2010 American Control Conference FrB17.

1
Marriott waterfront, Baltimore, MD, USA
June 30-July 02, 2010

Observer Design for Lipschitz Nonlinear

Systems Using Riccati Equations
1
G. Phanomchoeng and R. Rajamani
2

equivalent Lipschitz constant has to be chosen large due to the

non-Lipschitz nature of the nonlinearity, most existing
observer design results fail to provide a solution. This paper
Abstract-

develops a solution methodology that works for significantly

This paper presents a new observer design
technique for Lipschitz nonlinear systems. Necessary and

larger Lipschitz constants compared to exiting results.

sufficient conditions for existence of a stable observer gain are

Furthennore, the methodology requires only an algebraic

developed using a S-Procedure Lemma. The developed condition

Riccati equation in one variable to be solved and does not

is expressed in terms of the existence of a solution to an Algebraic

require solving a multi-variable LMI problem. Thus the design

Riccati Equation in one variable. Thus, the need to solve Linear

procedure is easier to implement.

Matrix Inequalities in multiple variables is eliminated. The
advantage of the developed approach is that it is significantly less
conservative than other previously published results for Lipschitz

II.
systems. It yields a stable observer for larger Lipschitz constants

I. This paper presents an efficient methodology for designing

than other techniques previously published in literature.
PROBLEM STATEMENT

For Linear Time Invariant (LTI) systems, the observer

INTRODUCTION
observers for the class of nonlinear systems described by
design problem is dual of the control system design problem. (1)
However, this is not the case for nonlinear systems. While the
x = Ax+ Bu + <ll(x,u)

introduction of geometric techniques has led to great success where x E is the state vector, uE is the input vector,
y=Cx

in the development of controllers for nonlinear systems and yE is the output measurement vector. AE ,
n
R RP

(lsidori, 1998, Khalil, 2001), it has not been possible to obtain and CE are appropriate matrices. The
n n
R x

results of wide applicability for state estimation. Early papers

Rm

functions <D(x,u) is a Lipschitz nonlinearity with a Lipschitz

n n

by Krener and Respondek (1985), Krener and Isidori (1983)

BE R xp, Rmx

and Xiao-Hua and Gao (1985) attempted to fmd a coordinate constant r i.e.
transfonnation so that the state estimation error dynamics (2)
were linear in the new coordinates. Necessary and sufficient
1I <D(x,u)-<ll(x,u)1 r l x -x l , "ix,x
conditions for the existence of such a coordinate To begin with, note that any nonlinear system of the fonn
�

transfonnation have been established but in practice are (3)

extremely difficult to satisfy. can be expressed in the fonn of equation (1), as long as
=f (x,u)
x

f (x,u) is differentiable with respect to x . Further, many

There has been work by authors to propose observers for
more specialized classes of nonlinear systems. These include
results on observers for Lipschitz nonlinear systems (Thau,
197 3, Kou, et. aI., 19 75, Raghavan and Hedrick, 1992,
Rajamani, 1998), results on sliding mode observers (Misawa,
et. aI., 1989), results for sector nonlinear systems (Arcak and
Kokotovic, 2001), and results on observer based control for a
fully linearizable nonlinear system (Esfandiari and Khalil
(1992». The extended Kalman Filter also continues to be used
frequently as a detenninistic observer for nonlinear systems
nonlinearities can be assumed to be Lipschitz, at least locally.
For instance, the sinusoidal tenns usually encountered in
many problems in robotics are all globally Lipschitz. Even
tenns like x 2 can be regarded as Lipschitz provided we know
that the operating range of x is bounded. Thus, the class of
systems being considered in this paper is fairly general, with
the linearity assumption being made only on the output vector
The observer will be assumed to be of the fonn
y.

