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

On: 25 March 2013, At: 00:58

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:
[Link]

Observer design for a class of nonlinear systems

a a
SEKHAR RAGHAVAN & J. KARL HEDRICK
a
Department of Mechanical Engineering, University of California, Berkeley, U.S.A.
Version of record first published: 11 Aug 2011.

To cite this article: SEKHAR RAGHAVAN & J. KARL HEDRICK (1994): Observer design for a class of nonlinear systems,
International Journal of Control, 59:2, 515-528

To link to this article: [Link]

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: [Link]

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.
 INT. J. CONTROL, 1994, VOL. 59, No.2, 515-528

Observer design for a class of nonlinear systems

SEKHAR RAGHAVANt:j: and J. KARL HEDRICKt

A viable design methodology to construct observers for a class of nonlinear

systems is developed. The proposed technique is based on the off-line solution
of a Riccati equation, and can be solved using commercially available software
packages. For globally valid results, the class of systems considered is charac-
terized by globally Lipschitz nonlinearities. Local results relax this assumption
to only a local requirement. For a more general description of nonlinear
Downloaded by [North Carolina State University] at 00:58 25 March 2013

systems, the methodology yields approximate observers, locally. The proposed

theory is used to design an observer for a single-link flexible joint robot. This
observer estimates the robot link variables based on the joint measurements.

1. Introduction
The problem of designing nonlinear observers has been studied for a long
time. Much of the effort has resulted in extensions of the linear Luenberger
observer (Kailath 1980): examples of this are the extended Kalman filter,
psuedo linearization techniques etc (see Misawa and Hedrick 1989 for a survey).
These techniques are valid in a small range around the operating point. They
also frequently require heavy real-time computation. Recently, geometric tech-
niques were proposed to build exact observers for a general description of
nonlinear systems (Krener and Isidori 1983, and Keller 1987). However, the
conditions that are required to be satisfied by these observers are extremely
stringent, making the applicable class very small.
This paper represents an effort towards developing viable methods to design
and implement observers for inherently nonlinear systems. We describe a new
computationally-viable design methodology for constructing observers for a class
of nonlinear systems. These systems are driven by nonlinear functions which are
Lipschitz in nature. The underlying theory makes use of the methods developed
for the quadratic stabilization of uncertain systems (Khargonekar et al. 1990,
Peterson 1987 and Peterson and Hollot, 1986). In particular, we develop dual
results applicable to the observer problem at hand.

2. System description and motivation

First, we define a Lipschitz function.
Definition: A (vector-valued) function f(x) is said to be globally Lipschitz if
there exists a constant y such that for all Xl, X2 E R", the following inequality
holds:
(1)

Received 8 June 1992. Revised 8 December1992.

t Department of Mechanical Engineering, University of California, Berkeley, U.S.A.
*
Currently with the Power Systems Research Department, General Motors Research,
Warren, Michigan, U.S.A.
002()'7179/94 $10.00 © 1994 Taylor & FrancisLtd
 516 S. Raghavan and J. K. Hedrick

In this case, y is said to be the Lipschitz constant of [, Here, and in the rest of
this paper, /·1 denotes euclidean norms. 0
If [ is a differentiable function with bounded partial derivatives, then y is
simply the upper bound on the norm of the jacobian matrix for the function, the
upper bound taken over the entire R", However, in general, a Lipschitz function
may not be differentiable.
The systems we consider have the form given by
x = Ax + g(t, U, y) + [(t, u, x)
(2)
y = Cx
where the state x E R"; A E R n x n . y E Rq represents the measurements of the
Downloaded by [North Carolina State University] at 00:58 25 March 2013

