Stability Analysis of Nonlinear Systems using

Frozen Stationary Linearization

Anders Hansson, Anders Helmersson and Torkel Glad
Division of Automatic Control
Department of Electrical Engineering
Link oping University
SE-581 83 Link oping, Sweden
[email protected]
AbstractIn this paper is discussed how to compute
stability regions for nonlinear systems with slowly varying
parameters using frozen stationary linearization. It is shown
that larger stability regions can be obtained as compared to
traditional approaches using recent stability results for linear
parameter-varying systems.
Key Words: Nonlinear Systems, Stability Analysis, Lya-
punov Stability, Linear Parameter-Varying Systems, Gain-
Scheduling.
I. INTRODUCTION
Stability analysis of nonlinear differential equations has
been an active research area since the pioneering work by
Lyapunov in 1892, e.g. [11]. Local stability analysis based
on linearization around a stationary point is taught in under-
graduate courses. Linearization around a nominal trajectory
and optimal control of the resulting time-varying system
was introduced in the early optimal control literature, e.g.
[5, Chapter 6.4], and [4]. The derivation of the time-
varying linear system is in general quite involved, since
it is required to nd the nominal solution. Especially this
is the case when investigating input-to-state stability. For
slowly varying systems this problem can be circumvented
by considering the input as a frozen parameter. This
has important applications in design and analysis of gain-
scheduled controllers.
Gain-scheduling is a very powerful control methodology
for control of systems with varying process dynamics that
can be predicted. It was originally used for ight control
systems, [19]. With digital implementations it s now in-
creasingly used also in process control. Other areas where
gain scheduling is applied are e.g. autopilots for ships and
combustion control for cars, [2]. Scheduling variables in
ight control systems are typically velocity, altitude and
angle of attack. In process industry a typical scheduling
variable is production rate.
Traditionally, the design of gain-scheduled controllers is
carried out by rst linearizing the system to be controlled
at a discrete number of operational points parameterized
by the scheduling variables. Then linear controllers are
designed for each of these operational points, and the overall
controller is obtained by interpolation. Even if each linear
controller is stable for its linearized system model, there
is no guarantee that the overall control scheme is stable
when the parameter that the scheduling is based on starts
to vary, [18]. However, if bounds on the rate of variation
of the parameter is imposed it is possible to show that
stability is recovered, [10], [16], [9], [17], [15]. Common to
these approaches is that they all assume that the linearized
system is exponentially stable, uniformly in the parameter.
Under a rate-constraint on the parameter they then typically
show that the nonlinear system is uniformly ultimately
bounded locally. Recently much attention has been given to
analyzing the Linear Parameter Varying (LPV) system, i.e.
the linearized system for time-varying parameters, e.g. [6],
[1], [7]. It has been shown that computationally attractive
schemes can be derived to construct parameter-dependent
Lyapunov-functions which prove global stability for the
LPV system. It is the scope of this paper to make of use of
these new results to obtain larger stability regions for the
nonlinear system.
The remaining part of the paper is organized as follows.
In Section 2 the nonlinear system is dened together with
the frozen stationary points. Also the linearized system
is derived. In Section 3 the stability result is presented.
It is shown that the solution is stable assuming that the
parameters vary slowly. In Section 4 the results of this paper
are related to previous work. In Section 5 it is discussed
how to compute Lyapunov functions for the LPV system.
In Section 6 an example is investigated, and it is shown
that the stability region is larger for the proposed approach
as compared to previous approaches. Finally, in Section 7
some conclusions and remarks about extensions are given.
II. PRELIMINARIES
Consider the nonlinear differential equation
x = f (x, ), t 0 (1)
where x(t) R
n
denotes the state vector, and where (t)
R
m
, t 0 is a parameter vector. We remark that any
explicit time-dependence in f can be taken care of by incor-
porating t as a component of . When analyzing the system
Proceeding of the 2004 American Control Conference
Boston, Massachusetts June 30 - July 2, 2004
0-7803-8335-4/04/$17.00 2004 AACC
WeA13.3
424
dened by (1) we linearize around a frozenstationary
point, [8, Chapter 5.7], i.e. x
0
should satisfy
f (x
0
(t), (t)) = 0, t 0 (2)
Notice that we can express x
0
(t) as a relation with (t),
which will only depend on the current time t. This is a
standard approach for slowly varying parameters . We
will assume that (2) denes x
0
(t) as a function of (t), i.e.
x
0
(t) =((t)). We will also assume that is differentiable
with respect to . The analysis will be based on the
linearized system
z = A()z (3)
where A() =
f
x
((), ). This system is what is called an
LPV system.
Dene the change of variables z = x x
0
. This will
transform the system into
z = g(z, ) x
0
(4)
where g(z, ) = f (z +(), ). Notice that
x
0
=

