reducing float test tolerance to 4ε
Invariants and eigenvalues computations require slightly larger tolerance.