This action might not be possible to undo. Are you sure you want to continue?
Stability of Rotational Motion
J.F. Vasconcelos, A. Rantzer, C. Silvestre and P. Oliveira
Abstract—Lyapunov methods and density functions provide
dual characterizations of the solutions of a nonlinear dynamic
system. This work exploits the idea of combining both tech
niques, to yield stability results that are valid for almost all the
solutions of the system. Based on the combination of Lyapunov
and density functions, analysis methods are proposed for the
derivation of almost inputtostate stability, and of almost global
stability in nonlinear systems. The techniques are illustrated for
an inertial attitude observer, where angular velocity readings
are corrupted by nonidealities.
I. INTRODUCTION
Global stability is usually a highly desirable property
in control and estimation algorithms. However, topological
obstacles arise due to the fact that a smooth vector ﬁeld can
have a global attractor only if the state space is homeomor
phic to R
n
[5]. As a consequence, continuous state feedback
on smooth manifolds will always produce some trajectories
that do not converge to the origin [2], [9]. Due to the presence
of unstable manifolds, stability analysis using Lyapunov’s
second theorem is more complex.
New analysis tools have been brought forward by adopting
the milder notion of almost global stability [1], [12]. In
this framework, an equilibrium is ”almost globally stable”
if stability is satisﬁed for all initial states outside a set of
zero measure. A dual to the Lyapunov second method is
developed in [12], [13], based on density functions, that
represent the stationary density of a substance that ﬂows
along the system trajectories [10], [11], [12]. Almost global
stability is obtained by verifying that, for a timeinvariant
density function, particles are generated almost everywhere
and hence must ﬂow to a sink, located at the origin.
A similar approach has been adopted for the analysis
of inputtostate stability (ISS), that has been extensively
developed in recent years, as presented in the comprehensive
survey [16]. The limitations to global stability on non
Euclidean spaces, and the fact that global stability is a
necessary condition for ISS, motivate the relaxation to almost
ISS proposed in [1]. The notions of robust and weakly
This work was partially supported by Fundac¸ ˜ ao para a Ciˆ encia e a
Tecnologia (ISR/IST plurianual funding) through the POS Conhecimento
Program that includes FEDER funds and by the project PTDC/EEA
ACR/72853/2006 HELICIM. The work of J.F. Vasconcelos was supported
by a PhD Student Scholarship, SFRH/BD/18954/2004, from the Portuguese
FCT POCTI programme.
J.F. Vasconcelos, C. Silvestre and P. Oliveira are with the Insti
tute for Systems and Robotics, Instituto Superior T´ ecnico, Lisbon, Por
tugal. {jfvasconcelos, cjs, pjcro}@isr.ist.utl.pt
Tel: (+351) 218418054, Fax: (+351) 218418291.
A. Rantzer is with the Department of Automatic Control,
Faculty of Engineering, Lund University, 221 00 Lund, Sweden.
rantzer@control.lth.se
almost ISS are proposed, and stability results using density
functions are derived. More important, it is suggested that
a combination of Lyapunov methods with density function
results, may be the right technique for proving almost ISS
in general. Surprisingly enough, this enriching insight seems
to have gone unnoticed in the subsequent literature.
This work develops the idea of combining Lyapunov
and density functions, for the stability analysis of nonlinear
systems. Results are formulated for the analysis of almost
global asymptotic stability, and of almost ISS of the origin.
In the proposed analysis techniques, the Lyapunov func
tion is adopted to characterize the system trajectories, how
ever the resulting analysis is limited by the existence of
unstable manifolds. Density functions are used to resolve
for the regions where the Lyapunov method is inconclusive,
yielding sufﬁcient conditions for instability of undesirable
equilibrium points, and for convergence of almost all solu
tions to the region where stability is guaranteed by the Lya
punov function. The techniques are illustrated for the error
dynamics of an attitude observer deﬁned on SO(3), where
angular velocity readings are corrupted by nonidealities,
such as bounded measurement noise and unknown bias.
This work is organized as follows. Section II describes
the attitude observer, adopted to illustrate the combination of
Lyapunov and density function methods. The derivation of
almost ISS for nonlinear systems, using the combination of
Lyapunov and density functions, is discussed in Section III.
The approach is applied to demonstrate the stability of the
attitude observer in the presence of inertial sensor noise. A
new result for almost global stability of nonlinear systems
is presented in Section IV, and is illustrated for a nonlinear
system, motivated by the attitude observer dynamics subject
to inertial sensor bias. Concluding remarks and future work
are discussed in Section V.
NOMENCLATURE
The notation adopted is fairly standard. The set of n ×
m matrices with real entries is denoted as M(n, m) and
M(n) := M(n, n). The set of special orthogonal matrices is
denoted as SO(n) := {R ∈ M(n) : R
R = I, det(R) = 1}.
The nominal, the measured, and the estimated quantity s
are denoted by ¯s, s
r
and ˆs, respectively, and s denotes
the Frobenius norm. The operator (s)
×
produces the skew
symmetric matrix deﬁned by the vector s ∈ R
3
such that
(s)
×
b = s × b, b ∈ R
3
, and (·)
⊗
is the unskew operator
such that
(s)
×
⊗
= s. The time dependence of the variables
will be omitted in general, but otherwise denoted for the sake
of clarity.
Joint 48th IEEE Conference on Decision and Control and
28th Chinese Control Conference
Shanghai, P.R. China, December 1618, 2009
ThCIn1.9
9781424438723/09/$25.00 ©2009 IEEE 5941
Authorized licensed use limited to: Anders Rantzer. Downloaded on February 23,2010 at 10:37:36 EST from IEEE Xplore. Restrictions apply.
II. ATTITUDE OBSERVER
This section introduces the attitude observer, that is
adopted to illustrate the stability analysis techniques pro
posed in the paper. The observer estimates the attitude of a
rigid body with respect to a ﬁxed inertial frame, by merging
angular velocity measurements, with vectors observations
obtained in body coordinates. The detailed formulation of
the observer is presented in [17], and similar observers can
be found in [6] and [8].
The rigid body kinematics are described by
˙
¯
R =
¯
R( ¯ ω)
×
,
where
¯
R is the rotation matrix from body frame to the
inertial frame coordinates, and ¯ ω is the body angular velocity
expressed in body coordinates.
The body angular velocity is measured by a rate gyro
sensor triad, and the measurement model is
ω
r
= ¯ ω +u
ω
, (1)
where u
ω
is a measurement disturbance.
The vector observations are a function of the rigid
body’s attitude. The vectors coordinates are known and
timeinvariant in inertial frame, e.g. Earth’s magnetic and
gravitational ﬁelds, and measured in body coordinates by on
board sensors such as magnetometers and pendulums, among
others. The vector readings are introduced in the observer by
means of a conveniently deﬁned linear coordinate transfor
mation [17], and the transformed vector measurements are
described by
X
r
=
¯
R
I
¯
X, (2)
where X
r
:=
x
r 1
. . . x
r n
,
I
¯
X :=
I
¯ x
1
. . .
I
¯ x
n
,
X
r
,
I
¯
X ∈ M(3, n), the leading superscript I denotes inertial
coordinates, i = 1..n is the vector index, and n is the
number of vector measuring sensors. In this work, the vector
transformation is deﬁned such that
I
¯
X
I
¯
X
= I, to shape
uniformly the directionality of the vector readings. Also, it
is assumed that there are at least two noncollinear
I
¯ x
i
, so
that all rotational degrees of freedom are observable, see [17]
for a detailed characterization of the observer.
The attitude kinematics of the observer are given by
˙
ˆ
R =
ˆ
R( ˆ ω)
×
, (3)
where
ˆ
R is the estimated attitude and ˆ ω is the feedback
term. The attitude estimate dynamics (3) are stabilized by
exploiting the nonideal angular velocity measurements (1)
and the vector observations (2) in the feedback term ˆ ω.
In this work, the combination of Lyapunov and density
functions is illustrated in the stability analysis of the attitude
observer for the cases where u
ω
is i) an unmodeled, bounded
sensor disturbance, and ii) an unknown but constant sensor
bias.
III. STABILITY IN THE PRESENCE OF
UNMODELED INPUTS
This section discusses and formulates the combination of
Lyapunov and density function techniques for the analysis
of inputtostate stability, in the presence of unknown inputs.
