Journal of Mathematical Analysis and Applications 237, 464᎐504 Ž1999.

Article ID jmaa.1999.6456, available online at http:rrwww.idealibrary.com on

Adaptive Pole Positioning in MIMO Linear Systems by

Periodic Multirate-Input Controllers
K. G. Arvanitis
View metadata, citation of
Department and similar papers
Electrical at core.ac.uk
and Computer Engineering, National Technical Uni¨ ersity of Athens, brought to you by CORE
Di¨ ision of Computer Science, Zographou 15773, Athens, Greece provided by Elsevier - Publisher Connector
E-mail: [email protected]

G. Kalogeropoulos

Department of Mathematics, Section of Mathematical Analysis, Uni¨ ersity of Athens,

Panepistimiopolis 15784, Athens, Greece

and

E. A. Santas

Ministry of Interior, Administration and Decentralization, Region of Attica,

60 Theras Str., 11262 Athens, Greece

Submitted by C. T. Leondes

Received July 31, 1998

In this paper, the certainty equivalence principle is used to combine the

identification method with a control structure derived from the pole placement
problem, which rely on periodic multirate-input controllers. The proposed adaptive
pole placers, contain a sampling mechanism with different sampling period to each
system input and rely on a periodically varying controller which suitably modulates
the sampled outputs and reference signals of the plant under control. Such a
control strategy allows us to arbitrarily assign the poles of the sampled closed-loop
system in desired locations and does not make assumptions on the plant other than
controllability and observability of the continuous and the sampled system, and the
knowledge of a set of structural indices, namely the locally minimum controllability
indices of the continuous-time plant. An indirect adaptive control scheme is
derived, which estimates the unknown plant parameters Žand consequently the
controller parameters. on-line, from sequential data of the input and outputs of
the plant, which are recursively updated within the time limit imposed by a
fundamental sampling period T0 . Using the proposed algorithm, the controller
determination is based on the transformation of the discrete analogous of the

464
0022-247Xr99 $30.00
Copyright 䊚 1999 by Academic Press
All rights of reproduction in any form reserved.
 ADAPTIVE POLE POSITIONING 465

system under control to a phase-variable canonical form, prior to the application of

the control design procedure. The solution of the problem can, then, be obtained
by a quite simple utilization of the concept of state similarity transformation.
Known indirect adaptive pole placement schemes usually resort to the computation
of dynamic controllers through the solution of a polynomial Diophantine equation,
thus introducing high order exogenous dynamics in the control loop. Moreover, in
many cases, the solution of the Diophantine equation for a desired set of closed-loop
eigenvalues might yield an unstable controller, and the overall adaptive pole
placement scheme is unstable with unstable compensators because their outputs
are unbounded. The proposed control strategy avoids these problems, since here
gain controllers are essentially needed to be designed. Moreover, persistency of
excitation and, therefore, parameter convergence, of the continuous-time plant is
provided without making any assumption either the richness of the reference
signals or on the existence of specific convex sets in which the estimated parame-
ters belong or, finally, on the coprimeness of the polynomials describing the
ARMA model, as in known adaptive pole placement schemes. 䊚 1999 Academic
Press

1. INTRODUCTION

Multirate sampling schemes have long been the focus of interest by

many control designers. There are several reasons to use such a sampling
scheme in digital control systems. First of all, in complex, multivariable
control systems, often it is unrealistic, or sometimes impossible, to sample
all physical signals uniformly at one single rate. In such situations, one is
forced to use multirate sampling. Furthermore, in general, one gets better
performance if one can sample and hold faster. But faster ArD and DrA
conversions mean higher cost in implementation. For signals with different
bandwidths, better tradeoffs between performance and implementation
cost can be obtained using ArD and DrA converters at different rates.
On the other hand, multirate controllers are in general time-varying. Thus
multirate control systems can achieve what singlerate cannot; e.g., gain
improvement, simultaneous stabilization, and decentralized control. Fi-
nally, multirate controllers are normally more complex than singlerate
ones; but often they are finite-dimensional and periodic in a certain sense
and hence can be implemented on microprocessors via difference equa-
tions with finitely many coefficients. Therefore, like singlerate controllers,
multirate controllers do not violate the finite memory constraint in micro-
processors.
The study of multirate systems has its origins in late 1950s w1x ᎐ w3x.
Recent interests are focused on stability issues w4x, stabilization and pole
assignment w5x ᎐ w8x, LQGrLQR designs w9x ᎐ w12x, H ⬁ control w13x ᎐ w15x,
decentralized control w16x, adaptive designs w17x, w18x, etc. w19x, w20x. In
particular, in their excellent work w5x, Araki and Hagiwara propose a
466 ARVANITIS, KALOGEROPOULOS, AND SANTAS

digital multirate-input controller ŽMRIC., which suitably modulates the

sampled outputs and discrete reference signals by a multirate periodically
varying matrix function, in order to solve the sampled pole placement
problem for linear time-invariant continuous-time systems. MRICs contain
a multirate sampling mechanism with different sampling period to each
system input. They can essentially be viewed as the special class of
m-input, p-output multirate sampled-data control systems, in which all
output samplers operate with multiplicities 1 and the input samplers with
multiplicities  N1 , . . . , Nm 4 . Note that MRICs are the dual of multirate-out-
put controllers ŽMROCs., presented in w6x, and subsequently used in w20x,
in which input and output samplers have the reverse operation. In w19x, the
MRIC based approach has been extended to the solution of the model
matching problem. A main feature of the results reported in w5x and w19x is
that the pole placement or the model matching is obtained without the
requirement of pole-zero cancellation.
The purpose of the present paper is to explore the possibility of
extending the MRIC based approach presented in w5x and subsequently
used in w19x, to the control of linear time-invariant continuous-time plants
with unknown parameters. In particular, we use the certainty equivalence
principle to combine the identification method with a control structure
derived from the pole placement problem. It is worth noticing at this point
that, although the inputs of the continuous-time plane are sampled in a
multirate fashion, our aim here is to achieve adaptive pole placement
control, only at the sampling instants kT0 , associated with the fundamental
period T0 , on the basis of which the output samplers operate. To the best
of the authors’ knowledge, there are no results in the literature concerning
the use of this type of multirate sampled-data controllers in order to
achieve adaptive pole placement control.
Adaptive pole placement is of particular interest, since the middle of the
1970s, for obvious reasons. Several techniques based on either direct or
indirect adaptive control schemes were presented to treat the problem and
a very large number of papers were reported on the subject; see w21x ᎐ w29x,
and references therein. The feedback strategies proposed to solve the
adaptive pole placement problem, are hitherto based on dynamic output
feedback, thus introducing high order exogenous dynamics in the control
loop. On the other hand, a common feature of these techniques is that
they reduce the solution of the problem to the solution of a polynomial
Diophantine equation. This approach, however, does not ensure that the
compensators obtained from the solution of the Diophantine equation are
necessarily stable. In the case of unstable solutions, the control scheme
composed by feedforward and feedback compensators is not stable and
thus is not useful. The control signals are calculated from two sets of
unbounded signals that are the outputs of the compensators. In a short
 ADAPTIVE POLE POSITIONING 467

time the system becomes unstable. It is worth noticing at this point, that
unstable solutions of the Diophantine equation, can occur even though,
the system under control possesses the parity interlacing property Žp.i.p..
w30x Žis strongly stabilizable .. A plant is said that it possesses the p.i.p. if
the number of its real poles between each pair of zeros in the unstable
domain is even. In this case, it is possible to obtain a stable controller from
these unstable solutions by using the approach presented in w31x, which is
based on an interpolation procedure. Unfortunately, as mentioned above,
this approach can be applied only in cases where the system under control
is strongly stabilizable. When the system under control contains unknown
parameters Žas in the case of adaptive pole placement control., this
information of crucial importance is not available to the designer. Thus, up
to now, the design of a stable and useful adaptive pole placement compen-
sator cannot be guaranteed.
The motivation for studying an adaptive version of the particular con-
troller structure presented in w5x and w19x, is manifold. First, since it does
not rely on pole-zero cancellation, it may be readily applicable for solving
the adaptive pole placement problem for nonstably invertible plants.
Furthermore, the degrees of freedom in the choice of the modulating
function, provide a solution to the problem of assuring persistency of
excitation of the continuous-time plant under control, without imposing
any special assumption either on the existence of special convex sets in
which the estimated parameters belong or on the coprimeness of the
polynomials describing the ARMA model, as in known techniques, or
finally on the richness of the reference signals Žexcept boundedness., as in
known adaptive pole placement techniques. The determination of the
MRIC based adaptive pole placers sought is mainly based on the transfor-
mation of the discrete analogue of the continuous-time system under
control to a phase variable canonical phorm, prior to the application of the
control design procedure. As a consequence of this fact, the solution of the
problem can be obtained by a quite simple utilization of the concept of
state similarity transformation. No Diophantine equation is needed to be
solved here as compared to known techniques. The designed MRIC based
adaptive pole placers are always stable, since gain controllers are needed
to be designed here, as compared to Žpossibly unstable . dynamic compen-
sators obtained by known techniques. Therefore, the proposed adaptive
scheme is readily applicable to plants which do not possess the p.i.p. As a
consequence of this design philosophy, a useful globally stable indirect
adaptive control scheme is derived, which estimates the unknown plant
parameters Žand consequently the controller parameters. on-line, from
sequential data of the inputs and the outputs of the plant, which are
recursively updated within the time limit imposed by a fundamental
sampling period T0 . Finally, it is remarked that the a priori knowledge
468 ARVANITIS, KALOGEROPOULOS, AND SANTAS

needed in order to implement the proposed adaptive pole placers is

controllability and observability of the continuous and the discretized plant
under control, its order, and a set of structural indices, namely the locally
minimum controllability indices of the continuous-time plant.

2. PRELIMINARIES AND DEFINITION OF THE PROBLEM

Consider the linear multi-input, multioutput system having the following

state-space representation:

˙x Ž t . s Ax Ž t . q Bu Ž t . , y Ž t . s Cx Ž t . Ž 2.1.

where xŽ t . g R n is the state vector, uŽ t . g R m is the input vector, and

yŽ t . g R p is the output vector and where the matrices A, B, and C have
appropriate dimensions.
With regard to system Ž2.1., we make the following two assumptions:
ASSUMPTION 2.1. Ža. System Ž2.1. is controllable and obser¨ able and of
known order n. Žb. There are known integers n i , i g Jm , Jm s  1, 2, . . . , m4 ,
which comprise a set of locally minimum controllability indices of the pair
ŽA, B..
ASSUMPTION 2.2. Let Ni , i g Jm be positi¨ e integers. Also let N s
lcm N1 , . . . , Nm 4 , where lcmU , . . . , U 4 denotes the least common multiplier of
the arguments quoted in the braces. Then, there is a sampling period T0 g Rq,
such that the discretized systems, obtained by sampling Ž2.1. with periods T0
and ␶ s T0rŽ6 n y 1. N and ha¨ ing the following matrix triplets:

˜ C . ' exp Ž AT0 . , HT0exp Ž A ␭ . B d ␭ , C ,

Ž ⌽ , B, ž 0
/
␶
Ž ⌽␶ , B␶ , C . ' exp Ž A␶ . , H exp Ž A ␭ . B d ␭ , C ,
ž /
0

respecti¨ ely, are controllable and obser¨ able.

