1188 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO.

11, NOVEMBER 1989

It is easy to see that by suitable choice of the parameter c,, . . ., c,-*, an

asymptotically stable motion can be obtained for the first r - 1 normal
coordinates. Thus, while a sliding motion locally takes place on k-’(O),
the original output y and its first r - 1 derivatives asymptotically IV. CONCLUSIONS
approach zero (i.e., the state vector of the system approaches the manifold
h-’(O), as initially desired). The corresponding equivalent control is now In this note, the relevance of the relative degree concept has been
locally given, in original coordinates, as the unique solution of [Lxh](x, examined in the analysis and design issues related to the creation of sliding
uEQ(x)) = 0. regimes for general nonliner systems. The results indicate that the
Example: For systems with affine vector fields X = f + U g and simplest possible structure at infinity must be exhibited by nonlinear
output y = h(x) which exhibit relative degree r on x o , the function w = systems undergoing sliding motions on the zero level set of the output
k ( x ) is given by k ( x ) = h(x) + ~ , - 2 L ; - ~ h ( x+) . .. + clLfh(x) feedback function. General necessary as well as necessary and sufficient
+ c a ( x ) and the equivalent control is computed as uEQ(x)= -Lfk/ conditions for the existence of sliding regimes have been presented. The
L,k = [Ljh + ~,-2L;-’h+ . . . + C I L+ ~~,&h]/L&;-’h. Locally, disturbance rejection properties of sliding mode control were also
on an open set M o f h-’(O), the last expression takes the form uEQ(x) = examined and a generalization of the matching condition was found.
Ljh/L&L;- ‘h. 0
The use of the auxiliary output w = k ( x ) implies the possibility of ACKNOWLEDGMENT
either being able to completely measure the original state variables and
proceed to use (3. ll), or else being able to generate r - 1 derivatives of The author is indebted to Prof. J. Grizzle of the University of
the original output function y . The last possibility is usually accomplished Michigan, for kindly providing the key reference [3], closely followed in
by means of a high gain, phase lead, “post-processor,” as proposed in the preparation of this work.
[ 3 ] , fed by the output signal y ( t ) . The transfer function of such a post-
processor is given by REFERENCES

- Kn(s)
V. Utkin, Sliding Modes and Their Application in Variable Structure
(3.14) Systems. Moscow, MIR, 1978.
(1 + Ts)‘ C. I. Byrnes and A. Isidori, “A frequency domain philosophy for nonlinear
systems with applications to stabilization and to adaptive control,” in Proc. 23rd
with T being a sufficiently small positive constant, K a sufficiently large IEEE Conf. Decision Contr., Las Vegas, 1984, pp. 1569-1573.
gain with, locally, the same sign as Lax/j&>-’h(x). n(s) is a stable A. Isidori, Lecture Notes on Nonlinear Control (Notes for a Course at the Carl
polynomial built as n(s) = sr-I + c , - ~ s ‘ -+~ * * * + cIs + c,.
Cram Gesellshaft). Aug. 1987.
-, Nonlinear Control Systems: A n Introduction (Notes on Information and
Control Systems, vol. 72). New York: Springer-Verlag, 1985.
C. Dkturbance Rejection Properties of Systems Undergoing Sliding H. Sira-Ramirez, “Differential geometric methods in variable structure control,”
Int. J. Contr., vol. 48, pp. 1359-1390, Oct. 1988.
Regimes -, “Sliding motions on slow manifolds of systems with fast actuators,” Int.
J. Syst. Sci., vol. 19, pp. 875-887, June 1988.
Consider a smooth nonlinear perturbed system of the form

dx/dt=X(x, U , W)

y=h(x) (3.15)