The proposed method is illustrated by analyzing the stability
of the attitude observer, in the case where the inertial sensor
reading is corrupted by a bounded disturbance.
A. Almost ISS using Lyapunov and Density Functions
This section studies the inputtostate stability of systems
in the form
˙ x = f(x, u), (4)
where x ∈ M is the state, M is a smooth manifold, and f :
M ×U →TM, is a locally Lipschitz manifold map which
satisﬁes f(x, u) ∈ T
x
M, for all x ∈ M and all u ∈ U ⊂ R
m
.
The limitations to global stability on nonEuclidean spaces
motivate the relaxation of the classical notion of ISS [16] to
that proposed in [1]. Denoted as almost ISS, it allows for
each input u to destabilize a zero measure set of trajectories,
outside of which all trajectories converge to a neighborhood
of the origin.
Deﬁnition 1 (Almost ISS, [1]): The system (4) is almost
ISS with respect to the origin, denoted as 0
M
, if 0
M
is locally
asymptotically stable and
∀u ∈ U ∀a.a. x(t
0
) ∈ M limsup
t→∞
x(t) ≤ γ(u
∞
), (5)
where γ is a class K function and · is the distance to the
origin.
In this work, a method to derive almost ISS is obtained by
combining the properties of Lyapunov and density functions.
The adopted methodology has been sketched in [1], where it
is motivated by means of examples, however it seems to have
been unnoticed in subsequent literature. This section provides
a contribution to the concept of combining Lyapunov and
density functions, by formulating the technique in explicit
mathematical statements, and characterizing the stability
result as the combination of two ISS properties, introduced
in the following.
Deﬁnition 2 (Local ISS): A system (4) is locally ISS with
respect to 0
M
, if 0
M
is locally asymptotically stable and
there exists r > 0 such that
∀u ∈ U ∀ x(t
0
) ≤ r limsup
t→∞
x(t) ≤ γ
1
(u
∞
), (6)
where γ
1
is a class K function.
Deﬁnition 3 (Weakly almost ISS, [1]): A system (4) is
weakly almost ISS with respect to 0
M
, if 0
M
is locally
asymptotically stable and
∀u ∈ U ∀a.a. x(t
0
) ∈ M liminf
t→∞
x(t) ≤ γ
2
(u
∞
), (7)
where γ
2
is a class K function.
Provided that these ISS properties are veriﬁed, the main
result of this section shows that almost ISS is attained.
ThCIn1.9
5942
Authorized licensed use limited to: Anders Rantzer. Downloaded on February 23,2010 at 10:37:36 EST from IEEE Xplore. Restrictions apply.
(a) Local ISS using Lyapunov function analysis: solutions emanating below
the bound r converge to the region bounded by γ
1
(u∞).
(b) Weakly almost ISS using density function analysis: the liminf property
of almost all solutions satisﬁes the bound γ
2
(u∞).
(c) Almost ISS using Lyapunov and density functions analysis: by the
liminf property, almost all trajectories enter the region below the bound
r, and converge to the region bounded by γ
1
(u∞).
Fig. 1. Combination of Lyapunov and density functions for almost ISS of
the origin.
Lemma 1 (Almost ISS): Assume that the system (4) is
locally ISS and weakly almost ISS, then, for all u ∈ {u ∈
U : γ
2
(u
∞
) < r}, the system is almost ISS with γ = γ
1
.
Proof: Weakly almost ISS, expressed in (7), implies
that, by the continuity of the solutions of (4), almost every
solution satisﬁes x(t) ≤ γ
2
(u
∞
) < r for some t, thus
entering the region where the trajectories eventually satisfy
the limsup condition expressed in (6), yielding (5).
Lemma 1 shows that almost ISS can be obtained by
combining local ISS with weakly almost ISS, for sufﬁciently
small inputs. The proposed ISS analysis technique is there
fore based on i) Lyapunov methods to attain local ISS [4],
ii) density function techniques to yield weakly almost ISS
[1], and iii) Lemma 1 to derive almost ISS.
The stability analysis technique is illustrated in Fig. 1.
Lyapunov techniques yield local ISS based on ultimate
boundedness and/or ISS results [4]. As shown in Fig. 1(a),
Lyapunov methods ﬁnd a region {x : γ
1
(u
∞
) < x(t) <
r} where the Lyapunov function V decreases along the
system trajectories (
˙
V < 0), and drives the solutions to set
{x : x(t) < γ
1
(u
∞
)}, that is positively invariant.
However, the Lyapunov function analysis is inconclusive
with respect to {x : x(t) ≥ r}, and density functions
techniques are adopted to guarantee that almost all solutions
enter {x : x(t) < r} for some time instant. This is obtained
by ﬁnding a density function ρ such that div(ρf) > 0 in
the region {x : x(t) > γ
2
(u
∞
)}, which yields weakly
almost ISS by [1, Theorem 4]. Hence, the trajectories of the
system are endowed with the liminf characteristic depicted
in Fig. 1(b), and enter the region {x : x(t) < r} in ﬁnite
time, as shown in Fig. 1(c), yielding almost ISS of the origin.
B. Stability of the Nonlinear Observer in the Presence of
Inertial Sensor Noise
The combination of Lyapunov techniques is illustrated for
the nonlinear observer described in Section II, in the presence
of bounded timevarying disturbances in the angular velocity
measurements. The considered set of valid disturbances u
ω
in (1) is described by U = {u ∈ R
3
: u
∞
≤ u
max
}, and
the observer kinematics are given by
˙
ˆ
R =
ˆ
R( ˆ ω)
×
, ˆ ω =
ˆ
R
I
¯
XX
r
ω
r
−k
ω
n
¸
i=1
(
ˆ
R
I
¯ x
i
) ×x
r i
,
where k
ω
∈ R
+
is the feedback gain, for more details and
a motivation see [17]. Deﬁning the attitude estimation error
as R :=
ˆ
R
¯
R, the closed loop error kinematics are given by
˙
R = −k
ω
R(R−R
) −R(u
ω
)
×
. (8)
Although the origin of the unforced system is almost
globally asymptotically stable (GAS) [17], it can be shown
that the origin of the system (8) is not ISS, namely by taking
u
ω
as the destabilizing feedback law for a R(t
0
) sufﬁciently
close to the unstable equilibrium points of the unforced
system. In this counterexample, u
ω
destabilizes only a given
R(t
0
), and hence almost ISS is not precluded.
Almost ISS of the system (8) is obtained using the stability
analysis technique proposed in Section IIIA and depicted in
Fig. 1. The following proposition shows that the system is
locally ISS, using Lyapunov stability theory.
Theorem 2: Let k
ω
>
umax
2
, then for any initial condition
R(t
0
) ∈ {R ∈ SO(3) : I −R
2
< r(u
ω
∞
)}, (9a)
where r(u) = 4
1 +
1 −
u
2
4k
2
ω
, there exists T, indepen
dent of t
0
, such that the trajectory of the system (8) satisﬁes
R(t) ∈ {R ∈ SO(3) : I −R
2
< γ
1
(u
ω
∞
)}, (9b)
for all t ≥ t
0
+T, where γ
1
(u) = 4
1 −
1 −
u
2
4k
2
ω
.
Proof: The proof is based on the derivation of bounded
ness for nonlinear systems presented in [4, Theorem 4.18],
using Lyapunov methods. The time derivative of the Lya
punov function V =
I−R
2
2
along the system trajectories
ThCIn1.9
5943
Authorized licensed use limited to: Anders Rantzer. Downloaded on February 23,2010 at 10:37:36 EST from IEEE Xplore. Restrictions apply.
is given by
˙
V = −k
ω
I−R
2
2
2
+ tr
R−R
2
(u
ω
)
×
, and
algebraic manipulation produces
˙
V ≤ −k
ω
I −R
2
1
2
I −R
2
−
u
ω
k
ω
√
2
,
where 2 tr(I − R
2
) = I − R
2
2
= R− R
2
was used.
It is immediate that
1
2
I − R
2
>
uω∞
kω
√
2
⇒
˙
V < 0.
Using I − R
2
2
=
1
2
8 −I −R
2
I − R
2
produces
1
2
I −R
2
>
uω∞
kω
√
2
⇔R∈{
¯
Ω
t
−Ω
t
}, where Ω
t
= {R :
V (R) ≤
1
2
γ
1
(u
ω
∞
)},
¯
Ω
t
= {R : V (R) ≤
1
2
r(u
ω
∞
)}.
