This article was downloaded by: [North Carolina State University]

On: 18 December 2012, At: 02:21

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Control

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/tcon20

Sliding mode stabilization of uncertain systems using

only output information
a a
CHRISTOPHER EDWARDS & SARAH K. SPURGEON
a
Control Systems Research, Department of Engineering, University of Leicester, Leicester,
LEI 7RH, U.K., Tel: +(44) (0)116 252 2559; Fax +(44) (0)116 252 2619.
Version of record first published: 24 Feb 2007.

To cite this article: CHRISTOPHER EDWARDS & SARAH K. SPURGEON (1995): Sliding mode stabilization of uncertain systems
using only output information, International Journal of Control, 62:5, 1129-1144

To link to this article: http://dx.doi.org/10.1080/00207179508921587

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to
anyone is expressly forbidden.

The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should
be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims,
proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in
connection with or arising out of the use of this material.
 Sliding mode stabilization of uncertain systems using only output
information
CHRISTOPHER E D W A R D S t and S A R A H K. S P U R G E O N t

A practical procedure is presented for synthesizing output feedback controllers

for uncertain systems based on sliding mode concepts. The class of systems to
which the results apply is identified and includes the requirement that the
nominal linear system is minimum phase. The procedure proposed for the
Downloaded by [North Carolina State University] at 02:21 18 December 2012

design of the sliding surface uses established output feedback eigenvalue

assignment results. It will be shown that all the assumptions imposed on the
system pertain to the design of the sliding surface. The proposed controller,
which guarantees attainment of a sliding mode despite the presence of
uncertainty requires no additional assumptions.

1. Introduction
Variable structure control, with a sliding mode, is an established method of
controlling uncertain dynamical systems. Traditional variable structure control
systems employ a discontinuous control action to force the system states to
remain on a surface within the state-space. This results in a reduced order,
surface dependent sliding motion. (For the definition of a sliding motion
employed in this work and associated sufficient conditions see the tutorial paper
by DeCarlo et al. 1988.) This motion is insensitive to disturbances implicit in the
input channel-the so-called marched uncertainty. A key issue is therefore the
selection of the surface to provide an appropriate and stable sliding motion.
Many proponents of theoretical developments in this area have found it
convenient to assume all internal systems states are available to the control law.
In practice, of course, this is often not the case. One way of overcoming this
difficulty is to use a dynamical system (an observer) to generate estimates of the
unavailable internal states. Sliding mode approaches to this problem were
originally put forward by Utkin (1992), Slotine er al. (1987), Walcott and ~ a k
(1986) and, more recently, by Edwards and Spurgeon (1994). A n alternative
approach is to design the surface in such a way that only output information is
required. For linear systems with no uncertainty this problem has been investig-
ated by White (1990) and more recently by El-Khazali and DeCarlo (1991,
1992). For uncertain systems a further approach has recently been reported by
Hui and ~ a (1993)
k and ~ a and
k Hui (1993) who proposed an algorithm for
output dependent hyperplane design that is based upon eigenvector methods.
This paper develops a framework for designing a regulator for uncertain
systems. The controller structure is similar to that employed by Ryan and
Corless (1984) and Spurgeon and Davies (1993) except only output information
is required. T h e approach is quite different to that of Hui and ~ a (1993)
k and
will be shown to be applicable to a wider class of systems than that considered

Received 15 May 1994. Revised 21 September 1994.

t Control Systems Research, Department of Engineering, University of Leicester. Leicester
LEI 7RH, U.K. Tel: +(44) (0)116 252 2559: Fax +(44) (0)116 252 2619.
M20-7179/9SSI0.W Q IWS Taylor & Francis Ltd
 1130 C. Edwards and S. K. Spurgeon

by El-Khazali and DeCarlo (1991, 1992); $ 2 develops a framework for the

design of the sliding mode regulator and identifies the class of systems for which
the results are valid; 8 3 provides some examples to demonstrate the practicality
of the approach.