where w is a scalar perturbation signal affecting the system behavior. Let

us assume that the system locally has relative degree 1 on xo, and suppose
Error Bounds in the Averaging of Hybrid Systems
the input w is assumed to have local relative degree higher than 1 with
respect to the output function h. In normal form coordinates, the J. EZZINE AND A. H. HADDAD
perturbed system is written as
Abstract- This note analyzes the errors introduced by the averaging
dzi/dt= [ L x h l ( W ’ ( z ) U, , w)’ [ L x h l ( W ’ ( z )U, ) of hybrid systems. These systems involve linear systems which can take
a number of different realizations based on the state of an underly-
dq/dt=q(z1, q , w); y=z1. (3.16) ing finite state process. The averaging technique (based on a formula
from Lie algebras known as the Backer-Campbell-Hausdorff (BCH)
If a sliding motion can be created on z I = 0, the equivalent control formula) provides a single system matrix as an approximation to the
uEQ(z),is clearly unaffected by the perturbation signal w, although the hybrid system. The two errors discussed are: 1) the error induced by the
ideal sliding motion may still be affected. The following lemma follows truncation of the BCH series expansion; and 2) the error between the
immediately. actual hybrid system and its average. A simple sufficient stability test is
Lemma 3.8: Let (X, h ) have local relative degree 1. The existence of a proposed to check the asymptotic behavior of this error. In addition,
local sliding motion on an open set M of h-I(O) is independent of the conditions are derived that allow the use of state feedback to arrive at
perturbation signal w if and only if w has local relative degree, at least, a time-invariant system matrix instead of averaging.
equal to 2, with respect to the output function h(x). 0
Theorem 3.9: The ideal sliding dynamics is totally unaffected by I. INTRODUCTION
AND PROBLEM
F~RMULATION
perturbation signals w,of any kind, if and only if the matching condition
Hybrid systems are a special class of piecewise constant time-varying
ax/& E range {aX/au} (3.17) systems. Such models switch at different time instants among a finite set
of linear time-invariant realizations. Systems of this type can be used
is satisfied.
Proof: If the zero dynamics is unaffected by w ,necessarily, La,y/jw$i Manuscript received April 4, 1988; revised July 25, 1988 and October 31, 1988. This
= 0 for i = 2, * . . , n. Since the 4;s were chosen to also satisfy Ljx/au$i work was supported by the U.S. Air Force under Grant AFOSR-87-0308.
= 0, for i = 2, . . ., n, it follows that ax/& is, at least locally, exactly J. Ezzine is with the School of Electrical Engineering, Georgia Institute of Technology,
Atlanta, GA 303324250,
in the range of aX/au. On the other hand, aX/aw is locally in the range A. H. Haddad is with the Department of Electrical Engineering and Computer Science,
of aX/au if and only if there exists a smooth function b ( x ) such that, Northwestern University, Evanston, IL 60208.
locally on M, ax/& = b ( x ) [aX/au].Hence, for i = 2, 3 , - .. , n we IEEE Log Number 8930717.