Except for this prior information, the matrix triplet ŽA, B, C. is arbitrary
and unknown. It is mentioned that no assumption is made here on the
relative degree of the plant or its stable invertibility.
For a controllable matrix pair ŽA, B., with B sˆ wb 1 b 2 ⭈⭈⭈ b m x, its locally
minimum controllability indices ŽLMCIs. are a collection of m integers
 ADAPTIVE POLE POSITIONING 469

 n1 ,n 2 , . . . , n m 4 , for which the following relationships simultaneously hold:

m
Ý ni s n and
is1

rank b i ⭈⭈⭈ An1y1 b 1 ⭈⭈⭈ bm ⭈⭈⭈ An m y1 b m s n

Note that, LMCI defined as above are also known as the ‘‘Kronecker
invariants’’ or ‘‘Kronecker indexes’’ of the pair ŽA, B..
Consider now applying to system Ž2.1., the multirate control strategy
depicted in Figure 1. With regard to the sampling mechanism, we assume
that all samplers start simultaneously at t s 0. The sampling periods Ti
have rational ratio, i.e., Ti s T0rNi , for i g Jm , where T0 is the common
sampling period, Ni g Zq are the input multiplicities of the sampling. The
hold circuits Hi and H0 are the zero order holds with holding times Ti and
T0 , respectively. Let

m
NU s Ý Ni , l i s NrNi , TN s T0rN
is1

The modulating matrix function FŽ t . g R m= p is assumed to be bounded,

integrable and T0-periodic, i.e.,

F Ž t q T0 . s F Ž t . for t g kT0 , Ž k q 1 . T0 . Ž 2.2.

FIG. 1. Control strategy in the nonadaptive case.

470 ARVANITIS, KALOGEROPOULOS, AND SANTAS

As it can be easily shown, the resulting closed-loop system is described by

the following state-space equations:

x Ž k q 1 . T0 s Ž ⌽ y K f C . x Ž kT0 . q K f w Ž kT0 . ,
y Ž kT0 . s Cx Ž kT0 . , kG0

where xŽ kT0 . g R n is a discrete measurement vector obtained by sampling

xŽ t . with sampling period T0 and where the matrix K f g R n= p is defined
as

T0
Kfs H0 exp A Ž T0 y ␭ . BF Ž ␭ . d ␭ Ž 2.3.

The adaptive pole placement problem treated in the present paper is as

follows: Find a periodic controller FŽ t ., which when applied to system Ž2.1.
drives the poles of the resulting closed-loop system Žalso called the
closed-loop monodromy eigenvalues., to new desired values ␭ ˆ1 , ␭ˆ2 , . . . , ␭ˆn ,
where complex poles appear in conjugate pairs.
To solve the above problem, an indirect adaptive control scheme is
exhibited in the sequel. In particular, we first solve the pole placement
problem, namely, the assignment of the poles of the sampled system to the
prespecified values ␭ ˆ1 , ␭ˆ2 , . . . , ␭ˆn , using period MRICs, for known systems.
This is done in Section 3. Next, using these results, the pole placement
problem is solved for the configuration of Figure 2, wherein the periodic
controller FŽ t . is with prespecified periodic behavior and persistent excita-
tion signals are introduced in the control loop for future identification
purposes. This is done in Section 4. It is remarked that the motivation for
modifying the control strategy as in Figure 2, is that it facilitates the
derivation of the indirect adaptive control scheme sought, which is pre-
sented in Section 5. In Section 5, the global stability of the proposed
scheme is also studied.

3. SOLUTION OF THE POLE PLACEMENT PROBLEM

VIA MRICS FOR KNOWN SYSTEMS

The procedure for stabilization through pole placement using MRICs,

consists in finding a periodic controller FŽ t ., such that

det Ž zI y ⌽ q K f C . ' ˆ
pŽ z . Ž 3.1a.
 ADAPTIVE POLE POSITIONING 471

FIG. 2. Structure of the adaptive control system.

where
n
ˆp Ž z . s Ł Ž z y ␭ˆi . s
ˆ z n q aˆ1 z ny1 q ⭈⭈⭈ qaˆny1 z q aˆn Ž 3.1b.
is1

Since, detŽ zI y ⌽ q K f C. ' detŽ zI y ⌽ T q C T K Tf ., relation Ž3.1a. is

equivalent to the relation

det Ž zI y ⌽ T q C T K Tf . s ˆ
pŽ z . Ž 3.2.
Consider now the following fictitious discrete time system:

˜x Ž k q 1 . T0 s ⌽ T ˜x Ž kT0 . q C T ˜u Ž kT0 . Ž 3.3.

Clearly, the pole placement problem via MRIC based control, defined in
relation Ž3.2., is equivalent to the problem of choosing the matrix K Tf in
the state feedback control law

˜u Ž kT0 . s yK Tf ˜x Ž kT0 . Ž 3.4.

such that Ž3.2. is satisfied.
We start our analysis to this equivalent state feedback pole placement
problem by first transforming system Ž3.3. to its equivalent input Luen-
berger canonical form. To this end, let ␦ i , i s 1, 2, . . . , m be the controlla-
bility indices of the pair Ž ⌽ T , C T . Žwhich obviously are the observability
472 ARVANITIS, KALOGEROPOULOS, AND SANTAS

indices of the pair Ž ⌽, C.. and let P g R n= n be the following matrix:

␦ 1 y1 T ␦ p y1 T
P s c T1 ⭈⭈⭈ Ž ⌽T . c1 ⭈⭈⭈ c Tp ⭈⭈⭈ Ž ⌽T . cp
where c Ti , i s 1, 2, . . . , p are the ordered columns of C T. Setting
j
␥j s Ý ␦␳ , j s 1, 2, . . . , p
␳s1

and defining hTj as the ␥ j th row of Py1 , it can be shown that under the
transformation ˜ zŽ kT0 . s Qx
˜Ž kT0 ., where Q g R n= n is the columnar stack
of ␦ 1 q . . . q␦ p Žs n. rows, defined by
hT1
..
.
␦ 1 y1
hT1 Ž ⌽ T .
..
Qs .
hTp
..
.
␦ p y1
hTp Ž ⌽ T .

system Ž3.3. can be written as

˜z Ž k q 1 . T0 s ⌽U ˜z Ž kT0 . q CU ˜u Ž kT0 . Ž 3.5.
where
⌽U s Q⌽ T Qy1 , CU s QC T
U U
and where the matrices ⌽ and C have the following respective forms:
⌽U11 ⭈⭈⭈ ⌽U1 p CU1
.. .. .. .
⌽U s . . . , CU s ..
⌽p1U
⭈⭈⭈ ⌽pUp CUp
where
0␦ iy1 I ␦ iy1
⌽Uii s g R ␦ i= ␦ i ,
yaTii

OŽ ␦ iy1.= ␦ j
⌽iTj s g R ␦ i= ␦ i Ž i / j. ,
yaTi j

OŽ ␦ iy1.=p
CUi s g R ␦ i=p
qTi
 ADAPTIVE POLE POSITIONING 473

where
aTii s Ž a ii . 0 Ž aii . 1 ⭈⭈⭈ Ž aii . ␦ iy1 ,

aTi j s Ž a i j . 0 Ž ai j . 1 ⭈⭈⭈ Ž ai j . ␦ jy1 Ž i / j.

qTi s 0 Tiy1 1 ˜qTi , ˜qTi s Ž c ii . iq1 Ž c ii . iq2 ⭈⭈⭈ Ž c ii . p

Here, 0 r , Or=q , and I r represent a zero r-dimensional vector, a zero
r = q matrix, and an r-dimensional identity matrix, respectively Žempty if r
or q is zero..
Now, let ¨˜Ž kT0 . be the set of inputs defined as follows:
˜u Ž kT0 . s L¨˜Ž kT0 .
where L is the following upper triangular nonsingular matrix:

1 q̃T1
0 1 q̃T2
..
Ls .
0 0 ⭈⭈⭈ 0 1 q̃Tpy 1
0 0 ⭈⭈⭈ 0 1

It is now obvious that

C̃U1
..
CU L s QC T L ' C
˜U s .
C̃Um
where
OŽ ␦ iy1.=p
C̃Ui s
0 Tiy1 1 0 Tpyi

and that system Ž3.5. can be transformed to the following form:

˜z Ž k q 1 . T0 s ⌽U ˜z Ž kT0 . q C
˜U ¨˜Ž kT0 . Ž 3.6.
which is the input Luenberger canonical form corresponding to system
Ž3.3.. In what follows, to system Ž3.6., we apply the following state feedback
law:
¨˜Ž kT0 . s FU ˜
z Ž kT0 . Ž 3.7.
474 ARVANITIS, KALOGEROPOULOS, AND SANTAS

in order to derive the eigenvalues of system Ž3.6. to desired positions

ˆ1 , ␭ˆ2 , . . . , ␭ˆn . Clearly, this is equivalent to the application of a state
␭
feedback law of the form Ž3.4., with
K Tf s yLFU Q Ž 3.8.
ˆi ,
to system Ž3.3., in order to drive its eigenvalues to the desired positions ␭
i s 1, 2, . . . , n.
From the above analysis, it is clear that in order to solve the pole
placement problem for system Ž3.3., under the control law Ž3.4., one can
equivalently solve the pole placement problem for system Ž3.6., under the
control law Ž3.7.. The solution of this later problem can be obtained as
follows: Observe first that the solution of this problem is equivalent to the
problem of selecting FU and a nonsingular transformation matrix T such
that
⌽U q C
˜U FU s T ⌸ Ty1 Ž 3.9.
where

ˆ1 , . . . , ␭ˆn
diag ␭ž / if the desired eigenvalues are distinct
⌸s
½ blockdiag Ž J1 , . . . , Js . if the desired eigenvalues are repeated
Ž 3.10.
with
ˆq
␭ 1 0 ⭈⭈⭈ 0
0 ˆq
␭ 1 ⭈⭈⭈ 0
Jq s 0 0 ˆq
␭ ⭈⭈⭈ 0 g R r q=r q Ž 3.11.
.. .. .. .. 1
. . . .
0 0 0 ⭈⭈⭈ ˆq
␭

and where the order r q of the qth Jordan block Jq is the multiplicity of the
ˆq .
eigenvalue ␭
To solve Ž3.9. for FU and T, partition the matrices ⌽Uii and ⌽Ui j as

ˆ Uii
⌽ ˆ Ui j
⌽
⌽Ui i s , ⌽Ui j s Ž i / j.
yaTii yaTi j

where
ˆ Ui i s 0␦ iy1
⌽ I ␦ iy1 , ˆ Ui j s OŽ ␦ 1y1.= ␦ j
⌽ Ž i / j.
 ADAPTIVE POLE POSITIONING 475

define the matrices

T11 ⭈⭈⭈ T1 p
.. .. .. T̂i j
Ts . . . , Ti j s ,
t̃ Ti j
Tp1 ⭈⭈⭈ Tp p

t Ti jl
..
ˆi j s
T . , ˜t Ti j s t Ti j, ␦ i
t Ti j, ␦ iy1

˜Ui as
partition C

C̃q
C̃Ui s
i

c̃ Ti

where

˜qi s OŽ ␦ iy1.=p ,
C ˜c Ti s 0 Tiy1 1 0 Tpy1

and define
⌸ s blockdiag  ⌸ 1 ⭈⭈⭈ ⌸ p4

where each matrix Ti j has the dimensionality of ⌽Ui j , and each matrix ⌸ i
has the dimensionality of ⌽Uii , and may have one of the forms given in
Ž3.10. Žor its combination..
Next, define ⌽ˆU, C ˜ U , C,
˜q, ⌽ ˜ and T
˜ as

ˆU s
⌽ U
ž ⌽ˆ /
i j is1, . . . , p , ˜qs Ž C
C ˜qi . is1, . . . , p ,
js1, . . . , p
Ž 3.12a.
ˆs
T ž Tˆ /
i j is1, . . . , p js1, . . . , p

