1

NONLINEAR OBSERVERS

A. J. Krener

University of California, Davis, CA , USA

4
5

Keywords: nonlinear observer, state estimation, nonlinear filtering, observability, high gain
observers, minimum energy estimation, H estimation, extended Kalman filter

Contents

1. Introduction

2. Observability

3. Construction of Observers by Linear Approximation

10

4. Construction of Observers by Error Linearization

11

5. High Gain Observers

12

6. Nonlinear Filtering

13

7. Minimum Energy and H Estimation

14

8. Multiple Extended Kalman Filters

15

9. Conclusion

16

Bibliography

17

Summary

18
19
20
21
22
23
24
25

This paper is a review of the existing methods for designing an observer for a system modeled by
nonlinear equations. We focus our attention on autonomous, finite dimensional systems described
by ordinary differential equations. The current condition of such a system is described by its state
variables about which we just have partial and possibly noisy measurements. The goal of the
observer is to process these measurements and any information regarding the initial state of the
system and to obtain an estimate of the current state of the system. This estimate should improve
with additional measurements and, ideally, converge to the true value in the absence of noise. The
observer does this by taking advantage of our a priori knowledge of the dynamics of the system.

26

1. Introduction

27
28
29
30
31
32

Systems are sets of components, physical or otherwise, which are connected in such a manner as to
form and act as entire units. A nonlinear system is described by a mathematical model consisting of
inputs, states, and outputs whose dynamics is given by nonlinear equations. Such models are used
to represent a wide variety of dynamic processes in the real world. The inputs are the way the
external world affects the system, the states are the internal memory of the system and the outputs
are the way the system affects the external world. An example of such a system is

33

x& ( t ) = f ( t , x ( t ) , u ( t ) )

(1)

34

y ( t ) = h ( t, x ( t ), u ( t ) )

(2)

E6-43-21-15-TXT_NT.doc

06/23/03

x ( 0 ) x 0

(3)

2
3
4

The input is the m vector u, the state is the n vector x and the output is the p vector y. The state of
the system at the initial time t = 0 is not known exactly but is approximately x 0 . Typically, the
dimensions of the input and output are less than that of the state.

A particular case is an autonomous linear system

x& = Ax + Bu

(4)

y = Cx + Du

(5)

x ( 0 ) x 0

(6)

Other examples include systems described by difference equations

10

x ( t + 1 ) = f ( t, x ( t ), u ( t ) )

(7)

11

y ( t ) = h ( t, x ( t ), u ( t ) )

(8)

12
13
14

and infinite dimensional systems described by partial differential and/or difference equations, delay
differential equations or integro-differential equations. This review will focus on finite dimensional
systems described by ordinary differential equations.

15
16
17
18

An observer is a method of estimating the state of the system from partial and possibly noisy
measurements of the inputs and outputs and inexact knowledge of the initial condition. More
precisely an observer is a causal mapping from any prior information about the initial state x0 and
from the past inputs and outputs

19

{( u ( ) , y ( ) ) : t

(9)

21
22
23
24

to an estimate x ( t ) of the current state x(t) or an estimate z ( t ) of some function z(t) = k(x(t)) of
the current state. Causality means that the estimate at time t does not depend on any information
about the inputs and outputs after time t. This restriction reflects the need to use the estimate in real
time to control the system. The essential requirement of an observer is that the estimate converges
to the true value as t gets large.

25
26
27

Sometimes it is not necessary to estimate the full state but only some function of it, say (t, x). For
example, if one wishes to use the feedback control u = (t, x). This article will focus on observers of
the full state.

28

The prototype of an observer is that of an autonomous linear system Eqs. (4) - (6). The system

20

29

x& = Ax + Bu + L ( y y )

(10)

30

y = Cx + Du

(11)

31

x ( 0 ) = x 0

(12)

32
33

is an observer where L is an n p matrix to be chosen by the designer. The dynamics of the error
x% = x x is given by

34

x&% = ( A LC ) x%

(13)

35

x% ( 0 ) = x 0 x 0

(14)

E6-43-21-15-TXT_NT.doc

06/23/03

1
2
3
4
5

If the spectrum of the matrix A LC lies in the open left half plane, then the error decays to zero
exponentially fast. In this way, the problem of designing an observer for an autonomous linear
system is reduced to the following problem. Given A, C, find L so that A LC is Hurwitz, i. e., the
spectrum of A LC is in the open left half plane. We discuss when L can be so chosen in the next
section (see Design Techniques for Time Varying Systems for further details.)

6
7

For nonlinear systems the distinction between nonautonomous Eqs. (1) - (3) and autonomous
systems

x& = f ( x, u )

(15)

y = h ( x, u )

(16)

x ( 0 ) = x 0

(17)

10

14
15
16
17
18
19

is frequently not important as one can add time as an extra state xn+1 = t t0 and thereby reduce the
former to the latter. Since an observer operates in real time, time is usually observable and so can be
added as an extra output also. Frequently models depend on parameters as in x& = f ( x, u, ) .
But in a nonlinear system the distinction between states and parameters is not always clearcut.
Parameters can always be treated as additional states by adding the differential equation & = 0 .
Therefore, the problem of real time parameter estimation reduces to the problem of real time state
estimation and may be solvable by an observer. If the state estimate is not going to be used in real
time, then one can collect data after time t to estimate x(t). This problem is sometimes called
nonlinear smoothing and is related to the identification of nonlinear systems (see 6.43.10) .

20
21
22
23

Another example of an observer is the extended Kalman filter described in more detail in (see State
Reconstruction by Extended Kalman Filter) and in the following statements. This is an observer for
a nonlinear, nonautonomous system Eqs. (1) - (3) which is derived using stochastic arguments. Two
quantities x ( t ) and P(t) are computed by the extended Kalman filter. The stochastic interpretation

11
12
13

24
25
26
27

is that the distribution of the true state x(t) is approximately Gaussian with mean x ( t ) and
covariance P(t).
Most observers are described recursively as a dynamical system whose input is the measured
u
variables and whose output is the state estimate x such as
y

28

z& = f ( t , z , u , y )

(18)

29

x = h ( t , z , u, y )

(19)

30
31
32

If the state of the observer, z, is of the same dimension as the state of the system, then it is called a
full order observer; if it is of greater dimension then it is called an expanded order observer, and if it
is of lesser dimension, then it is called a reduced order observer.