0018-9286/89/1100-1188$01.00 0 1989 IEEE

IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 1 1 , NOVEMBER 1989 1189
to model systems subject to known abrupt parameter variations such as U. DERIVATION OF THE PERTURBATION BOUNDS
synchronously switched linear systems [ 11, networks with periodically
varying switches [2], or to approximate some types of time-varying sys- This section addresses the accuracy of the averaging technique. First,
tems [3]. This is achieved by imposing a deterministic switching rule 141. upper bounds for the errors introduced by using a truncated BCH formula
To model unknown abrupt phenomena such as systems subject to failures are derived.
151, the switching can be modeled, for example, as a finite-state Markov The averaging methodology to be analyzed in this note is based on
chain (FSMC). An earlier review of hybrid systems may be found in [ 6 ] . a formula from Lie algebras known as the Backer-Campbell-Hausdorff
Averaging theory, which is used in a deterministic or probabilistic (BCH) formula [2], 1111, 1121: Given two real matrices A and B, there
context, is an approach to the approximation of such systems by a single is no guarantee that there exists a real matrix C such that
constant linear model. In the probabilistic case, averaging is introduced
Exp ( A )Exp (B) = Exp (C). (3)
in a natural way by taking expected values. In the deterministic case,
however, averaging is introduced via perturbation techniques. Brockett However, if \\A\\+ JJBJJ 5 In (2), then C exists [ll], and will be given
and Wood [2] used a deterministic averaging technique to analyze and by a convergent infinite expression (the BCH formula)
stabilize a class of bilinear systems which are very hard to analyze or
control otherwise. Geman [7] used probabilistic averaging techniques to C = A + B + ( 1 / 2 ) [ A , B I + ( 1 / 1 2 )
study the stability of random differential equations. His main interest
was to explore the relation between asymptotic stability in the average +
. [ [ A ,Bl, BI (1/12)11B, AI, AI + . ’ . (4)
equation and asymptotic stability in the random equation. Specifically, where the symbol [A, B] = A B - B A (i.e., the commutator product).
when does the first imply the second? Kosut et al. 181 applied the theory In the sequel we show how the BCH formula and related expressions
of averaging to the analysis of the stability of adaptive systems. Ezzine can be used to compute the average of N-form hybrid systems. However,
and Haddad [9] used an averaging technique very similar to the one used for notational simplicity, F2-systems (i.e., two-form hybrid systems) only
in 121 to analyze and stabilize hybrid systems via a nonswitching gain. are treated.
As a matter of fact, Mariton et al. [lo] showed that nonswitching control The averaging idea introduced in [2] and used in 191 to stabilize hybrid
gains may be preferable, in addition to the fact that they are much easier systems is outlined in the following section. Given an F2-system such
to implement. that
In this note the averaging procedure used in 121, [9] is considered
further. In [2] there was no attempt to analyze the errors introduced by AI for At, time units (t.u.)
the BCH formula and the averaging method. However, there are two
very important issues in using such a formula and averages derived from
it as mentioned in [9]. The first is the error introduced by truncating
A(t) =
{
A2
or
for 6tz t.u.,
the BCH expression while computing the average matrix. The second is then an approximation for Exp (Al s t , ) Exp (A2&) is desired. The BCH
(5)

the difference between the actual system, the F2-system in [9], or the formula provides the approximation as Exp (CO, where C is as in (4).
bilinear system in 121, and the average system used to approximate the Now if we require an expression for C independent of the order of the
average behavior of the system under consideration. This note addresses product, to agree with the order independence introduced in the definition
both issues by providing bounds on the resulting errors. Furthermore, the of F2-systems, then we must compute
note provides conditions under which the BCH formula can be avoided.
Instead of using the BCH formula to compute an average system matrix, +
In f {Exp ( A )Exp (B) Exp (B)Exp ( A ) }
state feedback is used to obtain a constant closed-loop matrix for the - A +B+(1/12)[lA, Bl, B1+(1/12)[1B,AI,Al
system.
The class of hybrid systems considered in this note are assumed to
=A + +
B (1/12)[[A, B1, B - A l . (6)
have the form Therefore, we obtain a series of approximate averages for F2-systems.
If the F2-system realization is E, = (A1, B , ) for 6tl t.u., andE, =
(A2, B2) for 6t2 t.u., then the first-order approximation is

=(aA1 + ( 1 -(Y)Az,aBI + ( 1 - a m (7)

Y ( 4 = C(r(t))x(t) (1b)
where x is the system state vector of dimension n, U is the control input with (Y 6 t l / T . Second-order terms may be included to obtain the
vector of dimension p , y is the system output vector of dimension m, approximation
and r(t) is the “form index” which is a deterministic scalar sequence
taking values in the finite index set N = { 1, 2,. . . , N } . Let the ith form +
= (IUA, + (1 - Q)AZ (1/12)(Y(1 - a)[IA1,A d ,
denote the realization E, = ( A , ,B, , C , ) associated with the ith form (1 - CY)& - OLAI], +
aB1 (1 - a)&). (8)
index (i.e., r ( t ) = i ) , for i E N . Higher order approximations may also be derived. Consequently, ifwe !et
It is assumed that any r(t) sequence is composed of a succession of C denote the approximating matrix, then C can be written as C = C + C ,
N-termed blocks. Every block is a permutation of the index set N . It is where C is the error due to the approximation. Therefore, the induced
important to note that the succession of the blocks is completely arbi- error in computing Exp (Ct) by using the matrix C is
trary (e.g., for N = 3, a possible r(t)-sequence is 123, 321, 213, 213,
312, . . .). The time interval during which r ( t ) = i is denoted by 6t, , and E(t) Exp (Ct) - Exp (et). (9)
the time interval for each N-termed block is denoted by
The solution formula for inhomogeneous differential equations can be
used to derive an exact expression of E 1131
N

T = Est,.
i=l

The following is an outline of the note. Section I1 begins with an

E(t) =
.I’{Exp(C(t -s))CExp((C +C)s)}ds.