˜U s
⌽ Ž yaTi j . is1, . . . , p , ˜ s Ž˜c Ti . is1, . . . , p ,
C
js1, . . . , p
Ž 3.12b.
˜ s Ž˜t i j . is1, . . . , p
T is1, . . . , p

where parentheses define a column of blocks and brackets w ⴢ x define a row

of block columns, and apply a linear transformation upon Ž3.9. to obtain

ˆU
⌽ ˜q U
C ˆ
T
U
q F s ⌸ Ty1 Ž 3.13.
˜
⌽ ˜
C ˜
T
476 ARVANITIS, KALOGEROPOULOS, AND SANTAS

˜qs OŽ nyp.=p and C

where it is noted that C ˜ s I p . From Ž3.13. we obtain
ˆU T s T
⌽ ˆ⌸ Ž 3.14.
FU s y⌽ ˜U q T ˜ ⌸ Ty1 Ž 3.15.
Equations 3.14 and 3.15 show that the problem of determining FU and
Ž . Ž .
T has been decoupled, i.e., one first finds T from Ž3.14. and then FU from
Ž3.15..
To find T, observe that
ˆ U11T11
⌽ ˆ 11T12
⌽ ⭈⭈⭈ ˆ U11T1 p
⌽
ˆ U22 T21
⌽ ˆ U22 T22
⌽ ⭈⭈⭈ ˆ U22 T2 p
⌽
⌽̂U T s .. .. .. ..
. . . .
U
ˆ p pTp1 U
ˆ p pTp2 U
ˆ p pTp p
⌽ ⌽ ⭈⭈⭈ ⌽
Ž 3.16.
T11 ⌸ 1 ˆ12 ⌸ 2
T ⭈⭈⭈ ˆ1 p ⌸ p
T
ˆ 21 ⌸ 1
T ˆ 22 ⌸ 2
T ⭈⭈⭈ ˆ2 p ⌸ p
T
T̂ ⌸ s .. .. .. ..
. . . .
ˆ p1 ⌸ 1
T ˆ p2 ⌸ 2
T ⭈⭈⭈ ˆp p ⌸ p
T

Hence, Ž3.14. reduces to

ˆ Ui j Ti j s T
⌽ ˆi j ⌸ j Ž i , j s 1, . . . , p . Ž 3.17.
As it can be shown, the solution of Ž3.17. with regard to Ti j has the form
␳ iTj
␳ iTj ⌸ j
Ti j s .. ˆ Ž ␳ iTj ⌸ kj . ks0, . . . , ␦ iy1
s Ž 3.18.
.
␳ i j ⌸ ␦j iy1
T

where ␳ iTj is a ␦ j-dimensional row vector with arbitrary elements for all
i, j s 1, 2, . . . , p. The general form of T will be
Ž ␳ 11
T
⌸ 1k . ks0 , . . . , ␦ 1y1 Ž ␳ 12
T
⌸ 2k . ks0 , . . . , ␦ 1y1 ⭈⭈⭈ Ž ␳ 1Tp ⌸ kp . ks0 , . . . , ␦ y1
1

Ž ␳ 21
T
⌸ lk . ks0 , . . . , ␦ 2y1 Ž ␳ 22
T
⌸ 2k . k s 0, . . . , ␦ 2 y 1 ⭈⭈⭈ Ž ␳ 2Tp ⌸ kp . ks0 , . . . , ␦ y1
2
Ts .. .. .. ..
. . . .
Ž ␳ p1
T
⌸ lk . ks0 , . . . , ␦ y1
p
Ž ␳ p2
T
⌸ 2k . ks0 , . . . , ␦ y1
p
⭈⭈⭈ Ž ␳ pTp ⌸ kp . ks0 , . . . , ␦ y1
p

Ž 3.19.
 ADAPTIVE POLE POSITIONING 477

In Ž3.19., all elements of the first row of each block of T are arbitrary and
hence we have a total number of arbitrary elements in T equal to n = p.
Note also that this arbitrariness is constrained by the requirement that T
must be invertible, i.e., det T / 0.
To find FU , observe that relations Ž3.12., Ž3.18., and Ž3.19. yield

␳ 11
T
⌸ 1␦ 1y1 ␳ 12
T
⌸ 2␦ 1y1 ⭈⭈⭈ ␳ 1Tp ⌸ ␦p 1y1
␳ 21
T
⌸ 1␦ 2y1 ␳ 22
T
⌸ 2␦ 2y1 ⭈⭈⭈ ␳ 2T p ⌸ ␦p 2y1
T̃ s .. .. .. .. Ž 3.20.
. . . .
␳ p1
T
⌸ 1␦ py1 ␳ p2
T
⌸ 2␦ py1 ⭈⭈⭈ ␳ pTp ⌸ ␦p py1

On the basis of Ž3.20., relation Ž3.15. yields

FU s y⌽
˜ U q R ⌸U Ty1 Ž 3.21.
⌸ ␦1
U ..
R s blockdiag  ␳ 1T , . . . , ␳ pT 4, ⌸ s . Ž 3.22.
␦p
⌸

with ␳ iT s w ␳ i1T ⭈⭈⭈ ␳ pT x. Note that, when ⌸ j is in Jordan form, Ti j can

take the form

Ž ␳iTj . 1 Ž ␳iTj . 2 ⭈⭈⭈ Ž ␳iTj . s j

Ž ␳iTj . 1 ŽJj . 1 Ž ␳iTj . 2 ŽJj . 2 ⭈⭈⭈ Ž ␳iTj . s ŽJj . s

j j
Ti j s .. .. .. .. Ž 3.23.
. . . .
Ž ␳iTj . 1 ŽJj . 1␦ y1 Ž ␳iTj . 2 ŽJj . 2␦ y1
i i
⭈⭈⭈ Ž ␳iTj . s ŽJj . ␦s y1
j j
i

⌸ j s blockdiag Ž J j . 1 , . . . , Ž J j . s j ,
½ 5 ␳ iTj s Ž ␳iTj . 1 ⭈⭈⭈ Ž ␳iTj . s j

in which Ž ␳ iTj .q is a row vector of dimensionality equal to that of ŽJ j .q . In

ˆj .q , then s j s ␦ j , Ž ␳ iTj .q s ␳ i jq ,
particular if ŽJ j .q s Ž ␭

␳ i j1 ␳ i j2 ⭈⭈⭈ ␳i j␦ j

␳ i j1 ␭ ˆj
ž / ␳ i j2 ␭ ˆj
ž / ⭈⭈⭈ ␳i j␦ j ␭
ž /ˆj sj
1 2
Ti j s .. .. .. .. Ž 3.24.
. . . .
␦ iy1 ␦ iy1 ␦ iy1
ˆj
␳ i j1 ␭ ž / ˆj
␳ i j2 ␭
ž / ⭈⭈⭈ ˆj
␳i j␦ j ␭
ž / ␦j
1 2
478 ARVANITIS, KALOGEROPOULOS, AND SANTAS

It is remarked that, if we choose all the arbitrary elements ␳ i j , for i / j

equal to 0, and all elements of each ␳ i j equal to 1, relation Ž3.21. can be
written as

FU s y⌽
˜ U q R s p ⌸U Tsy1
p Ž 3.25.
In this case, the open-loop poles contained in the subsystem determined by
⌽Ui j , when closing the feedback, are shifted to the desired poles involved in
the corresponding block ⌸ i .
In order to determine the matrix K f , substitute relation Ž3.21. in Ž3.8. to
yield
T
˜ U y R s p ⌸U Tsy1
K f s QT ⌽ ž p / LT Ž 3.26.

Using the matrix K f as specified by Ž3.26., we can readily determine the

controller matrix FŽ t ., by solving Ž2.3.. Under Assumption 2.1, on the
controllability of the pair ŽA, B., a solution of Ž2.3. is the following:

F Ž t . s BT exp AT Ž T0 y t . Wy1 Ž A, B, T0 . K f Ž 3.27.

where WŽA, B, T0 . is the controllability Grammian on w0, T0 x of the pair
ŽA, B., which has the form

T0
W Ž A, B, T0 . s H0 exp A Ž T0 y ␭ . BBT exp AT Ž T0 y ␭ . d ␭

Note that the controllability Grammian WŽA, B, T0 . is nonsingular and

hence a solution of Ž2.3. of the form Ž3.27. exists if the pair ŽA, B. is
controllable.
On the basis of Ž3.26. and Ž3.27., a solution of the pole placement
problem using MRIC based control is given by
T
ž˜ U y R s p ⌸U Tsy1
F Ž t . s BT exp AT Ž T0 y t . Wy1 Ž A, B, T0 . Q T ⌽ p / LT
Ž 3.28.

4. A SOLUTION OF THE POLE PLACEMENT PROBLEM

APPROPRIATE FOR THE ADAPTIVE CASE

In order to obtain a solution of the pole placement problem which will

be more appropriate for application in the case of systems with unknown
parameters, we slightly modify in the sequel the control strategy of Figure
1 as it is depicted in Figure 2. In particular, we focus our attention on the
 ADAPTIVE POLE POSITIONING 479

special class of the time-varying T0-periodic functions FŽ t ., for which every

element of FŽ t ., denoted by f i j Ž t ., is piecewise constant over intervals of
length Ti , i.e.,

f i j Ž t . s f i j, ␮ , ᭙ t g ␮Ti , Ž ␮ q 1 . Ti . , ␮ s 0, 1, . . . , Ni y 1 Ž 4.1.

Moreover, the persistent excitation signals ¨ i Ž t ., ᭙ i g Jm , are defined as

¨ i Ž t . s d Ti Ž t . vi , d Ti Ž t . s Ž d i . 0 Ž t . ⭈⭈⭈ Ž d i . N iy1 Ž t . Ž 4.2a.

Here, d i Ž t . is the Ti-periodic vector function with elements having the
form

Ž d i . q Ž t . s Ž d i . q , ␮ for t g ␮Ti , Ž ␮ q 1 . Ti . ,
q s 0, 1, . . . , Ni y 1, ␮ s 0, 1, . . . , Ni y 1 Ž 4.2b.
where Ž d i .q, ␮ are constant taking the following values:

1 for ␮ s q,
Ž di . q , ␮ s ½ 0 for ␮ / q
Ž 4.3.

and where vi is as yet unknown. It is worth noticing that the additive term
¨ i Ž t . s d Ti Ž t .vi , ᭙ i g Jm in each one of the inputs of the continuous-time
system are used only for identification purposes and as it will be shown
later, they are selected so that they will not influence the pole placement
problem.
We are now able to establish the following Lemma.
LEMMA 4.1. Consider the controllable and obser¨ able system of the form
Ž2.1., controlled by MRICs of the form Ž4.1.. Furthermore, consider that
persistent excitation signals of the forms Ž4.2. and Ž4.3. are introduced in each
input of the system. Then, the sample closed-loop system takes the form

ˆˆ Ž kT0 . q BU v,
ˆˆ . x Ž kT0 . q BFw
x Ž k q 1 . T0 s Ž ⌽ y BFC
Ž 4.4.
y Ž kT0 . s Cx Ž kT0 . for k G 0
where

ˆs
B ˆ ˆb 1 ⭈⭈⭈ ˆN1 1y1ˆb 1
A ⭈⭈⭈ ˆb m ⭈⭈⭈ ˆ Nmm y1ˆb m
A Ž 4.5a.

 i s
ˆ exp Ž ATi . ' exp Ž A l i TN . ,
Ti l iTN Ž 4.5b.
b̂ i s
ˆ H0 exp Ž A ␭ . b i d ␭ ' H0 exp Ž A ␭ . b i d ␭
480 ARVANITIS, KALOGEROPOULOS, AND SANTAS

M1 e␴ j
M2 e ␴ jy1
BU s B
ˆ M, Ms .. , Mj s .. Ž 4.5c.
. .
Mm e ␴ jyN jq1

U
ˆ and the column ¨ ector v g R N
and where the m = p block matrix F ha¨ e
the forms

ˆf 11 ⭈⭈⭈ ˆf 1 p f i j, N iy1
.. .. .. ..
ˆs
F . . . , ˆf i j s . Ž 4.6a.
ˆf m1 ⭈⭈⭈ ˆf m p f i j, 0

T
v s v1T v2T ⭈⭈⭈ vmT Ž 4.6b.
U
while ␴j s Ý␬j s1 N␬ , where in general, the ¨ ector e i g R N is the row vector
whose elements are zeros except for a unity appearing in the ith position.
Proof. To show that the sampled closed-loop system takes the form
Ž4.4., we start by discretizing system Ž2.1. with sampling period T0 . This
operation yields

HkTkq1 T exp A Ž k q 1. T
Ž .
x Ž k q 1 . T0 s ⌽ x Ž kT0 . q 0
0 y ␭ 4 Bu Ž ␭ . d ␭
0

Ž 4.7.

