Stability Theory For Nonlinear Systems: Giuseppe Oriolo

Stability Theory
for Nonlinear Systems
Giuseppe Oriolo
Sapienza University of Rome
Introduction
consider a nonlinear time-invariant dynamic system

ẋ = f (x, u)
y = g(x)
with state x ∈ IRn , input u ∈ IRp , output y ∈ IRq
typical problem
compute, given x0 = x(0) and u[0,t] , the state x(t) and/or the output y(t) for any t > 0
e.g., in linear systems, where f (x, u) = Ax + Bu, one has

Z t
x(t) = eAt x0 + eA(t−τ ) Bu(τ )dτ
0
however:
often, one is not interested in computing the explicit solution, but rather in studying some
properties such as boundedness, asymptotic behavior, . . .
⇒ qualitative theory of differential equations (Poincaré 1880, Lyapunov 1892, LaSalle and
Lefschetz 1947. . . )
Oriolo: Stability Theory for Nonlinear Systems 1

basic idea
study the qualitative behavior of the system under perturbations of the initial state and
of the input with respect to nominal values
denoting by x(t) the state solution corresponding to x0 and u[0,t] , we wonder:
• what happens if x0 → x0 + ∆x0 ?
• what happens if u[0,t] → u[0,t] + ∆u[0,t] ?
in particular:
• how close is the perturbed evolution to the nominal evolution?
• under which conditions the two solutions tend to coincide for t → ∞?
it seems natural to call
• stable a system in which small perturbations give rise to small variations
• unstable a system in which small perturbations give rise to large variations

stability theory consists of
definitions
stability properties (different kinds depending on system behavior or application needs)
conditions
that a system must satisfy to possess these various properties
criteria
to check whether these conditions hold or not, without computing explicitly the perturbed
solution of the system
e.g., in linear systems
• definition of stability, asymptotic stability, instability
• asymptotic stability condition: limt→∞ x(t)|u≡0 = limt→∞ eAt x0 = 0
• asymptotic stability criteria:

– all eigenvalues of A must have negative real part
– Routh criterion
– Nyquist criterion for feedback systems

typically one considers the behavior of systems in free evolution
ẋ = f (x)
with respect to perturbations of the initial state x0
motivation:
• consider ẋ = f (x, u) with a feedback control law u = h(x); the closed-loop dynamics
becomes
ẋ = f (x, h(x)) = f 0 (x)
that is, a (new) system in free evolution
• consider ẋ = f (x, u); if the input perturbation is non-persistent

u(t) + δ(t) t ∈ [0, t1 ]
ũ(t) =
u(t) t > t1
the problem can be recast as studying the effect of a perturbed initial state (i.e., x(t1 ))

Definitions
an important preliminary concept: equilibrium point
a state xe ∈ IRn is an equilibrium point of system ẋ = f (x) if setting x0 = xe implies

x(t) = xe , ∀t > 0 (a degenerate trajectory of the system)
hence
xe is an equilibrium point if f (xe ) = 0
i.e., equilibrium points are the zeros of the vector function f (x)
e.g., in linear systems ẋ = Ax, equilibrium points must satisfy
Axe = 0 i.e., xe ∈ N (A)
• if A is nonsingular, the only equilibrium point is the origin
• if A is singular, the equilibrium points are infinite and contiguous: geometrically,

they are hyperplanes passing through the origin (lines if dim(N (A)) = 1, planes if
dim(N (A)) = 2, . . . )

e.g., pendulum of length ` and mass m in the presence of viscous friction with coefficient d
m `2 θ̈ + d θ̇ + m g ` sin θ = 0
setting x = (x1 , x2 ) = (θ, θ̇), the state space equations are

ẋ1 = x2
g d
ẋ2 = − sin x1 − x2
` m`2
g
⇒ f (x) = (x2 − sin x1 − d 2 x2 )T ; a nonlinear system!
` m`
equilibrium points are characterized by x1 = jπ (j = 0, 1) and x2 = 0 (pendulum pointing

up/down and at rest)

here are the pendulum trajectories in the plane (x1 , x2 ) = (θ, θ̇) (phase plane)
5
x2
0
-1
-2
-3
-4
-5
-2 0 2 4 6 8
x1
e.g., another nonlinear system
ẋ1 = 1 − x31
ẋ2 = x1 − x22
equilibrium points are described by x1 = 1 and x2 = ±1
note: the equilibrium points of a nonlinear system can be finite (2 in the previous examples,
but any other number is possible, including zero) or infinite, and they can be isolated points
in state space
stability definitions [Lyapunov]
(in the following, | · | denotes any norm in IRn )
an equilibrium point xe is stable (S) if:

∀, ∃ δ() : |x0 − xe | < δ ⇒ |x(t) − xe | < , ∀t > 0
ǫ
δ
xe xe xe xe
x0 x0
∀ ∃ δ() |x0 − xe | < δ |x(t) − xe | < , ∀t > 0
an equilibrium point xe of a dynamic system is stable if it possible to keep the system

evolution arbitrarily close to xe by choosing the initial condition x0 sufficiently close to
xe ; that is, if it is possible to arbitrarily bound the solution in the neighborhood of xe by
suitably bounding the perturbation
obviously: an equilibrium point xe is unstable (U) if it is not stable

