On the accuracy of conservation of the adiabatic invariant