2. System description
Consider an uncertain dynamical system of the form
+ Bu(r) + f ( t , x , u:)
i ( t ) = Ax(!)
~ ( 1 =) J
where x E R n , 11 E R"' and y E RP with rn G p < n . Assume that the nominal
(1)
Downloaded by [North Carolina State University] at 02:21 18 December 2012

linear system ( A , B, C) is known and that the input and output matrices B and
C are both of full rank.
In addition, it is assumed that
(a) rank (CB) = m
( b ) the unknown function f : R+ x Rn x Rm + R n , which represents the
system nonlinearities plus any model uncertainties in the system, satis-
fies the matching condition
f ( t , x , U ) = B5(t, x , u ) (2)
where the bounded function : R+ X Rn X R" + R m satisfies

for some known function @ : R" -+ (W and positive constant k , < 1.

(Throughout the paper 1) . (1 indicates the euclidean norm for vectors.)
The reason for the restriction rank(CB) = tn will be made clear in the
exposition that follows. The intention is to develop a control law that induces a
sliding motion on the surface
9 = {x E Rn : FCx = 0) (4)
for some selected matrix F E R m X P The
. proposed control law has the form

where G , is a fixed gain matrix and the discontinuous vector v, depends only on
the outputs. A key design consideration is the choice of the matrix F so that the
associated sliding motion is stable.

2.1. Design framework

This section considers the general case when there are more outputs than
inputs. This situation turns out t o exhibit a richer mathematical structure than
the case of equal numbers of inputs and outputs. Necessary and sufficient
conditions will be developed for the existence of a control law of the form given
in ( 5 ) , which quadratically stabilizes the uncertain system given in (1). In
addition, it will be shown that a sliding mode is exhibited. The following lemma
provides a canonical form for the system triple ( A , B , C), which will be used
throughout the analysis and is similar to that given by Edwards and Spurgeon
(1994) for the design of sliding mode observers.
 Sliding mode stabilization of uncertain systems 1131

Lemma 1: Let ( A , B , C ) be a linear system with p > rn and rank (CB) = m .

Then a change of coordinates exists so that the system triple with respect to the
new coordinates has the following structure
( a ) the system matrix can be written as

and the sub-block A l l when partitioned has the structure

Downloaded by [North Carolina State University] at 02:21 18 December 2012

( b ) the input distriburion matrix has the form

B = [j2]where B2 E (WmX"' and is non-singular (8)

( c ) the output distribution matrix has the form
C = [0 TI where T E RPxp and is orthogonal. (9)
Proof: A pair of linear transformations will be demonstrated that bring about
the required canonical form. Without loss of generality assume
C = [O I,]
and the input distribution matrix is partitioned in a compatible way so that

B= [ill where E ( W ( " - P ) ~ ~and B2 E RPX"'

Then CB = B2 and so, by assumption, rank(B2) = m . Hence, in particular, the
left pseudo-inverse ~f = (BTB2)-'BT is well defined and there exists an
orthogonal matrix T E ( W p x P such that

where B2 E R"'X"' and is non-singular. Consequently the coordinate transforma-

tion

is non-singular and with respect to the new coordinates the input and ouput
distribution matrices are in the form of ( 8 ) and (9) respectively. Partition the
new system matrix as
 1132 C. Edwards and S. K. Spurgeon

where A l l E [w("-"~(("-"), A2,, E R ( P - " ' ) ~ ( " - P )and E R P X P . Let Tabs E

R("-P)X("-~ b e any matrix which puts ( A l 1 ,A211)into the following observabil-
ity canonical form

where E [ w ~ ~ ~E ,R ( P - ~ ) ~ ( ~ - P ,- ~the
) pair (Ai2, A$) is completely
observable and r 3 0 represents the number of unobservable states of
( A l l , A z l l ) It can easily be verified that changing coordinates with respect to
the non-singular transformation
Downloaded by [North Carolina State University] at 02:21 18 December 2012

provides a basis in which the system triple satisfies properties (a) (b) and (c) in
the lemma statement. 0
The proof of this lemma has been included because of its constructive nature-
which is a key consideration throughout the paper. In the analysis that follows it
is convenient to re-partition the orthogonal component of the output distribution
matrix s o that

Lemma 2: Let ( A , B, C ) be a given system in the carzonical form of Lemma I ,

!hen the invariant zeros of the system are the r eigenval~iesof A:,.