Consequently, R ∈{
¯
Ω
t
− Ω
t
} ⇒
˙
V < 0 and the set
¯
Ω
t
is a positively invariant set. The trajectories of the system
starting in
¯
Ω
t
enter Ω
t
in ﬁnite time, see [4, Section 4.8]
for a motivation of the level sets involved, and any solution
starting in Ω
t
will remain in the set since
˙
V < 0 in the
corresponding boundary. The initial conditions given by (9a)
satisfy R(t
0
) ∈
¯
Ω
t
; any R ∈ Ω
t
satisﬁes (9b), which
concludes the proof. The gain condition k
ω
>
umax
2
is
required so that {
¯
Ω
t
−Ω
t
} = ∅.
The results stated in Theorem 2 guarantee that any tra
jectory emanating from the region (9a) converges to the
bounded region (9b), whose boundaries are a function of
the noise to gain ratio
uω∞
kω
. Following the proposed
technique, a density function is adopted to show that almost
all trajectories of the system (8) satisfy a liminf condition.
Theorem 3: The system (8) is weakly almost ISS with
respect to I. Namely, the solutions verify
∀u
ω
∈ U ∀a.a.R(t
0
) ∈ SO(3)
liminf
t→∞
I −R(t)
2
< γ
2
(u
ω
∞
),
where γ
2
(u) =
8(
u
kω
)
2
1+(
u
kω
)
2
.
Proof: The result is obtained by satisfying the
conditions of [1, Theorem 4], with the density function
ρ(R) =
1
tr
2
(I−R)
. From the local ISS property obtained
in Theorem 2, it is immediate that R = I is a lo
cally stable equilibrium point for u
ω
= 0. The function
f := vec
kR(R
−R) −R(u
ω
)
×
is locally Lipschitz
over SO(3) and C
1
over SO(3) \ {I}. The density function
ρ(R) is of class C
1
over SO(3) \ {I} and, given that SO(3)
is compact, veriﬁes
SO(3)\U
ρ(R)dR < +∞, for all open
neighborhoods U of 0
M
.
The divergence is given by
div(ρf) =
k
ω
tr
3
(I −R)
I −R
2
+
2
k
ω
(R−R
)
⊗
u
ω
,
where div(ρf) = ρ div(f) +∇(ρ)
f, div(f) = −2k
ω
tr(R)
and ∇(ρ) = 2 tr(I − R)
−3
vec(I), for more details on the
computations of divergence and integrals in SO(3) see [3].
To attain the “density propagation inequality” [1, Theorem
4], given by
∀u ∈ U ∀x ∈ M x ≥ γ
2
(u) ⇒div(ρf) ≥ Q(x), (10)
with Q(x) > 0 for almost all x ∈ M, the sufﬁcient condition
I −R
2
+
2
k
ω
(R−R
)
⊗
u
ω
≥ ξ, ξ > 0,
0 1 2 3 4 5 6 7 8 9 10
10
−8
10
−6
10
−4
10
−2
10
0
10
2
Time (s)


I
−
R


2
I −R(t
0
)
2
<r(u
max
)
I −R(t
0
)
2
>r(u
max
)
r(u
max
)
γ
1
(u
max
)
Fig. 2. Simulation results of the attitude observer, illustrating almost ISS.
Convergence for I − R(t
0
)
2
≤ r(umax) is guaranteed by Lyapunov
methods; convergence for I − R(t
0
)
2
> r(umax) is guaranteed by
combining Lyapunov and density functions.
is analyzed, such that (10) is veriﬁed with Q(R) =
kω
tr
3
(I−R)
ξ > 0 for almost all R ∈ SO(3). The inequality
is satisﬁed if I −R
2
> 4
uω
kω
R−R
2
⊗
⇔
I − R
2
>
8
uω
kω
2
1+
uω
kω
2
= γ
2
(u
ω
). Since γ
2
(u) is a class
K function (γ
2
(0) = 0 and
d γ2(u)
du
> 0), the inequality
(10) is veriﬁed with I − R
2
> γ
2
(u
ω
) ⇒ div(ρf) ≥
kω
tr
3
(I−R)
ξ for some ξ > 0, producing the desired result.
As expressed in Lemma 1, the weakly almost ISS obtained
in Theorem 3 is used to guarantee that the solutions enter the
local ISS set derived in Theorem 2, producing almost ISS.
Theorem 4: Let k
ω
>
umax
√
3
. Then, the trajectories of the
system (8) satisfy
∀u
ω
∈ U ∀a.a.R(t
0
) ∈ SO(3)
limsup
t→∞
I −R(t)
2
< γ(u
ω
∞
),
where γ(u) = γ
1
(u) = 4
1 −
1 −
u
2
4k
2
ω
, i.e. the attitude
observer is almost ISS with respect to I.
Proof: The system (8) is locally ISS and weakly
almost ISS, by Theorems 2 and 3, respectively. The con
dition γ
2
(u
ω
∞
) < r(u
ω
∞
) formulated in Lemma 1 is
equivalent to
2(
uω∞
kω
)
2
1+(
uω∞
kω
)
2
< 1 +
1 −
uω∞
2kω
2
, that is
satisﬁed for
uω∞
kω
<
√
3, which yields almost ISS.
Simulation results of the observer estimation error are
depicted in Fig. 2. The exponential convergence for
γ
1
(u
max
) < I − R(t)
2
< r(u
max
) is justiﬁed by the fact
that
˙
V < −αV in that region, for some α ∈ R
+
.
IV. LOCAL STABILITY ANALYSIS USING
DENSITY FUNCTIONS
This section derives new results for local stability analysis
of equilibrium points other than the origin, using density
functions. By combining the proposed stability results with
LaSalle’s invariance principle, a new tool for global stability
analysis of the origin is obtained. The proposed technique
is illustrated for the case of a simple attitude observer with
biased inertial measurements.
ThCIn1.9
5944
Authorized licensed use limited to: Anders Rantzer. Downloaded on February 23,2010 at 10:37:36 EST from IEEE Xplore. Restrictions apply.
A. Stability using Density Functions and LaSalle’s Invari
ance Principle
The proposed stability analysis results are derived for
autonomous nonlinear systems of the form
˙ x = f(x), (11)
where f : R
n
→ R
n
is smooth, and the associated ﬂow
φ
t
: R
n
→ R
n
is deﬁned by φ
t
(x
0
) = x(t, x
0
), where
x(t, x
0
) denotes the solution of the system at time t with
initial condition x
0
. In the remainder of this work, it is
assumed that φ
t
is well deﬁned [15, Chapter 7].
Assumption 1: The ﬂow φ
t
is unique, continuous, and
exists for all t.
To formulate the stability results in the presence of
multiple equilibrium points, some concepts and results are
introduced, for more details the reader is referred to [15].
The values at time t and at the time interval t ∈ [τ
0
, τ]
of the trajectories starting in the set A, are respectively
denoted by φ
t
(A) = {x : x = φ
t
(x
0
), x
0
∈ A} and
φ
[τ0,τ]
(A) = {φ
t
(A) : t ∈ [τ
0
, τ]}. The local inset of x
u
is the set of all initial conditions inside a neighborhood U
of x
u
that converge to x
u
without leaving U, i.e. Z
U
(x
u
) =
{x ∈ U : ∀
∃
T
∀
t>T
φ
t
(x) − x
u
 < and ∀
t>0
φ
t
(x) ∈ U}.
The global inset of x
u
, denoted as W(x
u
), is deﬁned by
taking Z
U
(x
u
) with U = R
n
.
The following theorem is a new result in density function
methodologies, and provides sufﬁcient conditions to show
that an equilibrium point is not stable, given a suitable
density function. This property is of interest to exclude the
stability of equilibrium points other than the origin.
Theorem 5: Let x
u
∈ R
n
, and suppose there exists a non
negative ρ ∈ C
1
(R
n
\ {0}, R), integrable in a neighborhood
U of x
u
, and with div(ρf) > 0 in U. Then, the global inset
of x
u
has zero measure.
Proof: First, it is shown that the local inset, denoted
as Z
U
with a slight abuse of notation, has zero mea
sure. By Lemma 9 presented in Appendix A, the local
inset Z
U
is measurable. Using [12, Lemma A.1] with
D = U produces 0 ≥
φt(ZU)
ρ(x)dx −
ZU
ρ(z)dz =
t
0
φτ (ZU)
[div(ρf)] (x)dxdτ. Since div(ρf) > 0 in U,
then φ
t
(Z
U
) ⊂ Z
U
⊂ U has zero measure. The ﬂow φ
t
is a diffeomorphism and hence Z
U
has zero measure [7].