Using perturbation techniques [ 131 a useful approximate expression

(10)

overview of the averaging technique for hybrid systems. It also addrcsscs for E can be derived. Let C EC,,, where E is a scalar, then
the perturbations induced by the averaging procedure. Two important
perturbation errors are identified, and the first one is analyzed. Section 111 Exp ((C + C)t) = Exp (Ct)+ Exp (Ct)

..I’
E
discusses the second error, and also offers, where possible, an alternative
to averaging. Section IV concludes the note and points to additional open Exp (-Cs)C, Exp (Cs) ds + o(E’). (1 1)
questions.
1190 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 1 1 , NOVEMBER 1989

This is the Liouville-Neumann series solution for the integral equation. the second one is a sufficient test only [9], [4]. Because the first condition
It is a convergent perturbation series for all E [13]. However, this exact is computationally involved and is a generalization of a well-known result
expression, given as a series in E , does not lend much insight to qualitative (see [4]) we choose to present the second one. The second test is more
analysis questions. In that regard, the following result is given. general in the sense that it can be easily generalized to a larger class of
Proposition 1 (141: Let El denote the first-order approximation in E hybrid systems.
to E, then E, satisfies the following matrix differential equation: In order to state this sufficient condition, a brief introduction to the
notion of logarithmic norm is given.
Y = C Y + eC, Exp(Ct), Y(0) = 0. (12) The logarithmic norm (also known as the logarithmic derivative, the
measure of a matrix) was introduced in 1958 separately by Dahlquist [15]
Consequently, if all eigenvalues of C have negative real parts, then and Lozinskij [16] as a tool to study the growth of solutions to ordinary
E l ( t ) + 0 as t -+ 03. differential equations and the error growth in discretization methods for
Moreover, it is possible to derive a general explicit expression for their approximate solution. It is formally defined as follows.
E l . To do so we first recall the well-known result [12] Definition: The logarithmic norm associated with the matrix norm
M 11 . 11 is defined by
~ x p ( s ~ ) ~ ~ x p= s ~(). + / B
( -x i}! ) { ~ ~,
(13)
p ( A ) = lim (111+hAII - l)/h. (22)
I =O h -0-

with {Ao,B } = B and {A"+',B} = [A, {A", B}]. Using this identity Explicit expression for the logarithmic norm associated with the Eu-
the general explicit expression for E, will follow clidean norm is

00 p ( A ) = max { p : p E X((A +A*)/2)}. (23)

~ , ( t=) E ~ x p ( C t ) C ( r ' / i ! ) { ( - C ) ' - 'c,}.
, (14) Then the following double inequality is true:
,=I

At this point, we are ready to compute upper bounds for E, and E. EXP(-p(-AY) II I
L IIExP(4 Exp(~(A)t)- (24)
Proposition 2: Assume that IIExp(Ct)l( 5 M ( t )Exp(P(t)t), with P ( t )
Theorem I : For the null solution of the hybrid system (21) to be
a scalar function, then
uniformly asymptotically stable, it is necessary that
IIEi(t)II 5 ( I I ~ ~ ~ / ~ I I C I I ) M{P(t)f}(Exp
( ~ ) E ~ P {211Cllt} - 1) (15)
and with the added assumption that M(t) is monotonic, then
and sufficient to have
IIE(t)II I tllcllM'o EXP {(P(t) + M(t)IICll)t). (16)
ProoJ Using (13) in the evaluation of the integral in (17) yields
00

E l @ )= Exp(-At)x(t'/i!){A'-', B}. (17) where p , = 6ti / T , and 6Ai Ai - A,, for i E N .