Observing that u i Ž t . s ri Ž t . q d Ti Ž t .vi and taking into account the struc-

ture of the control system in Figure 2, we obtain

u i Ž t . s f iT Ž t . e Ž kT0 . q d Ti Ž t . vi for t g ␮Ti , Ž ␮ q 1 . Ti . Ž 4.8a .

where f iT Ž t . is the ith row of the controller matrix FŽ t . and eŽ kT0 . is given
by

e Ž kT0 . s w Ž kT0 . y y Ž kT0 . s w Ž kT0 . y Cx Ž kT0 . Ž 4.8b.

Combining relations Ž4.7. and Ž4.8., we obtain the following relationship:

x Ž k q 1 . T0 s Ž ⌽ y K f C . x Ž kT0 . q K f w Ž kT0 . q ⌫ v Ž 4.9.

 ADAPTIVE POLE POSITIONING 481

where

HkTkq1 T exp A Ž k q 1. T
Ž .
⌫s 0
0 y ␭ 4 BD Ž ␭ . d ␭ ,
0

DŽ t . s
ˆ blockdiag  d Ti Ž t . 4
igJ m

Now, partition ⌫ as follows:

⌫ s ⌫1 ⌫2 ⭈⭈⭈ ⌫m

Then, the Ž q q 1.th column of the matrix ⌫i , for i g Jm , denoted by

Ž ⌫i .qq1 for q s 0, 1, . . . , Ni y 1, can be expressed as

T
Ž ⌫i . qq 1 s H exp A Ž T0 y ␭ . b i Ž d i . q Ž ␭ . d ␭ for q s 0, 1, . . . , Ni y 1
0

0
Ž 4.10.

Introducing relations Ž4.2a. and Ž4.3. in Ž4.10., we obtain

Niy1

H␮T␮q1 T exp A Ž T
Ž .
Ž ⌫i . qq 1 s Ý i
0 y ␭. b i Ž di . q , ␮ d ␭
␮s0 i

for q s 0, 1, . . . , Ni y 1 Ž 4.11.
Relation Ž4.11. may further be written as
Niy1
Ti
Ž ⌫i . qq 1 s Ý Ž d i . q , ␮ exp A Ž Ni y ␮ y 1. Ti H0 exp A Ž Ti y ␭ . b i d ␭
␮ s0

Ni
s
ž Ý Ž d i . q , N y ␳ Aˆ i␳y1 ˆb i
i
/
␳s1

Making use of relation Ž4.3., we arrive at the following relationship:

ˆ Ni iyq y1ˆb i
Ž ⌫i . qq 1 s A
Clearly ⌫ ' BU . Application of the above algorithm to the first term of
ˆˆ Žsee w5x for details.. This completes the proof of the
Ž4.9. yields K f ' BF
Lemma.
Thus far, we have established that the pole placement controller matrix
K f is related to the matrix Fˆ via the relation K f s BF.
ˆˆ It remains to
482 ARVANITIS, KALOGEROPOULOS, AND SANTAS

determine F.ˆ To this end, we need the following result, whose proof is
given in w5x.
LEMMA 4.2. Let ŽA, B. be a controllable pair. Let also n i , i g Jm be a set
of locally minimum controllability indices of the pair ŽA, B.. Define an
analytic function ␺ ŽTN . by

␺ Ž TN . s det ˆb 1 ⭈⭈⭈ ˆn11y1ˆb 1

A ⭈⭈⭈ ˆb m ⭈⭈⭈ ˆnmm y1ˆb m
A

Then the set of zeros of ␺ ŽTN . does not ha¨ e any limiting points except
infinity, and therefore, ␺ ŽTN . is not equal to zero for almost all TN Ž i.e., in a
finite inter¨ al w TN1 , TN2 x, there are at most a finite number of points such that
␺ ŽTN . s 0..
ˆ of the form
Applying Lemma 4.2, we can conclude that the matrix S

ˆ s ˆb 1
S ⭈⭈⭈ ˆn11y1ˆb 1
A ⭈⭈⭈ ˆb m ⭈⭈⭈ ˆnmm y1ˆb m
A Ž 4.12.

is nonsingular for almost all TN g w TN1 , TN2 x. Furthermore, if the input

multiplicities of the sampling Ni are chosen such that Ni G n i , i g Jm
then, the matrices Bˆ and BU have full row rank n for almost all TN g
w TN , TN x.
1 2
U U
Now, let E g R N =N be the nonsingular permutation matrix with the
property Ey1 ' ET, having the form

T
E s E1 E2

where

E1 s w ␧ 1 ␧2 ⭈⭈⭈ ␧ n1 ␧ N1q1 ␧ N1q2 ⭈⭈⭈ ␧ N1qn 2 ⭈⭈⭈

=␧ N U yN m q1 ␧ N U yN m q2 ⭈⭈⭈ ␧ N U yN m qn m x

and

E 2 s w ␧ n1q1 ⭈⭈⭈ ␧ N1 ␧ N1qn 2q1 ⭈⭈⭈ ␧ N1qN 2 ⭈⭈⭈

=␧ N U yN m qn m q1 ⭈⭈⭈ ␧N U x
U
where, in general, ␧ j g R N is the column vector whose elements are zeros
except for a unity appearing in the jth position. Also, let

B噛 s ˆ y1 s Sˆ
ˆ BE ˆ
Q
 ADAPTIVE POLE POSITIONING 483

ˆ is defined by Ž4.13. and the matrix Q

where the matrix S ˆ is given by

ˆs A
Q ˆn11ˆb 1 ⭈⭈⭈ ˆN1 1y1ˆb 1
A ⭈⭈⭈ ˆnmmˆb m
A ⭈⭈⭈ ˆNmm y1ˆb m
A
U U
Furthermore, let ⌬ g R N =N be the nonsingular permutation matrix with
the property ⌬y1 ' ⌬T, having the form

T
⌬ s ⌬1 ⌬2 ⌬3

where

⌬ 1 s w ␧ N1yn1q1 ⭈⭈⭈ ␧ N1 ␧ N1qN 2yn 2q1 ⭈⭈⭈ ␧ N1qN 2 ⭈⭈⭈

=␧ N U yn m q1 ⭈⭈⭈ ␧N U x
⌬ 2 s w ␧ N1yn1 ␧ N1qN 2yn 2 ⭈⭈⭈ ␧ N U yn m x
⌬3 s w ␧1 ⭈⭈⭈ ␧ N1yn1y1 ␧ N1q1 ⭈⭈⭈ ␧ N1qN 2yn 2y1 ⭈⭈⭈

=␧ N U yN m q1 ⭈⭈⭈ ␧ N U yn m y1 x

Finally, let

B ˆ BU⌬y1 ' SˆU

˜噛 s ˆn11ˆb 1
A ⭈⭈⭈ ˆnmmˆb m
A ˆU
Q

where

ˆU s A
S ˆn11y1ˆb 1 ⭈⭈⭈ ˆb 1 ⭈⭈⭈ ˆnmm y1ˆb m
A ⭈⭈⭈ ˆb m Ž 4.13.
ˆU s A
Q ˆN1 1y1ˆb 1 ⭈⭈⭈ ˆn11q1ˆb 1
A ⭈⭈⭈ ˆ Nmm y1ˆb m
A ⭈⭈⭈ ˆnmm q1ˆb m
A

ˆ by mere inspec-
Using these definitions, it is plausible to determine F
tion, as

T
ˆy1 T ˜ U U y1
F̂ s ET S Q ⌽ y R s p ⌸ Ts p ž / LT
Ž 4.14.
0

It only remains to determine the appropriate vector v which guarantees

that the pole placement problem will not be dependent on the vector v. In
other words

v g ker BU or BU v s 0
484 ARVANITIS, KALOGEROPOULOS, AND SANTAS

An obvious selection of such v obtained also by inspection is the following:

ˆUy 1 A
yS ˆn11ˆb 1 q ⭈⭈⭈ A
ž ˆnmmˆb m /
vs⌬ T
␨ Ž 4.15.
0 N U ynym

where ␨ g R m is the column vector whose elements are all equal to 1.

It is noted that the N U-dimensional column vector v, even though does
not affects the discrete pole placement problem, it provides persistent
excitation useful for the consistent identification of the system, as will be
shown in the following section.
Clearly, the multirate controller matrix FŽ t . of Figure 2 can readily be
determined by making use of relations Ž4.1., Ž4.6a., Ž4.14., and Ž3.28.. More
precisely, the ith row f iT Ž t . of the matrix FŽ t . and the ith block row of the
matrix Fˆ are interrelated as

f iT Ž t . s f i1 Ž t . ⭈⭈⭈ fi p Ž t . s e N y␮ ˆ
f i1 ⭈⭈⭈ ˆf i p ,
i

␮ T0 Ž ␮ q 1 . T0
᭙ Ft- Ž 4.16.
Ni Ni

for i g Jm and for ␮ s 0, 1, . . . , Ni y 1, where e N iy ␮ g R N i is the row

vector defined as e N iy ␮ s ␧ NT iy ␮ . Note that the controller matrix FŽ t ., as
specified by Ž4.16., is largely affected by the multirate mechanism, while
the controller matrix FŽ t . as specified by relation Ž3.28. is not. Further-
more, the introduction of the excitation signals ¨ i Ž t . in the control loop
greatly facilitates the consistent estimation of the plant parameters in the
case of unknown systems. For these reasons, the control strategy of Figure
2 is more appropriate than the control strategy of Figure 1 for the
development of the indirect adaptive control scheme presented in the
following section.

5. CONTROL STRATEGY FOR THE ADAPTIVE CASE

The control scheme presented in Section 4 has a corresponding scheme

in the case where the system is unknown. For this case, the control strategy
is largely based on the computation of the matrix F ˆ and of the vector v
from estimates of the plant parameters, and results in a globally stable
closed-loop system whose poles are located to the prespecified values
ˆ1 , ␭ˆ2 , . . . , ␭ˆn .
␭
 ADAPTIVE POLE POSITIONING 485

5.1. Plant Parameters Estimation Algorithm

The algorithm proposed here for estimating the unknown plant parame-
ters is as follows: System Ž2.1., discretized with sampling period ␶ s
T0rŽ6 n y 1. N, takes the form

x Ž ␯ q 1 . ␶ s ⌽␶ x Ž ␯␶ . q B␶ u Ž ␯␶ . , y Ž ␯␶ . s Cx Ž ␯␶ . ,
␯G0 Ž 5.1.
where
␶
⌽␶ s exp Ž A␶ . , B␶ s H0 expŽ A ␭. B d ␭
Clearly, uŽ ␯␶ . takes constant values for ␯␶ g w ␳ TN , Ž ␳ q 1.TN ., ␳ G 0.
This can be easily shown by taking into account the action of the proposed
controller. Hence, iterating relation Ž5.1. 6 n y 1 times, we obtain

x Ž m q 1 . TN s ⌽ T N x Ž mTN . q BT N u Ž mTN . , mG0

where
6 ny2
⌽␶␳ B␶
6 ny1
⌽ T N s Ž ⌽␶ . , BT N s Ý Ž 5.2.
␳s0

Using the same argument, we can easily conclude that

l iy1
l
ˆ i s ⌽ Ti N ,
A ˆb i s Ý ⌽ T␳N Ž BTN . Ž 5.3.
i
␳s0

where ŽBT N . i is the ith column of the matrix BT N . Introducing relation Ž5.2.
in Ž5.3., yields
l iy1 Ž6 ny2 .
ˆ i s Ž ⌽␶ . Ž6 ny1. l i , ˆb i s Ý Ž ⌽␶ . Ž6 ny1. j ⌽␶␳ B␶
A
js0
žÝ ␳ s0
/ i
Ž 5.4.

Moreover, the matrix ⌽ can be written as

ˆNi i s ⌽ TNN s Ž ⌽␶ .
⌽sA
Ž6 ny1 . N
Ž 5.5.