33

For example, the prototype autonomous linear observer Eqs. (10) - (12) can be written as

34

u
z& = ( A LC ) z + [ B LD L ]
y

(20)

35

x = z

(21)

36

z ( 0 ) = x 0

(22)

E6-43-21-15-TXT_NT.doc

06/23/03

1
2
3

and hence is a full order observer. The state of extended Kalman filter discussed as follows is the
pair z = ( x , P ) , so it is an expanded order observer. We briefly discuss the Luenberger observer, a
reduced order observer for a linear autonomous system in the form

x&1 A11
x& = A
2 21

y = x1 + Du

(24)

x ( 0 ) = x0

(25)

A12 x1 B1
+
u
A22 x2 B2

(23)

The reduced order observer is given by

z& = ( A22 LA12 ) z + ( A21 LA11 ) + ( A22 LA12 ) L ( y Du )

(26)

x1 = y Du

(27)

x2 = z + L ( y Du )

(28)

10
11
12
13
14
15

where L is a design parameter. If the model is exact then x1 = x1 and

x&%2 = ( A22 LA12 ) x%2

(29)

so if the spectrum of the matrix A22 LA12 lies in the open left half plane then the error decays to
zero exponentially fast. We discuss when L can be so chosen in the next section. For more on
reduced order linear observers, (see (6.43.5.3)) .

20
21
22
23

The state z of the observer is some measure of the likely distribution of the state of the original
system given the past observations. If the observer is derived using stochastic arguments, the state
of the observer is typically the conditional density of the state of the system given the past
observations and the initial information. In the extended Kalman filter, the state z = ( x , P ) is the
mean and the covariance of the approximately Gaussian distribution of the true state. For the full
and reduced order linear observers described previously, which were derived by nonstochastic
arguments, one can view the conditional density as being singular and concentrated at a single
point, x ( t ) .

24

2. Observability

25
26
27
28
29
30
31
32

The question of whether an observer converges is of paramount importance. A more immediate

question is when a nonlinear system Eqs. (15) - (17) admits a convergent observer. This leads to the
concepts of observability and detectability which are discussed in (see 6.43.21.7) . Briefly two
states x01, x02 are said to be distinguishable by an input u(t) if the outputs y1(t), y2(t) of Eqs. (15) (17) satisfying the initial conditions x0 = x01, x0 = x02 differ at some time t 0. The system is said to
be observable if every pair x01, x02 can be distinguished by some input u(t). An input u(t) which
distinguishes every pair x01, x02 is said to be universal. A system where every input is universal is
said to be uniformly observable.

33

Consider a smooth autonomous nonlinear system without inputs

16
17
18
19

34

x& = f ( x )

(30)

35

y = h( x )

(31)

E6-43-21-15-TXT_NT.doc

06/23/03

1
2
3
4

x ( 0 ) = x0

(32)

At time t = 0 the output and its time derivatives are given by the iterated Lie derivatives
y ( 0 ) = h ( x0 )

(33)

y& ( 0 ) = L f ( h ) ( x 0 ) =
&&
y ( 0 ) = L2f ( h ) ( x 0 ) =

h 0
( x ) f ( x0 )
x
L f ( h )
x

(34)

(x ) f (x )
0

(35)

7
8
9
10
11
12
13

and so on. If the p-vector-valued functions h, Lf (h), L2f ( h ) , . . . distinguish points then clearly the
system is observable. For a real analytic system, this is a necessary and sufficient condition for
observability. This suggests a way of reconstructing the state of a system, differentiate the output
numerous times, and find the state which generates such values. One does not proceed in this
fashion because differentiation greatly accentuates the effect of the almost inevitable noise that is
present in the observations, and multiple differentiations greatly increase this problem. That is why
observers are usually dynamic systems driven by measurements. When such systems are
integrated, the effect of the noise is mitigated not enhanced.

14

For simplicity of exposition, suppose that n = kp. If the matrix

15
16

(h) 0
(x )

L f ( h ) 0
(x )

k 1

L f ( h ) x 0
x ( )
is invertible then the p-vector-valued functions

17

1 = h ( x )

18

2 = L f ( h )( x )

19

k = Lkf1 ( h )( x )

20

(36)

(37)
,...,

(38)
(39)
0

are local coordinates around x and in these coordinates the system Eqs. (30) - (32) becomes

21

y = 1

(40)

22

& 1 = 2

(41)

& 2 = 3
23
24

(42)

& k = f k ( )

E6-43-21-15-TXT_NT.doc

(43)

06/23/03

1
2
3

Each i is a p-vector. Such a system is said to be in observable form, since it is clearly observable.
Many algorithms for constructing observers start with the assumption that the system is in
observable form. The observable form of a n = kp system with inputs is

y = 1 + g 0 ( , u )

(44)

& 1 = 2 + g1 ( , u )

(45)

& 2 = 3 + g 2 ( , u )
6
7
8
9

M
& k = f k ( ) + g k ( , u )

(46)
.

(47)

where gi(, 0) = 0. Such a system is clearly observable as the input u(t) = 0 distinguishes every pair
of points, but it may not be uniformly observable. A system

10

y = 1 + g 0 ( u )

(48)

11

& 1 = 2 + g1 ( 1, u )

(49)

12

& i = i +1 + g 2 ( 1,...,i , u )

13

& k = f k ( ) + g k ( 1 ,..., k , u )

(50)

M
.

(51)

14
15

is said to be in uniformly observable form for it is clearly uniformly observable. From the
knowledge of u(t), y(t) we can determine 1(t), from the knowledge of u(t), y(t), & 1 ( t ) we can

16

determine 2(t), etc.

17

An autonomous linear system is observable if, and only if, the matrix

18

C
CA

O=
M
n1
CA

(52)

19
20
21
22

is of full column rank in which case C, A is said to be an observable pair. Moreover, for such
systems the spectrum of A LC can be set up arbitrarily to complex conjugation by choice of L. (As
a real matrix the spectrum of A LC is invariant with respect to complex conjugation.) (See
(6.43.5.3)).

23
24

A system Eqs. (15) - (17) is detectable, if whenever the outputs are equal y1(t) = y2(t) from the
initial states x01, x02 using the same control u(t), then the state trajectories converge x1(t) x2(t) 0.