system, and the input U E R'", The matrices A and C are constants, and, g:
R x R'" x Rq ~ R"; [: R x R'" x R" ~ R", The nonlinearity [(',', x) is
assumed to be globally Lipschitz in x with a Lipschitz constant y. i.e.
I[(t, U, x) - [(t, u, x)1 < ylx - xl. for all U E R'", t E R (3)
We also assume that the pair (A, C) is detectable.
The assumption that [ is Lipschitz globally may be relaxed to assume that [
is only locally Lipschitz. All the results in the ensuing sections will then be valid
in some local neighbourhood around a nominal point. (Hereafter, when we say
that [ is Lipschitz, we will mean that the inequality (3) is satisfied.)
Systems of type (2) are not uncommon: mechanical systems frequently
contain Lipschitz-type nonlinearities-trignometric nonlinearities which occur in
robotic applications, a nonlinear softening spring etc. Nonlinearities which are
square or cubic in nature are not globally Lipschitz; however, they are locally
so. Moreover, when such functions occur in physical systems, they frequently
have a saturation in their growth rate, making them globally Lipschitz functions.
Frequently, we make the system measurements part of the system state, so
the assumption that the output is linear in the state is justified. The assumption
that the system dynamics are linear in the input is usually true for mechanical
systems because the input is usually a torque or a force which enters the
dynamics linearly due to the form of Newton's laws. In summary, while the class
of systems we are addressing is not exhaustive, it is fairly large from an
engineering point-of-view.

3. A background result
Consider the system given by (2). For this system, we construct an observer
as follows:
£ = Ax + g(t, U, y) + [(t, u, x) + L(y - Cx) (4)
where the observed state is denoted by X, and L E R'":" is the observer gain
matrix. Then, the error in the state estimate, x := x - X, has the following
dynamics:
t = (A - LC)x + [(t, u, x) - [(t, u, x) (5)
The design problem is to choose the observer gain L so that (5) is made stable.
The following result which establishes a condition in this respect was established
by Thau (1973).
 Observer design for class of nonlinear systems 517

Theorem 1 (Thau 1973): Given the system given by (2) and its observer (4). If
Y < Amin(Q)/2Am.x(P), where P and Q are positive definite, symmetric matrices
satisfying the Lyapunov equation
(A - LC)Tp + P(A - LC) = -Q (6)
Then limr-oo x = O.
Thus, Theorem 1 provides a procedure to check the stability of the error
dynamics after having picked L. However, it does not suggest a method to
design a stable observer. Stated otherwise, given the Lipschitz constant of f to
be y, it is not apparent how one chooses the observer-gain matrix L so that we
satisfy y < Amin(Q)/2Ama.{P).
Downloaded by [North Carolina State University] at 00:58 25 March 2013

One might be tempted to think that an effective design technique may be

placing the poles of (A - LC) by a suitable choice of L. This line of thought
does not yield any results: for one thing, no clear relation between the
eigenvalues of (A - LC) and Amax(P) is evident. Secondly, Am a.{P) can be
influenced without necessarily affecting the eigenvalues of (A - LC). Both these
points are illustrated by the following example.

Example: Consider the matrix

A = [~ ~J (7)

and the solution of its Lyapunov equation, AT P + PA = - I. One can easily

verify that for e »» 1, Am.x( P) = e2/4. However, the eigenvalues of A remain
at 1. 0
The example above shows that it is not merely the eigenvalues of (A - LC),
but the entire structure of (A - LC) that influences Am a.{P), and thus our
conclusions about the stability of (5). The next section treats the design of
observers for such systems under this broader perspective.

4. A new design procedure

4.1. Main result
In this section, we will formulate a computationally viable procedure for
designing observers for systems described by (2). The theory is based on the
methods employed in quadratic stabilization of uncertain systems (Khargonekar
et al 1990, Peterson, 1987 and Peterson and Hollot 1986). In this paper, we state
and prove some dual results. The dual results are directly useful in formulating a
design technique. Essentially, our goal is to get a method to choose L so that
the error dynamics (5) is stabilized. To this end, we will first establish two
intermediate results, Lemmas 1 and 2. We then use these results to formulate a
sufficient condition (Theorem 2) for stabilizing (5). Then we outline a computa-
tional procedure to calculate L.