Proof: For the proof see Appendix A . 17

Suppose a controller exists that induces a stable sliding motion on the surface Y,
defined in (4). For a unique equivalent control to exist the matrix FCB E RmX"'
must have full rank (for comprehensive details see Chapter 2 in Utkin 1992).
This implies that rank(CB) = m since rank(F) S m . Therefore, in all the
analysis that follows it can be assumed, without loss of generality, that the
system is already in the canonical form of Lemma I . It will be useful to partition
the matrix F C in a compatible way to the canonical form. T o this end let
F C = [ F l F2]
where F, E [ ~ 4 " and- " R'nX"'.In this way FCB = F2B2 and, in particu-
F2 E' )
lar, the square matrix F2 is non-singular. By assumption f(.r, u , t) E Im B and
therefore the sliding motion is independent of the uncertainty. In addition,
because the canonical form of Lemma 1 can be viewed as a special case of the
'regular form' normally used in sliding mode controller design, it is well known
(see for example Zinober 1994) that the reduced-order sliding motion is
governed by a free motion with system matrix
A ; ~2 A l l - A ~ ~ F ; ' F ~ (14)
For a stable sliding motion, Asl must therefore be stable. If C1 E IW(P-'")~("-"')
is defined to be
 Sliding mode stabilization of uncertain system 1133

and K E I W " ' ~ ( ~ is

- ~defined
) as K = F;'FTI then it can easily be verified that
KCI = Fy1F1and so

The problem of designing a suitable hyperplane is therefore equivalent to an

output feedback pole assignment problem for the system (All, A,*, C,). It is
well established that invariant zeros play an important part in the sliding
motion-in particular, Hui and ~ a (1993) k identified that any invariant zeros of
( A , B, C) must appear in the sliding dynamics when only output information is
used. Intuitively, if the original system has invariant zeros, then by Lemma 2 the
eigenvalues of Ayl must appear in the spectrum of A i l and so the poles of
All - A12KCl cannot be assigned arbitrarily by choice of K . The intention is to
Downloaded by [North Carolina State University] at 02:21 18 December 2012

construct a new sub-system ( A l l , E l , el), which is both controllable and

observable, with the property that u(A;,) = o(A';,) U o(xll - El K c 1 ) . T o this
end partition the matrices A12 and A: as

A = [ and A = [
:]
I
m

122
where AIz2 E I W ( ~ - ' " - ' ) ~ ~ A;;~ ~ ( n - m - r ) x ( p - m ) and form a new sub-system
represented by the triple ( x l l , A,,,, e l ) where

Lemma3: The spectrum of A i l contains the invariant zeros of ( A , B , C ) and,

in particular
~ ( A l-
l A I ~ K C I )= U ( A ? ~U
) 4x11- ~ 1 2 2 K c 1 )
Proof: For the proof see Appendix A. 0
It follows directly that for a stable sliding motion the invariant zeros of the
system ( A , B , C ) must lie in the open left half-plane and the triple
( x l l , AIZ2, el) must be stabilizable with respect to output feedback. (The
phrase 'stabilizable with respect to output feedback' will indicate that for the
linear system ( A . B, C) there exists a fixed gain K such that the matrix
A - B K C is stable.) The matrix A122 is not necessarily full rank. Suppose
rank(AI2*) = tn', then it is possible to construct a matrix of elementary column
operations T,,, E (W"'Xnl such that
Ai22Tm, = [ E l 01 (19)
where z1 and
~(ff-nl-r)~"~ ' is of full rank. If K,,,c= T , ! K and K,,,. is
partitioned as

then
2'11 - A I ? ~ K=~ I - [ E l 0]Km.cI = xI1
- BlK1el
and ( x l l , e l ) is stabilizable by output feedback if and only if (All, E l , e l )
is stabilizable by output feedback. (The reason for the somewhat tortuous
 1134 C. Edwards and S. K. Spurgeon

argument to establish the sub-system ( A l l , E l , E l ) will be made clear in $2.3,

where the results of Davison and Kimura will be utilized.) As a result of the
preceding argument the 'only if' direction of the following proposition has been
established.

Proposition 1: There exists a matrix F defining a surface 9 which provides a

stable sliding motion with a unique equivalent control if and only if
(i) the invariant zeros of ( A , B, C) lie in the open left half-plane;
(ii) the triple ( A l l , E l , E l ) is stabilizable with respect to output feedback.