Proof:We start by showing that the sufficient condition holds. Using
,=I
[17, Theorem 271 one can write
Taking the norm of both sides in (17) and using the following inequality:

leads to
M

llEl(Oll I I1 Exp(-At)ll~~I'/i!)112A/1'~1
IIBII (19)
i=l Using [17, Theorem 5 (e, d)] and after some algebra, we get
which after simple algebra results in (15). Equation (16) follows in the
same manner by using the monotonicity property of M(t).

III. STABILITY
OF THE ERROR
DYNAMICS
In this section the dynamics of the error between the average sys-
tem and the actual hybrid system are derived. The stability of the error
dynamics is discussed and two stability criteria are introduced.
Given a homogeneous N-form hybrid system with state vector x(t),
and one of its time-invariant averages with state vector x, (t), the error with
is given as

e(t) = x ( t ) - x , ( t ) .
From this definition it is easy to see that the error dynamics are governed
(20) p i = lim (l/(t

which completes the proof of (25b).

1-02
- to))
I' u;(s)ds (28)

by the following hybrid system: The necessary condition (25a) is shown similarly by using the fact that
N Exp {-@(--AV} 5 ((ExpAtll.
This simple sufficient condition states that for a hybrid system to be
e ( t ) = C u , ( t ) ( A i- A,)e(t) (21) uniformly asymptotically stable the weighted average of the logarithmic
i=i norms of each realization has to be negative. Therefore, this sufficient
where the indicator functions ui (t) are defined by vi ( t ) = 1 when the condition allows for unstable forms. That is, as long as the stable forms
original hybrid system is described by the ith realization
0 otherwise.
ci
, and 9 ( t ) = dominate, the overall system is asymptotically stable. This domination
can occur in three ways: either the stable forms are strongly stable (i.e.,
At this point two stability criteria are introduced to check the stability highly negative logarithmic norms), or the time span of the stable forms
of (21). The first criterion is a necessary and sufficient condition [4] and is large relative to the time span of the unstable ones, or a combination
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 11, NOVEMBER 1989 1191
of both reasons. This stability property of hybrid systems was reported which is itself an algebraic equation with K; , i = 1, 2, as unknown and
in [5] via examples. The difference between (25b) and (25a) could be the condition given in the theorem is the one needed for the existence of
used as a measure of the conservativeness of (25b). both gains. Equation (34) is a compact way to write these equations, and
Sometimes it is possible for F2-systems to avoid the computation of an is also useful when numerical techniques are used to solve the problem.
average matrix and, consequently, minimize the errors induced by aver- Corollary: If
aging [9]. That is, under certain conditions, it is possible, via sfatefeed-
back, to make the A-matrix of the closed-loop F2-system time-invariant . Range [A, - A2IB, - B2] = Range [B, - B2] (41)
That is, given E, and E, satisfying appropriate conditions, one can
then
compute a feedback gain matrix
K = K , =K2 (42)
K = [ K IIKz1 (29)
which will make the A-matrices of both forms equal, so that the A- and the gain matrix is given by
matrix of the closed-loop hybrid system (using gain K, for x;
) becomes
a constant. Moreover, this constant A matrix is given by the following
s ( K ) = { V I-&It @ Z I , } W , -A,). (43)
equation: To illustrate the above results, consider the following example.
Example: Given the following F2-system:
A = A I - B , K , = A2 - B2K2. (30)
In [18], Mariton proposed a technique quite similar to this idea. He
showed that it is possible to solve the jump linear quadratic (JLQ) prob-
lem by making the performance index independent of the different real-
izations of the form-index r(t).His approach renders the cost incurred by
any realization of r(t) the same. In other words, it makes all realizations
equal in that sense.
Even though the goals seem similar, the approaches are not. In contrast it is easy to check that after closing the loop via the gain K = (1 2), the
to 1181, the present equalization is direct; the homogeneous parts of system A-matrix becomes the identity for i = 1 , 2.
the two forms are made equal via feedback. In this section a sufficient
condition is given for the existence of K and a simple computational IV. CONCLUSION
algorithm based on the Kronecker-product and the generalized-inverse In this note the errors introduced by averaging hybrid systems were
techniques is proposed. As stated above, the following results hold for
analyzed. Two errors were identified and discussed. The first is the error
N = 2 only.
induced by truncating the BCH expression while computing the average
Theorem 2: Given the F2-system
A-matrix. The second is the error between the actual hybrid system and
its average. The note provides two upper bounds for the first error along
+
x = Aix Bju, i = 1 , 2 (31) with a simple sufficient stability criterion for the second one.
such that It is important to note that the results of this note are independent of
the averaging technique used to compute the average system. Thus, the
Range [A, - A2 (B,1 - B2] = Range [B, \ - B2] (32) results can be used, for instance, when the switching is governed by a
stochastic process such as a Markov chain and the average system is the
then, there exists a minimum-norm G = [ K ,J K z ]such that probabilistic average of the hybrid system. However, the stability tests
Ai - B , K , = A2 - B2K2. used in the note have to be adjusted accordingly. As a matter of fact,
(33) imposing some ergodicity conditions on the dynamics of the hybrid sys-
Moreover, the G matrix is given by tem yields exactly the same sufficient stability condition when properly
interpreted.
s(G) = { [ B I1 -&It @Z,,}S(AI - A d (34) A different direction that might be helpful in deriving better sta-
where s(.) is the stacking operation, (.)t is the generalized inverse, and
bility conditions for such systems is to think of every system = E,
(A,,B, , C , ) with i E N as an operator acting on the state x during 6t, ,
@ is the Kronecker-product.
and these operators are applied in a successive manner; then this process
Proof Starting with the forms E,, i = 1, 2, the ith model is given
can be viewed as an iterative process [19]. Viewing a hybrid system as
bY an iterative process sheds some light on the complexity of the stability
X(t) = A;x + B,u. of such systems.
(35) It would be interesting to compare the present work and [9] to [20]
If we define A A; - 6A; the above system can be written as follows: where a different type of averaging hybrid systems is discussed.

