would the k$-$means algorithm still converge to a local optimum if instead of randomly assigning each item to a cluster, we instead initialized the algorithm by picking K centriod vectors? why