Lemma 1: Given a stable n X n matrix A, and y> O. There exists a positive

definite, symmetric matrix PI such that
ATp\ + PIA + y2 p\ p\ + 1<0 (8)
 518 S. Raghavan and J. K. Hedrick

if and only if there exists another positive definite, symmetric matrix Pz such that
(9)

Proof of Lemma 1: We will show the necessity. The sufficiency part is similar.
Suppose that
ATp[ + PIA + rPIPI + 1<0
is satisfied. Now, define Pz := (Plr)-I. Obviously, Pz is a positive-definite and
symmetric matrix. Now, substituting for PI in the above inequality as
PI := (pzr)-I, we have
AT(Pzr)-1 + (Pzr)-IA + yZ(pzr)-I(pzr)-1 + 1<0
Downloaded by [North Carolina State University] at 00:58 25 March 2013

Multiplying the above inequality on the left and right by (y Pz)* and y Pz
respectively (where * denotes complex conjugate transpose), and noting that Pz
is a real matrix, we get
PzAT + APz + I + rPzPz < 0
which is the same as (9). This proves the necessity. Proving sufficiency follows
similar steps. 0

Lemma 2: Given the system (2) and its observer (4). If P is any positive definite
and symmetric matrix, then
2x T P(f(t, u, x) - f(t, u, x» '"
rx T PPX + xTx (10)

Proof of Lemma 2: Referring to (f(t, u, x) - f(t, u, x» as 1, we have

2x T pI'" 21xT pli
'" 21 pxllll
'" 2yl pxl· Ixl
where, we have used (3) in the last step. Now, it is easy to verify that for real
numbers, a and b, the inequality
aZ 2
ab '" - + b
4
is true. Thus, we get the following required result:
2x T pl '" (rxT PPX) + (xTx) o
Now we are ready to state the main result of this section.

Theorem 2: Given the system (2) and its observer (4). If there exists an e > 0
such that the following Algebraic Riccati Equation (ARE) has a symmetric,
positive definite solution P

AP + PAT + p(rl - ~CTC)p + I + el = 0 (11)

then, the observer gain L = (1/2e) PC T stabilizes the error dynamics, given (5) for
all f with a Lipschitz constant y.
 Observer design for class of nonlinear systems 519

Proof of Theorem 2: Suppose an e > 0 that solves the Riccati equation is found.
Then substituting L = (l/2e)PC T in (11) gives that
(A - LC)P + P(A - LC)T + Y PP + I + eI =0
Since e > 0, it is clear that
(A - LC)P + P(A - LC? + YPp + I <0 (12)
where the inequality above denotes that the matrix sum on the left-hand side is
negative definite. From this, the following is true:
(A - LC)P + P(A - LC? < 0
Then (A - LC)T and, consequently, (A - LC) are stable matrices. Using
Downloaded by [North Carolina State University] at 00:58 25 March 2013

Lemma 1, (12) implies that there exists a positive definite and symmetric matrix
PI satisfying the inequality
(A - LC)TPI + PI(A - LC) + yPIPI + 1<0 (13)
Now, consider the Lyapunov function, V = x PIX. Differentiating V with
T

respect to time along the trajectories of the error dynamics (5), we get
V = xT«A - LC)T PI + PI(A - LC»x + 2x T Pj(f(t, u, x) - f(t, u, x»
Using Lemma 2, we see that
V"" xT«A - LC)TPI + PI(A - LC»x + YXTPIPIX + xTx
From (13), V is clearly negative definite. Hence, x -> 0 as t -> 00. 0
Theorem 2 thus establishes a sufficient condition for stabilizing the error
dynamics (5). We note here that the resulting design is an H observer design, 00

(Nagpal and Khargonekar 1991).