25
26
27
28
29

For an autonomous linear system, the kernel of the matrix Eq. (52) is the largest invariant subspace
of the matrix A contained in the kernel of C. It is not hard to show that the system is detectable if,
and only if, the spectrum of A restricted to the kernel of Eq. (52) is in the open left half plane.
Clearly, the spectrum of A LC on the kernel of Eq. (52) does not depend on L. The rest of the
spectrum of A LC can be set up arbitrarily to complex conjugation by choice of L.

30
31

Hence a linear system admits a convergent observer if, and only if, it is detectable. It is not hard to
show that the system Eq. (23) - (25) is detectable if, and only if, the reduced system is.
E6-43-21-15-TXT_NT.doc

06/23/03

x&2 = A22 x2

(53)

y = A12 x2

(54)

Hence a linear system admits a convergent reduced order observer if and only if it is detectable.

3. Construction of Observers by Linear Approximation

5
6
7
8

Consider an autonomous nonlinear system without inputs Eqs. (30) - (32) . If the system is known
to operate in a neighborhood of some fixed state, say x = 0 where f (0) = 0, h(0) = 0, then the
simplest approach to constructing an observer is to approximate the dynamics around this operating
condition by the linear autonomous system

x& = Ax

(55)

10

y = Cx

(56)

11

x ( 0 ) = x0 0

(57)

12
13
14
15

where
A=

f
(0)
x

(58)

C=

h
(0)
x

(59)

and use an observer for the latter,

16

x& = Ax + L ( y y )

(60)

17

y = Cx

(61)

18

x ( 0 ) = x 0

19
20
21

(62)

The error x% = x x dynamics is

x&% = ( A LC ) x% + f ( x ) Lh ( x )

(63)

where

22

f ( x ) = f ( x ) Ax

23

h ( x ) = h ( x ) Cx

(64)

(65)

24
25
26
27

If the linear system is detectable, then there are choices of L such that the spectrum of A LC lies in
the open left half plane. But the original system must be locally asymptotically stable to 0 for the
observer to converge. And, if it is locally asymptotically stable to 0, an observer may not be needed
as the estimate x = 0 is asymptotically correct.

28

A slightly more sophisticated approach is preferable. Define the observer to be

29

x& = f ( x ) + L ( y y )

(66)

30

y = h ( x )

(67)

E6-43-21-15-TXT_NT.doc

06/23/03

x ( 0 ) = x 0

(68)

2
3
4
5
6

Suppose the system Eqs. (30) - (32) is stable around x = 0 in the sense of Lyapunov; i. e., if the
system starts in a sufficiently small neighborhood of 0 it stays close to 0. If the original state x0 and
state estimate error x% 0 = x 0 x 0 are sufficiently small then the observer error converges to zero.
Hence, this is a local observer where local is meant in two senses. Both the original state x0 and the
original state estimate error x% 0 = x 0 x 0 must be close to 0 for guaranteed convergence of the error.

7
8

To see this, suppose the spectrum of A LC lies in the open left half plane, then there exists a
positive definite solution P1 to the Lyapunov equation

9
10

11
12

( A LC ) P 1 + P 1 ( A LC ) = I .

(69)

So, if x(t) satisfies Eqs. (30) - (32) and x ( t ) satisfies Eqs. (66) - (68), then
d

x% P 1 x% ) = x% x% + f% Lh% P 1 x% + x% P 1 f% Lh%
(
dt

(70)

where

13

f% = f% ( x, x% ) = f ( x ) f ( x x% ) Ax%

14

h% = h% ( x, x% ) = h ( x ) h ( x x% ) Cx%

(71)

(72)

15
16
17

Assuming that the system is sufficiently smooth, the last two terms on the right side are O(x)O( x% )2
and so are dominated by x% x% for small x, x% . Hence, the right side is negative and the error
converges to zero.

18

If the system has an input Eqs. (15) - (17) and the input is measurable, then

19

x& = f ( x, u ) + L ( y y )

(73)

20

y = h ( x, u )

(74)

21

x ( 0 ) = x 0

(75)

22
23
24

is an local observer. If the controlled system Eqs. (15) - (17) stays in a sufficiently small
neighborhood of the origin then the observer converges as before based on the analysis of Eq. (70)
but with

25

f% = f ( x, u ) f ( x x% , u ) Ax%

26

h% = h ( x, u ) h ( x x%, u ) Cx%

(76)

(77)

27
28
29
30
31
32

Frequently, the state estimate x is used in a feedback control law u = ( x ). For the observer Eqs.
(73) - (75) to converge with the open loop system Eqs. (15) - (17) need not be stable, but the closed
loop system should be. If the spectrum of A LC lies in the open left half plane, if the state
feedback u = (x) locally exponential stabilizes the system Eqs. (15) - (17), and if x0, x% 0 are
sufficiently small then the state estimate feedback u = ( x ) will also be locally exponential
stabilizing the system and the observer will converge locally.

33
34

The previously shown techniques require choosing L so that A LC is Hurwitz. Of course, if there
is one such L, there are many and the question is which one to choose. One reasonable way of

E6-43-21-15-TXT_NT.doc

06/23/03

1
2

choosing L is via an approximating Kalman filter (see Kalman Filters). Assume that the linear
approximating system Eqs. (55) - (57) is corrupted by noise,

x& = Ax + Gw

(78)

y = Cx + Jv

(79)

x ( 0 ) = x 0 + x% 0

(80)

6
7
8

where w, v are standard independent white Gaussian noises, and x% 0 is an independent Gaussian
initial condition. Let the system be detectable and Q = G G , R = JJ . If R is invertible, then the
long time, stationary Kalman filter for this system is
x& = Ax + L ( y y )

(81)

10

y = Cx

(82)

11

x ( 0 ) = x 0

(83)

12
13
14
15

The observer gain is

L = PC R 1

(84)

where P is the unique positive definite solution to the algebraic Riccati equation
0 = AP + PA + Q PC R 1CP

(85)

16

This observer gain L is used in the observer Eqs. (73) - (75).

17
18
19
20
21
22
23
24
25

The Kalman filtering approach in effect replaces the design parameter L by a pair of design
parameters Q, R. The tradeoff between these two parameters is roughly as follows. The smaller that
R is as compared to Q, the more weight the observer puts on the most recent observations in
arriving at its estimate. Making R smaller while holding Q constant tends to move the spectrum of
ALC further left. At first this might seem an unmitigated benefit, but the further left that the
spectrum is the more errors in the observations increase the errors in the estimate. An observer with
the spectrum far to the left is severely compromised by observation noise and even by driving noise
although to a lesser extent. The Kalman filter finds the optimal place to put the spectrum given the
relative magnitudes (covariances) of the noise.

