pygsti.tools.optools.relaxed_scalar_tolerance

relaxed_scalar_tolerance(exponent=0.05)

Temporarily lower the global __SCALAR_TOL_EXPONENT__ used by fidelity() (and any other consumer of that global) to gate NumericalDomainWarning emission.

The effective tolerance becomes np.finfo(dtype).eps ** exponent. Smaller exponent => looser tolerance. In float64, eps ** 0.05 ≈ 0.165, which gives roughly a 2× margin over the drift typically observed during gauge optimization over non-unitary gauge groups (largest seen: ~0.08 trace deviation, ~-0.06 minimum eigenvalue). Pass a larger exponent (e.g. 0.15 ≈ 0.017) when wrapping calls on already-converged inputs that should be cleaner.

Not thread-safe: mutates a module-level global. Safe under pytest-xdist (forked workers have independent module state).

Parameters:

exponent (float)