Computationally, one can follow the conceptual procedure below to evaluate
the necessary observer gain L. This procedure can be easily implemented on
any' commercial control systems software package.
Algorithm 1- Procedure to obtain the observer gain
Step 1. Set e to a positive value.
Step 2. Solve the following ARE

AP + PAT + P(YI - ~CTC)p + I + «t = 0

Step 3. If P is symmetric and positive definite, then setting L = (1/2e)PCT in

the observer (4) gives stable error dynamics (5).
Step 4. If not, set e = e/2, and go to Step 2. If e is below some precision value,
abandon method.
Remark 1: The iteration over e is a binary search. That this is enough to find
the required e when it exists becomes clear in the next section. 0
Remark 2: When Algorithm 1 fails, it means that there does not exist a single
pair (P, L), such that V < 0 along the error trajectories for every f of the form
(3), when this V is chosen as xT PX. We will elaborate on the significance of this
fact in Remark 4, leading to the development in § 4 for the case when a
straightforward application of Algorithm 1 fails. 0
 520 S. Raghavan and J. K. Hedrick

5. Quadratic stabilization error dynamics

In this section we will establish that a modification of Theorem 2 provides a
stronger result. The modification uses the following definition.
Definition: Given system (2) and its observer (4). We say that the observer
gain L quadratically stabilizes the error dynamics if there exists a Lyapunov
function, V = xT P'i , P positive definite and symmetric, such that V < 0 along
the trajectories of (5) for all nonlinearities f with a Lipschitz constant y. 0
The above definition stems from the usual definition for quadratic stability of
a family of systems (Khargonekar et al. 1990), and (Peterson 1987). We have
adapted it to suit the nonlinear observer problem. The stronger version of
Theorem 2 is given next.
Downloaded by [North Carolina State University] at 00:58 25 March 2013

Theorem 3: Given the system (2) and its observer (4). If there exists an e > 0,
and a positive-definite and symmetric matrix Q such that the ARE

AP + PAT + P(YI - ~CTC)p + I +Q =0

has a positive definite solution, then, the error dynamics (5) can be quadratically
stabilized.
Conversely, if the error dynamics (5) can be quadratically stabilized by some
choice of L, then there exists an e* > 0, and a positive-definite matrix Q such that
for all e E (0, s"), the ARE above admits a symmetric positive-definite solution.
Proof of Theorem 3: Sufficiency was proved in Theorem 2. A detailed proof of
the necessity is given in the Appendix. 0
Remark 3: In Theorem 2, a specific form of Q( = e I) was used. This reduces
the unknown variable to simply e in Algorithm 1. Also, if a certain e* solves the
ARE (11), then, by Theorem 3, all e E (0, E*) also solve (11). Thus, in
Algorithm 1, it is sufficient to carry out a binary search over e. 0
Remark 4: From Theorem 2, it is clear that Algorithm 1 tries to find a single
gain L that can lead to stable error dynamics for every member in a class of
nonlinear functions (that obey (3». Thus, it neglects the structure of the given
nonlinear function. This can potentially lead to conservative results; i.e. the
algorithm would fail when there indeed exists an L that would stabilize the error
dynamics for the given f. It would be useful to take advantage of our knowledge
on the structure of the function f( t, U, x), in addition to the fact its Lipschitz
constant is y. In the next section, we use a state-transformation procedure to
address this question. The transformation matrices that are picked are problem-
specific, and they reflect the structure of the given nonlinearity. Using these
transformations, we solve for the observer gain in the transformed coordinates.
o
6. State transformations
To illustrate the benefit of state transformations, we will consider a second-
order example. Let a system of the form (2) be given with
-2
A = [ 3 C = [0 1]
 Observer design for class of nonlinear systems 521

and

f(t, u, x) = [1'5Si~(Xl)J
Let us now consider an observer of the type (4) with the gain matrix as

L =DJ
Here, we have not followed any particular method to select L, but simply made
(A - LC) a stable matrix. Indeed, the error dynamics are of the form (5), with

A - LC = [-~ -n
Downloaded by [North Carolina State University] at 00:58 25 March 2013

Notice that the Lipschitz constant Y = 1·5. In order to prove the stability of
the resulting error dynamics, we can check if the condition imposed by Theorem
1 holds; i.e. whether 1/2YmaxCP) > Y where P is solved from
(A - LC)Tp + peA - LC) = -I

