Академический Документы
Профессиональный Документы
Культура Документы
Toshiyuki Ohtsuka
:=
9th IFAC Symposium on Nonlinear Control Systems
Toulouse, France, September 4-6, 2013
ThC3.4
Copyright 2013 IFAC 606
((x)/x) K
nn
. Let Di
n
K
(C) K
n
be the set of
K
n
such that rank
K
J
C
n
to M
,
where M
_
h(x)/x
L
f
h(x)/x
.
.
.
L
i
f
h(x)/x
_
_
, (6)
where L
0
f
h(x) = h(x) and L
i+1
f
h(x) := (L
i
f
h(x)/x)f(x)
(i N).
In general, the observability rank condition is a sucient
condition for local weak observability for all initial states
in C
n
. By restricting C
n
to an open and dense subset
M C
n
, the necessity is also guaranteed.
The observability rank condition has some applications in
analyzing observability of nonlinear systems. For example,
a system not satisfying rank condition (6) can be decom-
posed into an observable subsystem and an unobservable
subsystem (Conte et al. [2007], Nijmeijer and van der
Schaft [1990]).
Proposition 3.2. For system (1), let rank
K
O
n1
(x) = r <
n. Thus, there exists the coordinate transformation z =
(x) Di
n
K
(C) such that
Copyright 2013 IFAC 607
_
_
dz
1
/dt =
f
1
(z
1
, . . . , z
r
)
.
.
.
dz
r
/dt =
f
r
(z
1
, . . . , z
r
)
dz
r+1
/dt =
f
r+1
(z)
.
.
.
dz
n
/dt =
f
n
(z)
y = h(z
1
, . . . , z
r
)
(7)
holds.
3.2 PLT dened by the system
In this resarch, we study observability using PLT (Bron-
stein and Petkovsek [1996], Jacobson [1937], Leroy [1995]).
We give the denition of PLT. The derivation on the
eld K is an additive mapping : K K such that
(a + b) = (a) + (b), a, b K, (8)
(ab) = a (b) + b (a), a, b K. (9)
A eld K is a dierential eld if K is closed under a
derivation .
Let V be a vector space over K.
Denition 3.2. A mapping : V V is called PLT if
(u + v) = (u) + (v),
(au) = a(u) + (a)u
hold for any a K and u, v V .
We show a PLT dened by system (1). Let : K K be
the mapping
(a) :=
n
i=1
a
x
i
f
i
, a K. (10)
Note that (a) is the Lie derivative of a function a along
f in (1), which implies that depends on system (1). The
mapping satises conditions (8) and (9). Thus, is a
derivative of K, and K is a dierential eld because of
(a) K for any a K.
Next, let d : K X be the mapping
da =
n
i=1
a
x
i
dx
i
.
Finally, we dene the mapping s : X X.
s :=
n
i=1
((a
i
)dx
i
+ a
i
d(x
i
)) , =
n
i=1
a
i
dx
i
. (11)
For simplicity, we omit the symbol from (11). Note that
s is the Lie derivative of the one-form along f in
system (1). Thus, s depends on system (1).
From (11), for any a K and X, we have
s(a) = as + (a).
This equality implies that s : X X is a PLT.
For the PLT s : X X dened in (11), an eigenvalue and
an eigenvector are dened as follows (Leroy [1995], Lam
et al. [2008]).
Denition 3.3. K and X are called an eigenvalue
and an eigenvector of the PLT s : X X if s = holds.
Since the PLT dened in (11) depends on system (1), an
eigenvalue and eigenvector of the PLT s : X X are
determined by system (1).
Example 3.1. Consider the same system as Example 2.2.
For instance, x
1
x
2
and d(x
1
x
2
) are an eigenvalue and
eigenvector of the PLT dened by the system, respectively.
Actually, we have
sd(x
1
x
2
) =sdx
1
sdx
2
=d
_
x
2
1
x
2
2
2
_
d ((x
1
x
2
)x
2
)
=(x
1
dx
1
x
2
dx
2
) (x
2
dx
1
+ (x
1
2x
2
)dx
2
)
=(x
1
x
2
)d(x
1
x
2
).
In Example 2.2, an unobservable system does not satisfy
condition (3) at an eigenvalue = x
1
x
2
of PLT. In
Section 4.1, we clarify the relations between observability
and the eigenvalues of PLT.
4. OBSERVABILITY CONDITIONS
4.1 Nonlinear case
An eigenvalue of PLT (11) characterizes observability of
the nonlinear system.
