We consider different kinds of "pathological traps" for stochastic algorithms, thus extending a previous study on regular traps. An illustration is given by the complete proof of the convergence of a principal component analysis (PCA) algorithm when the eigenvalues are multiple.