A query about the paper

[wwma]

Thanks for your great work!
In Definition 3.1,

Given the top r smallest eigenvalues λ1, λ2, ..., λr and their corresponding eigenvectors

you are using the smallest eigenvalues building the weighted spectral embedding matrix V, while in the next part,due to the dominant singular components are hardly affected by adversarial attacks, the weighted spectral embedding is therefore also resistant to adversarial attacks,so it's safe to use the adv's weighted spectral embedding matrix V.

In the adversarial attacks,it will not change the dominant singular components ,but only the small singular components are getting changed? Does it means right? So why the V stay unchanged for using small eigenvalues ? because 1- λ1?