Remark 1: The approach presented in this section is more general than that of
Downloaded by [North Carolina State University] at 02:21 18 December 2012

El-Khazali and DeCarlo (1991) since it can be shown that their assumptions
require the nominal linear system ( A , B, C) to have no invariant zeros (for
details see Appendix C). Furthermore, the sub-block AI2 is required to fulfil
certain rank conditions for the algorithm for the hyperlane design given by
El-Khazali and DeCarlo (1991) to be valid. The framework developed here is
therefore applicable to a larger class of systems. 0

In the following section it will be demonstrated that the conditions given

above are sufficient for a control law of the type given in (5) to exist, which
induces a stable sliding motion on 9 ' and quadratically stabilizes the uncertain
system (1).

2.2. Controller fortnulation

Let ( A , B, C) be the nominal linear part of the uncertain system given in
(1). Suppose that p > m , rank(CB) = m and any invariant zeros of the nominal
linear system lie in the open left half-plane. Without loss of generality, the triple
(A, B, C) can be assumed to be in the canonical form of Lemma 1 and from
equations (17), (18) and (19) the sub-system ( x l l , E l , e l ) can be determined.
Assume this subsystem is o u t p t feedback stabilizable so there exists a
K , ~ m ' x ( p - m ) such that A l l - B ~ K , is~ stable.
, Let

and matrix T,,,. E R"'Xn'is defined in (19). Then it follows that

and so the matrix A I I - A12KCI is stable. Choose

where F2 E RnfX"'is non-singular and will be defined later. Introduce a non-sing-

ular state transformation x ++ ?x = z where

and C1 is defined in (15). In this new coordinate system the system triple
( 2 ,3, ~ 2 has) the property that
 Sliding mode stabilization of uncertain systems 1135

where All = A l l - A 1 2 K C I .The square system ( A , 3, FC) is relative degree

one and minimum phase because All is stable. It is well known that such
systems are output feedback stabilizable. Using the construction given by G u
(1990), the symmetric positive definite matrix

partitioned as in (22) and where P1 is a Lyapunov matrix for All satisfies the
structural constraint
Downloaded by [North Carolina State University] at 02:21 18 December 2012

p3 = ~ F T (24)
if the design matrix F2 2 BFP~.
The matrix P can be shown to be a Lyapunov
matrix for A. 2 A - y h ~ E for all y > yo. (The precise value of the lower
bound yo is given in Appendix B.) Define a variable structure control law
depending only on outputs by
u ( t ) = - y F y ( t ) - v, (25)
where y > yo and vy is the discontinuous vector given by

lo otherwise
and p ( y ) is the positive scalar function
P ( Y ) = (kl~IlFYll+ 46J)+ ~ 2 ) / ( 1- k l ) (27)
where y2 is a positive design scalar. In the new coordinate system the uncertain
system ( 1 ) can be written as
i(t) = Az(t) + 6 ( u ( t )+ a t , X , u)) (28)
Proposition 2: The variable structure control law above quadratically stabilizes
the uncertain system given in (28).
Proof: Consider as a candidate Lyapunov function the positive definite expres-
sion
V ( z ) f zTPz (29)
Taking derivatives along the system trajectory and using the structural constraint
from (24) gives
v = z T ( A T P + P A - 2 Y ( ~ ? ) T ~ c +) z2 z T P 6 ( < - v y )
+
= Z ~ L ( ~ ) 2yTFT(E
Z - vy)
T T
zs z T L ( y ) z - 2.Y F v, + 2lIFylI IISII
= z T U y ) z - 2~(y)llFyIl+ 211F~11l E11
< z T L ( y ) z - ~IIFYII(P(Y)
- klllull - 9 % ~ ) )
where L ( y ) = PAo + A:P.
 1136 C. Edwards and S. K. Spurgeon

But by definition

and so by rearranging

Using (30) in the inequality for the Lyapunov derivative

if z 0 and y > yo
Downloaded by [North Carolina State University] at 02:21 18 December 2012

V < z T ~ ( y ) z- 2y21(Fyll< 0 f

and therefore the system is quadratically stable.

