When a quantity is approximated instead of computed exactly, the mismatch between the actual and approximate values is measured by three related notions of error, all built on linear approximation (equation f(x0+Δx)−f(x0)≈f′(x0)Δx).
Definition (Absolute Error).
Absolute error=Actual value−Approximate value.
If h=x−x0 is the change in the input and E(h)=f(x0+h)−f(x0)−f′(x0)h, then E(h) is the absolute error made by using the linear approximation in place of the true value; E(0)=0 and h→0limhE(h)=0 whenever f is differentiable at x0, i.e. the error vanishes faster than h itself.
Absolute error alone doesn't say whether an approximation is good — an absolute error of 5 is trivial if the actual value is 10,000 but huge if the actual value is 6. This is fixed by dividing out the actual value:
Definition (Relative and Percentage Error). If the actual value is nonzero,
Relative error=Actual valueActual value−Approximate value,Percentage error=Relative error×100.
Absolute error carries the same units as the quantity being measured (cm, cm2, cm3, ...); relative error and percentage error are dimensionless (unit-free) ratios, which is exactly what makes them fit for comparing two very differently-sized approximations.
Working method for an error problem. Given a quantity Q=f(x) measured via x, and a measurement/change dx in x:
- Absolute error in Q ≈dQ=f′(x)dx (evaluated at the reference value of x).
- Relative error ≈QdQ=f(x)f′(x)dx.
- Percentage error =100× relative error. …