In case we only have one sample, the determinant of the Hessian (and thus at least one of its eigenvalues) becomes 0, so the loss function here is not strictly convex (but still convex)