Theorem 4.1. If system (1) is observable, then
_
_
(s )I
n
I
n
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
_
v = 0 (12)
holds for all Di
n
K
(C), K, v K
n
\ {0} and
X \ {0}.
Proof. We prove this by contraposition. If there exist
Di
n
K
(C), K, v K
n
\ {0} and X \ {0}
such that
_
_
(s )I
n
I
n
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
_
v = 0 (13)
holds, then we have
_
sI
n
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
v = 0.
Let u K
n
be J
1
v. The nonsingularity of J
implies
u = 0, and we have v = J
u. By substituting v = J
u
into the above equation, we obtain
_
J
(sI
n
(f(x)/x))
(h(x)/x)
_
u = 0,
and consequently, from the nonsingularity of J
,
_
sI
n
(f(x)/x)
(h(x)/x)
_
u = 0. (14)
Next, we show that equation (14) implies
Copyright 2013 IFAC 608
L
i
f
h
x
u = 0, i N (15)
by induction. Equation (14) yields
_
sI
n
f
x
_
u = 0, (16)
h
x
u = 0. (17)
When i = 0, equation (15) is nothing but (17).
Suppose that (15) holds when i = k, i.e.,
L
k
f
h
x
u = 0 (18)
holds. By premultiplying (L
k
f
h/x) by (16), we have
L
k
f
h
x
_
sI
n
f
x
_
u = 0. (19)
Also, by premultiplying s by (18), we have
s
L
k
f
h
x
u =
_
_
L
k
f
h
x
_
+
L
k
f
h
x
s
_
u = 0. (20)
By subtracting the left-hand side of (19) from the second
left-hand side of (20), we obtain
_
_
L
k
f
h
x
_
+
L
k
f
h
x
s
L
k
f
h
x
_
sI
n
f
x
_
_
u = 0
The left-hand side can be computed as follow
_
_
L
k
f
h
x
_
+
L
k
f
h
x
s
L
k
f
h
x
_
sI
n
f
x
_
_
u
=
_
2
L
k
f
h
x
2
f +
L
k
f
h
x
f
x
_
u
=
_
x
_
L
k
f
h
x
f
__
u =
L
k+1
f
h
x
u.
Thus, we have
L
k+1
f
h
x
u = 0.
Therefore, (15) holds.
Finally, we show that if (15) holds for u K
n
\ {0} and
X \{0}, then the observability rank condition does not
hold. Equation (15) implies that
O
n1
u = 0, (21)
where O
i
K
(i+1)n
is dened in (6). For any X,
there exists a vector a K
n
such that
= a
T
dx (22)
holds, where = 0 implies a
T
= 0 because {dx
1
, . . . , dx
n
}
is a basis of the K-vector space X. By substituting (22)
into (21), we have
O
n1
ua
T
dx = 0,
and consequently
O
n1
ua
T
= 0,
where u = 0 and a = 0 imply that (ua
T
) K
nn
is a nonzero matrix. Therefore, O
n1
is singular. That
is, the observability rank condition does not hold. From
Proposition 3.1, system (1) is not observable. 2
From Theorem 4.1, if nonlinear system (1) is not observ-
able, then (13) holds for some Di
n
K
(C), K,
v K
n
\ {0} and X \ {0}. Condition (13) can be
decomposed into
(s )v = 0 (23)
_
I
n
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
v = 0. (24)
Condition (23) implies that K and all v
i
X
(i = 1, . . . , n) are eigenvalues and eigenvectors of the
PLT s : X X. That is, Theorem 4.1 shows that
the eigenvalues of PLT play important roles when testing
observability of nonlinear systems. For linear systems, the
observability PBH test shows that the eigenvalues, in the
sense of linear algebra, of a system matrix characterize
observability. Therefore, in Theorem 4.1, the eigenvalues
of PLT operate like those of a system matrix.
In Section 4.2 below, it is shown that the condition of
Theorem 4.1 is equivalent to the observability PBH test
in the linear case. Thus, Theorem 4.1 can be viewed as a
generalization of the observability PBH test on nonlinear
systems.
Condition (24) is equivalent to
_
I
n
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
v = 0. (25)
Theorem 4.2 below shows that condition (25) also helps
in testing the observability of nonlinear system (1). Con-
dition (25) implies that K and v K
n
\ {0} are an
eigenvalue and right eigenvector of the matrix ((J
) +
J
(f(x)/x))J
1
K
nn
, respectively, in the linear
algebraic sense. Note that, the eigenvalues of ((J
) +
J
(f(x)/x))J
1
)+J
(f(x)/x))J
1
becomes an eigenvalue
of PLT.