• stability is a property of equilibrium points: a system may have both stable and
unstable equilibrium points (only happens in nonlinear systems, e.g., the pendulum)
• the definition of stability does not require the perturbed evolution to converge to xe
• on the other hand, instability does not mean that the perturbed evolution diverges
e.g., Van der Pol oscillator (mass-spring-damper with position-dependent damping)
ẋ1 = x2
ẋ2 = −x1 + (1 − x21 )x2
x0_2
1
regardless of the initial condition,
x0_1
trajectories converge to a limit cycle:
hence, it is impossible to bound
x2
arbitrarily the displacement from 0

-1
(e.g., for = 1 there esists no δ)
-2
x0_3
-3
-3 -2 -1 0 1 2 3
x1
⇒ the origin is an unstable equilibrium point for this system

in practice, often simple stability is not enough:
an equilibrium point xe is asymptotically stable (AS) if:
1. it is stable
2. ∃ δa : |x0 − xe | < δa ⇒ lim |x(t) − xe | = 0
t→∞
• in addition to stability, AS requires that the state converges to xe for initial conditions
sufficiently close to xe
• asymptotic stability is a local concept, i.e., convergence is guaranteed provided that

x0 belongs to the spherical neighborhood of xe of radius δa (basin of attraction); if
x0 is outside this neighborhood, x(t) may not converge or even diverge!
• 2. does not imply 1.; that is, one may have convergence without stability (equilibrium
points of this kind are sometimes called quasi-asymptotically stable, but they are
actually unstable)

e.g., [Vinograd]
x21 (x2 − x1 ) + x52
ẋ1 =
(x21 + x22 )(1 + (x21 + x22 )2 )
x22 (x2 − 2x1 )
ẋ2 =
(x21 + x22 )(1 + (x21 + x22 )2 )
regardless of how close x0 is to the origin,

if x1,0 < 0 the trajectory will converge only
after going a finite distance away from 0:
hence, it is not possible to bound at will
the solution around the origin
⇒ the origin is an unstable (quasi-asymptotically stable) equilibrium point for this system

in applications, it is useful to have an estimate of the time needed by the perturbed
evolution to return to a given neighborhood of xe
an equilibrium point xe is exponentially stable (ES) if there exist positive constants α, λ

and c such that:
|x(t) − xe | ≤ α|x0 − xe |e−λt , ∀t > 0, ∀|x0 − xe | < c
• ES requires that there exists a neighborhood of xe from which perturbed solutions

converge to xe with at least exponential speed; with respect to the definition of AS,
condition 1 may be omitted because it is implied by exponential convergence
• λ is called exponential convergence rate; setting α = eλτ0 , an easy computation

shows that after (τ0 + 1/λ) seconds the distance from xe is reduced to at least 1/e
(around 35%) of its initial value
• ES implies AS but the opposite is not true

e.g., the origin is asymptotically but not exponentially stable for system
ẋ = −x2
in fact, its solution is x(t) = x0 , whose convergence to zero is slower than any
1 + tx0
exponential function

asymptotic stability and exponential stability, which are intrinsically local properties, become
global when the domain of attraction coincides with IRn :
an equilibrium point xe is globally asymptotically stable (GAS) if it is stable and the state
converges to xe for any initial state
an equilibrium point xe is globally exponentially stable (GES) if the state converges

exponentially to xe for any initial state
summarizing, we have the following classification of stable equilibrium points
AS
GAS GES ES
note: a necessary condition for an xe to be GAS is that it is the only equilibrium point
Stability of Linear Systems
theorem
if a linear system admits multiple equilibrium points, stability (instability) of one of them
implies stability (instability) of all the others
proof it is enough to show that, if the generic equilibrium point xe is stable, then the origin
is stable, and vice versa
by hypothesis, we have: ∀, ∃ δ() : |x0 − xe | < δ ⇒ |x(t) − xe | < , ∀t > 0
x(t) − xe is the difference between the solutions starting from x0 and xe , respectively ⇒ due
to linearity, x(t) − xe is also the solution starting from x0 − xe = z0 , denoted by xz0 (t)
we have then: ∀, ∃ δ() : |z0 | < δ ⇒ |xz0 (t)| < , ∀t > 0; that is, the origin is stable
the vice versa is shown similarly
theorem
in a linear system:
1. only the origin can be AS, and only when there exist no other equilibrium points
2. if the origin is AS, it is also GAS
proof
1: trivial (see slide 5)
2: trivial for finite-dimensional time-invariant systems: local convergence of the free evo-
lution x(t) = eAt x0 requires the eigenvalues of A to have negative real part; but then,
convergence is global
theorem
in a linear system, the origin is ES if and only if it is AS
proof
necessity: trivial
sufficiency: trivial for finite-dimensional time-invariant systems, because if the origin is AS
then the free evolution is a combination of converging exponentials
summarizing, in linear systems:

• if the origin is the only equilibrium point, it can be S, AS (actually, ES), or U
• if there are multiple equilibrium points, they are infinite, contiguous, and they are either
all S or all U
• in any case, one may directly say that the system is S, AS (actually, ES), or U
the following stability criterion is immediate

theorem
a finite-dimensional time-invariant linear system is S if and only if
1. eigenvalues of A with geometric multiplicity equal to algebraic multiplicity have non-
positive real part
2. eigenvalues of A with geometric multiplicity lower than algebraic multiplicity have neg-
ative real part
the system is ES if and only if all the eigenvalues of A have negative real part
to avoid computing eigenvalues, apply Routh criterion to characteristic polynomial of A