Corollary: A sliding motion rakes place on the surface Y defined in (4) in the
donzain Q = { z E Rn : ( ( B ~ ' A ~ z (y 2( )< where the matrix sub-block representsA~L
the last p rows of the closed loop matrix Ao.
Proof: Let the 'switching functionls(z) = F ~ Zthen
, from (28) it follows that
s = F~A"Z + F2Bz(<- v,.)
Let the 'Lyapunov function' Vc : R"' 4 R be defined by
V,(s) = ~ S ~ ( F ; ' ) FT'S
~P~
Then using the fact that F:= P2B2 it follows that ( F ; ' ) ~ P ~ F ; ' F ~ A=~
6 ; ' ~ ; . Then, arguing as in Proposition 2, it can be verified that
= 2sTA;z + 2sT(E - v ~ ) ~ ) ( s I ( I I B ; ' A ~ z I I - 2y2))~l(
Therefore V ,< 0 if z E Q and from Theorem 1 in DeCarlo er al. (1988) a sliding
motion takes place in the domain 52. From Proposition 2 the states z(r) are
quadratically stab!e and so in finite time z(r) E Q and a sliding motion will be
attained. 0
The 'if' direction of Proposition 1 has been proved and hence, in conjunction
with the results of $2.1, the proof of Proposition 1 is complete.
Remark 2: The controller synthesis is more straightforward than that proposed
by El-Khazali and DeCarlo (1991) and does not require the structural constraint
on the controller imposed by ~ a and
k Hui (1993).
Remark3: In the special case when the number of outputs equals the number
of inputs the necessary and sufficient condition for the existence of a controller
is that the nominal linear system ( A , B, C) is minimum phase. In this case it can
be readily demonstrated that F has no effect on the dynamics of the sliding
motion. 0

2.3. Design considerations

This short section addresses the practical problem of designing the sliding
surface hyperplane matrix F and determining when the triple (2",E l , is cl)
output feedback stabilizable.
 Sliding mode srabilizarion of uncertain s y s t e m 1137

Lemma 4: If the pair ( A , B ) is cotnpletely conrrollable then the pair ( A l l , B1)

el)
is completely conrrollable and ( A l l , is completely observable
Proof: For the proof see Appendix A. 0
If ( A , B ) is controllable and the scalar inequality

holds then the 'Kimura-Davison' conditions are satisfied for the system
( A I IE, l , 2 . 1 )
since inequality (31) can be rewritten as ( n - m - r ) C ( p - m ) +
m ' - 1 and the system ( A l l E , l , 2')) with m' inputs and p - rn outputs is
controllable, observable and has input and output distribution matrices of full
rank. Under these circumstances the poles of All - El K ~ Z can , be assigned
Downloaded by [North Carolina State University] at 02:21 18 December 2012

arbitrarily by output feedback for 'almost all pairs ( E l , C , ) ' . For details see
Davison (1970), Davison and Wang (1975), Kimura (1975, 1977) and, more
recently, Misra and Patel (1989).

3. Examples
The following non-trivial examples substantiate the practicality of the pro-
posed design procedure.

3.1. Example I
Consider the system from Hui and ~ a (1993)
k

In the canonical form of Lemma 1 the system becomes

and
0 0.3417 -0.93981
cc= [0 0.9398 0.3417
from which B2 = -3.9016 and the orthogonal matrix
, [,:I,
= -0.93981
0.3417
can be identified. Also, it can be verified that the triple ( A l l E, l , 2 ' 1 ) is given by

Here, r = 0 and so the original systzm does not possess any invariant zeros.
Arbitrary placement of the poles of A l l - El K I C I where K 1 E R is clearly not
possible. However, if K l = -1.0556 then it can be verified that a ( A l l -
E I ~ l C 1= ){ - I , - 2 ) from which
 1138 C . Edwnrds and S. K . Spurgeon

where F2 is a non-zero scalar, which will be computed later. Transforming the

system into the control law design canonical form using f defined in (21)
generates the sub-block

21, = [-1.5816
1.4071 -1.4184
where o ( A l l ) = {-I, -2) by construction. It can be verified that
Downloaded by [North Carolina State University] at 02:21 18 December 2012

is a Lyanpunov matrix for All and that if the design parameter P2 = 1 then
F2 = -3.9016. From Lemma B.5 in Appendix B it can be checked that
yo = 0.2452 and substituting for F2 in (32) gives
F = [5.0741 2.53701
Note that if F2 = -1.5378 then F = [2 11 which is the solution of Hui and ~ a k
(1993).