Therefore, ⌽, A ˆ i and ˆb i Žwhich are the only matrices involved in

computing Fˆ and v. can be computed on the basis of ⌽␶ and B␶ . For this
reason, in what follows our aim will be the estimation of the matrix triplet
486 ARVANITIS, KALOGEROPOULOS, AND SANTAS

Ž ⌽␶ , B␶ , C.. To this end, let the matrix ⍀ be defined as

is1, 2, . . . , n
⍀ s  ⍀ i j 4 js1, 2, . . . , n , ⍀ i j s C⌽␶iqjy2 B␶ Ž 5.6.

Clearly, if one establishes estimates of the matrix ⍀, then one may

easily compute the desired matrix triplet Ž ⌽␶ , B␶ , C., using anyone of the
minimal realization algorithms reported in the literature Žsee, e.g., those
reported in w32x ᎐ w34x.. To estimate the matrix ⍀, one must resort to an
input᎐output representation Žalso called ARMA representation . of system
Ž5.1.. This representation is summarized in the following theorem:

THEOREM 5.1. Suppose that there is a sampling period T0 g Rq and

input multiplicities of the sampling Ni , i g Jm , such that system Ž5.1., ob-
tained by sampling the controllable and obser¨ able system Ž2.1., is also
controllable and obser¨ able. Then, an alternati¨ e representation of system Ž5.1.
is gi¨ en by

⌿ Ž ␯␶ . s J1⌿ Ž ␯ y 2 n . ␶ q J2W Ž ␯␶ . q VW Ž ␯ y n . ␶
q VU W Ž ␯ y 2 n. ␶ Ž 5.7.

where

y Ž ␯ y n q 1. ␶
y Ž ␯ y n q 2. ␶
⌿ Ž ␯␶ . s ..
.
y Ž ␯␶ .

y Ž ␯ y 3n q 1 . ␶
y Ž ␯ y 3n q 2 . ␶
⌿ Ž ␯ y 2 n. ␶ s .. Ž 5.8a.
.
y Ž ␯ y 2 n. ␶

u Ž ␯ y n q 1. ␶
u Ž ␯ y n q 2. ␶
W Ž ␯␶ . s ..
.
u Ž ␯␶ .
 ADAPTIVE POLE POSITIONING 487

u Ž ␯ y 2 n q 1. ␶
u Ž ␯ y 2 n q 2. ␶
W Ž ␯ y n. ␶ s .. ,
.
u Ž ␯ y n. ␶
Ž 5.8b.
u Ž ␯ y 3n q 1 . ␶
u Ž ␯ y 3n q 2 . ␶
W Ž ␯ y 2 n. ␶ s ..
.
u Ž ␯ y 2 n. ␶

0 ⭈⭈⭈ 0 0
CB␶ ⭈⭈⭈ 0 0
ˆ
J1 s ⌶Uy 1 J 0 ⌶U , J2 s .. .. .. .. ,
0 0 . . . . Ž 5.8c.
C⌽␶ny 2 B␶ ⭈⭈⭈ CB␶ 0

V s PU ⌺ and V U s ⌶Uy1 Vq 0

and where

C
C⌽␶
Ĵ s P1U ⌽␶2 n P1Uy1 , PU s .. ,
. Ž 5.9a.
C⌽␶ny 1

⌺ s ⌽␶ny 1 B␶ ⭈⭈⭈ ⌽␶ B␶ B␶

Vqs P1U ⌽␶n ⌺ y ⌽␶n PUy1 U1 Ž 5.9b.

while the nonsingular permutation matrix ⌶U g R n p=n p is such that

P1U
⌶U PU s Ž 5.10 .
0

where P1U g R n= n is the nonsingular matrix whose rows are the linearly
independent rows of the matrix PU . Finally, U1 g R n= n p is the matrix
containing the first n rows of the matrix

U s ⌶U J2 Ž 5.11.
Proof. In order to prove relation Ž5.7., we next generalize the approach
presented in w35x, to the multivariable case. More precisely, from relations
488 ARVANITIS, KALOGEROPOULOS, AND SANTAS

Ž5.1. we have

y Ž ␯ y n q 1 . ␶ s Cx Ž ␯ y n q 1 . ␶
y Ž ␯ y n q 2 . ␶ s C⌽␶ x Ž ␯ y n q 1 . ␶ q CB␶ u Ž ␯ y n q 1 . ␶
..
.
y Ž ␯␶ . s C⌽␶ny 1 x Ž ␯ y n q 1 . ␶
ny2
q Ý C⌽␶␳ B␶ u Ž ␯ y ␳ y 1. ␶
␳s0

or more compactly,
⌿ Ž ␯␶ . s PU x Ž ␯ y n q 1 . ␶ q J2W Ž ␯␶ . Ž 5.12.
U
where ⌿ Ž ␯␶ . and WŽ ␯␶ . are defined by Ž5.8a. and P and J2 are defined
by Ž5.9a. and Ž5.8c., respectively.
Since, by Assumption 2.2, the pair Ž ⌽␶ , C. is observable, the matrix PU
has full column rank. Hence, there exists a nonsingular permutation matrix
⌶U g R n p=n p , such that relation Ž5.10. holds, where, as already mentioned,
P1U g R n= n is the nonsingular matrix whose rows are the linearly indepen-
dent rows of the matrix PU . It is pointed out that matrix ⌶U can be defined
as a product of two nonsingular matrices ⌶ ˜ g R n p=n p and ⌶ˆ g R n p=n p
via the following chain of definitions:

e1
e2
.. e j1
. e j2
U en ..
˜ˆ
⌶ s ⌶⌶, ˜s
⌶ , ˆs
⌶ .
␻1
e jn
␻2
.. ⌶q
1
.
␻ n pyn

where ⌶q
1 gR
Ž n pyn.=n p
is the matrix produced by the nonsingular matrix
q
⌶ gR n p=n p
of the form
e1
e2
⌶qs ..
.
enp
 ADAPTIVE POLE POSITIONING 489

by dropping the row vectors e i , i s j1 , j2 , . . . , jn , where j1 , j2 , . . . , jn are the

indices of the n linearly independent rows of PU defined as pUj ␳T , ␳ s
1, 2, . . . , n, Note also that ␻ k g R n p , k s 1, 2, . . . , np y n is the column
vector of the form

␻ k s Ž ␭ j1 . k Ž ␭j . k
2
⭈⭈⭈ Ž ␭j . k
n
0 ⭈⭈⭈ 0 ^y1
`_ 0 ⭈⭈⭈ 0
Ž n q k . th
position

where Ž ␭ j ␳ . k , ␳ s 1, 2, . . . , n, k s 1, 2, . . . , np y n are the coefficients of

the following dependence relation holding for the rows of the matrix PU :
n
Ý Ž ␭ j . k pUj T y pUk T s 0,
␳ ␳
k f  j1 , j2 , . . . , jn 4
␳s1

where, pUk T , k f  j1 , j2 , . . . , jn4 is the kth row of the matrix PU .

Now, multiplying Ž5.12. from the left by ⌶U , yields

P1U
ZU Ž ␯␶ . s x Ž ␯ y n q 1 . ␶ q UW Ž ␯␶ .
0
where
ZU Ž ␯␶ . s ⌶U ⌿ Ž ␯␶ . Ž 5.13.
and where U is defined by Ž5.11.. Next, decompose ZU Ž ␯␶ . and U as
follows:

U
ZU1 Ž ␯␶ . U1
Z Ž ␯␶ . s , Us Ž 5.14.
ZU2 Ž ␯␶ . U2

where ZU1 Ž ␯␶ . g R n, ZU2 Ž ␯␶ . g R nŽ py1., U1 g R n=n p and U2 g R nŽ py1.=n p.

Clearly,

ZU1 Ž ␯␶ . s P1U␰ Ž ␯ y n q 1 . ␶ q U1W Ž ␯␶ . and ZU2 Ž ␯␶ . s U2W Ž ␯␶ .

Ž 5.15.
From Ž5.15., one may easily obtain the following relation:

x Ž ␯ y n q 1 . s P1Uy 1 ZU1 Ž ␯␶ . y U1W Ž ␯␶ . Ž 5.16.

Furthermore, as it can be easily shown, the following relationship holds:

x Ž ␯ y n q 1 . ␶ s ⌽␶2 n x Ž ␯ y 3n q 1 . ␶ q ⌽␶n ⌺W Ž ␯ y 2 n . ␶
q ⌺W Ž ␯ y n . ␶ Ž 5.17.
490 ARVANITIS, KALOGEROPOULOS, AND SANTAS

where WwŽ ␯ y n.␶ x and WwŽ ␯ y 2 n.␶ x are given by Ž5.8b., and where ⌺ is
defined by Ž5.9a.. Introducing appropriately relation Ž5.16. in relation
Ž5.17., after some algebraic manipulations, yields

ZU1 Ž ␯␶ . s U1W Ž ␯␶ . q ˆ
JZU1 Ž ␯ y 2 n . ␶ q V U W Ž ␯ y 2 n . ␶
q P1U ⌺W Ž ␯ y n . ␶ Ž 5.18.

where ˆJ and Vq are defined as by Ž5.9a. and Ž5.9b., respectively. Combin-

ing relations Ž5.11., Ž5.13. ᎐ Ž5.15., and Ž5.18., we readily obtain Ž5.7.. This
completes the proof of the Theorem.
It is remarked at this point that matrix V and matrix ⍀ are related
through the following relationship:

0 ⭈⭈⭈ 0 I
0 ⭈⭈⭈ I 0
⍀ s V⌳ , ⌳ s .. .. .. .. Ž 5.19.
. . . .
I ⭈⭈⭈ 0 0

Relation Ž5.7. will be used in the sequel for the identification of the
unknown matrices J1 , J2 , V, and V U . To this end, relation Ž5.7. is next
written in the linear regression form

⌿ Ž ␯␶ . s ⌰␾ Ž ␯␶ .
where

⌰ s J1 J2 V VU

is the true value of the plant parameter matrix, and where

␾ T Ž ␯␶ . s ⌿ T Ž ␯y2 n . ␶ W T Ž ␯␶ . W Ž ␯yn . ␶ W T Ž ␯y2 n . ␶

Next, define

Z Ž kT0 . s ␾ Ž kT0 . ␾ Ž kT0 y ␶ . ⭈⭈⭈ ␾ Ž k y 1 . T0

Y Ž kT0 . s ⌿ Ž kT0 . ⌿ Ž kT0 y ␶ . ⭈⭈⭈ ⌿ Ž k y 1 . T0

ˆ Ž kT0 . s J1 Ž kT0 .
⌰ J2 Ž kT0 . V Ž kT0 . V U Ž kT0 .

where J1Ž kT0 ., J2 Ž kT0 ., VŽ kT0 ., and V U Ž kT0 . are the matrices J1 , J2 , V, and
V U evaluated at kT0 , through the identification procedure. Clearly, the
 ADAPTIVE POLE POSITIONING 491

following relation holds:

Y Ž kT0 . s ⌰Z Ž kT0 .