(Reif and Unbehauen, 1996). (4)

This paper focuses on some new observer design results for The estimation error dynamics are then seen to be given by
£=AX+Bu +<D(x,u)+L(y-Cx ).

the class of Lipschitz nonlinear systems. A major limitation (5)

of the existing results for Lipschitz nonlinear systems is that where i=x -X
they work only for adequately small values of the Lipschitz
i=(A-LC)i+<D(x,u)-<D(x,u)

constant. When the Lipschitz constant is large or when the The Lyapunov function candidate for observer design is
defined as
•

(6)
where P and PE
V=iT pi

Its derivative is
l Gridsada Phanomchoeng and 2Rajesh Rajamani are with the University of n n.
> 0 R x
Minnesota, Minneapolis, MN 55455, Email: phano002@umn.edu or
rajamani@me.umn.edu, TEL 612-626-7961

978-1-4244-7427-1/10/$26.00 @2010AACC 6060

 [
V=XT (A-LCf P+P(A-LC) x ]
p[<I>(X)-<I>(X)]+[<I>(X)-<I>(X)Y PX. ( 7) [(x-xt (<I>(x,u)-<I>(x,u) {Jx
(16)
This can be rewritten in matrix form as [_�21 �][<I>(X'u�=!(X'U)]S;O
+XT

V=[ XT {<I>(x)-<I>(x)t]x
Thus the nonlinear function <I>(x,u) satisfies an inequality of
[(A-LC)T P+P(A-LC) P][<I>(x)-<I>( (8) the type
x)] x-x. ]:::;;0 (17 )
x
P 0 [(x-xl (<ll(X,U)-<ll(X,U){ ] M [<ll(x,u)-<ll (x,u)
III. where MER2nx2nis a symmetric matrix and is given by
(18)
BACKGROUND RESULTS

3.1
The S-Procedure Lemma is as follows [16].
The S-Procedure Lemma M=[_�2I �1
Let Va(x) and V; (x) be two arbitrary quadratic forms over Applying the S-Procedure Lemma to equations (8) and (1 7),
Lemma 1:

Rn. Then Va(x)<O for all X ERn-{O} satisfying V; (x)S; 0 if we fmd V<0 if and only if there exist s 0 such that
and only if there exist s 0 such that [(A-LCf P; P(A-LC) �]-SM<0. (19)
�

Va(x)< sV; (x), \ixERn-{O} . (9)

�

Hence the necessary and sufficient condition for observer

The Schur Inequality formula is as follows [16].
3.2 The Schur Inequality design in this case is given by 0 such that
The linear matrix inequality (LMI)
3s �

]<0 (20)
Lemma 2:
[(A-LCf P+P(A-LC)+sy21 P

[� �]<0, Q=QT andR=RT (1 0) P -sl

is equivalent to one of the followingI conditions. It is possible to write equation (14) in the matrix form
(11)
Less Conservative Lipschitz Condition

or
R<0, Q-SK ST <0
defined as
(12) 1I <I>(x,u)-<I>(x,u)1 S;I G(x-x)l , \ix,x (21)
Note that the matrix G in this case could be a sparsely
IV. populated matrix. Hence I G(x-x)11 can be much smaller than
NONLINEAR OBSERVER
the constant Ilx-xi used earlier in equation (2) for the same
For the class of systems and observer forms nonlinear function.
r
4.1 Observerfor Lipschitz Nonlinear Systems

described in equations (1)-(2) and (4), if an observer gain

Theorem 1:

matrix L can be chosen such that Illustrative Example for Less Conservative Lipschitz

[(A-LC)T P+P(A-LC)+sy21 P ]
<0 (13) Let <I>(x,u) given by
Condition

for some positive defmite symmetric matrix P , then this <I>(x,u)=[2 S�(XI) l (22)
P -sl