3.2. Exntnple 2
The following eighth-order linear system models the small-perturbation rigid
body motion of a helicopter about the hover condition. The states are given by
Pitch attitude (rad)
Roll attitude (rad)
Body roll rate (rad s-I)
Body pitch rate (rad s-I)
Body yaw rate (rads-')
Forward velocity (ft s-I)
Lateral velocity (ft s-')
Normal velocity (ft s-')
The system matrices are
 Sliding mode stabilization of uncertain systems 1139

with inputs
Main rotor colective (deg)
Longitudinal cyclic (deg)
Lateral cyclic (deg)
Tail rotor collective (deg)
Downloaded by [North Carolina State University] at 02:21 18 December 2012

and outputs
Heave velocity (ft s-')
Pitch attitude (rad)
Roll attitude (rad)
Heading rate (ft s-I)
Body pitch rate (rad s-')
Body roll rate (rads-')

In the canonical form of Lemma 1

from which r = p - m and so

It can be checked that a(Ayl) = (-0.0014, -0.0054) and so the method is

applicable. Also, from (18)

and

A122 = ~:~~~~ -0.0127

0.9919
1.0001
-0.0006 O.OO1Ol
-0.0298

A possible choice for El is

8, =[ -0.0026
1.0034
- 1.0013
O]
 1140 C. Edwards and S. K. Spurgeon

where

Because is the identity and ( i l l , E l ) is controllable the poles of

A l l - El K I el where K IE tRZx2can be assigned arbitrarily.

Remark4: It should be noted that this system does not satisfy the assumptions
of El-Khazali and DeCarlo (1991) because of the presence of the invariant zeros
at {-0.0014, -0.0054). 0
Downloaded by [North Carolina State University] at 02:21 18 December 2012

4. Conclusions
A new design procedure has been presented to synthesize robust output
feedback controllers for uncertain systems based on sliding mode concepts. The
class of systems to which the results apply has been identified and includes the
requirement that the nominal linear system is minimum phase. Emphasis has
been placed on the tractability of the associated design procedure. The design of
the sliding surface utilizes established output feedback eigenvalue assignment
results. I t has been demonstrated that all the assumptions imposed on the
system pertain to the design of the sliding surface. The proposed controller,
which guarantees attainment of a sliding mode despite the presence of uncer-
tainty, requires no additional assumptions. The practicality of the procedure has
been substantiated by two non-trivial design examples.

Financial support from the provision of a Research Scholarship by British

Gas plc is gratefully acknowledged. The access to helicopter models provided by
DRA (Bedford) is very much appreciated. The helpful comments of an
anonymous reviewer were very useful in refining the ideas presented in this
paper.

Appendix A
Proof of Lemma 2: Let P(s) denote Rosenbrock's system matrix:

The invariant zeros of ( A , B, C) are defined to be

{s E @ : P(s) loses normal rank}
By assumption ( A , B, C) is in the canonical form of Lemma 1 and so

P(s) loses rank - sl - A l l

[O -7-11
-A12
$1 - AZ2
-T2
i2]
0
loses rank

"- *11 loses rank

0 -TI, -T2
 Sliding mode srabilization of uncertain systems 1141

since det ( B 2 )# 0. Substituting for All from (7) and repartitioning gives

sl-All -Al2
[[O - T I ] -T2

where T is defined in (13) and * represents a matrix sub-block, which plays no

part in the analysis. Because T is of full rank, the pencil on the right, and hence
P(s), loses rank if and only if

[
sI - ~y~ -&20

I
Downloaded by [North Carolina State University] at 02:21 18 December 2012

SI - Az2
-A;l
loses rank.
By construction, the pair (A;2,A ; ~ is
) completely observable and, therefore,
from the PHB observability rank test

rank[
Sf - A;2
-AZ1 ] = n - p - r for all s E c
Therefore
P(s) loses rank o det (sI - ~ 7 = 0~ )
and so the zeros of ( A , B, C ) are the eigenvalues of A:,.
Proof of Lemma3: Using the partition in (7),by definition

A l l - A12KC1= - [O(n-m)x(n-p) A12K1

A:, [ A % I A ~ -IA I U K I ]
= [ O 711-A122~?1
Therefore, a ( A r 1 )= u ( A I I- A I 2KCI)= ~ J ( A ' ; ~U) ~ (- ~ 21 2 K2~
? ~ )as~claimed.
From Lemma 2 the spectrum of A:, represents the invariant zeros of ( A , B, C ) .
0
Proof of Lemma 4: Because the pair ( A , B ) is in the canonical form of Lemma
I, which is a special case of the 'regular form' used for sliding mode design, it is
well known that pair ( A , B) is completely controllable if and only if the pair
, 1 2 )is completely controllable. Therefore from the PHB rank test
( A I lA
rank [sf - A l l A I 2 ]= n - m for all s E @
Substituting for A l l from (7) and A12from (17) gives
S I - A : ~ [ A h AY~I]
rank[
s1 - X I ,
A '2'] = n - m
A122 for all s E c
 1142 C. Edwards and S. K. Spurgeon