Condition (25) is also important in its own right when
testing observability of system (1).
Theorem 4.2. System (1) is observable if
_
I
n
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
v = 0 (26)
holds for all Di
n
K
(C), K and v K
n
\ {0}.
Copyright 2013 IFAC 609
Proof. We prove this by contraposition. That is, we show
that if a system is not observable then there exist
Di
n
K
(C), K and v K
n
\ {0} such that (25) holds.
From Proposition 3.1, if system (1) is not observable,
then the observability rank condition does not hold. Let
rank
K
O
n1
(x) = r < n. Proposition 3.2 shows that
system (1) can be transformed into (7) by a coordinate
transformation z = (x) Di
n
K
(C). By choosing as ,
condition (25) becomes
_
_
I
r
(
f
1
( z
1
)/ z
1
) 0
(
f
2
(z)/ z
1
) I
nr
(
f
2
(z)/ z
2
)
(
h( z
1
)/ z
1
) 0
_
_
v = 0,
(27)
where z
1
= [z
1
, . . . , z
r
]
T
K
r
, z
2
= [z
r+1
, . . . , z
n
]
T
K
nr
,
f
1
= [
f
1
, . . . ,
f
r
]
T
K
r
and
f
2
= [
f
r+1
, . . . ,
f
n
]
T
K
nr
.
It suces to show the existence of K and v K
n
\{0}
satisfying condition (27). Let
K and v K
nr
\{0} be
an eigenvalue and eigenvector of the matrix (
f
2
(z)/ z
2
)
K
nn
in the sense of linear algebra. Then, for :=
K
and v := [0 v]
T
K
n
\ {0}, condition (27) holds. 2
The condition of Theorem 4.2 holds if and only if
rank
K
_
I
n
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
= n (28)
holds for all Di
n
K
(C) and K. When Di
n
K
(C)
is the identity mapping, i.e., J
K
nn
is the identity
matrix, condition (28) is nothing but condition (3). Thus,
condition (3) is a necessary condition for Theorem 4.2.
4.2 Linear Case
In the linear case, we show that the conditions of Theorems
4.1 and 4.2 are equivalent. The condition of Theorem
4.2 is a sucient condition for observability, and that of
Theorem 4.1 is a necessary condition. Thus, the condition
of Theorem 4.2 is a sucient condition for that of Theorem
4.1. Here, the converse is shown in the linear case.
Proposition 4.1. Suppose that f = Ax and h = c
T
x in (1),
where A C
nn
and c C
n
. If condition (12) holds for
all Di
n
K
(C), K, v K
n
\ {0} and X \ {0},
then condition (26) holds for all Di
n
K
(C), K and
v K
n
\ {0}.
Proof. We prove this by contraposition. That is, we show
that if there exist Di
n
K
(C), K and v K
n
\ {0}
satisfying condition (25), then there exists X\{0} such
that condition (13) holds for the same , and v.
In condition (25), let Di
n
K
(C) be the identity
mapping. Then, J
K
nn
is the identity matrix. In the
linear case, C K and v (C
n
\ {0}) (K
n
\
{0}) satisfying (25) are one of the eigenvalues and right
eigenvectors of the matrix A in the sense of linear algebra.
It suces to show the existence of X \ {0} such that
(13) holds for the above , and v. Condition (13) can
be decomposed into (23) and (25). Since (25) holds for the
above , and v, we show the existence of X \ {0}
satisfying (23) for the same , and v. Let X \ {0}
be a
T
dx, where a (C
n
\ {0}) (K
n
\ {0}) is a left
eigenvector of matrix A corresponding to the eigenvalue
. By substituting = a
T
dx into the left-hand side of
(23), we have
(s )va
T
dx = va
T
(AI
n
)dx = 0.
That is, condition (23) holds. 2
In the linear case, the conditions of Theorems 4.1 and
4.2 are equivalent, and the condition of Theorem 4.2 is
equivalent to the observability PBH test. Therefore, the
condition of Theorem 4.1 is equivalent to the observability
PBH test.
5. EXAMPLE
By using our results, we test the observability of the
following system.
_
_
dx
dt
=
_
_
x
2
+ x
2
3
x
2
1
x
3
x
2
1
/2
_
_
,
y = x
1
.
