“…Then, the gradients and Hessians of the two cost functions can be computed in a standard way by suitably filtering the collected data D. Further, it can be seen that the problem of ensuring the numerical stability of the computation of u.t / u ı  .t / and y.t/ y ı  .t / is analogous to the problem of ensuring the stability of the predictor in the OE method. In this respect, we note that, when the experiment is sufficiently informative, u u ı  and y y ı  (i.e., the data) is almost surely unbounded as N goes to infinity for any value of  corresponding to an unstable C 1 .d; Â/ Q.d / (e.g., see lemma 1 in [28]). In contrast, u u ı  and y y ı  always remain bounded when C 1 .d; Â/ Q.d / is stable.…”