Error Propagation in Scientific Computing

Is there any generic way to preserve loss of numerical accuracy for any given function?