The forward propagation of Z
U
is Z
U
itself, i.e. Z
U
=
φ
[0,∞)
(Z
U
). Therefore, φ
[0,∞)
(Z
U
) has zero measure, and
by Lemma 10 of Appendix A, the set φ
(−∞,∞)
(Z
U
) has
zero measure. The global inset of x
u
can be expressed as
W(x
u
) = φ
(−∞,∞)
(Z
U
) and therefore has zero measure.
The combination of Theorem 5 with LaSalle’s invariance
principle can be used to provide almost GAS of the origin.
The technique is based on using LaSalle’s invariance princi
ple to show that the trajectories approach a candidate set M
[4]; and then using the div(ρf) > 0 property for M \ {0},
to show that the set of trajectories converging to M \ {0} is
of zero measure, and hence that the origin is almost GAS.
Lemma 6: Consider the system (11). Let V : R
n
→ R
be a continuously differentiable function such that the level
sets {x : V (x) ≤ c} are bounded and
˙
V (x) ≤ 0. Let M be
the largest invariant set in {x :
˙
V (x) = 0}. Suppose that M
is a countable union of isolated points, and that there is a
density function that satisﬁes the conditions of Theorem 5
for all x
U
∈ M\{0}. Then the origin of (11) is almost GAS.
Proof: The conditions on V (x) satisfy LaSalle’s
invariance principle [4], [15], and hence guarantee
that the trajectories approach M as t → ∞, i.e.
∀x
0
∀
ε
∃
T>0
∀
t>T
inf
y∈M
φ
t
(x
0
)−y < ε. By the continuity
of φ
t
(x), choosing ε < min
x,y∈M
x − y shows that each
solution of (11) must converge to an isolated point x
u
∈ M.
By Theorem 5, the condition div(ρf) > 0 for a neighbor
hood U of every x
u
∈ M \ {0} guarantees that the global
inset of x
u
has zero measure. The set of initial conditions
that converge to M \ {0}, given by ∪
xu∈M\{0}
W(x
u
), is a
countable union of zero measured sets and hence has zero
measure. Consequently, almost all solutions converge to the
origin, and hence the origin is almost GAS.
Remark 1: The results of Theorem 5 provide an alterna
tive approach to the stability analysis based on Hartman
Grobman theorem [15]. Also, the derived stability result is
also valid for nonhyperbolic equilibria. Future work will
evaluate the contribution of Theorem 5 in this topic.
Remark 2: Lemma 6 is valid when the invariant set is a
countable union of isolated points. The extension for more
generic sets, based on generalizing the results adopted in the
proof of Theorem 5, will be addressed in future work.
B. Stability of the Nonlinear Observer in the Presence of
Biased Inertial Readings
In this section, the proposed stability analysis is illustrated
for the attitude observer, in a case where the disturbance in
(1) is a constant bias, i.e. ˙ u
ω
= 0. The observer kinematics
are augmented to estimate the bias, and are given by
˙
ˆ
R =
ˆ
R( ˆ ω)
×
,
˙
ˆ
b
ω
= k
bω
s
ω
, where ˆ ω =
ˆ
R
I
¯
XX
r
ω
r
−
ˆ
b
ω
−
k
ω
s
ω
, s
ω
=
¸
n
i=1
(
ˆ
R
I
¯ x
i
)×x
r i
, k
ω
, k
bω
∈ R
+
are feedback
gains, and
ˆ
b
ω
is the rate gyro bias estimate, for more details
on the adopted observer see [17]. The closed loop error
kinematics are given by
˙
R = R
k
ω
(R
−R) + (b
ω
)
×
,
˙
b
ω
= k
bω
(R
−R)
⊗
, (12)
where b
ω
=
ˆ
b
ω
− u
ω
is the bias estimation error. The
stability analysis technique is illustrated for the case where
initial bias and attitude estimation errors exist along the
zaxis, i.e. b(t
0
) =
0 0 b
0
, b
0
∈ R, and R(t
0
) =
exp(θ
0
(λ
0
)
×
), θ
0
∈ R, λ
0
=
0 0 1
. In this case,
the trajectories of (12) satisfy R(t) = exp(θ(t) (λ
0
)
×
),
b(t) =
0 0 b(t)
, and the dynamics can be reduced to
˙
θ = −sin(θ) +b,
˙
b = −sin(θ), (13)
with initial conditions θ(t
0
) = θ
0
, b(t
0
) = b
0
.
As discussed in Section IVA, the stability of the system
(13) is ﬁrst analyzed using LaSalle’s invariance principle to
derive an invariant set M.
Proposition 7: The trajectories of the system (13) ap
proach M = {(θ, b) : θ = πk, k ∈ Z, b = 0} as t →∞.
ThCIn1.9
5945
Authorized licensed use limited to: Anders Rantzer. Downloaded on February 23,2010 at 10:37:36 EST from IEEE Xplore. Restrictions apply.
Fig. 3. Phase portrait of the reduced order attitude observer. Using the
density function property div(ρf) > 0 in a neighborhood of the equilibrium
points (θ, b) = (π + 2πk, 0), k ∈ Z, shows that these are unstable.
Proof: The result is obtained by considering the Lya
punov function V = 2(1 −cos(θ)) +b
2
. The time derivative
is given by
˙
V = −2 sin
2
(θ) and the result is immediate from
LaSalle’s invariance principle.
The phase portrait of the system, depicted in Fig. 3,
suggests that the equilibrium points in the set E = {(θ, b) :
θ = 2πk + π, k ∈ Z, b = 0} ⊂ M are unstable. Using
the technique summarized in Lemma 6, the instability of E
is shown using a density function that satisﬁes Theorem 5,
yielding that almost all trajectories approach M \ E =
{(θ, b) : θ = 2πk, k ∈ Z, b = 0} as t →∞.
Proposition 8: Almost all trajectories of the system (13)
approach the set {(θ, b) : θ = 2πk, k ∈ Z, b = 0} as t →∞.
Proof: To exclude the points in the set E, the density
function ρ =
1
2(1−cos(θ))+b
2
is adopted. The divergence is
div(ρf) =
2(1−cos(θ))−cos(θ)b
2
(2(1−cos(θ))+b
2
)
2
, so there is a neighborhood
of every point in E where ρ is integrable and div(ρf) > 0.
By Theorem 5, the set of initial conditions converging to
each point in E has zero measure. The set E is a countable
union of points, and hence the set of initial conditions that
approach E has zero measure. Consequently, almost all the
trajectories approach M \ E.
V. CONCLUSIONS
This work addressed the combination of Lyapunov and
density functions, for stability analysis of nonlinear systems.
Almost ISS of the origin was formulated as the combination
of local ISS and weakly almost ISS, that can be derived using
Lyapunov and density functions, respectively. For the case
of autonomous systems, it was shown that global stability of
the origin can be obtained by combining LaSalle’s invariance
principle, with a density function that excludes the stability
of undesirable equilibrium points. The proposed stability
analysis techniques were illustrated for an attitude observer
with nonideal angular velocity readings. Future work will
address the topics discussed in Remarks 1 and 2.
APPENDIX
A. Set Measure Results
This section presents auxiliary results adopted in the paper.
Lemma 9: The local inset of an equilibrium point is
measurable under Assumption 1.
Proof: The set Z
U
can be written as the intersection of
a “stability” and a “convergence” sets, given by Z
U
= S ∩C
where S = {x ∈ U : φ
t
(x) ∈ U for all t ≥ 0} and C =
{x ∈ U : ∀
∃
T
∀
t>T
φ
t
(x) − x
u
< }. The set S can be
described by S =
¸
k∈N0
S
k
where S
k
= {x ∈ U : φ
t
(x) ∈
U for t ∈
kT kT +T
}. By the continuous dependence
of φ
t
(x) on the initial conditions [15], and on t, for each
x ∈ S
k
there exists δ sufﬁciently small, such that x−y <
δ ⇒φ
t
(y) ∈ U for the compact interval t ∈
kT kT +T
.