ˆ Ž kT0 . as
We now choose the recursive algorithm for the estimation of ⌰

ˆ Ž kT0 . s ⌰
⌰ ˆ Ž k y 1 . T0
ˆ Ž k y 1 . T0 Z Ž k y 1 . T0 y Y Ž k y 1 . T0
y ⌰
y1
= Z T Ž k y 1 . T0 ␣ I q Z Ž k y 1 . T0 Z T Ž k y 1 . T0
Ž 5.20.

where ␣ g Rq is arbitrary, ⌰ ˆ Ž kT0 . is the estimated parameter matrix ⌰ ˆ at

ˆ ˆ
time t s kT0 , and ⌰ 0 s ⌰Ž kT0 .< ks 0 is arbitrarily specified. It is pointed
out that the term ␣ I in Ž5.20. is added in order to avoid numerical ill
conditioning, arising in the identification procedure based on the usual
least-squares algorithm, when the determinant of the matrix ZwŽ k y
1.T0 xZ T wŽ k y 1.T0 x takes small values.
Commenting on the nature of the adaptive law Ž5.20., we point out that,
it describes an on-line estimation procedure which deals with sequential
data in which the parameter estimates are recursively updated within the
time limit imposed by the sampling period T0 . It is worth noticed, at this
point that, in the present case, it is presumed that, a complete block of
information needed for the estimation of the plant parameters, is not
available prior to analysis and control, as in several off-line estimation
procedures. Therefore, in our case, identification and control of the plant
are performed concurrently. In order to calculate the parameters of the
desired MRIC based pole placement controller, it is necessary here to
update the plant parameter estimates using Ž5.20. and then solve the
canonical equations of Sections 3 and 4 for every time step k Žsee the
following subsection for details.. This is in contrast, to the standard policy
followed in cases where identification and control of the plant are per-
formed separately, in which we solve equations for the plant and the
controller parameters once, after an appropriate minimum number of
observations on the basis of which, a fixed model for the controlled plant is
available for further analysis Žsee w36, 37x for a comparative study of the
two approaches..
It is worth noticing at this point that although exact solutions to the
equation schemes of the paper are possible, the convergence of the
identification procedure is crucial for our analysis. This is due to the fact
that the adaptive law Ž5.20. is chosen so that ⌰ ˆ Ž kT0 . will satisfy the
equation YŽ kT0 . s ⌰ZŽ kT0 . Ž k G 0. asymptotically with time, i.e., for
492 ARVANITIS, KALOGEROPOULOS, AND SANTAS

k ª ⬁, rather than at every time instant. In other words, in the early

stages of the on-line identification procedure, the estimated parameter
matrix ⌰ˆ Ž kT0 ., obtained by Ž5.20., is usually far from its true value ⌰ and
it is expected that the plant parameter estimates Žand consequently the
controller parameter estimates . converge to their true values, only as
k ª ⬁. Therefore, exact determination of the desired MRIC based pole
placement controller through the procedures presented in Sections 3 and
4, is expected here, only after a certain step of the overall control
procedure. Before this step, the calculated controllers are far from being
those which guarantee the desired performance of the closed-loop system.
However, it is a standard fact in all adaptive control schemes that conver-
gence of the parameter estimates to their true values depends on the
specific properties of the particular identification procedure used and
crucially affects the adaptation since in cases where convergence of the
estimated parameters to their true values is not guaranteed, either the
calculated controllers are not the admissible ones or they cannot be
computed Žfor instance, if ⌰ ˆ Ž kT0 ., as obtained by the identification, is
unbounded.. So the effectiveness of our method depends on the conver-
gence and the boundedness properties of the proposed identification
procedure. These properties are summarized in the following Proposition.
PROPOSITION 5.1. ˜ Ž kT0 . be the parameter estimation error, defined
Let ⌰
as

˜ Ž kT0 . s ⌰
⌰ ˆ T Ž kT0 . y ⌰T Ž 5.21.

Then, for the parameter estimation algorithm of the form Ž5.20., the
following properties hold:
Ža. 5⌰
ˆ Ž kT0 .5 F ␮ for some finite ␮ g Rq
Žb. ˆ Ž kT0 .
If lim k ª⬁ Ý k␳s0 ␭ minŽZŽ ␳ T0 .Z T Ž ␳ T0 .. s ⬁ then lim k ª⬁ ⌰
s⌰
where ␭minŽ ⴢ . denotes the minimum eigen¨ alue of a matrix.
Proof. Ža. Taking the transpose of both sides in Ž5.20., introducing
Ž5.21. in the resulting relation and taking into account the fact that
Z T Ž kT0 .⌰ T y Y T Ž kT0 . s 0, we readily obtain

y1
˜ Ž kT0 . s I y ␣ I q Z Ž k y 1 . T0 Z T Ž k y 1 . T0
⌰ ½
=Z Ž k y 1 . T0 Z T Ž k y 1 . T0 5 ⌰˜ Ž k y 1. T 0 Ž 5.22.
 ADAPTIVE POLE POSITIONING 493

On the basis of the matrix inversion lemma, relation Ž5.22. may further be
written as
y1
1
˜ Ž kT0 . s I q
⌰ ½ Z Ž k y 1 . T0 Z T Ž k y 1 . T0 5 ˜ Ž k y 1 . T0
⌰
␣
Ž 5.23.

Therefore,

˜ T Ž kT0 . ⌰
⌰ ˜ Ž kT0 . s ⌰
˜ T Ž k y 1 . T0
y2
1
= Iq½ ␣
Z Ž k y 1 . T0 Z T
Ž k y 1 . T0 5
˜ Ž k y 1 . T0
=⌰
y2
␭min Ž Z Ž k y 1 . T0 Z T Ž k y 1 . T0 .
F 1q
ž ␣ /
˜ T Ž k y 1 . T0 ⌰
=⌰ ˜ Ž k y 1 . T0 Ž 5.24.
By repeatedly using the above inequality, we obtain
y2
ky1 ␭min Ž Z Ž ␳ T0 . Z T Ž ␳ T0 . .
˜ T Ž kT0 . ⌰
⌰ ˜ Ž kT0 . F Ł 1 q ˜ T0 ⌰
⌰ ˜0
␳s0 ž ␣ /
y2
1 ky1
F 1q Ý ␭min Ž Z Ž ␳ T0 . Z Ž ␳ T0 . .
T ˜ T0 ⌰
⌰ ˜ 0 Ž 5.25.
␣ ␳s0

where ⌰ ˜ T0 y ⌰T. Hence, 5 ⌰

˜0 s ⌰ ˜ Ž kT0 .5 is uniformly bounded by 5 ⌰
˜ 0 5,
ˆ Ž kT0 . is also uniformly bounded by some finite
and since ⌰ is finite, ⌰
␮ g Rq.
Žb. If lim k ª⬁ Ý k␳s0 ␭minŽZŽ ␳ T0 .Z T Ž ␳ T0 .. s ⬁ then, from Ž5.24., it
˜ Ž kT0 . s 0, and therefore, lim k ª⬁ ⌰
follows that lim k ª⬁ ⌰ ˆ Ž kT0 . s ⌰.
Clearly, Proposition 5.1 states that for the convergence of the plant
parameters estimates ⌰ ˆ Ž kT0 . to their true values ⌰ it is sufficient that the
regression vector ZŽ kT0 . is persistently exciting to the amount that
k
lim Ý ␭min Ž Z Ž ␳ T0 . ZT Ž ␳ T0 . . s ⬁
kª⬁ ␳ s0
494 ARVANITIS, KALOGEROPOULOS, AND SANTAS

Therefore, since adaptation and stability of the adaptive scheme depend

on the convergence of the parameter estimates to their true values, it is
necessary to prove excitation of ZŽ kT0 .. This is done in Subsection 5.3 Žsee
Theorem 5.2, therein ..
Remark 5.1. It is pointed out that although controllability and observ-
ability of the sampled system Ž5.1. is instrumental for our analysis, no
assumption is made in the present paper on the canonical structure of the
triplet Ž ⌽␶ , B␶ , C.. This is in contrast to the standard policy of many known
adaptive schemes, in which controllability or observability canonical forms
are assumed for the matrix triplet involved in the estimation procedure
Žsee, for example, w38x, w39x.. The reason here for avoiding an assumption
on the canonical structure for the triplet Ž ⌽␶ , B␶ , C. is mainly due to the
fact that canonical forms for multivariable systems are interwoven with the
knowledge of a set of controllability or observability indices of the matrix
triplet sought Žfor example, in w38x, w39x a set of observability indices is
needed to be known.. As a consequence, when identification procedures
based on canonical structures are used, much more prior knowledge
relative to the structure of the controlled plant is necessary, as compared
to our approach.

5.2. Algorithm for the Synthesis of the Adapti¨ e Controller

On the basis of the estimated parameter matrix ⌰ ˆ Ž kT0 . obtained by
Ž5.20., as well as on the basis of the relations Ž5.4. ᎐ Ž5.6. and Ž5.19. and of
anyone of the algorithms reported in the literature for the construction of
a minimal realization, one can obtain the estimates which are necessary
for the computation of the unknown matrices A ˆi'A ˆ i Ž kT0 ., ⌽ ' ⌽ Ž kT0 .
and the unknown vector ˆb i ' ˆb i Ž kT0 . involved in the algorithms presented
in the previous sections. Moreover, since the matrices Q, ⌽ ˜ U , R s p , Ts p , L,
U
ˆ and Sˆ are constructed on the basis of A ˆ i Ž kT0 ., ⌽ Ž kT0 ., and ˆb i Ž kT0 .,
S,
then provided that the matrix triplet Ž ⌽ Ž kT0 ., B ˆ Ž kT0 ., CŽ kT0 .. is control-
lable and observable for any possible value of ⌰ ˆ Ž kT0 ., we can obtain the
following results sought:

ˆ'F
F ˆŽ⌰
ˆ Ž kT0 . . , ˆ Ž kT0 . .
v ' vŽ⌰ Ž 5.26.

whereas no update is taken otherwise.

Overall, the procedure for the synthesis of the adaptive MRIC based
adaptive pole placer, consists of the ten steps given below:
Step 1. Choose Ni G n i and the sampling period ␶ such that
␶ s T0rŽ6 n y 1. N.
Step 2. Update the estimates of the matrix V using relation Ž5.20..
 ADAPTIVE POLE POSITIONING 495

Step 3. Find the matrix ⍀ using relation Ž5.19..

Step 4. Obtain a minimal realization for the matrix triplet Ž ⌽␶ , B␶ ,
C. using any one of the minimal realization algorithms reported in the
literature Žsee, e.g., the algorithms in w32x ᎐ w34x..
Step 5. Find the matrices A ˆ i and the vectors ˆb i , as well as the
matrix ⌽ using relations Ž5.4. and Ž5.5., respectively.
Step 6. Use the algorithm presented in Section 3 to compute the
controllability indices ␦ i of the pair Ž ⌽ T , C T ., as well as the values of the
matrices L, Q, ⌽ ˜ U , R s p , and Ts p .
Step 7. Use Ž3.26. to compute the controller matrix K f .
Step 8. Find the matrices S ˆ and SˆU using relation Ž4.12. and Ž4.13.,
respectively.
Step 9. Find the matrix F ˆ and the vector v using relations Ž4.14. and
Ž4.15., respectively.
Step 10. Find the matrix FŽ t . of the MRIC based controller sought
and the persistent excitation signals vi Ž t . using relations Ž4.16. and Ž4.2a.,
Ž4.2b., and Ž4.3., respectively.

