The general idea is that you have a polynomial $p$ with one exact root $\alpha$. Due to numerical representation of the coefficients or the evaluation procedure the computer sees another polynomial that can, to some degree, be represented as $p(x)+\epsilon c(x)$. The basic idea of the perturbation calculation is now that this new polynomial still has a root close to $\alpha$, write it as $\alpha+\delta$. Now look at the Taylor expansion
$$
0+p'(\alpha)\delta+O(\delta^2)+\epsilon c(\alpha)+O(\epsilon\delta)
$$
If $p'(\alpha)$ is large against $\epsilon$ and $\delta$, then a root approximation for the perturbed polynomial can be found at $$\delta=-\epsilon\frac{c(\alpha)}{p'(\alpha)},$$ that is
$$
x=\alpha-\epsilon\frac{c(\alpha)}{p'(\alpha)}
$$
If you see that as the start of a Taylor expansion of the root curve $x(\epsilon)$, then this second term is the linear term of the expansion associate with the derivative $x'(0)$.
The derivative of $p(x)=\prod_{k=1}^n(x-\alpha_k)$ is
$$
p'(\alpha_m)=\lim_{x\to\alpha_m}\frac{p(x)-0}{x-\alpha_m}=\prod_{k\ne m}(\alpha_m-\alpha_k).
$$