sum constraint cannot be applied. A number of approaches have been adopted to mitigate this issue. The most general one uses Matrix Element (ME) integration for likelihood estimation, followed by invariant mass reconstruction as the value maximizing the likelihood. However, this method has a significant computational cost due to the need for integration over the phase space of the decay products. We propose an algorithm that reduces the computational cost by two orders of magnitude, while maintaining the resolution of the invariant mass reconstruction at a level comparable to the ME-based method. Moreover, we introduce additional features to estimate the uncertainty of the reconstructed mass and the kinematics of the initial τ leptons (e.g., their momenta).

💡 Research Summary

The paper addresses the long‑standing challenge of reconstructing the invariant mass of a τ‑pair (ττ) at hadron colliders, where at least two neutrinos escape detection and the longitudinal momentum balance cannot be imposed. Traditional approaches used by ATLAS and CMS rely on matrix‑element (ME) based maximum‑likelihood methods (SV‑Fit, cSV‑Fit) that integrate over the full decay phase space. While these techniques achieve excellent mass resolution, they are computationally intensive, typically requiring a few hundred milliseconds to seconds per event, which becomes prohibitive for analyses involving millions of events or systematic variations.

The authors propose a fast algorithm, dubbed fastMTT, that reduces the computational cost by roughly two orders of magnitude without sacrificing resolution. The key innovations are: (1) a collinear approximation (θ_GJ ≈ 0) that assumes the visible τ decay products and the invisible neutrinos are emitted along the τ direction, justified by the large boost (γ ≫ 1) of τ leptons from heavy resonances; this reduces each τ’s phase space from four dimensions to a single energy‑fraction variable x = E_vis/E_τ. (2) a simplified transfer function that treats the missing transverse energy (E_T^miss) as a Gaussian with an event‑by‑event covariance matrix, while assuming perfect measurement of the visible decay products.

Under these approximations the matrix element collapses to the τ Breit‑Wigner factor, and the δ‑function that enforces the test mass m_test = m_vis √(1/(x₁x₂)) allows the phase‑space integral to be performed analytically. The resulting integrals are simple logarithmic expressions (Eqs. 2.16‑2.18) that depend only on the limits of x₁ and x₂, which are themselves determined by kinematic constraints. The overall likelihood is then L = W(E_T^miss) × I(x₁,x₂), where W is the Gaussian transfer function and I is the analytic phase‑space term. Normalization constants are omitted because they cancel in the maximization.

To suppress the high‑mass tail caused by Z → ττ background, the authors introduce two regularization parameters: a scaling factor α that slightly reduces the test mass (α ≈ 1/1.1) and an exponent β that modifies the Jacobian term (m_vis m_test^β). These are tuned per decay channel (β = 6 for fully hadronic, 2 for semi‑leptonic, 3.5 for fully leptonic) to match the performance of cSV‑Fit.

The maximization is performed on a 100 × 100 grid in (x₁,x₂) space. For each grid point the analytic likelihood is evaluated, the maximum is located, and the corresponding τ four‑momenta and invariant mass are reconstructed using m_ττ = m_vis · √(1/(x₁x₂)). This grid scan takes about 2–3 ms per event, a ∼100× speed‑up over cSV‑Fit’s ∼0.25 s.

Uncertainty estimation exploits Wilks’ theorem: the likelihood map provides a χ² surface, and confidence intervals (68 %, 95 %, 99.7 %) are obtained by contouring at Δχ² = 2, 6, 9 respectively. This yields per‑event mass uncertainties without additional Monte Carlo sampling.

Performance studies using simulated Z → ττ and Higgs → ττ samples show that fastMTT achieves a mass resolution of roughly 10–12 % (σ/m), comparable to the full ME method, while dramatically reducing the high‑mass tail thanks to the β regularization. The algorithm scales well to datasets of order 10⁶ events, making it suitable for large‑scale LHC analyses and systematic studies.

In summary, fastMTT combines a physically motivated collinear approximation with a simplified transfer function and analytic phase‑space integration, delivering a fast, accurate, and uncertainty‑aware τ‑pair mass reconstruction. It offers a practical solution for current and future searches involving τ leptons, such as Higgs → ττ measurements and searches for heavy neutral resonances.