You don't. roots builds the companion matrix and computes the eigenvalues of that matrix.
I can't see the context before that section of that paper or assignment but I'm guessing the process was to repeatedly create a polynomal coefficient vector, use roots to solve it, and stop the repetition if one / the smallest of the roots of the current coefficient vector became "small enough." That uses the output of roots as part of the temination criteria for the process, but it doesn't try to control roots itself.