Direct Lyapunov Method
basic idea
if the total energy of a system (mechanical, electrical, . . . ) is continuously dissipated,

the system (linear or nonlinear) tends to an equilibrium ⇒ one may be able to prove
stability/instability by looking at the variation of a single scalar function
e.g., nonlinear mass-spring-damper system, state x = (z, ż)
nonlinear spring
m
mz̈ + bż|ż| + (k0 z + k1 z 3 ) = 0
nonlinear damper
z
one cannot study stability of the origin using the definition, because it is impossible to solve
the above differential equation in closed-form: let us look then at mechanical energy
Z z
1 1 1 1
V (x) = Vkin (ż) + Vpot(z) = mż 2 + (k0 ζ + k1 ζ 3 )dζ = mż 2 + k0 z 2 + k1 z 4
2 0 2 2 4

energy/stability relationships
• energy is zero only at the equilibrium point z = 0, ż = 0, i.e., the origin
• if energy (always) converges to zero, then the origin is (globally) asymptotically stable
• if energy diverges, then the origin is unstable
how does energy change when the system moves? it is sufficient to differentiate V with
respect to t (of which V is a composite function) and replace z̈ with its expression derived
from the dynamic model
V̇ (x) = mżz̈ + (k0 z + k1 z 3 )ż = −b|ż|3 ≤ 0
⇒ intuitively, energy is continuously dissipated until the system converges to a state with
zero velocity (ż = 0); moreover, since in any position different from z = 0 the mass would
be subject to a nonzero elastic force −k0 z − k1 z 3 , we may conclude that state trajectories
actually converge to the origin (z = 0, ż = 0)
the direct Lyapunov method is based on a generalization (and rigorous formalization) of

the above reasoning: one looks for a suitable energy-like scalar function for the nonlinear
system under consideration, and studies its evolution in time as the system moves

in the following, we refer to the nonlinear time-invariant systems in free evolution
ẋ = f (x) x ∈ IRn
and denote by xe the equilibrium point under study; thus, f (xe ) = 0
preliminary definitions: consider a scalar function V (x), continuously differentiable with

respect to x (i.e., V ∈ C 1 ), and a spherical neighborhood S(xe , r) of xe with radius r
• V (x) is positive definite (PD) in S(xe , r) if

a) V (xe ) = 0
b) V (x) > 0, ∀x ∈ S(xe , r), x 6= xe
• V (x)is positive semidefinite (PSD) in S(xe , r) if

a) V (xe ) = 0
b) V (x) ≥ 0, ∀x ∈ S(xe , r), x 6= xe
• V (x) is negative definite (ND) in S(xe , r) if −V (x) is positive definite, negative

semidefinite (NSD) in S(xe , r) if −V (x) is positive semidefinite
• V (x) is indefinite (I) in S(xe , r) if it is not DP, SDP, DN or SDN
note: V (x) PD (ND) in S(xe , r) ⇒ V (x) PSD (NSD) in S(xe , r)

case n = 2: local shape of a PD function V around xe
V x2
V=V3
V=V2
V=V1
V3 xe
V2
x2
V1
xe x1
x1 3D plot contour plot
e.g., in IR2 , function V (x) = xT x = x21 + x22 is PD in any neighborhood of the origin (all
level curves are closed)
e.g., in IR2 , function V (x) = x21 is PSD in any neighborhood of the origin (it is zero on all
points of the x2 axis; no level curves are closed)
e.g., in IR2 , function V (x) = x1 x2 is I in any neighborhood of the origin (there are always
neighborhood points where it is positive and neighborhood points where it is negative)
e.g., for the nonlinear mass-spring-damper, mechanical energy is PD in any neighborhood
of the origin
assume a function V (x) is given, and consider a solution x(t) of ẋ = f (x): one may consider
V (x(t)) as a composite function of t, continuously differentiable for any t; we have
n n
dV (x(t)) X ∂V ∂xi X ∂V
V̇ (t) = = = fi (x(t)) = V̇ (x)
dt i=1
∂x i ∂t i=1
∂x i
where fi (x(t)) is the i-th component of vector function f (x)
V̇ (x), regarded as a function of x, is the derivative of V along the system trajectories
V̇ (x) can then be positive (negative) definite, positive (negative) semidefinite, or indefinite
e.g., consider the system

ẋ1 = x1 x2 + x22
ẋ2 = −x21 − x1 x2 − x2
whose only equilibrium point is the origin, and let V = (x21 + x22 )/2, which is PD in any
neighborhood of the origin; we have
V̇ (x) = x1 ẋ1 + x2 ẋ2 = x21 x2 + x1 x22 − x21 x2 − x1 x22 − x22 = −x22
which is NSD in any neighborhood of the origin

theorem
an equilibrium point xe of ẋ = f (x) is stable if there exists a function V (x) ∈ C 1 such that
1. V (x) is PD in a neighborhood S(xe , r)
2. V̇ (x) is NSD in the same neighborhood
proof based on geometric arguments, case n = 2 (but valid in general)
note first that, since V (x) is PD in S(xe , r), the level curves Uk = {x ∈ IR2 : V (x) = k} are
closed for sufficiently small k; moreover, if k1 < k2 , Uk1 is inside Uk2
x2 S(xe ,r)
Uk 2
Uk 1
xe
x1