X(t) = AX + 6A,x + B;u. (36) REFERENCES

The next step is to compute a gain matrix K; such that T. L. Johnson, “Synchronous switching linear systems,” in Proc. 24th IEEE
Conf. Decision Contr., Ft. Lauderdale, FL, Dec. 1985, pp. 1699-1700.
SAix + B,u = 6A;x - B,Kix = (6A, - B;K,)x= 0 , for all x .
R. W. Brockett and 3 . R. Wood, “Electrical networks containing controlled
switches,” in Applications of Lie Groups Theory to Nonlinear Networks
Problems, Supplement to IEEE International Symposium on Circuit The-
(37) ory, San Francisco, CA, Apr. 1974, pp. 1 - 1 1 .
This is equivalent to 1. A. Richards, Analysis of Periodically Time-Varying Systems. New York:
Springer-Verlag, 1983.
B,K; = 6Ai (38) 1. Ezzine and A. H. Haddad, “On the controllability and observability of hybrid
systems,” in Proc. I988 Amer. Contr. Conf., Atlanta, GA, June 1988, pp.
which is an algebraic equation for the unknown K, . Subtracting the first 41-46.
equation from the second yields H. J . Chizeck, A . S. Willsky, and D . Castanon, “Discrete-time markovianjump
linear quadratic optimal control,” Int. J . Contr., vol. 43, no. 1, pp. 213-231,
1986.
AI -A2 = BIK, -B2K2 (39) D. D. Sworder, “Control of systems subject to sudden changes in character,”
which can be written as follows: Proc. ZEEE, vol. 64, no. 8, pp. 1219-1225, 1976.
S . Geman, “Some averaging and stability results for random differential equa-
tions,” SIAM J . Appl. Math., vol. 36, Feb. 1979.
[A, -A21 = [Bi, -BzI[K:, Ki1‘ (40) R. L. Kosut, B. D . 0. Anderson, and I. M. Y. Mareels, “Stability theory for
1192 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 34, NO. 11, NOVEMBER 1989