26
27
28
29

If the system Eqs. (1) - (3) is not operating in the neighborhood of some fixed state, then the
extended Kalman filtering approach can be used to construct an observer Eqs. (100)- (101). In
effect, the nonlinear system Eqs. (1) - (3) is approximated by a time varying linear system along the
estimate of the state trajectory with standard independent white Gaussian noises w, v,

30

x& = A ( t ) x + B ( t ) u + G ( t ) w

(86)

31

y = C ( t ) x + D ( t )u + J ( t ) v

(87)

32
33
34

where
A( t ) =

f
( t, x ( t ) , u ( t ) )
x
,

(88)

B(t ) =

f
( t , x ( t ) , u ( t ) )
u
,

(89)

E6-43-21-15-TXT_NT.doc

06/23/03

1
2
3

C (t ) =

h
( t , x ( t ) , u ( t ) )
x
,

(90)

D(t ) =

h
( t , x ( t ) , u ( t ) )
u
.

(91)

A Kalman filter for this linear system is

x& = A ( t ) x + B ( t ) u + L ( t )( y y )

(92)

y = C ( t ) x + D ( t ) u

(93)

P& ( t ) = A ( t ) P ( t ) + P ( t ) A ( t ) + Q ( t ) P ( t ) C ( t ) R 1 ( t ) C ( t ) P ( t )

(94)

x ( t 0 ) = x 0

(95)

P ( t 0 ) = P0

(96)

Q ( t ) = G ( t ) G ( t )

(97)

10

R ( t ) = J ( t ) J ( t )

(98)

11

L ( t ) = P ( t ) C ( t ) R 1 ( t )

(99)

12
13

The form of an extended Kalman filter is slightly different and obtained by changing the first two
equations as before

14

x& = f ( t , x, u ) + L ( t )( y y )

(100)

15

y = h ( t , x, u )

(101)

16
17
18
19
20

The matrices Q(t), R(t) are design parameters, the former represents the uncertainty in the system
dynamics (the driving noise covariance) and must be chosen to be nonnegative definite. The latter
represents the uncertainty in the system measurements (the measurement noise covariance) and
must be chosen to be positive definite. The initial state estimate x 0 and its covariance P0 describe
the prior knowledge of the true state at the beginning of the process.

21
22
23

The extended Kalman filter is the most widely used nonlinear observer. Its virtues are its relative
simplicity and its frequently good performance. Unfortunately, though it is not guaranteed to
converge, here is a simple example where it fails.

24

x& = f ( x ) = x ( 1 x 2 )

(102)

25

y = h ( x ) = x2 x 2

(103)

26
27
28
29
30

The system is observable as h, Lf (h), L2f ( h ) separate points. The dynamics has stable equilibria at
x = 1 and an unstable equilibrium at x = 0. Under certain conditions, the extended Kalman filter
fails to converge. Suppose the x0 = 1 so x(t) = 1 and y(t) = 1/2 for all t 0. But h(1/2) = 1/2 so if
x 0 1/2 the extended Kalman filter will not converge. To see this notice that when x (t) = 1/2,
the term y(t) y (t) = 0 so x& = f ( x (t)) = f (1/2) = 3/8. Therefore x (t) 1/2 for all t 0.

E6-43-21-15-TXT_NT.doc

10

06/23/03

1
2
3

It is not hard to see that any one dimensional observer will encounter the same difficulties as the
extended Kalman filter. One way around this difficulty might be to embed the system into a higher
dimensional system in observer form.

4. Construction of Observers by Error Linearization

5
6
7

There are several approaches that rely on finding a change of state coordinates that makes the
problem of constructing an observer easier. Perhaps, the simplest way is to try to find a change of
state and output coordinates

z = ( x )

(104)

w = ( y )

(105)

10
11

that transforms the nonlinear autonomous system Eqs. (15) - (17) into a linear autonomous system
with input output injection,

12

z& = Az + ( u , y )

13

w = Cz + ( u , y )

(106)
.

(107)

14
15

One would like the transformations to be global diffeomorphisms, but one may have to settle for
local diffeomorphisms around x0, y0 which map to z0 = 0, w0 = 0.

16

It is easy to construct an observer for the latter system,

17

z& = Az + ( u , y ) + L ( w w )

(108)

18

w = Cz + ( u , y )

(109)

19

z ( 0 ) = z 0 = ( x 0 )

(110)

20

x = 1 ( z )

(111)

21

with linear error dynamics

22

z&% = ( A LC ) z%

(112)

23

z% ( 0 ) = z 0 z 0

(113)

24
25
26

where z% = z z . If the nonlinear autonomous system Eqs. (15) - (17) is linearly observable at x0,
then C, A is an observable pair, so one can set the spectrum of A LC in the open left half plane and
the error will go to zero exponentially.

27
28

One can leave the observer in z , w coordinates or transform the observer Eqs. (108) - (111) back
into x , y coordinates,

29

x& = f ( x , u ) + ( x )
x

30

x ( 0 ) = x 0

31
32

( ( u, y ) ( u, h ( x ) ) + L ( ( y ) C ( x ) ( u, y ) ) )

(114)
(115)

The advantage of x , y coordinates is that they may be natural to the system. The advantage of z , w
coordinates is that the observer is a stable linear system driven by a signal that depends on the input
E6-43-21-15-TXT_NT.doc

11

06/23/03

1
2

and the output. If the signal is bounded, then the estimate remains bounded; even if the coordinate
transformations are only approximate.

3
4
5
6
7

The problem with this approach is that there are very few systems Eqs. (15) - (17) that can be
transformed into Eqs. (106) - (107). The functions Eqs. (104) - (105) must satisfy a first order
system of partial differential equations and must be at least local diffeomorphisms. To be solvable,
the system of partial differential equations must satisfy integrability conditions that are quite
restrictive. Also, the system Eqs. (15) - (17) needs to be linearly observable at x0.

8
9

This last condition can sometimes be avoided by allowing Eqs. (104) - (105) to be semidiffeomorphisms: that is, smooth functions with continuous inverses.

10
11

To get around this problem, one can broaden the class of systems Eqs. (106) - (107) to those in socalled state affine form

12

z& = A ( u, y ) z + ( u , y )

13

w = C ( u , y ) z + ( u, y )

14

(116)
.

