Someone asserts that because OLS guarantees that residuals would have zero correlation with x, i.e., corr(u, x) = 0, adding the quadratic term of x (i.e., x2) as another independent variable would not allow us to fit the data better. Do you agree