A quick calculation reveals that Ymax(P) = 0·36, so that 1/2Amax(P) = 1·38

which is less than y. Thus, we cannot guarantee observer stability at this stage.
Let us now consider the following state transformation, z = Tx, with
T = diag (3, 0·4)
The error-system represented in the new coordinates is given by
t = Aez + !(z) - le2)
where
Ae = T(A - LC)T- 1

1(,) = Tf(T-' ,) = [04 1:,;, (~ )]

x

It is seen that y = 0·2. Further, when the Lyapunov equation is solved for the
transformed (A - LC) matrix, we get Amax(P) = 1·47. It is immediate, then, to
verify that the stability condition holds in the transformed coordinates.
Thus, we see that the designed observer gives stable error dynamics for the
system at hand. The fact that the stability condition was not verified earlier
merely implies that there is at least one function f( t, U, x) (with a non-zero entry
in its first channel) having a Lipschitz constant y = 1·5, for which the stability of
its error dynamics cannot be established using the Lyapunov function candidate,
iT Pi. Thus, a straightforward calculation does not take into account the special
structure of the nonlinearity under consideration. The transformation used to
prove the stability takes advantage of the fact that one of the channels of f is
zero. If this were not true, the state-transformation may not have resulted in a
much lower value of Y (as it did in this case). Of course, for stability, not only
does the value of y matter but also how exactly the matrix A gets transformed,
in the sense that the structure of the transformed matrix manifests itself in the
form of the new Lyapunov solution P.
 522 S. Raghavan and f. K. Hedrick

Nonlinearities arising out of many physical systems have some special

structures associated with them. We would like to exploit this fact for construct-
ing stable observers. A systematic way of doing this is a topic for future
research. As a practical illustration of this procedure, in the next section we
design a nonlinear observer for a single-link robot with a flexible joint.

7. An observer for a single-link flexible joint robot

Experimental evidence (Sweet and Good 1984) has revealed that the
performance of a large class of robots is severely limited when joint flexibility
introduced by their transmission is not considered. The joint flexibility intro-
duces low-frequency resonance effects which, when unaccounted for, limit the
Downloaded by [North Carolina State University] at 00:58 25 March 2013

robot's performance range. With this in view, recent literature has focused on
the control of flexible joint robots, e.g. Spong (1987). In this paper, Spong
proposed the use of an I/O linearizing control law for these robots. This control
however, uses information on all states-which for a single-link flexible-joint
robot are the joint position and velocity, and the link position and velocity. For
physical reasons, while one can easily measure the motor position and velocity,
the measurement of the other states are not trivial. In the present paper, we
present a nonlinear observer that estimates the link position and velocity based
on the inputs of motor position and velocity.
Figure 1 shows the schematic of a laboratory model of a single-link flexible
joint robot. In the figure, f m represents the inertia of the actuator, a dc motor,
and f l represents the inertia of the controlled link. em and e, are angular
rotations of the motor and the link respectively, and, W m and CO) are their
angular velocities. In general, these will be different from each other as
functions of time, due to the torsional compliance, k.
A state-space description of this system is given next.
8m = Wm

rom = ~(e, - em) - !!-.-wm + Kru

i; s; i:
8, = CO)
6Jj = -~( e) - em) - mgh sin (e,)
f, f)

cot leal Shaft

Encoaers
Figure 1. Schematic of an elastic robot.
 Observer design for class of nonlinear systems 523

The length of the link is given by 2b. B represents the viscous friction in the
motor bearing and the back-e.m.f. effects. The following are the simulation
parameters used in the simulations. They are representative of a laboratory
model that can be used to model a flexible-joint robot.
The system model can be represented by the following equations.
i = Ax + f(x) + g(y)u (14)
with the following values for the matrices.
o
[ -4~'6 ~]. [2~'6]
1
-1·25 48·6
A = 0 o o 1 ' g = 0
o o
Downloaded by [North Carolina State University] at 00:58 25 March 2013

19·5 -19·5 0
and

f ~ ]
=[ -3'33~in(O,)
System (14) is clearly in the form described by (2): y = Xl, and z = Cx, with