choice of L leads to asymptotically stable estimates by the

observer (4) for the system (1). Likewise, if there exists any
symmetric positive defmite matrix P such that the derivative
of the Lyapunov function V in equation (8) is negative
definite, then there also exists a matrix L such that equation
(13) is satisfied. In this sense, equation (13) is both a
necessary and sufficient condition for observer stability.
Suppose there exist matrices L and P which satisfy
equation (13). Let this choice of L be used in the observer (4)
Proof:
for state estimation of the system given by (1).
Consider the traditional Lipschitz nonlinear system which
satisfies
1I <I>(x,u)-<I>(x,u)1 S; rllx-xll , \ix,x. (14)
In this case, we fmd
[<ll(x,u)-<ll(x,u)t [<ll(x,u)-<ll(x,u)] r2 [x-xt [x-x] (15)
Hence
S;
Then, apply the traditional Lipschitz condition for equation
(22).
1 2Sin(XI)�2Sin(XI)1 S;21 :� =:J \ix,x (23)
�(2 Sin(XI)-2sin(XI))2S;2�((XI _ XI)2+( X2 -X2)2) (24)
Next, apply the Less Conservative Lipschitz Condition for
equation (22).
1 2Sin(XI)�2Sin(XI)1 S;I [� �][:� =::1 \ix,x (25)
�(2Sin(XI)-2sin(XI))2S;2�(XI-xS (26)
It is clear thatr-
2�(XI- --xl-:"'t S;2�{(� -�)2+(X2-X2)2) . (27)
Thus the bound (21) is less conservative than the bound (2).

6061
 This result brings us to the corollary to theorem 1. Then equation (29) can be rewritten as
For the class of systems and ATP+PA+8GTG+!Pp-CT 8 J!P-PLC<0 (34)
observer forms described in equations (1), (4) and (21), if an which is the same as equation (28), thus proving the result.
Corollary to Theorem 1:

observer gain matrix L can be chosen such that

[(A-LCfP+P(A-LC)+8GTG P ]<0 (28) L
P -81 Solution ofLMI Inequality for P and L
v. NOTES ON COMPUTATION OF P AND

for some positive defmite symmetric matrix P , then this The previous section has shown the necessary and sufficient
5.1

choice of L leads to asymptotically stable estimates by the conditions for observer design. The observer design by using
observer (4) for the system (1). theorems 1-2 needs to search for P andL to satisfy inequality
(13) or (28). Solving the inequality by a numerical method
would be an obvious approach. However, the inequalities (13)
and (28) seem to be noncovex because each involves the
4.2 Reformulation of Observer Design Using Riccati

Equations (13) and (28) are LMIs involving two unknown

Equations
matrices L and P . The LMIs can be replaced by a Riccati
inequality in just one variable P . A necessary and sufficient
product of the two variables P andL . Therefore, a simple
change of variables that separates L from needs to be made.
Expand (13) to obtain
P

condition in term of the existence of a solution to a Riccati [ATP+PA-CTJ!P-PLC+ey2 1 P ]<0,P>0. (35)

inequality can then be obtained. The following theorem
summarizes the result.
There exists an observer of the type given by
equation (4) for the system given by equations (1) and (21)
Theorem 2:
P -el
Now let Y =PL . Then the inequality (13) becomes
[ATP+PA_CTYTP -YC+ey21 -81P ]<O,P>O. (36)

such that the error dynamics are quadratically stabilized if and
only if there exist 8>0 and 13 R such that the following
Riccati inequality has a symmetric positive defmite solution
P:
E
Searching for P and L satisfying (13) is equivalent to
searching for P and Y satisfying (36). Once a feasible set of
P and Y is found, L can be computed as L =p-IY . Note that

ATP+PA+8GTG+-PP-f32 1 CTC<0 (29) P>0 (or equivalently P is invertible) and that there is always
8 a one to one mapping from Y to L for a given P .
The observer gain can then be chosen as In the same way, we can apply the change of variables to
inequality (28). Then the inequality (28) becomes
L = 1322 p-1CT . (30) [ATP+PA_CTYTP -YC+8GTG -81P ]<0, P>O (37)
Applying the Schur Inequality (Lemma 2) to equation
(28), the necessary and sufficient condition for observer Hence, for a given 8 and or G , the inequality (36), or
Proof:

design is (37), is affme with respect to its two variables P and Y . The
r

(A-LCf P+P(A-LC)+8GTG+!Pp 8 <0 . (31) corresponding feasibility problem can be solved using many
standard convex optimization techniques or the TLAB
This can be re-written as LMI control toolbox.
MA

ATP+PA+8GTG+!Pp-CT 8 LTP-PLC<0 (32) Solution of Algebraic Riccati Equationfor P and L