() (5)
We assume that
_
_
_
_

_
_
_
_
2
(6)
. In case we know that certain parameters do not vary
with time we may remove the corresponding components
in and the corresponding columns in

in the expression
for x
0
. This may result in lower values of . In what follows
we will assume that g : D R
n
is such that solutions
of (4) are well dened t 0, . We will also assume
that the Jacobian matrix g/z is bounded and Lipschitz on
D, uniformly in t and , i.e.
_
_
_
_
g
i
z
(z
1
, )
g
i
z
(z
2
, )
_
_
_
_
2
L
1
z
1
z
2

2
(7)
z
1
, z
2
D, .
III. STABILITY ANALYSIS
We will now investigate stability of solutions x of (1)
or equivalently solutions z of (4). We will show stability
assuming that the LPV system admits a global Lyapunov
function. We have to require that x
0
(t)
2
, t 0,
, where is some constant. The stability concept
for the nonlinear system is that if z(0) E D, then z(t)
will converge to a region F E in nite time. Hence the
solution of (4) is uniformly ultimately bounded, e.g. [8,
Denition 5.1].
By the mean value theorem
g
i
(z, ) = g
i
(0, ) +
g
z
(
i
, )z
where
i
is a point on the line segment connecting z to the
origin. Since g
i
(0, ) = 0 by construction, we have
g
i
(z, ) =
g
z
(
i
, )z =
g
z
(0, )z
+
_
g
z
(
i
, )
g
z
(0, )
_
z
Hence
g(z, ) = A()z +h(z, ) (8)
where
h
i
(z, ) =
_
g
i
z
(
i
, )
g
i
z
(0, )
_
z
The function h(z, ) satises
h(z, )
2
Lz
2
2
(9)
where L =

nL
1
by (7). Now assume that there is a contin-
uous, differentiable, bounded, positive denite, symmetric
matrix P() such that
0 < c
1
I P() c
2
I (10)
and

P() +P()A() +A
T
()P() c
3
I < 0 (11)
t 0, , where

P() =
m
i=1
P

i
()
i
. We remark
that it is trivial to show that this is a global Lyapunov
function for the LPV system in (3). We will now use
V(z, ) =z
T
P()z as a Lyapunov function candidate for the
nonlinear system (4). The derivative of V(z, ) along the
trajectories of the system is given by

V(z, ) = z
T
P()[g(z, ) x
0
]
+[g(z, ) x
0
]
T
P()z +z
T

P()z
= z
T
_
P()A() +A
T
()P() +

P()

z
+2z
T
P()[h(z, ) x
0
]
c
3
z
2
2
+2z
T
P()[h(z, ) x
0
]
c
3
z
2
2
+2c
2
Lz
3
2
+2c
2
z
2
=
_
c
3
+2c
2
_
Lz
2
+

z
2
__
z
2
2
if x
0
(t)
2
, t 0, .
The region E \F in which

V(z, ) <0 is given by the set
of z such that z D and
c
3
+2c
2
_
Lz
2
+

z
2
_
< 0
which is equivalent to
_
z
2

c
3
4c
2
L
_
2

c
2
3
16c
2
2
L
2
+

L
< 0
We rst notice that there is a solution to this inequality if
and only if
<
c
2
3
16c
2
2
L
. (12)
425
The sets E and F are hence given by
E =
_
z

z
2

c
3
4c
2
L
+

c
2
3
16c
2
2
L
2

L
_
D (13)
F =
_
z

z
2

c
3
4c
2
L

c
2
3
16c
2
2
L
2

L
_
D (14)
Notice that because of (6) we may replace the condition on
x
0
with a condition on , i.e.