5.3. Stability Analysis of the Adapti¨ e Control Scheme

We now investigate the stability of the closed-loop system for arbitrary
initial conditions on the plant. To this end, the following fundamental
result can be established.
THEOREM 5.2. In the closed-loop adapti¨ e control system the regressor
sequence ␾ Ž ␯␶ . is persistently exciting, i.e., there is ␦ ) 0, such that
Ž6 ny1 . N
Z Ž kT0 . Z T Ž kT0 . s Ý ␾ Ž kT0 y ␯␶ . ␾ T Ž kT0 y ␯␶ . G ␦ I Ž 5.27.
␯s0

Proof. In order to prove relation Ž5.27., we work as follows: Set

u i t . s d Ti Ž t .vi . Then relation Ž5.7. yields
Ž
ny1
yi Ž ␯␶ . s Ý Ž J1 . Ž ny 1. pqi , Ž ny ␳y1. pqi yi Ž ␯ y 2 n y ␳ . ␶
␳s0
p ny1
q Ý Ý Ž J1 . Ž ny 1. pqi , Ž ny ␳y1. pq ␬ y␬ Ž ␯ y 2 n y ␳ . ␶
␬s1 ␳ s0
␬ /i
m ny2
q Ý Ý Ž J2 . Ž ny 1. pqi , Ž ny ␳y2. mqj u j Ž ␯ y ␳ y 1. ␶
js1 ␳ s0
496 ARVANITIS, KALOGEROPOULOS, AND SANTAS

m ny1
q Ý Ý Ž V. Ž ny1. pqi , Ž ny ␳y1. mqj u j Ž ␯ y n y ␳ . ␶
js1 ␳ s0

m ny1
q Ý Ý Ž V U . Ž ny1. pqi , Ž ny ␳y1. mqj u j Ž ␯ y 2 n y ␳ . ␶ Ž 5.28.
js1 ␳ s0

where in general ŽJ1 . r q , ŽJ2 . r q , ŽV. r q , and ŽV U . r q are the r y q elements of

the matrices J1 , J2 , V, and V U , respectively. Introducing the pseudovari-
ables ␤i, u jŽ ␯␶ ., j g Jm and ␤i, y␬Ž ␯␶ ., ␬ s 1, 2, . . . , p, ␬ / i, relation Ž5.28.
can be decomposed as follows:

ny1
␤i , u jŽ ␯␶ . y Ý Ž J1 . Ž ny 1. pqi , Ž ny ␳y1. pqi ␤i , u Ž ␯ y 2 n y ␳ . ␶
j
s u j Ž ␯␶ .
␳s0

Ž 5.29a.
ny2
yi , u jŽ ␯␶ . s Ý Ž J2 . Ž ny 1. pqi , Ž ny ␳y2. mqj ␤i , u Ž ␯ y ␳ y 1. ␶
j
␳s0

ny1
q Ý Ž V. Ž ny1. pqi , Ž ny ␳y1. mqj ␤i , u Ž ␯ y n y ␳ . ␶
j
␳s0

ny1
q Ý Ž V U . Ž ny1. pqi , Ž ny ␳y1. mqj ␤i , u Ž ␯ y 2 n y ␳ . ␶
j
␳s0

for j g Jm Ž 5.29b.
ny1
␤i , y␬Ž ␯␶ . y Ý Ž J1 . Ž ny 1. pqi , Ž ny ␳y1. pqi ␤i , y Ž ␯ y 2 n y ␳ . ␶
␬
s y␬ Ž ␯␶ .
␳s0

Ž 5.29c.
ny1
yi , y␬Ž ␯␶ . s Ý Ž J1 . Ž ny 1. pqi , Ž ny ␳y1. pq ␬ ␤i , y Ž ␯ y 2 n y ␳ . ␶
␬
␳s0

for ␬ s 1, 2, . . . , p, ␬ / i Ž 5.29d.

while

m p
yi Ž ␯␶ . s Ý yi , u jŽ ␯␶ . q Ý yi , y␬ Ž ␯␶ . Ž 5.29e.
js1 ␬ s1
␬ /i
 ADAPTIVE POLE POSITIONING 497

From relations Ž5.29b. ᎐ Ž5.29e., we obtain

¡
~ Ý Ý ŽJ .
m ny2
yi Ž ␯␶ . s Ž 1rp .
¢ js1
½ ␳ s0
2 Ž ny 1 . pqi , Ž ny ␳ y2 . mqj ␤i , u j Ž ␯ y ␳ y 1 . ␶

ny1
q Ý Ž V. Ž ny1. pqi , Ž ny ␳y1. mqj ␤i , u Ž ␯ y n y ␳ . ␶
j
␳s0

ny1
q Ý Ž V U . Ž ny1. pqi , Ž ny ␳y1. mqj ␤i , u Ž ␯ y 2 n y ␳ . ␶
␳s0
j 5
p ny1
q Ý Ý Ž J1 . Ž ny 1. pqi , Ž ny ␳y1. pq ␬ ␤i , y Ž ␯ y 2 n y ␳ . ␶
␬
␬s1 ␳ s0
␬ /i
p ny1
q
␬s1
Ý
␬ /i
½ ␤␬ , y iŽ ␯␶ . y Ý Ž J1 . Ž ny 1. pq ␬ , Ž ny ␳y1. pq ␬
␳ s0

¦¥
=␤␬ , y i Ž ␯ y 2 n y ␳ . ␶
§
5 Ž 5.30.

whereas relation Ž5.29a. yields

p ny1
u j Ž ␯␶ . s Ž 1rp . Ý
is1
½ ␤i , u jŽ ␯␶ . y Ý Ž J1 . Ž ny 1. pqi , Ž ny ␳y1. pqi
␳ s0

=␤i , u j Ž ␯ y 2 n y ␳ . ␶
5 Ž 5.31.

On the basis of relations Ž5.7., Ž5.30., and Ž5.31., the regressor vector
␾ Ž ␯␶ . can also be expressed as
␾ Ž ␯␶ . s ˆ
⌺ ␤ˆŽ ␯␶ .
where
␤ˆT Ž ␯␶ . s ␤˜Ž ␯␶ . ⭈⭈⭈ ␤˜ Ž ␯ y 6 n q 2 . ␶

␤˜Ž ␳␶ . s ␤˜u1Ž ␳␶ . ⭈⭈⭈ ␤˜u mŽ ␳␶ . ␤˜y 1Ž ␳␶ . ⭈⭈⭈ ␤ y pŽ ␳␶ . ,

␳ s ␯ y 6 n q 2, . . . , ␯

␤˜u jŽ ␳␶ . s ␤ 1, u jŽ ␳␶ . ⭈⭈⭈ ␤ p , u jŽ ␳␶ . ,

␳ s ␯ y 6 n q 2, . . . , ␯ , j g Jm
498 ARVANITIS, KALOGEROPOULOS, AND SANTAS

␤˜y 1Ž ␳␶ . s ␤ 2, y 1Ž ␳␶ . ⭈⭈⭈ ␤ p , y 1Ž ␳␶ . , ␳ s ␯ y 6 n q 2, . . . , ␯

␤˜y␬ Ž ␳␶ . s ␤ 1, y␬Ž ␳␶ . ⭈⭈⭈ ␤ py1 , y␬ Ž ␳␶ . ,

␳ s ␯ y 6 n q 2, . . . , ␯ , ␬ s 2, 3, . . . , p

and where ˆ ⌺ g RŽ3 n mqn p.=Ž6 ny1. pŽ pqmy1. is a full row rank matrix. Clearly,
the vector ␾ Ž ␯␶ . is persistently exciting if ␤ˆŽ ␯␶ . is also persistently
exciting. So, in what follows, it suffices to investigate excitation of ␤ˆŽ ␯␶ ..
To this end, observe that Ž5.31. can be written as

u j Ž ␯␶ . s ␺ jT␤ˆŽ ␯␶ . Ž 5.32.

where ␺ jT g RŽ6 ny1. mŽ pqmy1. is a row vector whose elements are known.
In order to prove excitation of ␤ˆŽ ␯␶ ., it suffices to prove that the
following relationship holds:

T0r ␶

Ý ␤ˆŽ kT0 q ␯␶ . ␤ˆT Ž kT0 q ␯␶ . G ␧ I Ž 5.33.

␯s1

for some ␧ ) 0. To this end, observe that from relation Ž5.32., we can
easily obtain

T0r ␶ T 0r ␶

Ý
␯s1
u 2j Ž kT0 q ␯␶ . s ␺ jT ½Ý ␯ s1
5
␤ˆŽ kT0 q ␯␶ . ␤ˆT Ž kT0 q ␯␶ . ␺ j Ž 5.34.

Observe also that the following relation holds:

¡0 if ␯ s 1, 2, . . . , Ž 6 n y 1 . Ž Nj y n j y 1 . l j y 1
u j Ž kT0 q ␯␶ . s 1 ~ if ␯ s Ž 6 n y 1 . Ž Nj y n j y 1 . l j , . . . ,
¢ Ž 6 n y 1 . Ž Nj y n j . l j y 1

Hence, relation Ž5.34. can also be written as

T0r ␶
Ž 6 n y 1. l j q Ý u 2j Ž kT0 q ␯␶ .
␯ s Ž6 ny1 .Ž Njyn j . l j

T0r ␶
s ␺ jT ½Ý
␯s1
␤ˆŽ kT0 q ␯␶ . ␤ˆT Ž kT0 q ␯␶ . ␺ j 5
 ADAPTIVE POLE POSITIONING 499

We can then conclude that

T0r ␶
␺ jT ½Ý
␯s1
5
␤ˆŽ kT0 q ␯␶ . ␤ˆT Ž kT0 q ␯␶ . ␺ j G Ž 6 n y 1 . l j

and that
T0r ␶
␺ jT ␺j Ž 6 n y 1. l j
½ 5½
5 ␺j 5 Ý ␤ˆŽ kT0 q ␯␶ . ␤ˆT Ž kT0 q ␯␶ .
␯s1
5½ 55 ␺j 5
G
5 ␺j 5 2

It is now clear that the vector ␺ jr5 ␺ j 5 is a vector whose norm equals
unity. Hence there is a unity norm vector such that
T0r ␶
Ž 6 n y 1. l j
␹T ½Ý␯s1
␤ˆŽ kT0 q ␯␶ . ␤ˆT Ž kT0 q ␯␶ . ␹ y5 5 ␺j 5 2
G0

In conclusion, relation Ž5.33. holds. As a consequence, the vector ␤ˆŽ ␯␶ .

is persistently exciting. Therefore, ␾ Ž ␯␶ . is also persistently exciting and
hence there is a ␦ ) 0 Žwhich, in general, depends on the matrix ˆ ⌺ ., such
that relation Ž5.27. holds. This completes the proof of the Theorem.
We are now able to establish the stability of the adaptive control system.
PROPOSITION 5.2. The closed-loop adapti¨ e control system presented abo¨ e
is globally stable, i.e., for arbitrary finite initial conditions all states are
uniformly bounded, and pole placement control is asymptotically attained.
Furthermore, the proposed adapti¨ e scheme pro¨ ides exponential con¨ ergence
of the estimated parameters.
Proof. Since, according to Theorem 5.2, the regressor sequence is
persistently exciting, then the difference ⌰ ˆ Ž kT0 . y ⌰ converges to zero.
That is, the plant parameter estimates converge to their true values. As a
consequence of this and of the fact that ⌰ ˆ Ž kT0 . is uniformly bounded, the
controller parameter estimates 5.26 also converge to their true values.
Ž .
Therefore, at the sampling instants uniform boundedness of all states and
discrete pole placement follow on the basis of Ž4.4.. Uniform boundedness
of uŽ t . and xŽ t . then follows from Ž2.1., Ž4.8a., Ž4.8b., and Ž4.16. and from
the fact that wŽ kT0 . is bounded by assumption. Finally, exponential con-
vergence of the plant parameter estimates follows from Ž5.23., which
together with Ž5.27. ensures that ⌰ ˆ Ž kT0 . ª ⌰ exponentially as k ª ⬁.
Remark 5.2. Commenting on the assumptions needed here, in order to
implement the MRIC based adaptive pole placer presented above, we
point out the following:
Assumption 2.1a, on the controllability and observability of the continu-
ous-time plant as well as on the knowledge of its order, is a standard
500 ARVANITIS, KALOGEROPOULOS, AND SANTAS

assumption in the area of adaptive control. It is worth noticing that here,

controllability of the pair ŽA, B. is also necessary for obtaining a solution of
the integral equation Ž3.2., with respect to the controller matrix FŽ t .. Note
also that uncontrollability Žandror unobservability. of the pair ŽA, B.
implies uncontrollability Žandror unobservability. of the plants obtained
from Ž2.1., by discretizing with sampling periods T0 , TN , and ␶ . From the
previous analysis, however, it becomes clear that for the implementation of
the adaptive control scheme, these discretized plants must be controllable
and observable.
Assumption 2.1b, on the knowledge of a set of LMCI indices of the pair
ŽA, B., is instrumental for the implementation of the proposed adaptive
scheme, since, on the one hand, the elements of the MRIC of the form
Ž4.1. and the persistent excitation signals Ž4.2a., Ž4.2b., and Ž4.3. depend on
the LMCI used, and on the other hand, the control strategy in the case of
unknown systems is based on the fundamental sampling period ␶ , which
also depends of the knowledge of a set of LMCI. Note also that, whenever
Assumption 2.1b is not fulfilled, one can readily compute a set of LMCI by
estimating the continuous-time system matrices A and B. This can be done
either using a continuous-time counterpart of the identification procedure
presented in Section 5.1, or following the structural identification ap-
proach proposed in w38x. For the sake of simplicity, we assume here that
the initial information about a set of LMCI of the pair ŽA, B. is available.
Assumption 2.2 on the existence of a sampling period T0 , for which
controllability and observability of the matrix triplets Ž ⌽, B, ˆ C. and
Ž ⌽␶ , B␶ , C. are guaranteed, is also instrumental for our analysis. In particu-
lar, observability of the pair Ž ⌽, C. must be guaranteed for being able to
transform the pair Ž ⌽ T , C T . in its input Luenberger canonical form and for
obtaining a solution of the pole placement control problem, in the case of
known systems. On the other hand, controllability and observability of the
matrix triplet Ž ⌽␶ , B␶ , C. is necessary for resorting to the equivalent
input᎐output representation Ž5.7., for the state space system of the form
Ž5.1., as well as for being able to apply any one of the minimal realization
algorithms presented in w32x ᎐ w34x, which are needed here to obtain the
estimates of the triplet Ž ⌽␶ , B␶ , C.. Note that for ensuring controllability
and observability of the triplets Ž ⌽, B, ˆ C. and Ž ⌽␶ , B␶ , C., the fundamental
sampling period T0 must be selected such that simultaneously