Consequently, the set S
k
is open, thus measurable, and the
set S is measurable. The set C can be described by C =
¸
n∈N
¸
k∈N0
C
n,k
, where C
n,k
= {x ∈ U : ∃
T
∀
t≥T
φ
t
(x) −
x
u
<
1
n
for all t ≥ k}. The set C
n,k
is measurable, by the
same arguments used for the measurability of S, and C is a
countable union and intersection of measurable sets and is
measurable, which concludes the proof.
Lemma 10: Under Assumption 1, the set φ
[τ0,τ]
(A) has
zero measure if and only if φ
(−∞,∞)
(A) has zero measure.
Proof: (⇐) Immediate from φ
[τ0,τ]
(A) ⊂
φ
(−∞,∞)
(A). (⇒) f is smooth, then φ
t
is a diffeomorphism
for each t [15], and φ
t
(φ
[τ0,τ]
(A)) has zero measure
[7]. Let t
k
= (τ − τ
0
)k, k ∈ Z, for autonomous
systems φ
t
(φ
[τ0,τ]
(A)) = φ
[τ0+t,τ+t]
(A) and hence
φ
(−∞,∞)
(A) =
¸
k∈Z
φ
[τ0+t
k
,τ+t
k
]
(A) is a countable union
of zero measure sets, and thus has zero measure [14].
REFERENCES
[1] D. Angeli. An almost global notion of inputtostate stability. IEEE
Transactions on Automatic Control, 49(6):866–874, 2004.
[2] S. P. Bhat and D. S. Bernstein. A topological obstruction to con
tinuous global stabilization of rotational motion and the unwinding
phenomenon. Systems and Control Letters, 39(1):63–70, Jan. 2000.
[3] G. S. Chirikjian and A. B. Kyatkin. Engineering Applications of
Noncommutative Harmonic Analysis: With Emphasis on Rotation and
Motion Groups. CRC, 2000.
[4] H. K. Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2001.
[5] D. E. Koditschek. The application of total energy as a Lyapunov func
tion for mechanical control systems. Control Theory and Multibody
Systems, 97:131–151, 1989.
[6] C. Lageman, R. Mahony, and J. Trumpf. State observers for invariant
dynamics on a lie group. In 18th International Symposium on
Mathematical Theory of Networks and Systems, 2008.
[7] J.M. Lee. Introduction to Smooth Manifolds. Springer, 2002.
[8] R. Mahony, T. Hamel, and J.M. Pﬂimlin. Nonlinear complementary
ﬁlters on the special orthogonal group. IEEE Transactions on Auto
matic Control, 53(5):1203–1218, June 2008.
[9] M. Malisoff, M. Krichman, and E. Sontag. Global stabilization for
systems evolving on manifolds. Journal of Dynamical and Control
Systems, 12(2):161–184, April 2006.
[10] G. Meinsma. On Rantzer’s density function. In 25th Benelux Meeting,
Heeze, The Netherlands, March 2006.
[11] P. Monzon. On necessary conditions for almost global stability. IEEE
Transactions on Automatic Control, 48(4):631–634, April 2003.
[12] A. Rantzer. A dual to Lyapunov’s stability theorem. Systems and
Control Letters, 42(3):161–168, March 2000.
[13] A. Rantzer. An converse theorem for density functions. In 41st IEEE
Conference on Decision and Control, Dec. 2002.
[14] Halsey Royden. Real Analysis. Prentice Hall, 1988.
[15] S. Sastry. Nonlinear Systems: Analysis, Stability and Control. Springer,
1999.
[16] E.D. Sontag. Input to state stability: Basic concepts and results.
In P. Nistri and G. Stefani, editors, Nonlinear and Optimal Control
Theory, pages 163–220. SpringerVerlag, Berlin, 2007.
[17] J.F. Vasconcelos, C. Silvestre, and P. Oliveira. Nonlinear observer
for rigid body attitude estimation using vector observations. In
Proceedings of the 17th IFAC World Congress, 2008.
ThCIn1.9
5946
Authorized licensed use limited to: Anders Rantzer. Downloaded on February 23,2010 at 10:37:36 EST from IEEE Xplore. Restrictions apply.
I X := I x1 . however it seems to have been unnoticed in subsequent literature. ATTITUDE OBSERVER This section introduces the attitude observer. it allows for each input u to destabilize a zero measure set of trajectories. A. The rigid body kinematics are described by ˙ ¯ ¯ ¯ R = R (ω)× . (6) ˆ ˆ where R is the estimated attitude and ω is the feedback term. the leading superscript I denotes inertial coordinates. the combination of Lyapunov and density functions is illustrated in the stability analysis of the attitude observer for the cases where uω is i) an unmodeled. Earth’s magnetic and gravitational ﬁelds. (7) where γ2 is a class K function. and f : M × U → T M . and similar observers can be found in [6] and [8]. (5) ¯ ¯ ¯ where Xr := xr 1 . and the measurement model is ¯ ω r = ω + uω . ˙ (4) where uω is a measurement disturbance. The vector readings are introduced in the observer by means of a conveniently deﬁned linear coordinate transformation [17]. . (2) where x ∈ M is the state. [1]): A system (4) is weakly almost ISS with respect to 0M .a. outside of which all trajectories converge to a neighborhood of the origin. i = 1. Denoted as almost ISS. [1]): The system (4) is almost ISS with respect to the origin. Provided that these ISS properties are veriﬁed. Almost ISS using Lyapunov and Density Functions This section studies the inputtostate stability of systems in the form x = f (x. u). u) ∈ Tx M . a method to derive almost ISS is obtained by combining the properties of Lyapunov and density functions.n is the vector index. M is a smooth manifold. and n is the number of vector measuring sensors. STABILITY IN THE PRESENCE OF UNMODELED INPUTS This section discusses and formulates the combination of Lyapunov and density function techniques for the analysis of inputtostate stability. denoted as 0M . Also.2010 at 10:37:36 EST from IEEE Xplore. and measured in body coordinates by onboard sensors such as magnetometers and pendulums. (3) where γ is a class K function and · is the distance to the origin. e. where γ1 is a class K function. to shape uniformly the directionality of the vector readings. I xn . (1) III. if 0M is locally asymptotically stable and ∀u ∈ U ∀a. n). and the transformed vector measurements are described by ¯ ¯ Xr = R I X. The limitations to global stability on nonEuclidean spaces motivate the relaxation of the classical notion of ISS [16] to that proposed in [1].. Deﬁnition 2 (Local ISS): A system (4) is locally ISS with respect to 0M . I ¯ Xr . among others. and ii) an unknown but constant sensor bias. and characterizing the stability result as the combination of two ISS properties. The body angular velocity is measured by a rate gyro sensor triad. In this work. . The adopted methodology has been sketched in [1]. ¯ where R is the rotation matrix from body frame to the ¯ inertial frame coordinates. x(t0 ) ∈ M lim sup x(t) ≤ γ( u t→∞ ∞ ). the main result of this section shows that almost ISS is attained. so that all rotational degrees of freedom are observable. This section provides a contribution to the concept of combining Lyapunov and density functions. The vectors coordinates are known and timeinvariant in inertial frame. Deﬁnition 1 (Almost ISS. with vectors observations obtained in body coordinates. by formulating the technique in explicit mathematical statements. in the presence of unknown inputs. by merging angular velocity measurements. see [17] for a detailed characterization of the observer. The observer estimates the attitude of a rigid body with respect to a ﬁxed inertial frame. The vector observations are a function of the rigid body’s attitude. X ∈ M(3. In this work. is a locally Lipschitz manifold map which satisﬁes f (x. . and ω is the body angular velocity expressed in body coordinates. if 0M is locally asymptotically stable and there exists r > 0 such that ∀u ∈ U ∀ x(t0 ) ≤ r lim sup x(t) ≤ γ1 ( u t→∞ ∞ ). xr n . it ¯ is assumed that there are at least two noncollinear I xi .ThCIn1. Downloaded on February 23. The attitude estimate dynamics (3) are stabilized by exploiting the nonideal angular velocity measurements (1) ˆ and the vector observations (2) in the feedback term ω. . 5942 Authorized licensed use limited to: Anders Rantzer. introduced in the following. for all x ∈ M and all u ∈ U ⊂ Rm .g. Restrictions apply. The attitude kinematics of the observer are given by ˙ ˆ ˆ ˆ R = R (ω)× . The proposed method is illustrated by analyzing the stability of the attitude observer. . x(t0 ) ∈ M lim inf x(t) ≤ γ2 ( u t→∞ ∞ ). that is adopted to illustrate the stability analysis techniques proposed in the paper. Deﬁnition 3 (Weakly almost ISS.a. if 0M is locally asymptotically stable and ∀u ∈ U ∀a. In this work.9 II. in the case where the inertial sensor reading is corrupted by a bounded disturbance. the vector ¯ ¯ transformation is deﬁned such that I X I X = I. where it is motivated by means of examples. The detailed formulation of the observer is presented in [17]. bounded sensor disturbance.
