Precision in units of machine epsilon & translation to significant digits

Question

I am a bit confused when quantifying the loss of precision in terms of machine epsilon. Suppose I am working with doubles (so eps=2^-52) and suppose that an app