2 ␳␲ j
Ž a. ,
T0

␳ s 0, 1, . . . Ž j s 'y 1 . is not the difference of

Ž 5.35a.
any two eigenvalues of the matrix A.
 ADAPTIVE POLE POSITIONING 501

2 Ž 6 n y 1 . N␳␲ j
Ž b. ,
T0
␳ s 0, 1, . . . is the difference of
Ž 5.35b.
any two eigenvalues of the matrix A.
Ž c . ␺ Ž TN . / 0. Ž 5.35c.

This implies that in the multirate adaptive case treated here, certain
sampling frequencies must be avoided, as compared to the nonadaptive
nonmultirate case. It is pointed out that conditions Ž5.35a. and Ž5.35b. are
standard conditions for the selection of a regular sampling period, in order
to avoid loss of controllability and observability under sampling Žsee w40x
for a detailed analysis of this issue..

6. CONCLUSIONS

The adaptive pole placement problem of linear time-invariant continu-

ous-time multi-input, multioutput, systems has been investigated and an
indirect adaptive control scheme based on periodic multirate-input con-
trollers has been presented for the first time. The proposed control
strategy has, as compared to known related techniques, the following main
advantages:
Ža. It is readily applicable to nostably invertible systems having
arbitrary poles and zeros and relative degree. This is due to the fact that
the approach used here to solve the adaptive pole placement problem does
not rely on pole-zero cancellations.
Žb. Following the proposed technique a gain controller is essentially
needed to be designed, as compared to dynamic compensators or state
observers needed by known indirect adaptive pole placement techniques.
Consequently, the present approach avoids the problems of known adap-
tive pole placement techniques, interwoven with the possibly unstable
solutions of the Diophantine equation. Moreover, no exogenous dynamics
are introduced in the control loop by our technique, whereas in many
known techniques the dynamics introduced are of high order. This fact
improves the computational aspect of the problem, since the proposed
technique does not require many on-line computations and its practical
implementation requires computer memory only for storing the modulat-
ing matrix function FŽ t . over one period of time.
Žc. It offers a solution to the problem of ensuring persistency of
excitation of the continuous-time plant under control, without imposing
502 ARVANITIS, KALOGEROPOULOS, AND SANTAS

any special requirement on the reference signal wŽ kT0 . Žexcept bounded-

ness. and without making any assumption concerning either the existence
of specific convex sets in which the estimated parameters belong or the
coprimeness of the polynomials describing the ARMA model.
The present paper gives some new insights to the adaptive pole place-
ment problem of linear systems. The present results can be extended to
solve other adaptive control problems, as, for example, the problems of
model reference adaptive control and adaptive decoupling using
multirate-input controllers or multirate generalized sampled-data hold
functions. Adaptive control schemes based on alternative parameter esti-
mation algorithms Žas, for example, the algorithm proposed in w39x. and
without the need of persistent excitation signals are currently under
investigation.

REFERENCES

1. G. M. Kranc, Input-output analysis of multirate feedback systems, IRE Trans. Autom.

Control 3 Ž1957., 21᎐28.
2. E. I. Jury and F. J. Mullin, The analysis of sampled-data control systems with a
periodically time-varying sampling rate, IRE Trans. Autom. Control 4 Ž1959., 15᎐21.
3. R. E. Kalman and J. E. Bertram, A unified approach to the theory of sampling systems, J.
Franklin Inst. 267 Ž1959., 405᎐436.
4. M. Araki and K. Yamamoto, Multivariate multirate sampled-data systems: State-space
description, transfer characteristics and Nyquist criterion, IEEE Trans. Autom. Control
AC-30 Ž1986., 145᎐154.
5. M. Araki and T. Hagiwara, Pole assignment by multirate sampled data output feedback,
Int. J. Control 44 Ž1986., 1661᎐1673.
6. T. Hagiwara and M. Arakim, Design of a stable state feedback controller based on the
multirate sampling of the plant output, IEEE Trans. Autom. Control AC-33 Ž1988.,
812᎐819.
7. D. G. Meyer, A parametrization of stabilizing controllers for multirate sampled-data
systems, IEEE Trans. Autom. Control AC-35 Ž1990., 233᎐236.
8. P. Colaneri, R. Scattolini, and N. Sciavoni, Stabilization of multirate sampled-data linear
systems, Automatica 27 Ž1990., 337᎐380.
9. M. C. Berg, N. Amit, and J. Powell, Multirate digital control system design, IEEE Trans.
Autom. Control AC-33 Ž1988., 1139᎐1150.
10. H. M. Al-Rahmani and G. F. Franklin, A new optimal multirate control of linear periodic
and time-invariant systems, IEEE Trans. Autom. Control AC-35 Ž1990., 406᎐415.
11. T. Chen and B. A. Francis, H 2-optimal sampled-data control, IEEE Trans. Autom.
Control AC-36 Ž1991., 387᎐397.
12. D. G. Meyer, Cost translation and a lifting approach to the multirate LQG problem,
IEEE Trans. Autom. Control AC-37 Ž1992., 1411᎐1415.
13. P. G. Voulgaris and B. Bamieh, Optimal H ⬁ and H 2 control of hybrid multirate systems,
Syst. Control Lett. 20 Ž1993., 249᎐261.
14. T. Chen and L. Qiu, H ⬁ design of general multirate sampled-data control systems,
Automatica J. IFAC 30 Ž1994., 1139᎐1152.
 ADAPTIVE POLE POSITIONING 503

15. K. G. Arvanitis and P. N. Paraskevopoulos, Sampled-data minimum H ⬁ -norm regulation

of linear continuous-time systems using multirate-output controller, J. Optim. Theory
Appl. 87 Ž1995., 235᎐267.
16. M. E. Sezer and D. D. Siljak, Decentralized multirate control, IEEE Trans. Autom.
Control AC-35 Ž1990., 60᎐65.
17. K. G. Arvanitis, An indirect model reference adaptive controller based on the multirate
sampling of the plant output, Int. J. Adapti¨ e Contr. Sign. Process. 10 Ž1996., 673᎐705.
18. K. G. Arvanitis, Adaptive LQ regulation by multirate-output controllers, Found. Comput-
ing Dec. Sci. 21 Ž1996., 183᎐213.
19. K. G. Arvanitis and P. N. Paraskevopoulos, Exact model matching of linear systems using
multirate digital controllers, Proc. 2nd E.C.C., Groningen, The Netherlands, 28 June᎐1
July 1993, Vol. 3, pp. 1648᎐1652.
20. K. G. Arvanitis and G. Kalogeropoulos, Stability robustness of LQ optimal regulators
designed on the basis of the multirate sampling of the plant output, J. Optim. Theory
Appl. 97 Ž1998., 299᎐337.
21. P. E. Wellstead, J. M. Edmunds, D. Prager, and P. Zanker, Self-tuning polerzero
assignment regulators, Int. J. Control 30 Ž1979., 1᎐26.
22. K. J. Astrom and B. Wittenmark, Self-tuning controllers based on pole-zero placement,
Proc. IEE Part D 127 Ž1980., 120᎐130.
23. H. Elliott, Direct adaptive pole placement with application to nonminimum phase
systems, IEEE Trans. Autom. Control AC-27 Ž1982., 720᎐722.
24. H. Elliott, W. A. Wolovich, and M. Das, Arbitrary adaptive pole placement for linear
multivariable systems, IEEE Trans. Autom. Control AC-29 Ž1984., 221᎐229.
25. B. D. O. Anderson and R. M. Johnstone, Global adaptive pole positioning, IEEE Trans.
Autom. Control AC-30 Ž1985., 11᎐22.
26. H. Elliott, R. Cristi, and M. Das, Global stability adaptive pole placement algorithms,
IEEE Trans. Autom. Control AC-30 Ž1985., 348᎐356.
27. R. Lozano-Leal and G. C. Goodwin, A globally convergent adaptive pole placement
algorithm without a persistency of excitation requirement, IEEE Trans. Autom. Control
AC-30 Ž1985., 795᎐798.
28. F. Giri, M. M’Saad, L. Dugard, and J. M. Dion, Robust pole placement indirect adaptive
controller, Int. J. Adapti¨ e Contr. Sign Proc. 2 Ž1988., 33᎐47.
29. L. Mo and M. M. Bayoumi, A novel approach to the explicit pole assignment self-tuning
controller design, IEEE Trans. Autom. Control AC-34 Ž1989., 359᎐363.
30. D. C. Youla, J. J. Bongiorno, Jr., and C. N. Lu, Single-loop feedback stabilization of
linear multivariable dynamical systems, Automatica 10 J. IFAC Ž1974., 159᎐173.
31. M. Kinnaert and V. Blondel, Discrete-time pole placement with stable controller,
Automatica 28 Ž1992., 935᎐943.
32. B. L. Ho and R. E. Kalman, Effective construction of linear state variable models from
inputroutput functions, Proc. 3rd Allerton Conf., pp. 449᎐459; Regelungstechnik, Vol. 14,
pp. 545᎐548, 1966.
33. L. M. Silverman, Realization of linear dynamical systems, IEEE Trans. Aut. Control
AC-16 Ž1971., 554᎐567.
34. R. Eising and M. L. J. Hautus, Realizations algorithms for systems over a principal ideal
domain, Memorandum COSOR 78-25, Eindhoven University of Technology, Dept. of
Mathematics, Eindhoven, 1978.
35. R. Lozano-Leal, Robust adaptive regulation without persistent excitation, IEEE Trans.
Autom. Control AC-34 Ž1989., 1260᎐1267.
36. G. C. Goodwin and K. S. Sin, ‘‘Adaptive Filtering, Prediction and Control,’’ Prentice Hall
International, Englewood Cliffs, NJ, 1984.
504 ARVANITIS, KALOGEROPOULOS, AND SANTAS

37. S. Sastry and M. Bodson, ‘‘Adaptive Control: Stability, Convergence and Robustness,’’
Prentice Hall International, Englewood Cliffs, NJ, 1989.
38. R. Guidorzi, Canonical structures in the identification of multivariable systems, Automat-
ica J. IFAC 11 Ž1975., 361᎐374.
39. J. K. McElveen, K. R. Lee, and J. E. Bennett, Identification of multivariable linear
systems from input-output measurements, IEEE Trans. Ind. Electr. 39 Ž1992., 189᎐193.
40. R. E. Kalman, Y. C. Ho, and K. S. Narendra, Controllability of linear dynamical systems,
Contrib. Diff. Equations 1 Ž1962., 189᎐213.