x2
S(xe ,r1 )
Uk
S(xe ,r2 )
x0
xe
x1
choose r1 such that 0 < r1 ≤ r: then, there exists certainly a k such that Uk is inside S(xe , r1 )
(just take the minimum of V along the boundary of S(xe , r1 ), and a k smaller than such
minimum); hence, Uk is closed
moreover, since Uk is a closed curve that contains xe , it is always possible to find an r2 such
that S(xe , r2 ) is inside Uk
now consider a trajectory starting from x0 ∈ S(xe , r2 ); we have V (x0 ) < k and, since V̇ is
negative or zero along the system trajectories contained in S(xe , r), V (t) is non-increasing
along the trajectory; hence, we have V (t) < k, ∀t > 0
⇒ the trajectory x(t) will always remain in S(xe , r1 )
wrapping up, for arbitrarily small r1 we can always find a sufficiently small r2 such that
|x0 − xe | < r2 ⇒ |x(t) − xe | < r1 , ∀t > 0
• a V (x) with the properties required by the previous theorem (i.e., such that V , V̇ are
respectively PD, NSD in a neighborhood of xe ) is called a Lyapunov function
• the theorem states that the existence of a Lyapunov function is a sufficient condition
for stability; actually, for finite-dimensional time-invariant system one may show that
this condition is also necessary
• applying the theorem requires 2 phases, possibly to be repeated:

1. choose a V (x) that is PD in a neighborhood of xe (called Lyapunov candidate)
2. compute V̇ along the system trajectories and check whether it is NSD in the same
neighborhood
note: if the chosen V (x) does not result to be a Lyapunov function, no conclusion can
be drawn; another V 0 (x) may exist that is a Lyapunov function
• if a system admits a Lyapunov function V (x), then it admits an infinity of them; e.g.,
all the following
V 0 (x) = βV γ (x) β > 0, γ > 1
• coming up with ‘good’ Lyapunov candidates is obviously essential: total energy is

usually a reasonable choice in mechanical or electrical systems, but in general better
alternatives may exist without a clear physical meaning

e.g., pendulum (for simplicity, m = 1, d = 1, ` = 1)
the state vector is x = (x1 , x2 ) = (θ, θ̇)
ẋ1 = x2
ẋ2 = −g sin x1 − x2
for the equilibrium point xdown

e = (0, 0), let us choose mechanical energy as a Lyapunov
candidate
1 2
V (x) = Vkin (x) + Vpot(x) = x2 + g(1 − cos x1 ) PD in S(0, 2π − )
2
we find
V̇ (x) = x2 ẋ2 + g sin x1 ẋ1 = −x22 NSD in the same neighborhood (actually, in all IR2 )
Rt
hence xdown
e is a stable equilibrium point (and 0
V̇ dτ is the dissipated energy)
physical intuition suggests that, in the presence of friction, the origin is an asymptotically
stable equilibrium point for the pendulum ⇒ we need a stronger result to prove it

theorem
an equilibrium point xe of ẋ = f (x) is asymptotically stable if there exists a function

V (x) ∈ C 1 such that
1. V (x) is PD in a neighborhood S(xe , r)
2. V̇ (x) is ND in the same neighborhood
proof first, xe is certainly stable; in particular, if x0 ∈ S(xe , r2 ) (see previous proof) the
trajectory will always remain in S(xe , r1 ) ⇒ V (t) along the trajectory tends to a limit value
V̄ ≥ 0 (because V̇ < 0 and V is bounded below)
now suppose that V̄ > 0; since V is continuous and only zero at xe , then there would exist
a neighborhood S(xe , σ) in which the trajectory never enters; since V̇ too is continuous and
only zero at xe , there would also exist an α > 0 such that V̇ ≤ −α indefinitely
but then we could write
Z t
V (t) = V (0) + V̇ (τ )dτ ≤ V (0) − αt
0
and thus V would become negative in finite time, contradicting the hypothesis V̄ > 0
hence, if x0 ∈ S(xe , r2 ) we have limt→∞ V (t) = V̄ = 0; therefore, being V (x) zero only at
x = xe , we conclude that limt→∞ x(t) = xe
note: extrapolating the properties of S(xe , r2 ) in the above proof, one may infer that any neighborhood of
xe contained in UV ∗ (where V ∗ is the minimum of V along the boundary of S(xe , r)) is a lower estimate of
the basin of attraction of xe
ẋ1 = x1 (x21 + x22 − 1) − x2
ẋ2 = x1 + x2 (x21 + x22 − 1)
for which the origin is an equilibrium point
choosing
1 2
V (x) = (x1 + x22 ) PD in any neighborhood of the origin
2
we find
V̇ (x) = (x21 + x22 )(x21 + x22 − 1) ND for x : x21 + x22 < 1, i.e., in S(0, 1− )
the origin is then asymptotically stable for the system
to estimate the basin of attraction:
let UV ∗ = {x ∈ IR2 : V (x) ≤ 1/2} = S(0, 1− ); for any choice of ρ ∈ (0, 1), the neighbor-
hood S(0, ρ) is contained in UV ∗ , and hence it represents a lower estimate of the basin of
attraction of the origin