We check the necessary condition of Theorem 4.1 for this
system. Here, we consider nding Di
n
K
(C), K,
v K
n
\ {0} and X \ {0} such that condition (13)
holds. First, we nd an eigenvalue and an eigenvector of
the PLT dened by the system. One of the eigenvalues
and one of the eigenvectors of such PLT are 0 K and
d(x
2
+ x
2
3
) X, respectively.
Next, we nd a coordinate transformation Di
n
K
(C)
such that an eigenvalue of the matrix ((J
)+J
(f/x))
J
1
:=
_
1 0 0
0 1 2x
3
0 0 1
_
, (J
) :=
_
_
0 0 0
0 0 x
2
1
0 0 1
_
_
,
and thus
_
(J
) + J
f(x)
x
_
J
1
=
_
0 1 0
0 0 0
x
1
0 0
_
.
Then, 0 is an eigenvalue of ((J
) + J
(f/x))J
1
, and
one of its right eigenvectors is v := [0 0 1]
T
.
Finally, we check condition (12) for := [x
1
x
2
+ x
2
3
x
3
],
:= 0, v := [0 0 1]
T
and := d(x
2
+ x
2
3
). We obtain
_
_
(s )I
3
I
3
((J
) + J
(f(x)/x))J
1
(h(x)/x)J
1
_
_
v
=
_
_
sI
3
_
0 1 0
0 0 0
x
1
0 0
_
[ 1 0 0 ]
_
_
_
0
0
1
_
d(x
2
+ x
2
3
) = 0.
Therefore, condition (12) does not hold. From Theorem
4.1, the system is not observable.
Copyright 2013 IFAC 610
6. CONCLUSION
In this paper, we have derived two observability conditions
for the nonlinear system: a necessary condition and a
sucient condition. Our necessary condition shows that
observability of the nonlinear system is characterized by
the eigenvalues of the PLT dened by the system. In
the linear case, each eigenvalue of the PLT is nothing
but an eigenvalue, in the sense of linear algebra, of the
system matrix. Both our conditions are equivalent to the
observability PBH (Popov-Belevitch-Hautus) test in the
linear case. Therefore, our necessary condition can be
viewed as a generalization of the observability PBH test
for the nonlinear system.
REFERENCES
E. Aranda-Bricaire and C.H. Moog. Invariant codistribu-
tions and the feedforward form for discrete-time nonlin-
ear systems. Systems & Control Letters, 52(2):113122,
2004.
M. Bronstein and M. Petkovsek. An introduction to
pseudo-linear algebra. Theoretical Computer Science,
157:333, 1996.
G. Conte, C.H. Moog, and A.M. Perdon. Algebraic Meth-
ods for Nonlinear Control Systems. Springer-Verlag,
London, 2007.
K. Fujimoto and J.M.A. Scherpen. Nonlinear input-
normal realizations based on the dierential eigenstruc-
ture of Hankel operators. IEEE Transactions on Auto-
matic Control, 50(1):218, 2005.
M. Halas. An algebraic framework generalizing the concept
of transfer functions to nonlinear systems. Automatica,
44(5):11811190, 2008.
M. Halas and
U. Kotta. Transfer functions of discrete-
time nonlinear control systems. Proc. Estonian Acad.
Sci. Phys. Math., 56(4):322335, 2007.
R. Hermann and A.J. Krener. Nonlinear controllability
and observability. IEEE Transactions on Automatic
Control, 22(5):728740, 1977.
N. Jacobson. Pseudo-linear transformation. Annals of
Mathematics, 38(2):484507, 1937.
T.Y. Lam, A. Leroy, and A. Ozturk. Wedderburn polyno-
mials over division rings, ii. Proceedings of the Contem-
porary Mathematics, pages 7398, 2008.
A. Leroy. Pseudo linear transformations and evaluation in
ore extensions. Bull. Belg. Math. Soc., 2:321347, 1995.
J. Levine. On necessary and sucient conditions for
dierential atness. Applicable Algebra in Engineering,
Communication and Computing, 22(1):4790, 2011.
H. Nijmeijer and A.J. van der Schaft. Nonlinear Dynamical
Control Systems. Springer-Verlag, New York, 1990.
J.M.A. Scherpen. Balancing for nonlinear systems. Sys-
tems & Control Letters, 21(2):143153, 1993.
Y. Zheng, J.C. Willems, and C. Zhang. A polynomial
approach to nonlinear system. IEEE Transactions on
Automatic Control, 46(11):17821788, 2011.
Copyright 2013 IFAC 611