This implies
rank [sl - All A I z 2 ]= n - m - r for all s E C
and therefore ( A I IA, 1 2 2 ) is completely controllable by the PHB test. By
construction ( A l l A, l z 2 ) is controllable if and only if (Al1,B1) is controllable
and s o the first part of the lemma is established. Applying the PHB observabil-
ity rank test to the pair (. A.. c,)
l , , ..
rank = rankr-jt'2

-
-Az1
-A

4-rn
;;I,
~ 1- A Z 2

t 2] + ( p - m )
I for all s E C
Downloaded by [North Carolina State University] at 02:21 18 December 2012

= -nk[

Now ( A ; , A $ ) is observable and s o by the PHB observability test

and therefore
rnnk[., =n - m - r for all s E C

and hence ( A l l E, l ) is completely obervable. 0

Appendix B
Let P be the symmetric positive definite matrix defined in (23) where the
sub-block P2 is a design matrix and PI satisfies the Lyapunov equation
plAIl + TI PI = -Ql ( B 1)
for some symmetric positive define matrix Q l . For notational convenience let
Q2 2 P,A,, + A:~P, (B2 a)
Q3 p2AZ2+ 2T2p2 (B2 b)
and let
YO A L x ( ( ~ ; 1 ) T ( ~+3 Q:Q;'Q~)F;') (B 3 )
This scalar is well defined since the matrix o n the right is symmetric and
therefore has n o complex eigenvalues.
LemmaB.5: The symmetric matrir L ( y ) 2 PAo + A ~ where
P A. = 2 -y 6 ~ e
is negative definite if and only i f y > yo.
Proof: Using the definition of A. and the structural constraint (24) it follows
that
L ( Y ) 2 PA^ + A ~ = P PA + ;iTp - Y P ~ F - C yi2T~T6T~
= PA + A T p - 2 y ( ~ e ) ~ F e
Using the partitions of 2 , F e and P from (22) and (23) respectively, and the
definitions of Q2 and Q3 from ( B 2 a ) and ( B 2 b ) , it can be verified that
 Sliding mode stabilization of uncertain sysrems 1143

The matrix L(y) < 0 if and only if T ' L ( ~ ) T < 0 where T is any non-singular

which is clearly non-singular and evaluating T T ~ ( y ) T

implies
Downloaded by [North Carolina State University] at 02:21 18 December 2012

and the proof is complete. 0

Appendix C
This appendix demonstrates that the results of El-Khazali and DeCarlo
(1991, 1992) do not apply to systems (with more outputs than inputs) which
have invariant zeros. Assume as in El-Khazali and DeCarlo (1991,1992) that the
triple (91,933, %) is in 'regular form'

where dl, E I W ( ~ - " ' ) ~ ( " - ~ ) a2E R m x m , E I W P ~ ( ~ - " ' ) and q2E RPX'". El-
Khazali and De Carlo (i991) require that the pair ( d l l rg l ) is completely
observable where
dII2 dl, - d 1 2 % ; L ~ I and 9M ~ % ~
where M E RpX(p-m)is a full rank left annihilator of q2 and %iL E RmXPis a
left pseudo-inverse of %,. Assume that the system given in (C 1) is, in fact, in
the canonical form of Lemma 1 (which is just a special case of the regular form),
then recalling the notation of 5 2 it can be shown that

where E R r x r for r 3 0, TI E Rpx(p-"'), T2 E Rpxm and T = [TI T2] is an

orthogonal matrix. As a result of the orthogonality conditions
T:(e2 2 TTT2 = 0 and rank(T1) = p -m
so M = TI is a choice for the full rank left annihilator of %,. Also, the
orthogonality conditions imply
T;g2 9 T T T ~ = I,,,
and so %iL = TT is a left pseudo-inverse of %,. Therefore
GI1= dll1 - d I 2 % ; k l = dl, - dll2TT[0 Tl] = dI1 ( c 3)
 1144 Sliding mode stabiliznrion of uncertain systems