u = / implies that
x
0

2
by (5).
It is clear from the bound on that it is desirable to
have c
2
as small as possible and c
3
as large as possible
in order to allow for rapid variations in x
0
. This will also
maximize the region E \F. Since c
2
and c
3
only enters as
a function of the quotient c
3
/c
2
, the obvious approach is to
try to nd P() and c
i
, i=1,2,3, that satises (1011) and
maximizes c
3
/c
2
. We will later on discuss how to do this
in a systematic way.
IV. RELATION TO PREVIOUS WORK
The difference between the approach taken in this work as
compared to what has previously been done is that typically
it is assumed that there is a matrix P(), which satises (10)
and
P()A() +A
T
()P() c
3
I < 0 (15)
m

i=1

i
()
2
c
4
(16)
instead of (11) , e.g. [8, Chapter 5.7]. Notice that if
A() is Hurwitz uniformly in , then by [8, Lemma 5.12]
there is always a solution to these inequalities . Then
the derivative of V(z, ) along the trajectories of the system
is given by

V(z, ) = z
T
P()[g(z, ) x
0
]
+[g(z, ) x
0
]
T
P()z +z
T

P()z
= z
T
_
P()A() +A
T
()P()

z
+z
T

P()z +2z
T
P()[h(z, ) x
0
]
c
3
z
2
2
+z
T
m

i=1
P

i
()
i
z
+2z
T
P()[h(z, ) x
0
]
c
3
z
2
2
+c
4
uz
2
2
+2c
2
Lz
3
2
+2c
2
z
2
=
_
c
3
+c
4
u+2c
2
_
Lz
2
+

z
2
__
z
2
2
if (t)

u, t 0, . The region E \F in which

V(z, ) <0 is now given by the set of z such that z D and

c
3
+c
4
u+2c
2
_
Lz
2
+

z
2
_
< 0
which is equivalent to
_
z
2

c
3
c
4
u
4c
2
L
_
2

(c
3
c
4
u)
2
16c
2
2
L
2
+

L
< 0
We rst notice that there is a solution to this inequality if
and only if
<
c
2
3
16c
2
2
L
where c
3
= c
3
c
4
u. We then notice that the bound is
increasing in c
3
/c
2
. The sets E and F are given by
E =
_
z

z
2

c
3
4c
2
L
+

c
2
3
16c
2
2
L
2

L
_
D (17)
F =
_
z

z
2

c
3
4c
2
L

c
2
3
16c
2
2
L
2

L
_
D (18)
We conclude this section by noting that there are two
obvious approaches for obtaining large stability regions.
The rst one is to maximize c
3
/c
2
with respect to (10)
and (15) and then compute c
4
as the smallest value which
satises (16). The second approach is to maximize c
3
/c
2
=
(c
3
c
4
u)/c
2
with respect to (10), (15) and (16) for a xed
value of u.
V. LYAPUNOV FUNCTIONS FOR THE LPV SYSTEM
In this section we will discuss how to nd solutions
to (1011). First we notice that we can normalize these
inequalities with c
2
to obtain
0 < c
1
I

P() I

P() +

P()A() +A
T
()

P() c
3
I < 0
where c
1
= c
1
/c
2
, c
3
= c
3
/c
2
and

P() = P()/c
2
. Clearly
we should just x c
1
to some small number. Then we
should look for c
3
=c
3
/c
2
and

P() that satises the matrix
inequalities and which maximizes c
3
. This will as already
mentioned both maximize the bound on and maximize
the region E \F. The optimization problem is however not
tractable as it stands, since we have to search for general

P(). Several different approaches for how to circumvent

this problem has been proposed in the literature, see e.g.
[3], [6], [1], [7]. One of the more appealing approaches,
[7], assumes that A() and P() are rational functions of
given in Linear Fractional Transformation (LFT) form. This
is the approach we will pursue in this paper.
The LPV system is not always such that the A-matrix is
rational in . However, any rational matrix-valued function
can be realized as an LFT, e.g. [12]. On a compact domain
any function can be arbitrarily well approximated by a
rational function. Therefore for any >0 it will be possible
to nd an LFT-realization of a rational A
r
() such that