(117)

For these systems one can use a Kalman filtering approach,

15

z& = A ( u, y ) z + ( u, y ) + L ( t )( w w )

(118)

16

w = C ( u, y ) x + ( u, y )

(119)

17

P& ( t ) = A ( t ) P ( t ) + P ( t ) A ( t ) + Q ( t ) P ( t ) C ( t ) R 1 ( t ) C ( t ) P ( t )

(120)

18

Q ( t ) = G ( t ) G ( t )

(121)

19

R ( t ) = J ( t ) J ( t )

(122)

20

L ( t ) = P ( t ) C ( t ) R 1 ( t )

(123)

21
22
23

The partial differential equations for these state affine transformations are more complicated, but
the integrability conditions are less stringent. These techniques have been successful employed for
low dimensional problems but the calculations grow in complexity as the dimensions increase.

24
25
26
27

If one assumes that the input u(t) and/or the output y(t) is differentiable, then one can allow A, C, ,
to depend on their derivatives. The partial differential equations for , get even more
complicated, but the integrability conditions become even less stringent. Of course, this approach is
not advisable if there is noise present.

28

Recently, a simpler approach has been introduced for real analytic systems without inputs,

29

2
3
x& = f ( x ) = Fx + f [ ] ( x ) + f [ ] ( x ) + ...

(124)

30

y = h ( x ) = Hx + h[ 2 ] ( x ) + h[ 3 ] ( x ) + ...

(125)

31

x ( 0 ) x0 = 0

(126)

32

where f [ d ] ( x ) , h[ d ] ( x ) denote the degree d terms in the Taylor series expansion of f (x), h(x). One

33
34

seeks a local diffeomorphism z = (x) and an output injection (y) that transforms Eqs. (30) - (32)
into

35

z& = Az ( y )
E6-43-21-15-TXT_NT.doc

(127)
12

06/23/03

where A is Hurwitz. If these can be found, then the observer

z& = Az ( y )

(128)

x = 1 ( z )

(129)

4
5
6
7
8
9
10

has exponentially stable linear error dynamics, with linear error dynamics
z&% = Az% .

Necessary conditions for the existence of such a local diffeomorphism and output injection are that
H, F be an observable pair, and that there are no resonances between the eigenvalues of F and those
of A. Suppose the spectrum of F is (1, . . . , n) and the spectrum of A is (1,. . . , n). A resonance
occurs when there is a nonzero vector k = (k1, . . . , kn) of nonnegative integers and some j, 1 j n,
such that
n

11

(130)

k
i

i =1

= j

(131)

12
13
14
15
16
17
18

As originally proposed, a sufficient condition for the existence of such a local diffeomorphism and
output injection was that the spectrum of F be either in the open left half plane or the open right half
plane. This ruled out many interesting cases. The former implies that the system is exponentially
stable, and so x (t) = 0 is a convergent observer. The latter implies that the system is exponentially
unstable, and so a local observer is not of much use. But, recently, a much weaker sufficient
condition has been found, and the spectrum of F can be arbitrary. This method has not been
extended to systems with inputs as yet.

19
20
21
22

All of the above approaches lend themselves to power series methods for finding the desired
transformations term by term up to any degree of accuracy. For brevity, we illustrate this for only
the last method. The Hurwitz matrix A, the local diffeomorphism z = (x) and the output injection
(y) must satisfy the first order partial differential equation

23
24
25

( x ) f ( x ) = A ( x ) ( y )
x
.

We expand in a power series assuming without loss of generality that the linear part of (x) is the
identity,

26

( x ) = x + [

27

( y ) = BHx + [

28
29

(132)

2]

( x ) + [ 3 ] ( x ) + ...
2]

(133)

( y ) + [ 3 ] ( y ) + ...

(134)

The linear part of Eq. (132) is

A = F + BH .

(135)

30

and if H, F is an observable pair the spectrum of A can be set arbitrarily by choice of B.

31

The quadratic part of Eq. (132) is

32
33
34
35

[ 2 ]
( x ) Fx A[ 2 ] ( x ) = f [ 2 ] ( x ) [ 2 ] ( y ) .
x

(136)

If there is no resonance, then this equation has a unique solution, [ 2 ] (x), for any right side
2
2
equation. The unknown [ ] ( y ) can be chosen to keep [ ] ( x ) close to 0 so (x) remains a
diffeomorphism over a wide region.
E6-43-21-15-TXT_NT.doc

13

06/23/03

1
2
3

The degree d part of Eq. (132) is

d 1
[ d ]
[ i ]
( x ) Fx A[ d ] ( x ) =
( x ) f [ d i ] ( x ) [ d ] ( y ) .
x
i =1 x

If there is no resonance, then this equation has a unique solution, [

[d ]

[d ]

the unknown ( y ) can be chosen to keep

over a wide region.

If we stop at degree d

(x)

(137)
d]

( x ) , for any right side. Again

close to 0 so (x) remains a diffeomorphism

( x ) = x + [ 2 ] ( x ) + ... + [ d ] ( x )

(138)

( y ) = BHx + [ 2 ] ( y ) + ... + [

(139)

9
10

d]

( y)

the observer Eq. (128) - (129) has approximately linear error dynamics
d +1
z&% = Az% + O ( z )

(140)

11

5. High Gain Observers

12
13

A high gain observer can be constructed for a system in uniformly observable form Eqs. (48) - (51)
which satisfies certain Lipschitz conditions. For scalar output systems, it takes the form

14

y = 1 + g 0 ( u )

(141)

&
1 = 2 + g1 1 , u + L1 ( y y )
15

(142)

&
i = i +1 + g 2 1 ,..., i , u + Li ( y y )
16
17
18

19
20
21
22
23
24
25
26
27
28
29

(143)

( )

&
n = f + g n 1 ,..., n , u + Ln ( y y )

(144)

The gain is chosen as

n
Li = i
i

(145)

where is a constant that depends on the Lipschitz constant of the system. It has been proven that
if is sufficiently large, then the observer estimate is globally asymptotically convergent to the
system state. Of course, this assumes that there is no noise in the dynamics nor the observations.
High gain observers converge to differentiators as the gain increases and therefore become very
sensitive to noise. If one uses a high gain observer, then even small noise can lead to substantial
degradation in the performance of the observer. In most of the examples in the literature, the gain
parameter is chosen small without regard to global convergence. The resulting observers seem to
perform well in simulations, but this is probably due to linear rather than nonlinear effects. There is
no theoretical explanation of why this should happen. When the gain is set too low, there is no
guarantee of convergence even if there is no noise.