C = [b ~ ~ ~J
An attempt to solve the Riccati equation in Theorem 2 using Algorithm 1 to
obtain the observer gain in original coordinates was unsuccessful. However,
noticing that f has a zero entry in three of its channels, we take advantage of
this fact using a transformation of coordinates z = Tx, where
T = diag[l, 1,4,0·1]
Under this transformation, the new Lipschitz constant is ji = 0·083. Algorithm 1
was then employed to solve for the observer-gain [ in the transformed
coordinates. L in the original coordinates was then calculated to be L = r : L.
0.91
0·38 0,38]
6·71
L = 0.89 (15)
[ 1·45
1·02 4·16
The structure of the observer is given by
~ = Ax + g(u) + f(8,) + Ly (16)

System parameter (Units) Value

Motor inertia, L« (kg m2 ) 3·7 X 10- 3
Link inertia, J) (kg m2 ) 9·3 X 10- 3
Pointer mass, m (kg) 2·1 X 10- 1
Link length, 2b (rn) 3·0 X 10- 1
Torsional spring constant, k (Nm rad ") 1·8 X 10- 1
Viscous friction coefficient, B (Nm V-I) 4·6 X 10- 2
Amplifier gain, K, (NmV-l) 8 X 10- 2
 524 S. Raghavan and J. K. Hedrick

where,

y = ([~:J - cx) (17)

and, x = [Om' W m, OJ, W)]T represents the state-vector of the robot. The perform-
ance of the observer is shown in Figs 2 and 3, where the convergence of
estimation errors in 8, and ~ are seen. The input was such that the link position
tracks a sinusoidal trajectory.

8. Extensions to general nonlinear systems

While the main thrust of this paper has been to formulate a method to
design observers for the class of systems given by (2), a straightforward
Downloaded by [North Carolina State University] at 00:58 25 March 2013

extension leads to approximate observers for more general descriptions of

nonlinear systems. Consider a system such as
i = f(x) + g(X)u}
(18)
y = hex)
where f(·), g(') and h(') are continuously differentiable functions, with
f(O) = h(O) = O. Let us denote

A:=(::t=o; c:=(:;L=o
and
B := g(O)
Then the given system can be expanded as
i = Ax + Bu + fleX) + gl(X)U}
(19)
y = Cx + hl(x)

2~'':

...
o :

o I 2 3 4 '5 6
Timc(sec::)

Figure 2. Observer performance: link position (solid curve), and its estimate (broken curve)
in radians.

l:~://
·10 : :'

:::v
0 1 2 3 4 5 6
Time(sec)

Figure 3. Observer performance: link velocity (solid curve), and its estimate (broken curve)
rad s':".
 Observer design for class of nonlinear systems 525

where flex) is obtained from expanding f(x) in a Taylor series about x = 0 as

f(x) = f(O) + Ax + flex). The functions gj(x) and hl(x) are also obtained in a
similar manner.
For the observer design, we will neglect the functions gl (x) and h, (x).
However, we will take into account flex). The observer so designed will be
approximate, even locally, since we neglect the higher-order terms of g and h.
Since f(·) is continuous, so is fl(·). Therefore, for every bounded region
Q S R", of radius p around x = 0, we can find a constant y such that

If(xt) - f(x2)1 < ylXI - x21

for all Xl> X2 E Q. In other words, the function fl(·) is Lipschitz with a constant
y within the region Q.
Downloaded by [North Carolina State University] at 00:58 25 March 2013

One can now find the highest value of y for which an observer design is
possible for the following system.
x = Ax + Bu + h(X)}
(20)
y = Cx
This can be done by iterating with higher and higher values of y in Algorithm 1.
The highest possible value of y directly determines Q. Then, the observer for
the original nonlinear system is implemented as follows.
i = (A - LC)x + Bu + hex) + gl(x)u + Ly (21)
As mentioned before, the observer so designed will be approximate, even
locally, since we have neglected the higher-order terms of g and h. But we can
expect it to be better than a design that neglects all the higher-order terms, and
uses only the linearized version of the system.

