TripletCritic Module for Track Fitting

• TripletCritic module is a computational framework that fits three spatial hits using analytical closed-form equations, accounting for hit uncertainties, multiple scattering, and energy loss.
• It leverages kink angle linearization and regime-dependent optimizations to deliver rapid and accurate track fits across diverse detector geometries and magnetic field configurations.
• The module’s highly parallelizable design supports real-time processing in high-rate experiments, with iterative bias mitigation and scalability on GPUs and multicore CPUs.

The TripletCritic module is a specialized computational framework for three-dimensional particle track fitting based on hit triplets in tracking detectors, incorporating spatial hit uncertainties, multiple Coulomb scattering (MS), and optional energy loss effects. All fitting procedures rely on analytical closed-form equations, enabling a generic approach for arbitrary detector geometries and magnetic field configurations. The TripletCritic workflow is organized for maximal parallelism, early quality filtering, and regime-dependent optimal fit strategies, delivering track fits that are both rapid and adaptable across detector types (Schöning, 2024).

1. Analytical Closed-Form Triplet Fit

The module operates on precisely three spatial hits $\mathbf{x}_0$, $\mathbf{x}_1$, $\mathbf{x}_2$ within a known local magnetic field $B$, fitting a single circular helix segment through these points. The fit incorporates a single scattering "kink" at $\mathbf{x}_1$ and, if relevant, a small energy loss $\Delta E$. The primary fit parameters are the 3D curvature $\kappa = q/p$, residual shifts $\delta\mathbf{x}_0$, $\delta\mathbf{x}_1$, $\delta\mathbf{x}_2$ of the hit positions, and optionally the energy loss deviation $\mathbf{x}_1$0.

The triplet geometry is described by kink angles at layer 1 in polar and azimuthal projections: $\mathbf{x}_1$1 These transcendental functions are linearized about a reference curvature $\mathbf{x}_1$2 such that $\mathbf{x}_1$3, yielding: $\mathbf{x}_1$4 The quartet $\mathbf{x}_1$5 are the fundamental triplet parameters, uniquely derived from hit locations and field values.

Measurement-induced kink corrections follow: $\mathbf{x}_1$6 where $\mathbf{x}_1$7 etc. are numerically determined gradients.

Uncertainties from MS at layer 1 are assessed using Highland’s formula, with $\mathbf{x}_1$8: $\mathbf{x}_1$9 Hit covariances are denoted $\mathbf{x}_2$0. The full fit chi-squared reads: $\mathbf{x}_2$1 Minimization yields closed-form curvature and its variance: $\mathbf{x}_2$2 with $\mathbf{x}_2$3, $\mathbf{x}_2$4, $\mathbf{x}_2$5.

When hit errors are negligible, the solution reduces analytically: $\mathbf{x}_2$6 Energy loss $\mathbf{x}_2$7 shifts the reference parameters: $\mathbf{x}_2$8 with analogous modifications for $\mathbf{x}_2$9.

2. Triplet Fit Quality Metric

The fit quality is assessed locally via

$B$0

where the numerator encodes mismatch between polar and transverse kinks, and the denominator aggregates MS variances.

For general configurations incorporating hit and MS uncertainties, the global triplet-block quality metric is: $B$1 The local analytic formula for $B$2 is employed for rapid pre-filtering of candidate triplets, excising poorly reconstructed segments before costly matrix algebra.

3. TripletCritic Module Algorithm

Computation is batched across triplets with full GPU/CPU parallelism. The high-level steps for each triplet $B$3 are:

1. Compute Triplet Parameters:
   • Uniform Solenoidal Field: Apply analytic forms (Eqs. 49–56) for $B$4.
   • Simple Dipole Spectrometer: Use small-bend approximation (Eqs. 69–75) or numerically solve (Eqs. 64–67).
   • Arbitrary Magnetic Field: Execute four segment extrapolations (Appendix A.3), apply Newton refinement (Eq. A.5), then numerically differentiate to obtain parameters.
2. Hit Gradients (Optional):
   • Compute $B$5, $B$6 using Eqs. 13–14 and Appendix A.4.
3. Local Fit and Filtering:
   • Evaluate $B$7 (Eq. 35).
   • Reject triplets exceeding the quality threshold $B$8.
4. Result Storage:
   • For triplets passing the filter, store $B$9, $\mathbf{x}_1$0, and $\mathbf{x}_1$1.