A()
2
, , where

A() = A() A
r
(). This can
be taken into account in the stability analysis in Section 3
by redening (8), i.e. we write
g(z, ) = A
r
()z +h
r
(z, )
where h
r
(z, ) =h(z, )+

A()z. Then we do the rest of the
analysis based on A
r
and h
r
instead of A and h, where we
make use of the fact that h
r
by (9) satises the bound
h
r
(z, )
2
h(z, )
2
+

A()z
2
Lz
2
2
+z
2
426
It turns out that if c
3
in (11) is replaced with c
3
+2c
2

and A() with A

r
(), then the constraint on and the
region E \ F will remain the same. We also have to add
the constraint that c
3
> 0. Thus we have shown how to
use an approximation A
r
of A and obtain similar results as
for rational A assuming that we tighten the constraint in
(11). Notice that after scaling with c
2
we will again have
linear matrix inequalities, since the modication of c
3
is
proportional to c
2
. To summarize, the inequalities now read
0 < c
1
I

P() I

P() +

P()A
r
() +A
T
r
()

P() + c
3
I 2I < 0
together with c
3
> 0. Before concluding this section we
remark that there are alternative ways of taking care of the
approximation error

A. One can consider an extended LFT
description of A
r
by augmenting with entries that will take

A into account, e.g. [12]. This will increase the dimension of

the LFT-description, and hence increase the computational
complexity. Note that both approaches will result in a global
Lyapunov function for the original LPV system. Therefore
we do not believe it to be advantageous to use the latter
approach.
VI. EXAMPLE
Let
x
1
= (x
1
1)x
1
+2(+1)x
2
+

2
400
(10081
2
)
x
2
= 2(1)x
1
x
2
where || 0.5 and | | u. The Lipschitz constant may be
taken as L = 2. The linearized system becomes
z
1
= (2x
10
1)z
1
+2(+1)z
2
z
2
= 2(1)z
1
z
2
where we will look at frozen stationary solutions given
by x
10
=
2
/20 and x
20
= 2(1)x
10
. Hence,
_
_
_
_

_
_
_
_
2
=

10
_
_
_
_
_
1
32
__
_
_
_
2
=

10
_
9
2
12+5
The LPV system is given by
z
1
= (
2
/21)z
1
+2(+1)z
2
z
2
= 2(1)z
1
z
2
which is stable for constant [
_
50/41,
_
50/41].
In the relevant interval ( [0.5, 0.5]), we have =
max

2
= 0.1820.
The A-matrix can be written as an LFT in :
A =
_

2
10
1 2(+1)
2(1) 1
_
=
_
_
1 2+2

10
2 1 2
1 0 0
_
_

=
_

_
1 2 0 2
2 1 2 0
1 0 0 0
0 1
1
20
0
_

_
0
0
_
. .

=
_
A
0
B
0
C
0
D
0
_
= A
0
+B
0
(I D
0
)
1
C
0
where denotes the Redheffer star product, e.g. [14]. We
will here assume that the Lyapunov function is quadratic
and parameterized by V(x) = x
T
P()x, where
P() =
_
I
(I D
0
)
1
C
0
_
T
P
0
_
I
(I D
0
)
1
C
0
_
Maximizing c
3
subject to 10
9
I <

P() < I and (15)
for all [0.5, 0.5], yields c
3
= 1.0895 and c
4
=
max

2
= 0.957. The stability region in terms of c
3

u c
4
with respect to u is shown in Figure 1 as the dash-dotted
line.
We can increase the stability region by maximizing
c
3
(u) u c
4
(u) for each value of u subject to (15) and (16).
The result is shown as the dashed line in Figure 1.
If we instead use (11) where | | u for maximizing c
3
(u)
we obtain an even larger stability region, see solid line in
Figure 1. The stability regions are in this case dened by
(13) and (14). The only difference compared to (17) and
(18) is that c
3
c
4
u is substituted for c
3
(u). As can be seen
in Figure 1, the stability region is extended in our approach
compared to previous work.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
0
0.2
0.4
0.6
0.8
1
PSfrag replacements

c
3
,

c
3

c
4
c
3
(u) u c
4
(u)
c
3
(u)
c
3
u c
4
u
c
3
, c
3
u c
4
Fig. 1. Stability region
If we let the rate of be bounded by | | u, we get the
stability regions as dened in (13) and (14). The limit on
427
according to (12) is 0.0317 which is obtained using (11)
with u = 0.1744.
Using the related constants (c
3
= 1.0079 and c
2
= 1) in
(13) and (14), the sets E and F can be depicted in as in
Figure 2.
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
PSfrag replacements

