Typically, the non-rigorous way is just to show:
1. Round-off error is logarithmic to the number of operations.
2. After X iterations, the error is Y.
3. Extrapolate that to Z.
4. Therefore, it's not possible for the error to exceed W in this calculation.