E6-43-21-15-TXT_NT.doc

14

06/23/03

6. Nonlinear Filtering

2
3

The stochastic approach to nonlinear estimation is to assume that the nonlinear system is described
by Itos stochastic differential equations

dx = f ( x, u ) dt + g ( x ) dw

(146)

dy = h ( x, u ) dt + dv

(147)

x ( 0 ) = x0

(148)

7
8
9
10

corrupted by observation and driving noises. Here w(t) and v(t) are standard Wiener processes and
the initial condition x0 is assumed to have density p0(x). It can be shown that the unnormalized
density q(x, t) of x(t) conditioned on the past controls and observations satisfies a stochastic partial
differential equation called the Zakai equation,
n

11
12
13

dq ( x, t ) =
i =1

p
n

2
f i ( x, u ) q ( x , t ) ) +
aij ( x ) q ( x, t ) ) + q ( x, t ) hi ( x, u )dyi
(
(
xi
i , j =1 xi x j
i =1

q ( x, 0 ) = p 0 ( x )

(149)
(150)

where
n

14
15

16
17

18
19
20

aij ( x ) = gil ( x ) glj ( x )

l =1

(151)

The conditional density p(x, t) is obtained by normalizing q(x, t)

p ( x, t ) =

q ( x, t )
q ( , t ) d

(152)

and the conditional mean is

x ( t ) =

q ( , t ) d
q ( , t ) d

(153)

One can also take the conditional mode as the estimate,

x ( t ) = arg max x q ( x, t )

(154)

21
22
23
24
25
26
27

The Zakai equation is a stochastic parabolic partial differential equation in the Ito sense and its
numerical integration is quite delicate. It is theoretically quite important but of limited practical use.
It is driven by the observation process y(t), so it must be solved in real time. This is generally not
possible when the state dimension is greater than 1 and is difficult even when it is 1 because the
accuracy of the solution is dictated by the step sizes in x and t. Hence, the numerical integration of
the Zakai equation is generally not a practical approach to estimating the state of a nonlinear
system.

28
29
30

Notice when thought of as an observer, the Zakai Eq. (149) and state estimate Eqs. (153) or (154)
is an expanded order observer Eqs. (18) - (19) with the state z(t) being q(, t). The state of the
observer is infinite dimensional.

E6-43-21-15-TXT_NT.doc

15

06/23/03

7. Minimum Energy and H Estimation

2
3

An alternative to the stochastic approach is to assume that the noises w(t), v(t) are not stochastic but
unknown L2 functions corrupting the system,

x& = f ( x, u ) + g ( x ) w

(155)

y = h ( x, u ) + v

(156)

x ( 0 ) = x 0 + x% 0

(157)

7
8
9
10
11
12
13

14
15
16
17

where x% 0 is an unknown error in the initial condition. We seek the initial state error x% 0 and noises

{( w ( ) , v ( ) ) : 0 t }

(158)

of minimum energy
2
2
1 0 2 1 t
x% + w ( ) + v ( ) d
2
2 0

(159)

which are consistent with the initial estimate x 0 and the past controls and observations

{( u ( ) , y ( ) ) : 0 t }

(160)

and the system Eqs. (155) - (157). This is an optimal control problem. We consider the optimal cost
2
2
1 0 2 1 t

%
Q ( x, t ) = inf
x
+
w

+
y

h
x

,
u

d
(
)
(
)
(
)
(
)
(
)

0
0
x% , w( ) 2
2

(161)

subject to Eqs. (155) - (157) and x(t) = x. The optimal estimate is then given by
x ( t ) = arg min x Q ( x, t )

(162)

The dynamic programming approach yields a partial differential equation for Q(x, t),
0=

18

n
2
Q
Q
1 n Q
Q
1
x
,
t
+
x
,
t
f
x
,
u
+
( ) ( ) i ( ) ( x, t ) aij ( x ) ( x, t ) y ( t ) h ( x, u )
t
2 i , j =1 xi
x j
2
i =1 xi

(163)

19
20
21
22
23
24

It is of the Hamilton-Jacobian type, first order, nonlinear and again driven by the observations. As
with the Zakai equation, it is very difficult to compute an accurate solution in real time. Moreover,
it may not admit a smooth solution so the Eq. (163) must be interpreted in the viscosity sense. This
is an infinite dimensional observer with state Q(,t) evolving according to Eq. (163) with the state
estimate given by Eq. (162). Hence it is of limited practical use. In some sense, it is the larger
deviations limit of the Zakai observer.

25
26
27
28
29

The H approach to nonlinear estimation is an extension of this. If there were no noise and the
initial conditions were known exactly, then the estimation of x(t) would be relatively easy, just
integrate the differential equation. If there are disturbances, i.e., driving and observation noises and
an unknown initial condition, we would like the gain from these to the estimation error to be as
small as possible. Hence, our estimate z ( ) should satisfy the

30

2
2
t
2
2

x ( ) x ( ) 2 x% 0 + w ( ) + v ( ) d
0

E6-43-21-15-TXT_NT.doc

16

(164)

06/23/03

1
2
3

for as small as possible. Finding the minimal is a difficult problem to solve directly, so instead,
we choose a gain level , and see if we can construct an observer that achieves it. If this is
possible, then we try to do it for a smaller , if not we try a larger , etc.

We seek an estimator such that if

5
6
7

2 0 2 2
Q ( x, t ) = inf
x% +

x% 0 , w( )
2
2

w( ) + y ( ) h ( x ( ), u ( ) ) d
2

1 t

x
x
d
(
)
(
)

0
2

(165)

subject to Eqs. (155) - (157) and x(t) = x then Q(x, t) 0. If such a Q(x, t) exists then our estimate is
given by Eq. (162) and Q satisfies in the viscosity sense
0=

n
Q
Q
1 n Q
Q
x
,
t
+
x
,
t
f
x
,
u
+
( ) ( ) i( )
( x, t ) aij ( x ) ( x, t )
2
t
2 i , j =1 xi
x j
i =1 xi

(166)

2
2
2
1

y ( t ) h ( x, u ) + x z ( ) .
2
2

9
10

Again this is an infinite dimensional observer and hence of limited practical use. There is a
stochastic version of the H observer, but it too is infinite dimensional.