Subsequently, surviving triplets are linked into longer track candidates, followed by a global track fit over ordered triplets:

• MS-Dominated Regime ($\mathbf{x}_1$2): Employ fast summation ($\mathbf{x}_1$3 from weighted sums, $\mathbf{x}_1$4 includes curvature-consistency).
• General Regime: Assemble $\mathbf{x}_1$5 from block matrices, invert, compute global $\mathbf{x}_1$6, variance, and residuals via closed-form expressions.

Energy loss is incorporated through parameter substitution (Eqs. 56–59), or by solving the complete system (Eqs. 60–64). Regime classification is enabled by:

• Tracking Scale: $\mathbf{x}_1$7; use MS-only for $\mathbf{x}_1$8.
• Curvature Significance: $\mathbf{x}_1$9; strong-bend for $\Delta E$0.

4. Track Fit Biases and Mitigation Strategies

Two principal biases affect naive MS-dominated triplet fitting:

• MS Normalisation Bias:

  $\Delta E$1

  This bias manifests as a systematic offset in the local pull mean when uncertainties are underestimated.

• Weighting Bias in Global Triplet Combination:

  $\Delta E$2

  affecting aggregate curvature estimation for $\Delta E$3 identical triplets.

Mitigation involves:

1. Iterative Reweighting: After an initial fit, update $\Delta E$4 and refit.
2. Regularised MS Fit: Linearize $\Delta E$5 within the chi-squared term:

   $\Delta E$6

   yielding bias-free pulls and residual bias scaling as $\Delta E$7.

5. Parallelization, Implementation, and Detector Adaptivity

Local fits for each triplet are entirely decoupled and ideally suited to large-scale parallelization on both GPU and multicore CPU architectures. For arbitrary field configurations, only four extrapolations per segment are necessary; uniform and dipole fields admit purely algebraic derivations.

Early filtering via $\Delta E$8 effectively prunes approximately 90% of fake triplets prior to any matrix assembly. In the MS-dominated global fit ($\Delta E$9), computation reduces to weighted summations resulting in $\kappa = q/p$0 arithmetic per triplet; for general scenarios, inversion of small ($\kappa = q/p$1) block matrices is required, or a banded (%%%%49$\mathbf{x}_2$750%%%%20) solver for longer tracks.

Empirical benchmarks (e.g., Mu3e online filter) demonstrate that a single modern GPU can sustain $\kappa = q/p$4 MS-only triplet fits per second, supporting MHz-rate real-time tracking. Triplet parameters are optimally stored in a structure-of-arrays for coalesced memory access, with each thread handling one triplet; packed parameter lists include $\kappa = q/p$5.

AVX-512 vectorization on CPUs achieves $\kappa = q/p$6100 million triplets/s per core, providing a viable emulation path.

6. Detector-Specific Parameterization and Regime Classification

The TripletCritic module is detector-independent except for the triplet parameter calculation step, which depends on the magnetic field configuration. For solenoidal and simple dipole detectors, explicit analytic equations are provided. For arbitrary magnetic fields, extrapolation and Newtonian matching are utilized to derive the requisite triplet parameters.

Regime classification utilizes derived tracking scales and curvature significance to select the appropriate optimal fit formula. This minimizes computational overhead and optimizes reconstruction quality depending on track geometry and error composition.

7. Significance and Applicability

The TripletCritic module enables efficient, generic, and bias-mitigated track reconstruction across a wide variety of detector types and magnetic field environments, with minimal change to fitting code or performance parameters. Its modular, parallelizable architecture and analytical foundation permit high-throughput, regime-adaptive tracking, crucial for applications demanding rapid and robust event filtering (e.g., online selection in high-rate particle physics experiments) (Schöning, 2024). A plausible implication is widespread adoption for future detector designs emphasizing parallel hardware and real-time processing.