9. Conclusions
This paper described a viable computational procedure to design observers
for a class of nonlinear systems. These systems are characterized by nonlinear
functions which are Lipschitz in nature. Nonlinear observer design is still a field
in its infancy, and we hope this paper represents a fruitful step forward. The
suggested methodology is based on the dual of the results from the theory of
stabilization of uncertain systems. We have applied the present methodology to
the design of an observer for the single-link flexible joint robot. This observer
estimates the link variables based on the estimates of the joint variables. We
also offer some insights into exploiting the special structure of the given
nonlinear systems through state transformations.
On some relevant literature, the problem of designing observers for systems
with Lipschitz nonlinearities was first considered by Thau (1973). Related work
can be found in the work of Kou et at. (1975), and Zak (1990). Another paper
by the present authors (Raghavan and Hedrick 1992) discusses the efficient use
of nonlinear observers in designing compensators.

ACKNOWLEDGMENTS

The authors are very grateful to Professor Andy Packard for his invaluable
help and advice. Thanks are also due to the anonymous referee whose
comments helped provide a shorter proof for Lemma 1.
 526 S. Raghavan and J. K. Hedrick

Appendix
Proof of Theorem 3: Sufficiency was proved in Theorem 2. We will prove the
necessity here. The proof given below uses the methodology given by Khargone-
kar et af. (1990) and Peterson (1987) for the case of quadratic stabilization of
uncertain systems.
Suppose the error dynamics are quadratically stabilizable by some L E R"?",
For this choice of L, there exists a positive definite matrix P such that (with I
defined as in the Proof of Lemma 2)
xT«A - LC)T P + P(A - LC»x + 2x T e] < 0
for all x'" 0 and for all f obeying (3.3) i.e.
Downloaded by [North Carolina State University] at 00:58 25 March 2013

xT«A - LC)T P + P(A - LC»x < - max {(2x T ph ITI < Ilx!} ,,;;; 0
It follows that
(xT«A - LC)T P + P(A - LC»X)2 > 4 max {(x T pJ)2: ITI < Ilx!}
Then, from Lemma A.l (see below), we get that
(xT«A - LC)T P + P(A - LC»x)2 - 4(y2xT PPX)(xTx) > 0
By Lemma A.2, this means there exists a A> 0 such that
XT(A2y2PP + A«A - LC)T P + P(A - LC» + I)x < 0
Calling P := AP, we have that
xT«(A - LC)Tp + P(A - LC» + Ipp + I)x < 0
Then, by Lemma 1, there exists another positive definite matrix PI such that
xT«(A - LC)P1 + P\(A - LC)T) + Ip\p\ + I)x < 0
i.e.
xT(AP\ + PIAT)x - 2x T LCP\x + IxTp\p\x + xTx < 0
for all x'" O. Thus, we have
xT(AP\ + PIAT)x + IxTp\p\x + xTx < 0
for all x'" 0 and such that CPlx = O. It follows from Lemma A.3 that there
exists a a* > 0 such that
Q* := a*PICTCP\ - API - P\A T - Ip\P1 - I
is positive definite. This implies that
API + P\A T + P\(II - a*CTC)p\ + I + Q* = 0
Calling f* = l/a* gives the required ARE. Also, notice that when o" is replaced
by a> a*, the ARE holds for a different Q > Q*. Calling e = 1/a then proves
the Theorem. 0
LemmaA.l: Given the system (2)-(3) and its observer (4), let 1:= f(t, u, x)-
f(t, 1I, x). If P is a positive definite and symmetric matrix, then
max {(x T PJ)2: FI < Ilxl} = (y2xT PPX)· (xTx)
 Observer design for class of nonlinear systems 527

Proof of Lemma A.I: Notice first that (11) directly gives

(x T pj)2 ~ (rxT PPX)· (xTx)
We will now show there exists a specific f such that the equality holds. Choose
f(t, u, x) = yVW Tx where V and Ware unitary matrices chosen in the
following fashion. Set v(t) := Px(t), and lJ(t) := x(t), and define