ThCIn1. yielding almost ISS of the origin. such that the trajectory of the system (8) satisﬁes R(t) ∈ {R ∈ SO(3) : I − R 2 < γ1 ( uω ∞ )}. and drives the solutions to set {x : x(t) < γ1 ( u ∞ )}. for more details and a motivation see [17]. and density functions techniques are adopted to guarantee that almost all solutions enter {x : x(t) < r} for some time instant. then. Almost ISS of the system (8) is obtained using the stability analysis technique proposed in Section IIIA and depicted in Fig. Theorem 4]. in the presence of bounded timevarying disturbances in the angular velocity measurements.2010 at 10:37:36 EST from IEEE Xplore. This is obtained by ﬁnding a density function ρ such that div(ρf ) > 0 in the region {x : x(t) > γ2 ( u ∞ )}. In this counterexample. Lemma 1 shows that almost ISS can be obtained by combining local ISS with weakly almost ISS. As shown in Fig. The time derivative of the Lya2 punov function V = I−R along the system trajectories 2 5943 Authorized licensed use limited to: Anders Rantzer. (8) Although the origin of the unforced system is almost globally asymptotically stable (GAS) [17]. Theorem 2: Let kω > umax . The stability analysis technique is illustrated in Fig. the system is almost ISS with γ = γ1 . Proof: Weakly almost ISS. 2 (9b) u for all t ≥ t0 + T . yielding (5). and hence almost ISS is not precluded. Lyapunov methods ﬁnd a region {x : γ1 ( u ∞ ) < x(t) < r} where the Lyapunov function V decreases along the ˙ system trajectories (V < 0). and iii) Lemma 1 to derive almost ISS. B. where kω ∈ R is the feedback gain. then for any initial condition 2 R(t0 ) ∈ {R ∈ SO(3) : I − R 2 (c) Almost ISS using Lyapunov and density functions analysis: by the lim inf property. almost all trajectories enter the region below the bound r. Hence. Deﬁning the attitude estimation error ˆ ¯ as R := R R. there exists T . 1(c). Downloaded on February 23. The proposed ISS analysis technique is therefore based on i) Lyapunov methods to attain local ISS [4]. implies that. . 1(b). Restrictions apply. 1. The following proposition shows that the system is locally ISS. almost every solution satisﬁes x(t) ≤ γ2 ( u ∞ ) < r for some t. 1. Fig. Combination of Lyapunov and density functions for almost ISS of the origin. where γ1 (u) = 4 1 − 1 − 4k2 . ii) density function techniques to yield weakly almost ISS [1]. indepenω dent of t0 . i=1 ˆ ¯ (R I xi ) × xr i . The considered set of valid disturbances uω in (1) is described by U = {u ∈ R3 : u ∞ ≤ umax }. Theorem 4. Stability of the Nonlinear Observer in the Presence of Inertial Sensor Noise The combination of Lyapunov techniques is illustrated for the nonlinear observer described in Section II.18]. expressed in (7). for sufﬁciently small inputs. Lyapunov techniques yield local ISS based on ultimate 2 < r( uω ∞ )}. and enter the region {x : x(t) < r} in ﬁnite time. 1(a). the closed loop error kinematics are given by ˙ R = −kω R(R −R ) − R (uω )× . which yields weakly almost ISS by [1. it can be shown that the origin of the system (8) is not ISS. (b) Weakly almost ISS using density function analysis: the lim inf property of almost all solutions satisﬁes the bound γ2 ( u ∞ ). using Lyapunov methods. and the observer kinematics are given by ˙ ˆ ˆ ˆ ˆ ¯ ˆ R = R (ω)× . uω destabilizes only a given R(t0 ). ω = R I XXr ω r − kω + n (a) Local ISS using Lyapunov function analysis: solutions emanating below the bound r converge to the region bounded by γ1 ( u ∞ ). (9a) u where r(u) = 4 1 + 1 − 4k2 . that is positively invariant. the trajectories of the system are endowed with the lim inf characteristic depicted in Fig. using Lyapunov stability theory. and converge to the region bounded by γ1 ( u ∞ ). by the continuity of the solutions of (4). the Lyapunov function analysis is inconclusive with respect to {x : x(t) ≥ r}. for all u ∈ {u ∈ U : γ2 ( u ∞ ) < r}. as shown in Fig. However.9 boundedness and/or ISS results [4]. 1. namely by taking uω as the destabilizing feedback law for a R(t0 ) sufﬁciently close to the unstable equilibrium points of the unforced system. Lemma 1 (Almost ISS): Assume that the system (4) is locally ISS and weakly almost ISS. ω Proof: The proof is based on the derivation of boundedness for nonlinear systems presented in [4. thus entering the region where the trajectories eventually satisfy the lim sup condition expressed in (6).
Convergence for I − R(t0 ) 2 ≤ r(umax ) is guaranteed by Lyapunov methods. 2. i. 2 ω 2 1 2 2 2 2 Using I − R = 2 8− I−R I − R produces u 1 ¯ I − R2 > k ω√∞ ⇔ R∈{ Ωt − Ωt }. see [4. illustrating almost ISS. The exponential convergence for γ1 (umax ) < I − R(t) 2 < r(umax ) is justiﬁed by the fact . veriﬁes SO(3)\U ρ(R)dR < +∞. Namely.ThCIn1.9 ˙ is given by V = −kω I−R + tr 2 algebraic manipulation produces ˙ V ≤ −kω I − R2 2 2 R−R 2 (uω )× . 2. it is immediate that R = I is a locally stable equilibrium point for uω = 0. the sufﬁcient condition 2 (R −R )⊗ uω ≥ ξ. LOCAL STABILITY ANALYSIS USING DENSITY FUNCTIONS This section derives new results for local stability analysis of equilibrium points other than the origin. 2 ˙ ¯ ¯ t − Ω } ⇒ V < 0 and the set Ωt Consequently. The initial conditions given by (9a) ¯ satisfy R(t0 ) ∈ Ωt . producing the desired result. the solutions verify ∀uω ∈ U ∀a. From the local ISS property obtained in Theorem 2. Theorem 3: The system (8) is weakly almost ISS with respect to I. that is equivalent to uω ∞ 2 2k 1+( kω ) √ ω satisﬁed for ukω ∞ < 3. Theorem 4]. which concludes the proof.2010 at 10:37:36 EST from IEEE Xplore. Restrictions apply. the trajectories of the system (8) satisfy uω 2 kω uω 2 kω ⊗ ⇔ = γ2 ( uω ). 5944 Authorized licensed use limited to: Anders Rantzer. The density function ρ(R) is of class C 1 over SO(3) \ {I} and. (10) kω tr (I − R) 3 I−R 2 + 2 (R − R )⊗ uω kω u where γ(u) = γ1 (u) = 4 1 − 1 − 4k2 . Proof: The system (8) is locally ISS and weakly almost ISS.e. a density function is adopted to show that almost all trajectories of the system (8) satisfy a lim inf condition. div(ρf ) = where div(ρf ) = ρ div(f ) + (ρ) f . ξ > 0. Since γ2 (u) is a class lim inf I − R(t) 2 < γ2 ( uω ∞ ). ) Proof: The result is obtained by satisfying the conditions of [1. max Theorem 4: Let kω > u√3 . such that (10) is veriﬁed with Q(R) = kω tr3 (I−R) ξ > 0 for almost all R ∈ SO(3). is analyzed. with the density function 1 ρ(R) = tr2 (I−R) . using density functions. . the weakly almost ISS obtained in Theorem 3 is used to guarantee that the solutions enter the local ISS set derived in Theorem 2. The trajectories of the system ¯ starting in Ωt enter Ωt in ﬁnite time. Downloaded on February 23. any R ∈ Ωt satisﬁes (9b). By combining the proposed stability results with LaSalle’s invariance principle. Following the proposed technique. 