11
12
13
14
15
16

From a theoretical point of view there are two additional problems with the minimum energy and
H observers. The first is that the criteria weights the distant past as much as the present, and the
other is that Q(x, t) tends to grow. These dont arise in the Zakai estimator because the second order
terms in the Zakai partial differential equation tend to diffuse away the past, and the solution can
always be renormalized Eq. (152), and the partial differential equation restarted. One way around
these difficulties is to introduce a forgetting factor 0. For the H observer it takes the form

17
18

2
2 t 0 2 2 t ( t )

2
x% + e
w( ) + y ( ) h ( x ( ), u ( ) ) d
e
2 0
2

Q ( x, t ) = inf
0
2
x% , w( )
1 t ( t )

x ( ) x ( ) d
0 e

and the partial differential equation becomes

0 = Q ( x , t ) +

19

(167)

n
Q
Q
1
( x, t ) + ( x , t ) f i ( x , u ) + 2
t
2
i =1 xi

Q
Q
x ( x, t ) a ( x ) x ( x, t )
n

ij

i , j =1

2
2
2
1
y ( t ) h ( x, u ) + x z ( ) .
2
2

(168)

20
21
22
23
24
25

In H estimation one tries to find an observer that minimizes the induced L2 gain from disturbances
to estimation error. An alternate is to use a different Lp norm on the disturbances and errors.
Probably the most useful norm from a practical point view is the L norm, so that bounded
disturbances produce bounded estimation errors. Unfortunately, the mathematics is not so nice and
the resulting partial differential equation is even more difficult to state and solve. And it is still an
infinite dimensional observer.

26
27
28
29
30

As the reader may surmise, it is not hard, theoretically to construct infinite dimensional observers
for finite dimensional nonlinear systems. In fact every nonlinear observer has an infinite
dimensional realization as we shall show in a moment. The hard part is finding a finite dimensional
realization. Considerable effort has been expended on this topic for the Zakai observer with only
limited success.

31

Recall that an observer is a causal functional

E6-43-21-15-TXT_NT.doc

17

06/23/03

x 0

u ( ) a x ( t ) , 0 t
y ( )

(169)

from the initial state estimate, x 0 and the past control and observation, u ( ) , y ( ) , 0 t , to the

current state estimate, x ( t ) such that x ( t ) x ( t ) 0 as t .

For such an observer x ( t ) , there exists a function Q(x, t) which is a causal functional of the initial

5
6
7

state estimate x 0 and the past controls and observations, {u ( ) , y ( ) , 0 t } such that Q(x, t)
has a minimum at x = x ( t ) and along any state trajectory x(t), input u(t), observation y(t) and
noises w(t), v(t) are consistent with the system
t2

8
9

10
11

Q ( x ( t2 ) , t2 ) Q ( x ( t1 ) , t1 ) + w ( ) + v ( ) d
2

(170)

t1

Define Q(x, t) as
2
t
2
2

Q ( x, t ) = inf z ( 1 ) x ( 1 ) + w ( ) + v ( ) d : 0 1 t
1

(171)

where the infimum is over all z(), w(), v() satisfying

12

d
z ( ) = f ( z ( ), u ( ) ) + g ( z ( ) ) w ( )
d

(172)

13

y( ) = h( z( ))+ v( )

(173)

14

z(t ) = x

(174)

15
16

and u(), y() are the control and observation. This can be viewed as an infinite dimensional
observer with state Q(, t) and output Eq. (162).

17

8. Multiple Extended Kalman Filters

18
19
20
21
22
23

There is a hybrid approach to the nonlinear estimation. One can compute the solution of one of the
partial differential equations described previously on a very coarse spatial and temporal grid. Of
course, this will not lead to a very accurate estimate. But at each local maximum of the coarse
solution to the Zakai equation one can initiate an extended Kalman filter, and this should quickly
evolve to the true mode if it is nearby. For the minimum energy and H partial differential
equations, one initiates the extended Kalman filter at the local minima.

24
25

The relative accuracy of the different Kalman filters can be assessed by how well they explain the
most recent observations. If the observation is y(t) and an estimate is x ( t ) , then a measure of its
recent accuracy is q(t) where

26
27
28
29

q& ( t ) = q ( t ) +

2
1
y ( t ) h ( x ( t ) , u ( t ) )
2

(175)

where 0 is a forgetting factor. The extended Kalman filter estimate x ( t ) with the smallest
current value of q(t) is taken as true. This leads to extremely fast transitions between estimates as

E6-43-21-15-TXT_NT.doc

18

06/23/03

1
2

different q(t) cross. The bookkeeping of creating and merging extended Kalman filters is
substantial, but the computation of their updates is relatively trivial.

3
4
5

One can eliminate the step of solving the partial differential equation and instead continuously
initiate extended Kalman filters at likely spots in the state space. The practicality of these
approaches has yet to be demonstrated.

9. Conclusion

7
8
9
10
11
12
13

We have surveyed some but not all of the ways of constructing an observer for a nonlinear system.
The high gain observer is a theoretical finite dimensional solution to a broad class of noise free
problems, but performs poorly when noise is present. The Zakai, minimum energy, and H
observers are theoretically infinite dimensional solutions to broad classes of noisy problems. But
none of these are practically implementable. The linearization techniques give local and sometimes
only approximate solutions for narrower classes of problems, and sometimes they are very hard to
implement.

14
15
16

The extended Kalman filter is probably still the most robust and practical approach for most
problems. If there are substantial nonlinearities, e.g., multiple stable equilibria and/or stable limit
cycles, then the use of multiple extended Kalman filters is probably the preferred approach.

17
18
19
20
21

All these methods rely on the linear observability of the system to insure convergence. Other than
the method of semi-diffeomorphisms, little is known about constructing observers around an
equilibrium state where the system is not linearly observable. In summary, much is known about
state reconstruction by observers but much remains to be done to find implementable solutions for
broad classes of nonlinear systems.

22

Glossary

23
24
25
26
27
28
29
30

System:

A system is a set of components, physical or otherwise, which are connected in

such a manner as to form and act as an entire unit. Frequently, we use the term to
describe a mathematical model of a physical system. A system has state variables
which are its internal memory, input variables by which the external world
affects the system, and output variables by which it affects the external world.
The input variables that can be chosen or measured by the operator are called
controls. Other input variables are frequently viewed as noise. The output
variables that can be measured are called measurements.