Now, choose vectors V2, .•. , V n and '72, ... , IJn so that
{VI, ... , vn} and {IJI,.··, IJn}
Downloaded by [North Carolina State University] at 00:58 25 March 2013

represent two orthonormal bases for R", Then

V := [VI, ... , vn] and W:= [IJI> ... , IJn]
represent two unitary matrices. Then
IT! = rxTWVTVWTx = r x Tx
Thus, f obeys (3). Now, along the lines of Lemma 3.1, of Peterson (1987), it is
easy to show that

Therefore,
(x Tpj)2 = (v TyVWTx)2
= (VT YVbL)2
Ivi
= r(vTV)(1/T1/)
= r(x T PPX)(xTx)

Hence the proof. 0

Lemma A.2 (Peterson and Hollot 1986): If X, Y and Z are n x n symmetric
matrices such that X ;;;. 0, Y < 0 and Z ;;;. 0, and furthermore
(x TYx)2 - 4(x T Xxx T Zx) > 0
for all non-zero x ERn, then there exists a constant A> 0 such that the matrix
A2 X + AY + Z is negative definite.
Lemma A.3 (Finsler's Theorem-Peterson and Hollot 1986): If X and Yare
n x n symmetric matrices such that X;;;. 0 and x Tyx > 0 for all non-zero x E R"
such that x T Xx = 0, then there exists a constant a> 0 such that the matrix
Y + o X is positive definite.

REFERENCES
KAILATH,T., 1980, Linear Systems (Englewood Cliffs, NJ: Prentice Hall).
KELLER, H., 1987, Nonlinear observer design by transformations into a generalized observer
canonical form. International Journal of Control, 46, 1915-1930.
 528 Observer design for class of nonlinear systems

KHAROONEKAR, P. P., PETERSON, I. R., and ZHOW, K., 1990, Robust stabilization of uncertain
linear systems: quadratic stabilizability and H~ control theory. IEEE Transactions on
Automatic Control, 35, 356-361.
KRENER, A. J., and ISIDORI, A., 1983, Linearization by output injection and non-linear
observers. Systems and Control Letters, 3, 47-52.
Kou, S. R., ELLIOT, D. L., and TARN, T. J., 1975, Exponential observers for nonlinear
dynamic systems. Information and Control, 29, 204-216.
M,SAWA, E. A., and HEDRICK, J. K., 1989, Nonlinear observers: A state-of-the-art survey.
ASME Journal of Dynamic Systems and Measurement, 11,344-352.
NAGPAL, K. M., and KHARODNEKAR, P. P., 1991, Filtering and smoothing in an H oo selling.
IEEE Transactions on Automatic Control, 36, 152-166.
PETERSON, I. R., 1987, A stabilization algorithm for a class of uncertain linear systems.
Systems and Control Letters, 7, 351-357.
PETERSON, I. H., and HOLLOT, C. Y., 1986, A Riccati equation approach to the stabilization
Downloaded by [North Carolina State University] at 00:58 25 March 2013

of uncertain linear systems. Automatica, 22, 397-411.

RAoHAvAN, S., and HEDRICK, J. K., 1992, Observer-based controllers for non-linear systems.
Presented at the 1992 ASME, Winter Annual Meeting, Anaheim, California.
SPONO, M., 1987, Modeling and control of elastic joint robots. ASME Journal of Dynamic
Systems. Measurement and Control, 109,310-319.
SWEET, L., and GOOD, M, C.; 1984, Re-definition of the robot motion control problem; effects
of plant dynamics, drive system constraints, and user requirements, Proceedings of
23rd IEEE Conference on Decision and Control.
THAu, F. E., 1973, Observing the state of nonlinear dynamic systems. International Journal of
Control, 17,471-479.
ZAK S. H., 1990, On the stabilization and observation of nonlinear/uncertain dynamic systems.
IEEE Transactions on Automatic Control, 35, 604-607.