e.g., pendulum; consider the following Lyapunov candidate (no physical meaning)
1 2 1
V (x) = x2 + 2g(1 − cos x1 ) + (x1 + x2 )2 PD in any neighborhood of the origin
2 2
we get
V̇ (x) = −x22 −g x1 sin x1 ND in S(0, π − )
hence xdown
e is an asymptotically stable equilibrium point for the pendulum
domain of attraction: convergence to the origin is guaranteed from initial states in-
side level curves that are completely contained in the region where V̇ is DN; if
the initial state is inside a level curve that leaves such region, divergence may occur

what happens if we try to apply the previous theorems to the pendulum equilibrium point
xup
e = (π, 0)? physical intuition tells us that xup e is unstable, but the necessary (and
sufficient) condition for stability is the existence of a Lyapunov function, which we cannot
exclude a priori ⇒ an instability criterion may be useful
theorem [Cetaev]
an equilibrium point xe of ẋ = f (x) is unstable if there exists a V (x) ∈ C 1 such that

1. xe is an accumulation point for set P = {x : V (x) > 0}
2. V̇ (x) is PD in U = P ∩ S(xe , r), for some r > 0
e.g., the equilibrium point xe = (0, 0) is unstable for the system

ẋ1 = x1 + x22
ẋ2 = −x2
1 2
in fact, consider V (x) = x
2 1
− 12 x22 , positive in P = {x : |x1 | > |x2 |}, of which xe is an
accumulation point
x2
we have
U U
xe 1 x1
V̇ (x) = x21 + x1 x22 + x22 = x21 + x22 (1 + x1 )
that is clearly PD in U = P ∩ S(xe , 1)

there also exists a criterion for global asymptotic stability
theorem
an equilibrium point xe of ẋ = f (x) is globally asymptotically stable if there exists a
function V (x) ∈ C 1 such that
1. V (x) is PD in any neighborhood of xe
2. V̇ (x) is ND in any neighborhood of xe
3. V (x) is radially unbounded, i.e.., lim V (x) = ∞
|x−xe |→∞
e.g.
x21 2
V = 2 + x2 V = x21 + x22
1 + x1
x2 x2
x1 x1
not radially unbounded radially unbounded
proof as in the local case, having observed that the radial unboundedness of V , together
with the fact that V̇ is ND in all IRn , implies that for any initial condition x0 the trajectory
will remain within the limited region defined by V (x) ≤ V (x0 )
note: if V is not radially unbounded, level curves far from xe are not closed; hence, the state
trajectory may diverge from xe and still remain within the region defined by V (x) ≤ V (x0 ),
while actually crossing level curves that correspond to decreasing values of V
⇒ when x0 is sufficiently far from xe , x(t) may not converge to xe

e.g., consider the family of nonlinear systems described by
ẋ = −c(x), where xc(x) > 0, ∀x 6= 0, and c(0) = 0
and the Lyapunov candidate
1 2
xV (x) =
2
that is PD in any neighborhood of xe = 0 and radially unbounded
we find
V̇ (x) = xẋ = −xc(x)
i.e., V̇ (x) is ND in any neighborhood of xe = 0
⇒ xe is globally asymptotically stable
summarizing, the direct Lyapunov criterion provides the following results:
xe is S xe is AS xe is GAS xe is unstable
PD in any S(xe , r) xe accumulation pt.
V (x) PD in S(xe , r) PD in S(xe , r)
and radially unbounded of P = {x : V (x) > 0}
V̇ (x) NSD in S(xe , r) ND in S(xe , r) ND in any S(xe , r) PD in P ∩ S(xe , r)

Building Lyapunov Functions
the main difficulty in applying the direct Lyapunov method for studying an equilibrium
point xe of a nonlinear system ẋ = f (x) consists in choosing the candidate function V (x);
sometimes the physics of the problem provides an inspiration, but in general it may be
useful to proceed in a systematic way
a strategy which is often effective is to define V (x) as a quadratic form of the kind
1
V (x) = (x − xe )T Q(x − xe )
2
where the n × n matrix Q is symmetric and positive definite (i.e., such that wT Qw > 0,
∀w 6= 0)
to guarantee that Q is positive definite one may use the following necessary and sufficient
Sylvester condition

Q11 Q12 Q13
Q11 Q12
Q11 > 0, > 0, Q12 Q22 Q23 > 0, ... det(Q) > 0
Q12 Q22
Q13 Q23 Q33
since Q is symmetric, V̇ (x) is computed as

1 T 1 1 1
V̇ (x) = ẋ Q(x − xe ) + (x − xe )T Q̇(x − xe ) + (x − xe )T Qẋ = (x − xe )T Qẋ + (x − xe )T Q̇(x − xe )
2 2 2 2
ẋ1 = k2 x1
ẋ2 = −x32 + k1 x3
ẋ3 = −2x2 − x33
with k2 < 0 and k1 > 0; the origin is the only equilibrium point
• choose
1 1 1
V (x) =(x − xe )T I3×3 (x − xe ) = xT x = (x21 + x22 + x23 )
2 2 2
which is PD in any neighborhood of the origin and radially unbounded
we get
V̇ (x) = xT ẋ = x1 ẋ1 + x2 ẋ2 + x3 ẋ3 = k2 x21 − x42 + (k1 − 2)x2 x3 − x43
in the special case k1 = 2, V̇ (x) is ND in any neighborhood of the origin, which is

then GAS
• a much more general result is obtained by choosing Q = diag(1, 2 , 1)

k1
1 T 1 2 2 2 4
V (x) = x Qx = (x21 + x2 + x23 ) ⇒ V̇ (x) = k2 x21 − x2 − x43
2 2 k1 k1
that, for k2 < 0 and k1 > 0, is always ND ⇒ the origin is always GAS

