Worst-case asymptotic properties of /spl Hscr//sub /spl infin// identification

This paper studies asymptotic properties of /spl Hscr//sub /spl infin// identification problems and algorithms. The sample complexity of time- and frequency-domain /spl Hscr//sub /spl infin// identification problems is estimated, which exhibits a polynomial growth requirement on the input observation duration for the time-domain /spl Hscr//sub /spl infin// identification problem, and a linear growth rate of frequency response samples required for the frequency-domain /spl Hscr//sub /spl infin// identification problem. The divergence behavior is also established for linear algorithms for the time- and frequency-domain problems. The results extend previous work to more restricted sets of linear time-invariant systems with more refined a priori information, specifically imposed on the stability degree and the steady-state gain of the systems, thus demonstrating that no robustly convergent linear algorithms can exist even for a small set of exponentially stable systems.