Tao 1997

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 1

698 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 42, NO.

5, MAY 1997

number of data and model dimension. Though a Hilbert space setting A Simple Alternative to the Barbălat Lemma
is considered, least square algorithm is far from being optimal, and
its identification error may diverge as the model dimension increases Gang Tao
(“peaking effect”). Computation of the optimal algorithm appears to
be hard, but a linear almost-optimal algorithm is presented, which is
easy to be computed, is not affected by the peaking effect, and is Abstract— This paper gives a simple proof of the property that if a
asymptotically optimal. signal is square integrable and has a bounded derivative, then the signal
converges to zero asymptotically.
Index Terms—Asymptotic stability.
REFERENCES
[1] A. Vicino and M. Milanese, “Optimal estimation theory for dynamic The Barbălat lemma has been commonly employed in the adaptive
systems with set membership uncertainty: An overview,” Automatica, control and estimation literature to conclude that a bounded error
vol. 27, pp. 997–1009, 1991.
[2] M. Milanese and A. Vicino, “Information-based complexity and non- signal converges to zero asymptotically if the error signal is square
parametric worst-case system identification,” J. Complexity, vol. 9, pp. integrable and has a bounded derivative. In this note, as an alter-
427-445, 1993. native to the Barbălat lemma, we present a direct proof of such a
[3] R. S. Smith and M. Dahleh, Eds., Modeling of Uncertainty in Control convergence property.
Systems. New York: Springer Verlag, 1994.
As in an adaptive system, an estimation error (t) has the desired
properties: (t) 2 L1 \ L2 and (t) 2 L1. From the Barbălat lemma
[4] M. Milanese, J. Norton, H. Piet-Lahanier, and E. Walter, Eds., Bounded
Approaches to System Identification. New York: Plenum, 1996. _
[5] P. M. Mäkilä and J. R. Partington, “Robust approximation and identifi- [2] (which states that if a function f (t) is uniformly continuous and
H
cation in 1 ,” IJC, vol. 58, pp. 683–685, 1993. t
limt!1 0 f ( )d exists and is finite, then limt!1 f (t) = 0), with
[6] L. Lin, L. Wang, and G. Zames, “Uncertainty principles and identi- 2
f (t) =  (t), all of the above properties of (t), and the fact that
fication n-widths for LTI and slowly varying systems,” IEEE Trans. 1
f (t) is uniformly continuous because f_(t) 2 L , it follows that
Automat. Contr., vol. 39, pp. 1827–1838, 1994.
[7] B. Wahlberg and P. M. Mäkilä, “On approximation of stable linear dy- limt!1 (t) = 0.
namical systems using orthonormal functions,” Royal Inst. Technology, We now give a direct proof of limt!1 (t) = 0 under a set of
Internal Rep., Stockholm, 1993.
[8] L. Giarré, M. Milanese, and M. Taragna, “Robust control oriented 1H slightly weaker conditions.
Lemma 1: If (t) 2 L2 and (t) _ 2 L1 ,1then limt!1 (t) = 0.
Proof: Since (t) 2 L and (t) 2 L , there exists a constant
identification with mixed perturbation models,” in Proc. IFAC Symp. 2
Robust Contr. Design, 1994, pp. 323–328. _
H
[9] L. Giarré and M. Milanese, “ 1 identification and model structure c0 > 0 such that
selection,” Int. J. Robust Nonlinear Contr., vol. 6, pp. 367–377, 1996.
t 1
[10] L. Giarré, B. Kacewicz, and M. Milanese, “Model quality evaluation in 2
j
 ( ) (j  sup j(t)
_ ) d _ j
2
 ( )d  c0 ; 8t  0:
t0
set membership identification,” in Proc. ECC, 1995, vol. 2, pp. 937–942.
[11] B. Z. Kacewicz, M. Milanese, and A. Vicino, “Conditionally optimal 0 0
algorithms and estimation of reduced order models,” J. Complexity, vol.
This inequality implies that limt!1 0 2 ( )j( _ )jd exists and is
4, pp. 73–85, 1988. t
t
[12] M. Canale, S. Malan, and M. Milanese, “Model quality evaluation in finite. This property then implies that limt!1 0 2 ( )(_ )d exists
identification for control,” in Proc. Automat. Contr. Conf., 1995, pp.
3060–3064. and is finite (an absolutely integrable function f (t) is also integrable
[13] P. Van den Hof, P. Heuberger, and J. Bokor, “System identification [1, pp. 389–390] because both the negative and positive parts of such
with generalized orthonormal basis function,” Automatica, vol. 31, no. an f (t) are integrable). Hence, we conclude that
12, pp. 1821–1834, 1995.
[14] B. Whalberg, “System Identification using Kautz models,” IEEE Trans. t
Automat. Contr., vol. 39, pp. 6, 1276–1281, 1994.
[15] J. Traub, G. Wasilkowski, and H. Woźniakowski, Information-Based
2 3
lim  (t) = lim  (t)
t!1 t!1
j j = lim 3
t!1
2
 ( )(
3
_ )d +  (0)
0
Complexity. New York: Academic, 1988.
exists, is finite, and must be zero because (t) 2 L2 ; that is, we have
[16] A. Pinkus, n-Widths in Approximation Theory. New York: Springer
Verlag, 1985.
limt!1 (t) = 0.
[17] M. Melkman and C. A. Micchelli, “Optimal estimation of linear operator
in Hilbert spaces from inaccurate data,” SIAM J. Numer. Anal. , vol. 16,
1
Lemma 1 holds for a vector signal (t) if 0 T (t)(t)dt  c1
pp. 87–105, 1979. T
and supt0 _ (t)(t)_  c2 for some constants c1 > 0 and c2 > 0,
[18] F. C. Schweppe, Uncertain Dynamic Systems. Englewood Cliffs, NJ:
then limt!1 (t) = 0.
Prentice Hall, 1973.
[19] J. Partington and P. M. Mäkilä, “Analysis of linear methods for robust
l
identification in 1 ,” in Proc. IFAC-SYSID’94, vol. 2, pp. 79–84. REFERENCES
[20] H. Akcay and H. Hjalmarson, “The least-squares identification of FIR
systems subject to worst-case noise,” Syst. Contr. Lett., vol. 24, pp. [1] A. F. Bermant, A Course of Mathematical Analysis: Part I. New York:
329–338, 1994. Macmillan, 1963.
[21] M. Milanese, “Properties of least square estimates in set membership [2] V. M. Popov, Hyperstability of Control Systems. New York: Springer-
identification,” Automatica, vol. 31, pp. 321–332, 1995. Verlag, 1973.

Manuscript received October 2, 1995; revised March 20, 1996 and March
25, 1996.
The author is with the Department of Electrical Engineering, University of
Virginia, Charlottesville, VA 22903 USA (e-mail: jt9s@virginia.edu).
Publisher Item Identifier S 0018-9286(97)03383-7.

0018–9286/97$10.00  1997 IEEE

You might also like