Invariant Set Theorem
in the application of the direct method, one often finds that the time derivative V̇ (x) of
the chosen Lyapunov function is only NSD (rather than ND); in these conditions, one may
infer stability but not asymptotic stability of xe (e.g., see the first Lyapunov function for
the pendulum)
in this situation, the invariant set theorem may allow to analyze stability in more detail
without changing V (x)
a subset G ⊂ IRn of the state space is an invariant set for ẋ = f (x) if any trajectory x(t)
starting from a point x0 ∈ G always stays in G
it is a generalization of the concept of equilibrium point; examples of invariant sets are
• equilibrium points
• the basin of attraction of an asymptotically stable equilibrium point
• any trajectory of the system (as long as the system is time-invariant)
• IRn itself

basic idea
if V (x) is PD (i.e., V (x) > 0) and V̇ (x) is NSD (i.e., V̇ (x) ≤ 0) in a neighborhood of xe , V (x)
must tend to a limit value ⇒ V̇ (x) should tend to zero, at least under certain conditions
local invariant set theorem [LaSalle]
for a system ẋ = f (x), assume that there exists a function V (x) ∈ C 1 such that:
1. region Ωα = {x ∈ IRn : V (x) ≤ α} is bounded, for some α > 0
2. V̇ (x) ≤ 0 in Ωα
and define P , the set of points of Ωα where V̇ = 0; then, any trajectory of the system that
starts in Ωα tends asymptotically to M , the largest invariant set contained in P
here: largest invariant set contained in P = union of all invariant subsets of P
V x2
Ωα
M
xe
x0
α P
x2
xe x1
x1

an immediate consequence is the following
corollary
an equilibrium point xe of ẋ = f (x) is asymptotically stable if there exists a function

V (x) ∈ C 1 such that
1. V (x) is PD in a set D that contains xe
2. V̇ (x) is NSD in the same set
3. the largest invariant set M in P (the subset of D where V̇ = 0) consists of xe only
moreover, denoting by Ω the largest region defined by V (x) ≤ α, α > 0 and contained in D,
we have that Ω is an estimate of the basin of attraction of xe
• condition 1 of the corollary implies condition 1 of the local invariant set theorem
• compared with the direct Lyapunov criterion for AS, this corollary ‘relaxes’ condition 2
(ND → NSD) but adds condition 3
• the direct Lyapunov criterion for AS is a special case of this result (P = xe )
• in itself, D is not an estimate of the basin of attraction (in fact, some of the level
curves crossing D may be open, and in that case D would not be invariant)

e.g., let us consider again the pendulum with the first Lyapunov function
1 2
V (x) = x + g(1 − cos x1 ) PD in S(0, 2π − )
2 2
we have
V̇ (x) = x2 ẋ2 + g sin x1 ẋ1 = −x22 NSD in the same neighborhood (actually in all IR2 )
therefore, xdown
e = (0, 0) is a stable equilibrium point for the pendulum; but the invariant
set theorem tells us more
set P consists of points for which V̇ = 0, i.e., states such that x2 = 0; which is the largest
invariant set M contained in P ?
the dynamics of the system in P is

ẋ1 = 0
ẋ2 = −g sin x1
if x1 6= 0, we would get ẋ2 6= 0 and thus x2 would change, driving x(t) outside set P
⇒ set M consists only of the origin
hence, xdown
note: the basin of attraction of the origin can be estimated as follows: compute V ∗ , the minimum of V along
the boundary of S(0, 2π − ), and identify UV ∗ , the level curve of V corresponding to V ∗ ; a lower estimate of the
basin of attraction is the region of IR2 contained in UV ∗
there exists a global version of the invariant set theorem
global invariant set theorem [LaSalle]
for a system ẋ = f (x), assume that there exists a function V (x) ∈ C 1 such that:
1. V (x) is radially unbounded
2. V̇ (x) ≤ 0 in IRn
then, any trajectory of the system tends asymptotically to the set M , the largest invariant
set contained in P , the set of points of Ωα where V̇ = 0
note: radial unboundedness of V (x) guarantees that any region Ωα = {x ∈ IRn : V (x) < α}, α > 0, is bounded
and the corresponding
corollary
an equilibrium point xe of ẋ = f (x) is globally asymptotically stable if there exists a

function V (x) ∈ C 1 such that
1. V (x) is PD in any neighborhood of xe and radially unbounded
2. V̇ (x) is NSD in any neighborhood of xe
3. the largest invariant set M in P (the subset of D where V̇ = 0) consists of xe only

e.g., consider the family of second-order nonlinear systems described by
z̈ + b(ż) + c(z) = 0
where functions b(·) and c(·) are continuous and such that
ż b(ż) > 0, ∀ż 6= 0 z c(z) > 0, ∀z 6= 0
note that these conditions, combined with continuity, imply that b(0) = 0, c(0) = 0
.
b(z) c(z)
.
z z
this family includes mass-spring-damper mechanical systems (with nonlinear elastic force
c(z) and viscous friction b(ż)) and RLC electrical systems (with nonlinear inductance c(z)
and resistance b(ż))