31
32
33
34

Observer:

A system which accepts as inputs the measured inputs and outputs of another
system and returns an estimate of the other systems state. The observer is said to
be of full, reduced, or expanded order when its state is of the same, smaller, or
larger dimension than the other system.

35
36

Autonomous
system:

A system whose behavior is independent of time.

37

Causality:

A property of systems, future inputs cannot affect past outputs.

38

Controls:

Those inputs that can be chosen by the operator.

39
40

Detectable
system:

A system where the measurements eventually determine the state.

41
42
43

Error
linearization:

A technique for constructing an observer which results in linear error dynamics

in some coordinate system

E6-43-21-15-TXT_NT.doc

19

06/23/03

1
2

Extended Kalman
filter:
An extension of the Kalman filter to nonlinear systems.

3
4
5

H
estimation:

6
7

Hurwitz
matrix:

A square matrix with all its eigenvalues in the open left half plane.

Kalman filter:

A linear observer constructed using stochastic methods.

An approach to estimation that attempts to minimize the gain between the size of
noise and the size of the estimation error.

9
10

Linear approximating
system:
The linear part of a nonlinear system around an equilibrium point.

11

Linear system:

A system described by linear equations.

12
13

Linear
observer:

A linear system that is an observer for another system.

14
15

Luenberger
observer:

A reduced order linear observer for a linear system.

16
17
18

Minimum energy
estimation:
An approach to estimation that minimizes the size of the noise necessary to
produce the observations

19
20

Nonautonomous
system:
A system whose behavior varies with time.

21
22
23

Nonlinear
filter:

24
25

Nonlinear
system:

A system described by nonlinear equations.

26
27

Nonlinear
observer:

A nonlinear system that is an observer for another nonlinear system.

28
29
30

Observable
form:

31
32

Observable
system:

33

Bibliography

34
35

Bestle D. and Zeitz M. (1983). Canonical form observer design for non-linear time-variable systems. Internat. J.
Control 38, 419-431. [One of the first papers to treat nonlinear observer design by error linearization.]

36
37
38
39

Davis M.H.A. and Marcus S.I. (1981). An introduction to nonlinear filtering. In M. Hazewinkel, J.C. Willems, eds.,
Stochastic Systems: The Mathematics of Filtering and Identification and Applications, pp. 53-76. Dordrecht: D. Reidel
Publishing. [This paper contains a rigorous derivation of the Zakai equation for the evolution of the conditional density
of a nonlinear filter.]

40
41
42

Gauthier J.P., Hammouri H. and Othman S. (1992). A simple observer for nonlinear systems with applications to
bioreactors. IEEE Trans. Auto. Contr. 37, 875-880. [This paper introduces the high gain observer and proves its
convergence if the gain is high enough and there is no noise.]

43
44

Gelb A. (1989). Applied Optimal Estimation. Cambridge, MA: MIT Press. [An excellent introduction to Kalman
filtering, extended Kalman filtering and their extensions.]

An infinite dimensional, nonlinear observer constructed using stochastic methods

described by a stochastic partial differential equation, called the Zakai equation.

A mathematical way of representing a system using the measurement function

and its derivatives as state coordinates.
A system where the measurements and their time derivatives determine the state.

E6-43-21-15-TXT_NT.doc

20

06/23/03

1
2

Isidori A. (1985). Nonlinear Control Systems. New York: Springer-Verlag. [A basic reference on nonlinear control
systems.]

3
4

Kalman R.E. (1960). A new approach to linear filtering and prediction problems. Trans. of ASME, Part D, J. of Basic
Engineering 82, 35-45. [The introduction of the Kalman filter.]

5
6

Kalman R.E. and Bucy R.S. (1961). New results in linear filtering and prediction theory. Trans. of ASME, Part D, J. of
Basic Engineering 83, 95-108. [More on the Kalman filter.]

7
8
9

Kazantzis N. and Kravaris C. (1998). Nonlinear observer design using Lyapunovs auxiliary theorem. Systems and
Control Letters 34, 241-247. [This paper introduces an improved method of nonlinear observer design via error
linearization that is applicable to a wider class of nonlinear systems.]

10
11

Krener A.J. (1997). Necessary and sufficient conditions for nonlinear worst case (H-infinity) control and estimation.
Journal of Mathematical Systems, Estimation, and Control 7, 81-106. [An introduction to H-infinity observer design.]

12
13
14

Krener A.J. and Duarte A. (1996). A hybrid computational approach to nonlinear estimation. In Proc. of 35th
Conference on Decision and Control, pp. 1815-1819. Kobe, Japan. [This paper presents the multiple extended Kalman
filtering method.]

15
16

Krener A.J. and Isidori A. (1983). Linearization by output injection and nonlinear observers. Systems and Control
Letters 3, 47-52. [One of the first papers to treat nonlinear observer design by error linearization.]

17
18
19
20

Krener A.J. and Respondek W. (1985). Nonlinear observers with linearizable error dynamics. SIAM J. Control and
Optimization 23, 197-216. [This paper extends the error linearization method to systems with vector observations].
Luenberger D.G. (1964). Observing the state of a linear system. IEEE Trans. on Military Electronics 8, 74-80. [This
paper presents a reduced order linear observer for a linear system.]

21
22

Misawa E.A. and Hedrick J.K. (1989). Nonlinear observers, a state of the art survey. Trans. of ASME, J. of Dynamic
Systems, Measurement and Control 111, 344-352. [A survey of observer design methods.]

23
24

Muske K.R. and Edgar T.F. (1997). Nonlinear state estimation. In M.A. Henson, D.E. Seborg, eds., Nonlinear Process
Control, Upper Saddle River, NJ: Prentice Hall. [Another survey of observer design methods.]

25
26

Nijmeijer H. and Fossen T. (1999). New Directions in Nonlinear Observer Design. New York: Springer-Verlag. [The
proceedings of a conference on nonlinear observer design.]

27
28

Nijmeijer H. and van der Schaft A. (1990). Nonlinear Dynamical Control Systems, 467pp. New York: Springer-Verlag.
[Another basic reference on nonlinear control systems.]

29
30

Xia X. and Zeitz M. (1997). On nonlinear continuous observers. Internat. J. Control 66, 943-954. [Introduces the use of
semi-diffeomorphisms to observer design.]

E6-43-21-15-TXT_NT.doc

21

06/23/03