Convergence properties of the membership set (Q1298292)

The authors study the convergence properties of the membership set for system identification in the presence of unknown but bounded noise. The system considered is a discrete-time scalar system of the form \[ y_i= \phi^\tau_i\theta+ v_i\quad (i= 1,\dots, N). \] The noise models considered are: Pointwise bound \[ {\mathcal V}_p(\varepsilon)= \Biggl\{v: \max_{1\leq i\leq N}|v_i|\leq \varepsilon\Biggr\}, \] Window bound \[ {\mathcal V}_w(\varepsilon)= \Biggl\{v: \Biggl({1\over T} \sum^{i_0+ T-1}_{i= i_0} v^2_i\Biggr)^{1/2}\leq \varepsilon\Biggr\} \] with some \(T>0\), and Average bound \[ {\mathcal V}_w(\varepsilon)= \Biggl\{v: \Biggl({1\over t} \sum^t_{i= 1} v^2_i\Biggr)^{1/2}\leq \varepsilon: t_0\leq t\leq N\Biggr\}. \] Two different scenarios are considered: (1) the bound is known; (2) the bound is unknown and has to be estimated from measurement data. Under some probabilistic assumptions on the noise sequence and with the persistent excitation regressor, in the first case, the authors proved that the diameter of the associated membership sets converges to zero with probability one if the noise bound is tight. If the bound is overestimated, the upper bounds for the diameter is computed which are stated in terms of the difference between the true and the overestimated bound. In the second case, it is shown that the estimates of the bound converge to the true but unknown noise bound.