Note that equation (32) implies for all x such that CX =0 , we The observer design by using theorem 2 involves solving
5.2

for one unknown matrix, P , by providing 8 , G , and 13.

must havexT(ATP+PA+8GTG+!PP)x<0. 8 It also Since P is the only unknown, it is easier to solve the Riccati
therefore follows that there must 13 R sufficiently large such inequality than to solve a LMI inequality.
Equation (29) can be modified as follows. Assume
that xT (ATP+PA+8GTG+!PP)x-f32 xTCTCX<0 for all 1 CTC =-J11<0 (38)
E

8 ATp+PA+8GTG+-PP-f32 8
x(t) . where J1>0, J1 is a small scalar.
Equation (29) is therefore a necessary condition. The fact that 1
equation (29) is a sufficient condition can be proved as ATP+PA+-PP+8GTG-f32
8 CTC+J11 =0 (39)
follows: For given 8 , G , 13, and J1 , the Riccati equality (39) can be
Let there exist a positive defmite solution P to equation (29). easily solved. This equation may be solved with the TLAB
Let the observer gain be chosen as command "ARE':
MA

L =�2 p-ICT . (33) = ARE(a,b,c) (40)

This returns the stabilizing solution (if it exists) to the
x

6062
continuous-time Riccati equation: However, this control law requires measurement of all the
(41) states. Physically, one can measure the motor position and
with b being symmetric and nonnegative defmite and c being velocity, but the measurement of the other states is non-trivial.
a'X +Xa-XbX +c=O

symmetric.
Rearrange the equation (39) to appropriate equation (41) as In this section we present an observer for the above robotic
follows:
6.2 Observer design for ajlexible link robot

(42) system with guaranteed convergence of the state estimates.

Using Corollary to Theorem 1 on LMI based solution
(_AT)P+ P(-A) -P('!I)P+ (-sGTG+ [32CTC - pI), 0

The Matlab function "ARE' can now be used to fmd the with =1 , an observer gain is found to be
=

'-v--'
a
s�
a
' �

solution.
b

L=[�:!��� :��;:: l'

2 (46)
VI. 5.4689 12.8094
ILLUSTRATIVE EXAMPLE
4.1360 -13.1102
6,1
Consider a one link manipulator with revolute joints
Dynamic modelfor ajlexible link robot
The eigenvalues of (A-LC) are
actuated by a DC motor. The elasticity of the joint can be 7 85 4
well-modeled by a linear tensional spring [12]. The elastic 417 l'
coupling of the motor shaft to the link introduces an additional eig(A LC)=[�:'�
-5.0855 (47)
degree of freedom. The states of this system are motor
position and velocity, and the link position and velocity. -0.9525
_

The corresponding state-space model is Thus, the eigenvalues are well damped in this case. The
response will show no transient oscillations. However, there is
one small eigenvalue. So the estimated state will converge
slowly to the actual state.
Om =O)m
The robotic system with the above observer gain matrix was
(43) simulated under an open-loop excitation with being a
. k B K,
0)m =-Jm (B. - B)
m - J 0) m +-u
Jm
sinusoid at 1 Hz. Fig. 1 shows a comparison of actual and
-

m u

estimated link angular position. The actual system has an

O. =0).
) mgh
J sin(B ) initial condition equal to 1 radian while the observer starts
from an initial value of zero. Fig. 2 shows actual and observer
dJ• =-.!5....(B Bm
with Jm being the inertia of the motor; J. the inertia of the
J
estimated link angular velocities. Both the estimated states
_ _
•
• •

link; Bm the angular rotation of the motor; B. the angular converge to the correct values.
position of the link; O)m the angular velocity of the motor; and
the angular velocity of the link.
Actual and obser\er estimated link angular position

Thus the system dynamics are nonlinear and of the form

0).

