you can compute the squared error
norm(y-y_original,2)
or the variants thereof like mean squared error or root mean squared error if you have a random element.
If you have the Signal Processing Toolbox see the help for rms() where you would use the difference vector.
A smaller squared error means a better approximation.