adaptive systems: Method of averaging and persistency of excitation,” IEEE reference models, differing by a priori assumptions about the regulator
Tmns. Automat. Contr., vol. AC-32, pp. 26-34, Jan. 1987. parameters.
J. Ezzine and A. H. Haddad, “On the stabilization of two-form hybrid systems via
averaging,” in Proc. 22nd Annual Conf. Inform. Sei. Syst., Princeton Univ.,
The analysis is first concerned with global stability and asymptotic
Mar. 1988, pp. 579-584. optimality. The attention is focused on the general properties of the se-
M. Mariton and P. Bertrand, “Non-switching control strategies for continuous- quence of parameter estimates; some new results are derived by using
time jump linear quadratic systems,’’ in Proc. 24th IEEE Conf. Decision Kolmogorov’s inequality. Starting from the time-varying model of the
Contr.. Ft. Lauderdale, FL, Dec. 1985, pp. 916-921.
J. G. F. Belinfante. “Explicit version of the Campbell-Baker-Hausdorff formula:
closed-loop system, convergence of the adaptive control law to the op-
integral representation for In eXeY,” unpublished. timal one in a Cesaro sense is proved for both algorithms. This is a
R. M. Wilcox, “Exponential operators and parameter differentiation in quantum generalization of the results presented in [7]. Introducing additional as-
physics,” J. Math. Phys., vol. 8, pp. 962-982, 1967. sumptions concerning the structure of the adaptive regulator and the
R. Bellman, Introduclion to Matrix Analysis. New York: McGraw-Hill, 1960.
J. Ezzine and C. D. Johnson, “Analysis of continuous-discrete model parameter
properties of the optimal regulator transfer function, it is proved that
sensitivity via a perturbation technique,” in P m . 18th Southeastern Symp. Syst. the algorithm in which the a priori knowledge about the sign of one of
Theory, 1986, pp. 545-550. the regulator parameters is incorporated provide the strong consistency
C. Van Loan, “The sensitivity of the matrix exponential,” SIAM J. Numer. of the parameter estimates. This is, to the authors’ knowledge, the first
Anal., vol. 14, Dec. 1977.
T. Strom, “On logarithmic norms,’’ SIAMJ. Numer. Anal., vol. 12, Oct. 1975.
strong consistency result for self-tuning algorithms in the general delay
C. A. Desoer and M. Vidyasagar, Feedbock Systems: Input-Oulput Properties. case. When no prior knowledge is available, i.e., when all the adaptive
New York: Academic, 1975. regulator parameters are estimated, it is proved that each recursion in
M. Mariton, “The equalizing solution of the JLQ problem,” in Proc. 26th IEEE the estimation algorithm converges to a multiple of the optimal param-
Conf. Decision Contr., Los Angeles, CA, Dec. 1987.
J. N. Tsitsiklis, “On the stability of asynchronous iterative processes,” in Proc.
eter vector. This result shows that the main conclusions in [7] can be
25th IEEE Conf. Decision Contr., Athens, Greece, Dec. 1986, pp. 1617-1621. extended to the tracking problem and the general delay case. An anal-
D. A. Castanon et al., “Asymptotic analysis, approximation and aggregation ysis of the persistence of excitation condition is also given. It is shown
methods for stochastic hybrid systems,” in Prm. 1980 JACC, San Francisco, that this condition holds only for the algorithms providing the strong
CA, 1980, paper TA3-D.
consistency.

