Statistics of fracture fatigue are derived based on essentially general models of crack propagation and initiation. In fact, the form of the distribution is invariant with respect to the crack propagation law, however, the parameters do depend on the details of the phenomena. When this dependence is known, i.e. extracted from experimental data, our results can be applied even in the case of changing environments. Thus, such problems as projections to use conditions and the efficiency of burn in procedures can be precisely formulated and resolved, often without the requirements of further testing. These advantages derive from our use of physical models in deriving the failure distribution. This was not previously the case where for the almost universally used log-normal distribution each new situation requires the establishment of new rules. In fact, the log-normal is usually pessimistic and requires ad-hoc rectification to account for curvature when plotted on log-normal paper!, and generally does not relate to the fatigue process.