consider the only equilibrium point xe = (ze , że ) = (0, 0) and, as a Lyapunov candidate, the
total energy of the system (e.g., kinetic + potential)
Z z
1
V (x) = ż 2 + c(y)dy
2 0
which is PD in any neighborhood of xe (but not necessarily radially unbounded); then
V̇ (z) = żz̈ + c(z)ż = −żb(ż) − żc(z) + c(z)ż = −żb(ż)
is NSD in any neighborhood of xe = 0 in view of the assumptions
set P consists of points for which V̇ = 0, i.e., states such that ż = 0; which is the largest
invariant set M contained in P ? the dynamics of the system in P is
z̈ = −c(z)
if z 6= 0, we would get z̈ 6= 0 and thus ż would change, driving x(t) outside set P ⇒ set M
consists only of the origin
hence, the origin is an AS equilibrium point for any system of this family; if, in addition,
Z z
lim c(y)dy = ∞
z→∞ 0
then V is radially unbounded and the origin is GAS

Barbalat Lemma
invariant set theorems only apply to time-invariant systems!
the following result may sometimes prove useful for time-varying systems
lemma [Barbalat]
for a system ẋ = f (x, t), consider a function V (x, t) ∈ C 1 such that:

1. V (x) is lower bounded
2. V̇ (x, t) ≤ 0
3. V̇ (x, t) is uniformly continuous
then V̇ (x, t) converges to zero along the trajectories of the system
• conditions 1 and 2 guarantee that V (x, t) tends to a finite limit
• condition 3 is usually replaced by the (stronger) condition

3. V̈ (x, t) is bounded
• this lemma can also be useful for time-invariant systems, because it relaxes some
conditions (e.g., V is not required to be PD)

Indirect Lyapunov Method
basic idea
by analyzing the stability of the linear approximation of a nonlinear system around an

equilibrium point xe , it may be possibile to infer conclusions about the stability of xe for
the original system
consider a nonlinear system ẋ = f (x), with equilibrium point xe , that is, f (xe ) = 0
if f ∈ C ∞ , Taylor expansion around xe provides

df
f (x) = f (xe ) + (x − xe ) + h(x − xe ) = J(xe )(x − xe ) + h(x − xe )
dx xe
where h(x−xe ) collects the (infinite) terms of degree higher than 1 and J(xe ) is the Jacobian
matrix of f with respect to x, computed at xe
in the new coordinates ξ = x − xe , the dynamics is described by
ξ̇ = ẋ = f (x) = J(xe )ξ + h(ξ)
in the vicinity of the equilibrium point xe , higher-order terms may be neglected ⇒ we obtain
the following linear approximation
ξ̇ = J(xe )ξ
whose accuracy in describing the original system is higher as the state is closer to xe
the analysys of the linear approximation ξ̇ = J(xe )ξ leads to interesting conclusions for the
original nonlinear system ẋ = f (x)
theorem
if matrix J(xe ) is nonsingular, xe is an isolated equilibrium point of the nonlinear system
proof by contradiction: if the thesis were false, in any neighborhood of xe there would be
at least one x0e such that f (x0e ) = f (xe ) = 0; we would have then
f (x0e ) = f (xe ) + J(xe )(x0e − xe ) + h(x0e − xe ) = 0 ⇒ J(xe )(x0e − xe ) + h(x0e − xe ) = 0
since x0e changes, this condition requires both terms to be zero; in particular, it must be
J(xe )(x0e − xe ) = 0, which however contradicts the fact that J(xe ) is non singular
the opposite is not true; it may happen that xe is an isolated equilibrium and J(xe ) is singular
e.g., consider the nonlinear system

ẋ1 = x21
ẋ2 = x2
whose only equilibrium point is the origin; however

2x1 0
0 0
J(xe ) = =
0 1 x 0 1
e
is singular

the stronger result is the following
theorem (indirect Lyapunov criterion)
consider the linear approximation ξ̇ = J(xe )ξ of a nonlinear system ẋ = f (x) around an

equilibrium point xe
1. if all the eigenvalues of J(xe ) have negative real part (i.e., if the linear approximation
is AS), xe is an asymptotically stable equilibrium point for the nonlinear system
2. if at least one of the eigenvalues of J(xe ) has positive real part (i.e., if the linear
approximation is U), xe is an unstable equilibrium point for the nonlinear system
proof based on the application of direct Lyapunov method: in particular, one shows that
there exists a Lyapunov function for the linear approximation which is also a Lyapunov
function for the nonlinear system
• asymptotic stability of the origin for the linear approximation (which is always global)
only implies local asymptotic stability of xe for the nonlinear system
• if no eigenvalue of J(xe ) has positive real part, but some of them have zero real part (i.e.,
the linear approximation is S — not AS — or U, depending on the relationship between
algebraic and geometric multiplicity for these eigenvalues) we are in the critical case:
no conclusion can be drawn on the stability of xe for the nonlinear system (higher-order
terms are decisive)

e.g., consider the pendulum again
ẋ1 = x2
ẋ2 = −g sin x1 − x2
the Jacobian matrix is