11. PROBLEM
FORMULATION
Let the process be represented by a single-input single-output
ARMAX model
On the Convergence of Self-Tuning Stochastic
A(q-’)y(i) = q - d B ( q - ‘ ) u ( i ) + c ( q - ’ ) w ( i ) (i 2 1) (1)
Servo Algorithms Based on
Stochastic Approximation where {Y(J]}, {u(z]}, and {w(z]} are output, input, and stochastic dis-
turbance sequences, respectively, q-I stands for the unit delay opera-
SRDJAN s. STANKOVIC AND MILOJE S. RADENKOVIC tor, d for the pure process time-delay, while A ( q - l ) = 1 UI 4-l + +
. . . + ~ , , q - ~ ~ , B ( =q b- o~ +) b l q - ’ +...+
b , , q - “ ~ ( b ,# 0), and
Abstmct- Algorithms of the stochastic approximation type for self- + + +
C t q - ’ ) = 1 CI 4-1 . . , c,c q -”, .
tuning tracking of stochastic references in the general delay case are Let the reference sequence {Y*(J]} be generated by an ARMA model
proposed. Global stability, asymptotic optimality, convergence of the
adaptive control law in a Cesaro sense, and the strong consistency of the P(q-’)y*(i) = Q(q-’)W (2)
parameter estimates are proved. The persistence of excitation condition
where {U@} is a stochastic sequence independent of {w(z]}, while
is also analyzed. +
P ( q - ’ ) = 1 + p 1 q p 1 . . . +p,,q-“p and Q(4-I) = 1 q1q-I + +
I. INTRODUCTION ’ ’ ’ + qnaq-“Q .
Equations (1) and (2) are taken together with their initial con-
The problem of self-tuning control of linear stochastic systems has ditions x , = { y ” ( O ) , y ” ( - l ) , ~ ~ ~ , y “-( lk ) , u ” ( l - d),...,uD(l-
received considerable attention, starting from the pioneering contribution k), e(O),...,e(l - k ) } , w h e r e y o ( i r = b ( i ) : y * ( i ) ]u0(OT
, = [u(i):O],
due to h o m and Wittenmark 111. The problem of convergence of self- and = [w(i):u(i)], while k = max {na , n~ + d , ne, n p , n~ }. The
tuning control algorithms has been studied using different methodologies, process { x , , e(l), e(2), . . .} is defined on the probability space { R, F,
e.g., [2]-[5]. Goodwin, Ramadge, and Caines [6] have proved the global PI.
stability and asymptotic optimality for some algorithms. Becker, Kumar, Denote by F, the o-algebra generated by x, , and by F; the o-algebras
and Wei 171 have demonstrated convergence of the adaptive control law generated by { x , , e(l), . . . , e ( i ) } .Introduce also the following assump-
to the optimal one for the regulation problem in the unit delay case. tions concerning the process and the reference.
The available strong consistency results cover some specific aspects of A I : All finite dimensional distributions of .U, and { e(z]} are absolutely
the problem. Caines and Lafortune [8] studied an identifier working in continuous with respect to the Lebesgue measure.
parallel with the self-tuning control loop. Chen [9], Chen and Guo [lo], A2:
Kumar and Praly [ll], and Lai and Wei [13] have considered the tracking
E { e ( i ) l F i p 1=
} 0; E { e ( i ) e ( i ) T I F , - l }
problem with deterministic references in the unit delay case.
In this note we shall analyze asymptotic properties of direct self-
tuning control algorithms of the stochastic approximation type for track-
ing stochastic ARMA references in the case of ARMAX processes with
= [ut
0
‘1
U;
(as.) (3)

general delay. The motivation for such a control problem formulation can
be found in numerous applications where reference signals are neither , -k~, } < m; E { v ( ~ ) ~ J FI
E { w ( ~ ) ~ J F5 , -k~, }< m;
known nor reconstructable ahead of time [12]. We shall construct two (U,,,, U, < CO; i 2 1).
self-tuning control algorithms providing adaptation to both process and
A3: Polynomials B(z),C(z),P(z),and Q(z)have zeros strictly outside
the unit disk.
Manuscript received October 9, 1986; revised April 6, 1988. Paper recommended by Assume that the admissible control action u(z] is measurable with re-
Past Associate Editor, L. Valavani. specttotheo-algebrageneratedby { y ( l ) , . . ,. y ( i ) ,y * ( l ) , . . . J ( 0 , u(1)
The authors are with the Faculty of Electrical Engineering, University of Belgrade,
Belgrade, Yugoslavia. ,. . . , u ( i - l)}, which is, in general, smaller than F ; . The op-
IEEE Log Number 8930719.

0018-9286/89/1100-1192$01.00 0 1989 IEEE