= +Bu +<1>( x )
x Ax
(44)
with x =[Bm O)m B. O).t .
y=Cx

A=[
-4:.6 -1�25 4�.6 �l B=[2�'6l
0 0 0 ' 1 0'
19.5 0 -19.5 0 (45)
o - 1:-
0 -='--- 0.5--= ----O-
- 7:
1.,---
5 -: - - ::
2 '"=
.5 -

time (sec)
:----:'3. 5-=- ----0-4
- 7:
4 .,-
5 ---:

( )=[ �
l' =[� c
Fig. I Actual and observer estimated link angular position (Corollary

-3.33sin( x J
<I> X
to Theorem I)

The value of the Lipschitz constant for this system

is =3.33 . The above parameters for the system are typical
and have been taken from Spong [12].
r

[12] presents a nonlinear linearizing control law for this

system. The control law guarantees closed-loop stability and
1/0

tracking of any desired trajectory by the robotic link.

6063
 Actual and obserwr estimated link angular velocity
Actual and obserwr estimated link angular velocity

-2
-3
-3
-4
-4

S S
- ':--
O
---:O
: -.
':----'''--
S -: 1 S.
-':----:-: - -::' -::-
2.S----: S.
-: - -::'3.:S---:- ----:'c
4:--- -!
' -,
- O--- � 1--=S. �-�2S--=. �-�3'
S. '''---
� ----
-
O � ,...-
S. ----�-4 S. --J
time (sec)
time (sec)

Fig. 4 Actual and observer estimated link angular velocity (theorem

Fig. 2 Actual and observer estimated link angular velocity (Corollary
2)

Using theorem 2 on the Riccati based solution with =1

to Theorem I)

andf3=180 , an observer gain matrix is found to be This section will compare observer design techniques in
6.3 Comparison with previous observer design techniques

term of how conservative they are. We will examine which

I>

=� ���: :��� observer design techniques can deal with a larger Lipshitz
L =[ :
1� l. (48) constant. The problem in section 6.1 is reconsidered and we
10.0309 91.3947 will keep the same plant equations. However, the nonlinearity
14.2679 51.5085 will be scaled to continuously increase the value of the
Lipschitz constant for this system until the observer can no
The eigenvalues of (A-LC) are longer solve the problem when the Lipschitz constant is too
61 large.
·
-65.4;�; : 3 :'5212il. (49) First, we fmd an observer gain by Theorem 1 which needs
eig(A-LC) = [
the solution of LMI Inequalities. Then, we use Theorem 2 to
-65.4807 -34.5212 fmd an observer gain matrix by solving the corresponding
-1.0914 Algebraic Riccati equation. Finally, the traditional standard
Thus, the eigenvalues have lower damping in this case. LMI observer [16] as described below is used to fmd an
However, the overall eigenvalues are large. So the estimated observer gain matrix for comparison.
states rapidly converge to the actual state and transient
oscillations are not present. The transient performance is For the class of systems and observer forms described in
better than the previous LMI solution.
Standard LMI Observer

equations (1) and (4), if an observer gain matrix L can be

chosen such that
[(A-LCf P+P(A-LC)+I P ]<0 (50)
Actual and obserwr estimated link angular �ition
1.SI��-��-��-r====il

for some positive defmite symmetric matrix P , then this

P -Ily2

choice of L leads to asymptotically stable estimates by the

observer (4) for the system (1).
Table 1: Comparison of maximum Lipschitz constant for
. bserver
-ll.S
Method
varIOus
Corollary to
· techn·lques
deSlgn
0
Theorem 2 Standard
Theorem 1 LMI

48.5 48.4 0.99

1
- 0:---::
0.S----':---:'
': 1::- 2
.::--
.S----::---::' 3s:---C-
S ---:-7c
. -4 .s:----O
7c Observer
time (sec)

Ymax

Table 1 shows that the observer design techniques from

Fig. 3 Actual and observer estimated link angular position (theorem

corollary to theorem 1, and from theorem 2 can deal with a

2)

larger Lipshitz constant. They can solve the problem with a

Lipshitz constant up to 48.4. On the other hand, the standard
LMI observer can deal with the problem only for values of
Lipshitz constant up to 0.99. Thus, the new observer design
6064
techniques are significantly less conservative than the
traditional LMI observer design technique.
[18] X. Xiao-Hua and W. Gao, 1989, "Nonlinear observer design by
observer error linearization", SIAM Journal of Control and