and
G I = ~ ~ ' ( 2 =, T:[O TI] = [O [(p-t,,)] (C 4)
If (A,%,%) has invariant zeros then from Lemma 2 the zeros are the
eigenvalues of A:, and r > 0. Examining the structure of d I 1from (C 2) and 4,
from (C 4) it can be seen that
~ ( A Y ~ C) a(.dll- G e l ) for all G E I W ( " - ' " ) ~ P
and so (.all,4,) is not completely observable.
REFERENCES
DAVISON, E. J . , 1970. On pole assignment in linear systems with incomplete state feedback.
IEEE Transactions on A~ttomaticControl, 15, 348-351.
Downloaded by [North Carolina State University] at 02:21 18 December 2012

DAVISON, E. J . . and WANG,S. H., 1975. On pole assignment in linear multivariable systems
using output feedback. IEEE Transactions on Automatic Conrrol, 20, 516-518.
DECARLO.R. A,, ZAK, S. H., and Mnri-HEWS.G . P., 1988, Variable structure control of
nonlinear multivariable systems. Proceedings of the Instiritre of Electrical and Elecrron-
ics Etlgitreers, 76, 2 12-232.
€1.-KHAZAI.~,R., and DECAHI-0,R. A,. 1991, Variable structure output feedback control:
switching surface design. Proceedings of the 29th Allerron Conference on Commimica-
rions. Control and Contpuring, Monticello, Illinois. U.S.A., pp. 430-439; 1992,
Variable structure output feedback control. Proceedings of rhe American Corttrol
Corrference, Chicago, Illinois, U.S.A., pp. 871-875.
EDWARDS, C., and SPURGEON, S. K., 1994. On the development of discontinuous observers.
Itrrernatiotral Jorrrrral of Conrrol. 59, 1211-1229.
HUI.S.. and ZAK,S. H.. 1993, Robust output feedback stabilization of uncertain systems with
bounded controllers. Iriternarional Jortrnul of Robust and Nonlirrmr Conrrol, 1990, 3 ,
115-132.
G u , G., 1990. Stabilizability conditions of multivariable uncertain systems via output feedback
control. IEEE Transactions on Auronlatic Control, 35, 924-927.
KIMURA, H., 1975, Pole assignment by gain output feedback. IEEE Transactions on Automatic
Control. 20. 509-516; 1977, A further result on the problem of pole assignment by
output feedback. Ibid.. 22. 458-463.
MISRA,P., and PATEL, R. V.. 1989. Numerical algorithms for eigenvalue assignment by
constant and dynamic output feedback. IEEE Transactions on Automatic Conrrol, 34,
579-588.
R Y A NE.
, P., and CORLESS, M., 1984. Ultimate boundedness and asymptotic stability of a class
of uncertain dynamical systems via continuous and discontinuous feedback control.
/MA Journul of Mathenlatics and Control Information, 1, 223-242.
SI.OTINE,.I-J. E., HEDRICK, J. K., and MISAWA,E. A.. 1987, On sliding observers for
nonlinear systems. Jourtrul of Dynamic Sysrenls. Meusuremenr and Conrrol, 109,
245-252.
SPURGEON. S. K., and DAVIES,R., 1993, A nonlinear control strategy for robust sliding mode
performance in the presence of unmatched uncertainty. Inrernational Journal of
Conrrol. 57. 1107-1 123.
UTKIN,V. I.. 1992, Slieng Modes in Control Optimisation (New York: Springer-Verlag).
W~r.corr. B. L., and ZAK. S. H.. 1986, Observation of dynamical systems in the presence of
bounded nonlinearities/uncertainties. Proceedings of 25th Conference on Decision and
Conrrol, Athens, Greece, pp. 961-966.
WIIITE.B. A , , 1990, Applications of output feedback in variable structure control systems. In
Deterministic Control of Uncertain Systems, edited by A . S. I. Zinober (Stevenage.
U.K.: Peter Peregrinus).
ZAK, S. H., and Hur. S., 1993, Output feedback variable structure controllers and state
estimators for uncertain/nonlinear dynamic systems. Proceedings of the lnsiitrrrion of
Electrical Engineers, Pt D , 140, 41-50.
ZINOUER. A. S. I., 1994. An introduction to sliding mode variable structure control. In
Variuble Structure and Lyapunov Control edited by A. S. I. Zinober (London, U.K.:
Springer-Verlag).