2 kω 2 where 2 tr(I − R2 ) = I − R2 2 = R − R 2 was used. Ωt = {R : V (R) ≤ 1 r( uω ∞ )}.R(t0 ) ∈ SO(3) t→∞ 2 I −R2 10 −2 10 −4 10 −6 10 −8 0 1 2 3 4 5 6 7 8 9 10 Time (s) Fig. whose boundaries are a function of ω the noise to gain ratio ukω ∞ . respectively.a. by Theorems 2 and 3. convergence for I − R(t0 ) 2 > r(umax ) is guaranteed by combining Lyapunov and density functions. I−R 2+ kω ∀u ∈ U ∀x ∈ M x ≥ γ2 (u) ⇒ div(ρf ) ≥ Q(x). and 10 2 I −R(t0 )2 < max ) r(u I −R(t0 )2 > max ) r(u r(u ) max γ (u ) 1 max 10 0 uω 1 I − R2 − √ . Section 4. the inequality 2 (10) is veriﬁed with I − R > γ2 ( uω ) ⇒ div(ρf ) ≥ kω tr3 (I−R) ξ for some ξ > 0. for some α ∈ R+ . To attain the “density propagation inequality” [1.a. where Ωt = {R : 2 ω 2 1 ¯ V (R) ≤ 2 γ1 ( uω ∞ )}. The condition γ2 ( uω ∞ ) < r( uω ∞ ) formulated in Lemma 1 is 2 uω 2 2( k ω ∞ ) < 1 + 1 − uω ω∞ .8] for a motivation of the level sets involved. As expressed in Lemma 1. The proposed technique is illustrated for the case of a simple attitude observer with biased inertial measurements. The results stated in Theorem 2 guarantee that any trajectory emanating from the region (9a) converges to the bounded region (9b). the attitude ω observer is almost ISS with respect to I. producing almost ISS. R ∈{ Ω t is a positively invariant set. given by with Q(x) > 0 for almost all x ∈ M . and any solution ˙ starting in Ωt will remain in the set since V < 0 in the corresponding boundary. u ˙ It is immediate that 1 I − R2 > k ω√∞ ⇒ V < 0. a new tool for global stability analysis of the origin is obtained. ˙ that V < −αV in that region. for all open neighborhoods U of 0M . The function f := vec kR(R − R) − R (uω )× is locally Lipschitz over SO(3) and C 1 over SO(3) \ {I}. lim sup I − R(t) 2 2 < γ( uω ∞ ). which yields almost ISS. Simulation results of the observer estimation error are depicted in Fig. Theorem 4]. for more details on the computations of divergence and integrals in SO(3) see [3]. Then. IV. given that SO(3) is compact. div(f ) = −2kω tr(R) and (ρ) = 2 tr(I − R)−3 vec(I). The inequality is satisﬁed if I − R I−R 2 2 >4 uω kω R−R 2 > 8 1+ 2 (u) K function (γ2 (0) = 0 and d γdu > 0). The divergence is given by u kω where γ2 (u) = u 8( k ω ) 2 ∀uω ∈ U ∀a. Simulation results of the attitude observer.R(t0 ) ∈ SO(3) t→∞ 1+( . The gain condition kω > umax is 2 ¯ required so that {Ωt − Ωt } = ∅.
the stability of the system (13) is ﬁrst analyzed using LaSalle’s invariance principle to derive an invariant set M . and hence guarantee that the trajectories approach M as t → ∞. Chapter 7]. Lemma A.y∈M x − y shows that each solution of (11) must converge to an isolated point xu ∈ M . Proposition 7: The trajectories of the system (13) approach M = {(θ. The global inset of xu can be expressed as W(xu ) = φ(−∞. τ ] of the trajectories starting in the set A. By the continuity of φt (x). kω . the derived stability result is also valid for nonhyperbolic equilibria.∞) (ZU ) has zero measure. Since div(ρf ) > 0 in U . θ0 ∈ R. then φt (ZU ) ⊂ ZU ⊂ U has zero measure. for more details the reader is referred to [15]. The set of initial conditions that converge to M \ {0}. and by Lemma 10 of Appendix A. i. In the remainder of this work. and hence that the origin is almost GAS. choosing ε < minx. given by ∪xu ∈M \{0} W(xu ). and hence the origin is almost GAS. and with div(ρf ) > 0 in U . Suppose that M the largest invariant set in {x : V is a countable union of isolated points. i. To formulate the stability results in the presence of multiple equilibrium points. R). where ω = R I XXr ω r − bω − n ˆ ¯ kω sω .2010 at 10:37:36 EST from IEEE Xplore. The extension for more generic sets. λ0 = 0 0 1 .e. i. Therefore. Stability using Density Functions and LaSalle’s Invariance Principle The proposed stability analysis results are derived for autonomous nonlinear systems of the form x = f (x). almost all solutions converge to the origin. the set φ(−∞. . Then. The technique is based on using LaSalle’s invariance principle to show that the trajectories approach a candidate set M [4]. uω = 0. Proof: The conditions on V (x) satisfy LaSalle’s invariance principle [4]. bω = kbω (R − R)⊗ . The closed loop error kinematics are given by ˙ ˙ (12) R = R kω (R − R) + (bω )× . and then using the div(ρf ) > 0 property for M \ {0}. Downloaded on February 23. is deﬁned by taking ZU (xu ) with U = Rn . are respectively denoted by φt (A) = {x : x = φt (x0 ). ZU = φ[0. where x(t. The ﬂow φt is a diffeomorphism and hence ZU has zero measure [7]. The following theorem is a new result in density function methodologies. The forward propagation of ZU is ZU itself. in a case where the disturbance in ˙ (1) is a constant bias. τ ]}. Using [12. B. the condition div(ρf ) > 0 for a neighborhood U of every xu ∈ M \ {0} guarantees that the global inset of xu has zero measure. x0 ) denotes the solution of the system at time t with initial condition x0 . integrable in a neighborhood U of xu .∞) (ZU ) and therefore has zero measure. and the dynamics can be reduced to ˙ θ = − sin(θ) + b. sω = i=1 (R I xi )×xr i . (13) with initial conditions θ(t0 ) = θ0 . b = 0} as t → ∞. The global inset of xu . has zero measure. b) : θ = πk. As discussed in Section IVA.e. [15]. ˙ (11) ˙ sets {x : V (x) ≤ c} are bounded and V (x) ≤ 0. and are given by R = ˙ ˆ ˆ ˆ ˆ ˆ ¯ ˆ R (ω)× . The values at time t and at the time interval t ∈ [τ0 . it is assumed that φt is well deﬁned [15.∞) (ZU ). kbω ∈ R+ are feedback ˆ ω is the rate gyro bias estimate. x0 ∈ A} and φ[τ0 . some concepts and results are introduced. ˙ b = − sin(θ).∞) (ZU ) has zero measure. b0 ∈ R. and provides sufﬁcient conditions to show that an equilibrium point is not stable. will be addressed in future work. given a suitable density function. In this case. x0 ). and b on the adopted observer see [17]. and the associated ﬂow φt : Rn → Rn is deﬁned by φt (x0 ) = x(t. Assumption 1: The ﬂow φt is unique. continuous. for more details gains. b(t0 ) = b0 .ThCIn1. denoted as W(xu ). denoted as ZU with a slight abuse of notation. and suppose there exists a nonnegative ρ ∈ C 1 (Rn \ {0}. i. Then the origin of (11) is almost GAS. bω = kbω sω . Lemma 6: Consider the system (11).e.1] with D = U produces 0 ≥ φt (ZU ) ρ(x)dx − ZU ρ(z)dz = t 0 φτ (ZU ) ˆ where bω = bω − uω is the bias estimation error. φ[0. the proposed stability analysis is illustrated for the attitude observer. it is shown that the local inset. b(t) = 0 0 b(t) . Let V : Rn → R be a continuously differentiable function such that the level where f : Rn → Rn is smooth.9 A. ZU (xu ) = {x ∈ U : ∀ ∃T ∀t>T φt (x) − xu  < and ∀t>0 φt (x) ∈ U }.e. Remark 2: Lemma 6 is valid when the invariant set is a countable union of isolated points. Remark 1: The results of Theorem 5 provide an alternative approach to the stability analysis based on HartmanGrobman theorem [15]. By Theorem 5. is a countable union of zero measured sets and hence has zero measure. and exists for all t. and R(t0 ) = exp(θ0 (λ0 )× ). to show that the set of trajectories converging to M \ {0} is of zero measure. ∀x0 ∀ε ∃T >0 ∀t>T inf y∈M φt (x0 )−y < ε. Restrictions apply.e. the local inset ZU is measurable. Also. This property is of interest to exclude the stability of equilibrium points other than the origin.τ ] (A) = {φt (A) : t ∈ [τ0 . Let M be ˙ (x) = 0}. By Lemma 9 presented in Appendix A. The stability analysis technique is illustrated for the case where initial bias and attitude estimation errors exist along the zaxis. Proof: First. The observer kinematics ˙ ˆ are augmented to estimate the bias. Stability of the Nonlinear Observer in the Presence of Biased Inertial Readings In this section. [div(ρf )] (x)dxdτ . 5945 Authorized licensed use limited to: Anders Rantzer. k ∈ Z. Consequently. The combination of Theorem 5 with LaSalle’s invariance principle can be used to provide almost GAS of the origin. the global inset of xu has zero measure. Theorem 5: Let xu ∈ Rn . b(t0 ) = 0 0 b0 . the trajectories of (12) satisfy R(t) = exp(θ(t) (λ0 )× ). and that there is a density function that satisﬁes the conditions of Theorem 5 for all xU ∈ M \{0}. The local inset of xu is the set of all initial conditions inside a neighborhood U of xu that converge to xu without leaving U . Future work will evaluate the contribution of Theorem 5 in this topic. based on generalizing the results adopted in the proof of Theorem 5. i.