E \F
Fig. 2. Convergence set E \ F as a function of for c
3
= 1.0079 and
c
2
=1.
VII. CONCLUSIONS
In this paper we have analyzed stability of nonlinear
systems for slowly varying parameters. We have shown that
recent results for analysis of LPV systems can be used
to obtain larger regions of stability for nonlinear systems
as compared to previous approaches which only consider
the linear time invariant system obtained for each xed
parameter value.
The framework presented admits extensions to consider
robust models. It is also possible to take into account
approximate descriptions of the function relating the param-
eter and the frozen state. This can be done by introducing
bounds on the difference between the true system and a
nominal system. This difference can therefore be treated
similarly as the derivative of the nominal state.
ACKNOWLEDGMENTS
The authors gratefully acknowledge nancial support
from the Swedish Research Council under contract No. 271-
2000-770.
REFERENCES
[1] P. Apkarian and H. D. Tuan. Parameterized LMIs in control theory.
SIAM J. Control and Optimization, 38(4):12411264, 2000.
[2] K. J.

Astr om and B. Wittenmark. Adaptive Control. Addison-Wesley,
Reading, Massachusetts, 1995.
[3] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear Matrix
Inequalities in System and Control Theory. SIAM, Philadelphia,
1994.
[4] J. V. Breakwell, J.L. Speyer, and A. E. Bryson. Optimization and
control of nonlinear systems using the second variation. SIAM J.
Control A.1., page 193, 1963.
[5] A. E. Bryson and Y.-C. Ho. Applied Optimal Control. Ginn and
Company, 1969.
[6] P. Gahinet, P. Apkarian, and M. Chilali. Afne paremeter-dependent
Lyapunov functions and real parameteric uncertainty. IEEE Trans-
actions on Automatic Control, 41(3):437442, 1996.
[7] T. Iwasaki and G. Shibata. LPV systemanalysis using quadratic
separator for uncertain implicit systems. IEEE Transactions on
Automatic Control, AC-46:11951208, 2001.
[8] H. K. Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River,
1996.
[9] H. K. Khalil and P. V. Kokotovic. On stability properties of nonlinear
systems with slowly varying inputs. IEEE Transactions on Automatic
Control, AC-36:229, 1991.
[10] D. A. Lawrence and W. J. Rugh. On a stability theorem for nonlinear
systems with slowly varying inputs. IEEE Transactions on Automatic
Control, AC-35:860864, 1990.
[11] A. M. Lyapunov. The general problem of the stability of motion
(translated into English by A. T. Fuller). International Journal of
Control, 55:531773, 1992.
[12] J. F. Magni. Linear fractional representations with a toolbox for
use with MATLAB. Technical Report TR 240/01 DCSD, Dept. of
Systems Control and Flight Dynamic, ONERA, 2001.
[13] Y. Nesterov and A. Nemirovsky. Interior-Point Polynomial Methods
in Convex Programming. SIAM, Philadelphia, 1994.
[14] R. Redheffer. Inequalities for a matrix Riccati equation. Journal of
Mathematics and Mechanics, 8, 1959.
[15] W. J. Rugh. Analytical framework for gain scheduling. IEEE Control
Systems Magazine, 11(1):7989, 1991.
[16] J. S. Shamma and M. Athans. Anaysis of gain scheduled control
for nonlinear plants. IEEE Transactions on Automatic Control, AC-
35(8):898907, 1990.
[17] J. S. Shamma and M. Athans. Guaranteed properties of gain
scheduled control for linear parameter-varying plants. Automatica,
27(3):559564, 1991.
[18] J. S. Shamma and M. Athans. Gain scheduling: Potential hazards and
possible remedies. IEEE Control Systems Magazine, 12(3):101107,
1992.
[19] G. Stein. Adaptive ight control: A pragmatic view. In K. S.
Narendra and R. V. Monopoli, editors, Applications of Adaptive
Control. Academic Press, New York, 1980.
428