Solvers

gtracr provides three integration methods for the relativistic Lorentz force equation, all implemented in C++.

Comparison

Solver solver= B-field evals/step Step size Key properties
Frozen-field RK4 "rk4" 1 Fixed Default; O(h^4) with frozen-field O(h^2) error
Boris pusher "boris" 1 Fixed Symplectic; ~30% faster than RK4; 2x better momentum conservation
Adaptive RK45 "rk45" ~6 (FSAL) Adaptive Dormand-Prince; uses atol/rtol; ~100x fewer steps for escaping trajectories

When to use each solver

Single trajectory visualization

Use "rk4" or "boris" with bfield_type="igrf":

  • Fixed time steps produce regularly spaced output points, good for plotting.
  • All three solvers advance the state in Cartesian (ECEF) coordinates, which are regular at the geographic poles, and report the spherical six-vector.
  • RK4 is the safe default.

GMRC cutoff scanning

Use "rk45" with bfield_type="table":

  • Adaptive stepping takes far fewer steps for escaping trajectories.
  • Combined with the tabulated IGRF (~7x faster field evaluation), this gives the highest throughput (~35k trajectories/s in batch mode).
  • Default tolerances (atol=0, rtol=1e-7) resolve cutoffs to the 0.5 % ladder rung (see below).

High-precision work

Tighten the RK45 tolerance:

traj = Trajectory(..., solver="rk45", rtol=1e-9)

RK45 error norm and tolerances

The RK45 step error is measured on the dimensionless state. Position errors are lengths in units of R_E, with tolerance atol + rtol·r/R_E. Momentum errors are in units of |p|, with tolerance atol + rtol. The norm is the RMS over the six components, applied to the error vectors, so it does not depend on the coordinate system.

The error is measured on the dimensionless state (position / R_E, momentum / |p_0|), so atol = 1e-3 allows ~6 km in position and 0.1 % in |p| per step.

atol rtol Use case
0 1e-6 Coarser tolerance
0 1e-7 Default tolerance
0 1e-10 Reference solutions

Benchmarking solvers

Run the solver comparison script to compare accuracy and performance:

python examples/eval_solver_comparison.py --n-perf 100