Future work will address the topics discussed in Remarks 1 and 2. For the case of autonomous systems. Nonlinear and Optimal Control Theory. for each x ∈ Sk there exists δ sufﬁciently small. pages 163–220. such that x − y < δ ⇒ φt (y) ∈ U for the compact interval t ∈ kT kT + T . editors. Nistri and G. 2008. Consequently. 2002. Jan. Lemma 10: Under Assumption 1. Prentice Hall. S. 2000. b = 0} as t → ∞.M. Proof: (⇐) Immediate from φ[τ0 . [17] J. suggests that the equilibrium points in the set E = {(θ. Bhat and D. Introduction to Smooth Manifolds. the set of initial conditions converging to each point in E has zero measure. respectively. and hence the set of initial conditions that approach E has zero measure. b) = (π + 2πk. On necessary conditions for almost global stability. Control Theory and Multibody Systems.ThCIn1. Kyatkin. which concludes the proof. In 25th Benelux Meeting. B. Real Analysis. T. Nonlinear complementary ﬁlters on the special orthogonal group. Stability and Control. Global stabilization for systems evolving on manifolds. [5] D. 2002. The set C can be described by C = n∈N k∈N0 Cn. for autonomous systems φt (φ[τ0 . IEEE Transactions on Automatic Control. On Rantzer’s density function. Angeli. Using the density function property div(ρf ) > 0 in a neighborhood of the equilibrium points (θ. and φt (φ[τ0 . k ∈ Z.τ ] (A)) = φ[τ0 +t. [14] Halsey Royden. and C is a countable union and intersection of measurable sets and is measurable. By the continuous dependence of φt (x) on the initial conditions [15].k = {x ∈ U : ∃T ∀t≥T φt (x) − 1 xu < n for all t ≥ k}. [8] R. IEEE Transactions on Automatic Control. April 2003. Bernstein. An almost global notion of inputtostate stability. Hamel. Nonlinear observer for rigid body attitude estimation using vector observations. Pﬂimlin. thus measurable. Rantzer. Set Measure Results This section presents auxiliary results adopted in the paper. The time derivative ˙ is given by V = −2 sin2 (θ) and the result is immediate from LaSalle’s invariance principle. 53(5):1203–1218.k . Nonlinear Systems: Analysis. and thus has zero measure [14]. In Proceedings of the 17th IFAC World Congress. [15] S. the set Sk is open. April 2006. it was shown that global stability of the origin can be obtained by combining LaSalle’s invariance principle. 2008.∞) (A) has zero measure. b) : θ = 2πk + π. A PPENDIX A. The set Cn.2010 at 10:37:36 EST from IEEE Xplore. 3. Vasconcelos.F. Proof: The result is obtained by considering the Lyapunov function V = 2(1 − cos(θ)) + b2 . almost all the trajectories approach M \ E. 2 5946 Authorized licensed use limited to: Anders Rantzer. 42(3):161–168. [7] J. Koditschek. depicted in Fig. A dual to Lyapunov’s stability theorem.τ +tk ] (A) is a countable union of zero measure sets. Khalil. Sontag. [10] G. [6] C. 2004. 1999. b) : θ = 2πk. Lageman. 97:131–151. 49(6):866–874. M. Springer. The set E is a countable union of points. Sontag. k ∈ Z. E. (⇒) f is smooth. Phase portrait of the reduced order attitude observer. and E.∞) (A). Trumpf. Engineering Applications of Noncommutative Harmonic Analysis: With Emphasis on Rotation and Motion Groups. for stability analysis of nonlinear systems. K. The set S can be described by S = k∈N0 Sk where Sk = {x ∈ U : φt (x) ∈ U for t ∈ kT kT + T }. Mahony. Chirikjian and A. Consequently. 3. 0). The application of total energy as a Lyapunov function for mechanical control systems.k is measurable. Mahony. R EFERENCES [1] D. 1989. Heeze. [4] H. b) : θ = 2πk. and on t. so there is a neighborhood (2(1−cos(θ))+b2 )2 of every point in E where ρ is integrable and div(ρf ) > 0. Sastry. P. [3] G. The divergence is div(ρf ) = 2(1−cos(θ))−cos(θ)b . 2007. 3rd edition. Proof: To exclude the points in the set E.τ ] (A) has zero measure if and only if φ(−∞. R. .D. Proposition 8: Almost all trajectories of the system (13) approach the set {(θ. An converse theorem for density functions.τ ] (A)) has zero measure [7]. The Netherlands. 48(4):631–634.τ +t] (A) and hence φ(−∞. Malisoff. Lee. 2001. Nonlinear Systems. CONCLUSIONS This work addressed the combination of Lyapunov and density functions. [16] E. March 2006. the density 1 function ρ = 2(1−cos(θ))+b2 is adopted. [9] M.9 Proof: The set ZU can be written as the intersection of a “stability” and a “convergence” sets. shows that these are unstable. then φt is a diffeomorphism for each t [15]. The proposed stability analysis techniques were illustrated for an attitude observer with nonideal angular velocity readings. [11] P.M. Oliveira. Lemma 9: The local inset of an equilibrium point is measurable under Assumption 1. given by ZU = S ∩ C where S = {x ∈ U : φt (x) ∈ U for all t ≥ 0} and C = {x ∈ U : ∀ ∃T ∀t>T φt (x) − xu < }. Dec. and the set S is measurable. Krichman. [12] A. where Cn. Systems and Control Letters. In 41st IEEE Conference on Decision and Control. b = 0} as t → ∞. that can be derived using Lyapunov and density functions. March 2000. and P. Systems and Control Letters. Prentice Hall. by the same arguments used for the measurability of S. the set φ[τ0 . with a density function that excludes the stability of undesirable equilibrium points. Input to state stability: Basic concepts and results. Rantzer. Stefani. k ∈ Z. b = 0} ⊂ M are unstable. the instability of E is shown using a density function that satisﬁes Theorem 5. By Theorem 5. 12(2):161–184. yielding that almost all trajectories approach M \ E = {(θ. June 2008. S. and J. Silvestre. k ∈ Z. 2000. Springer. and J. In 18th International Symposium on Mathematical Theory of Networks and Systems. SpringerVerlag. CRC. The phase portrait of the system.∞) (A) = k∈Z φ[τ0 +tk . Journal of Dynamical and Control Systems. A topological obstruction to continuous global stabilization of rotational motion and the unwinding phenomenon. In P. Let tk = (τ − τ0 )k. [2] S. Berlin. Downloaded on February 23. Monzon. Using the technique summarized in Lemma 6. V. k ∈ Z.τ ] (A) ⊂ φ(−∞. [13] A. Restrictions apply. 39(1):63–70. C. Almost ISS of the origin was formulated as the combination of local ISS and weakly almost ISS. Fig. State observers for invariant dynamics on a lie group. 1988. Meinsma. IEEE Transactions on Automatic Control.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.