df 0 1
J(x) = =
dx −g cos x1 −1
• around the equilibrium point xdown

e = (0, 0) it is

df 0 1
Jedown (x) =

=
dx xdown −g −1
e
whose characteristic polynomial is λ2 + λ + g; therefore, the linear approximation of the

pendulum around xdown
e is AS
⇒ xdown
• around the equilibrium point xup

e = (π, 0) it is

df = 0 1
Jeup (x) =

dx xup g −1
e
whose characteristic polynomial is λ2 + λ − g; therefore, the linear approximation of the

pendulum around xupe is U
⇒ xup
e is an unstable equilibrium point for the pendulum

the application of the indirect criterion is inconclusive in the critical case; in this situation,
one needs to resort to the direct criterion, which is more powerful (and may also allow to
determine the basin of attraction, which cannot be analyzed with the indirect method)
e.g., consider the nonlinear system

ẋ = −x3
whose only equilibrium point is xe = 0; in this case, the Jacobian reduces to a scalar

df 2

J(xe ) = = −3x xe
=0
dx xe
and the linear approximation of the system around xe is ξ̇ = 0 ⇒ critical case
consider the following Lyapunov candidate

1 2
V (x) = x PD in any neighborhood of the origin, and radially unbounded
2
we have
V̇ (x) = −x4 ND in any neighborhood of the origin
⇒ xe is then GAS for the nonlinear system

Stability Theory For Nonlinear Systems: Giuseppe Oriolo

Uploaded by

Copyright:

Available Formats

Stability Theory For Nonlinear Systems: Giuseppe Oriolo

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Stability Theory For Nonlinear Systems: Giuseppe Oriolo

Uploaded by

Copyright:

Available Formats

Stability Theory

for Nonlinear Systems

consider a nonlinear time-invariant dynamic system

e.g., in linear systems, where f (x, u) = Ax + Bu, one has

Oriolo: Stability Theory for Nonlinear Systems 1

denoting by x(t) the state solution corresponding to x0 and u[0,t] , we wonder:

• what happens if x0 → x0 + ∆x0 ?

• what happens if u[0,t] → u[0,t] + ∆u[0,t] ?

• how close is the perturbed evolution to the nominal evolution?

• under which conditions the two solutions tend to coincide for t → ∞?

it seems natural to call

• stable a system in which small perturbations give rise to small variations

• unstable a system in which small perturbations give rise to large variations

Oriolo: Stability Theory for Nonlinear Systems 2

e.g., in linear systems

• definition of stability, asymptotic stability, instability

• asymptotic stability condition: limt→∞ x(t)|u≡0 = limt→∞ eAt x0 = 0

• asymptotic stability criteria:

Oriolo: Stability Theory for Nonlinear Systems 3

with respect to perturbations of the initial state x0

• consider ẋ = f (x, u); if the input perturbation is non-persistent

Oriolo: Stability Theory for Nonlinear Systems 4

an important preliminary concept: equilibrium point

a state xe ∈ IRn is an equilibrium point of system ẋ = f (x) if setting x0 = xe implies

e.g., in linear systems ẋ = Ax, equilibrium points must satisfy

Axe = 0 i.e., xe ∈ N (A)

• if A is nonsingular, the only equilibrium point is the origin

• if A is singular, the equilibrium points are infinite and contiguous: geometrically,

Oriolo: Stability Theory for Nonlinear Systems 5

setting x = (x1 , x2 ) = (θ, θ̇), the state space equations are

equilibrium points are characterized by x1 = jπ (j = 0, 1) and x2 = 0 (pendulum pointing

Oriolo: Stability Theory for Nonlinear Systems 6

e.g., another nonlinear system

equilibrium points are described by x1 = 1 and x2 = ±1

an equilibrium point xe is stable (S) if:

∀ ∃ δ() |x0 − xe | < δ |x(t) − xe | < , ∀t > 0

an equilibrium point xe of a dynamic system is stable if it possible to keep the system

obviously: an equilibrium point xe is unstable (U) if it is not stable

Oriolo: Stability Theory for Nonlinear Systems 8

arbitrarily the displacement from 0

⇒ the origin is an unstable equilibrium point for this system

Oriolo: Stability Theory for Nonlinear Systems 9

an equilibrium point xe is asymptotically stable (AS) if:

• asymptotic stability is a local concept, i.e., convergence is guaranteed provided that

Oriolo: Stability Theory for Nonlinear Systems 10

regardless of how close x0 is to the origin,

Oriolo: Stability Theory for Nonlinear Systems 11

an equilibrium point xe is exponentially stable (ES) if there exist positive constants α, λ

• ES requires that there exists a neighborhood of xe from which perturbed solutions

• λ is called exponential convergence rate; setting α = eλτ0 , an easy computation

• ES implies AS but the opposite is not true

Oriolo: Stability Theory for Nonlinear Systems 12

an equilibrium point xe is globally exponentially stable (GES) if the state converges

summarizing, we have the following classification of stable equilibrium points

summarizing, in linear systems:

the following stability criterion is immediate

to avoid computing eigenvalues, apply Routh criterion to characteristic polynomial of A

if the total energy of a system (mechanical, electrical, . . . ) is continuously dissipated,

∀ ∃ δ() |x0 − xe | < δ |x(t) − xe | < , ∀t > 0