VII.
Optimization, vol. 27, pp.199 - 216, 1989.

This paper presented a new observer design technique for

CONCLUSIONS

Lipschitz nonlinear systems. Necessary and sufficient

conditions for existence of a stable observer gain were
developed using a S-Procedure Lemma. The developed
condition was then expressed in terms of the existence of a
solution to an Algebraic Riccati Equation in one variable.
Thus, the need to solve Linear Matrix Inequalities in multiple
variables was eliminated. The advantage of the developed
approach is that it is significantly less conservative than other
previously published results for Lipschitz systems. Using an
illustrative example of a flexible joint robot, the developed
observer design technique was found to yield a stable observer
for much larger Lipschitz constants than a traditional LMI
design technique.
References
[I] A.J. Krener and W. Respondek, "Nonlinear observers with linearizable
error dynamics", SIAM Journal of Control and Optimization, vol. 23,
pp.197 - 216, 1985.
[2] A.J. Krener and A. Isidori, "Linearization by output injection and
nonlinear observers", Systems & Control Letters, vol. 3, pp.47 - 52,
1983.
[3] C. Edwards, S. K. Spurgeon and R.I. Patton, "Sliding mode observers
for fault detection and isolation", Automatica, vo1.36, pp.541-553, 2000.
[4] C. Edwards and S. K. Spurgeon, "On the development of discontinuous
observers", International Journal of Control, vol. 59, no. 5, pp.1211-
1229, 1994.
[5] F. Esfandiari and H. K. Khalil, "Output feedback stabilization of full
linearizable systems", International Journal of Control, vol. 56, pp.1007
- 1037 , 1992.
[6] F. E. Thau, "Observing the state of nonlinear dynamic systems", Int. J.
Contr., vol. 17, no. 3, 1973.
[7] J.J.E. Siotine and W. Li, Applied Nonlinear Control, Prentice-Hall, New
Jersey, 1991.
[8] K. Reif and R. Unbehauen, "Linearization along trajectories and the
extended kalman filter", Proc. 13th IFAC World Congr., vol. H, pp.509--
514, 1996.
[9] H.K. Khalil, Nonlinear Systems, Prentice Hall, New Jersey, 2001,
edition 3.
[10] M. Arcak and P. Kokotovic, "Observer-based control of systems with
slope-restricted nonlinearities". IEEE Transactions on Automatic
Control, 46(7), pp. 1146-1151, 2001.
[II] M. Arcak and P. Kokotovic, "Nonlinear Observers: A circle criterion
design", In Proceedings of the 38th IEEE Conference on Decision and
Control, Phoenix, AZ, pp. 4872-4876, 1999.
[12] M. Spong, "Modeling and control of elastic joint robots", ASME Journal
of Dynamic Systems, Measurement and Control, Vol. 109, pp.310-319,
1987.
[13] R. Rajamani, "Observers for Lipschitz Nonlinear Systems", IEEE
Transactions on Automatic Control, Vol. 43, No. 3, pp. 397-401, March
1998.
[14] R. Perla and S. Mukhopadhyay, "Observer Design for Lipschitz
Nonlinear Systems with State Dependency and Nonlinearity
Distribution", INDICON, 2005 Annual IEEE, pp.437 - 441, Dec. 2005.
[15] R. Rajamani and Y. Cho, "Observer Design for Nonlinear Systems:
Stability and Convergence", Decision and Control, Proceedings of the
34th IEEE Conference on, New Orleans, LA, vol. I, pp.93 - 94, Dec.
1995.
[16] S. Boyd, L.E. Ghaoui, E. Feron, and V. Balakrishnan, "linear matrix
inequalities in systems and control theory", Society for Industrial and
Applied Mathematics, SIAM Studies in Applied Mathematics, 1994.
[17] S. Raghavan, "Observers and compensators for nonlinear systems with
application to flexible-joint robots", Ph.D. Dissertation, University of
California at Berkeley, Berkeley, CA 94720, U.S.A, 1992.

6065