Lecture Notes for Math 524

Dr. Yingfei Yi

These notes are based on the lecture notes of Professor James S. Muldowney, the books of Hale,
Copple, Coddington and Levinson, and Perko. They are for the use of students in my graduate
ODE class.

Chapter 3. Stability Theory

3.1. Definitions and examples

Consider an ODE in Rn
x0 = f (t, x) (1)
where f ∈ C([0, ∞) × D → Rn ), D ⊂ Rn open. A vector x ∈ Rn is an equilibrium of (1)
if f (t, x) = 0, 0 ≤ t < ω. In this case, x(t) ≡ 0 is a solution of (1). The change of variables
x̃(t) = x(t) − x transforms (1) to

x̃0 = f˜(t, x̃) = f (t, x̃ + x) (2)

and x̃(t) ≡ 0 is an equilibrium of (2), corresponding to the equilibrium x of (1). We thus will
assume that f (t, 0) = 0, 0 ≤ t < ω.

Definitions.

(1) The equilibrium 0 is Lyapunov stable (or simply stable) if, for each ² > 0 and t0 ≥ 0, there
exists δ = δ(t0 , ²) > 0 such that |x(t0 )| < δ =⇒ x(t) exists for all t ≥ t0 and |x(t)| < ².

(2) The equilibrium 0 is uniformly stable if δ = δ(²).

(3) The equilibrium 0 is unstable if it is not stable.

(4) The equilibrium 0 is asymptotically stable if

(a) 0 is stable; and

(b) ∃δ0 = δ0 (t0 ) > 0 such that |x(t0 )| < δ0 =⇒ |x(t)| → 0 as t → ∞.

(5) The equilibrium 0 is uniformly asymptotically stable if

(a) 0 is uniformly stable; and

(b) ∃δ0 > 0 such that, for each ² > 0, there exists T = T (²) > 0 such that |x(t0 )| <
δ0 , t ≥ t0 + T =⇒ |x(t)| < ².

1
Remark. The discussion of stability of equilibria is easily extended to stability of any particular
solution x(t) of (1) which exists on [t0 , ∞), by the change of variables x̃(t) = x(t) − x(t). Then

x̃0 (t) = x0 (t) − x0 (t) = f (t, x) − f (t, x) = f (t, x̃(t) + x(t)) − f (t, x(t))
:= f˜(t, x̃(t)).

Thus x̃ satisfies
x̃0 = f˜(t, x̃)
and x̃ = 0 is an equilibrium.

To illustrate the differences among various notions of stability, we examine the following exam-
ples.

Example 1. Integrating
−x
x0 = , t ∈ (−1, ∞)
1+t
we get x(t) = x(t0 ) 1+t
1+t . Thus
0

|x(t)| ≤ |x(t0 )|, for t ≥ t0 .

Therefore x = 0 is uniformly stable (choose δ(t0 , ²) = ² ), and x = 0 is asymptotically stable since

limt→∞ |x(t)| = 0. However, x(t) is not uniformly asymptotically stable, since for any T > 0,
1 + t0
x(t0 + T ) = x(t0 )
1 + t0 + T
1+t0
and 1+t0 +T → 1 as t0 → ∞.

Example 2. Integrating x0 = 0 we get x(t) = x(t0 ). Thus x = 0 is uniformly stable but not
asymptotically stable.

Example 3. Integrating x0 = x we get x(t) = x(t0 )et−t0 . x = 0 is unstable.

Example 4. Integrating x0 = −x we get x(t) = x(t0 )e−(t−t0 ) . x = 0 is uniformly stable since

|x(t)| ≤ |x(t0 )| for t ≥ t0 . Moreover, x = 0 is uniformly asymptotically stable, since, by choosing
δ0 = 1, T (²) = t − t0 sufficiently large such that e−T < ², we can ensure |x(t)| < ² for all
|x(t0 )| < δ0 and t ≥ t0 + T.
Rt
a(s)ds
Example 5. Integrating x0 = a(t)x we get x(t) = x(t0 )e t0 .
Rt
(a) x = 0 is stable ⇐⇒ a(s)ds ≤ K < ∞, ∀ t > 0.
0
Rt
(b) x = 0 is asymptotically stable ⇐⇒ limt→∞ 0 a(s)ds = −∞.
Rt
(c) x = 0 is uniformly stable ⇐⇒ t12 a(s)ds ≤ K < ∞, ∀ 0 ≤ t1 < t2 < ∞.
Rt
(d) x = 0 is uniformly asymptotically stable ⇐⇒ t0 a(s)ds ≤ −α(t − t0 ) + β. (α > 0)

2
Example 6. Considering x0 = a(t)x with
√
a(t) = sin(log t) + cos(log t) − α, 1<α< 2.

From Z t
a(s)ds = t sin(log t) − t0 sin(log t0 ) − α (t − t0 ) → −∞
t0
as t → ∞, we know x = 0 is asymptotically
√ stable. We will show that x = 0 is NOT uniformly
stable. For this, choose α < β < 2 and θ1 < π/4 < θ2 such that

sin θ + cos θ ≥ β > α θ ∈ [θ1 , θ2 ].

Set
tn = e2nπ+θ2 , t0,n = e2nπ+θ1 .
Then
Z tn Z e2nπ+θ2
a(s)ds = [sin(log t) + cos(log t) − α]dt (2nπ + θ = log t)
t0,n e2nπ+θ1
Z θ2
= [sin θ + cos θ − α]e2nπ+θ dθ
θ1
Z θ2
2nπ
≥ (β − α) e eθ dθ → ∞
θ1

as n → ∞.

Remarks.

(1) asymptotic stability 6⇒ uniform asymptotic stability (Example 1).

(2) asymptotic stability 6⇒ uniform stability (Example 6).

(3) uniform stability + asymptotic stability 6⇒ uniform asymptotic stability (Example 1).

3.2. Stability of linear systems

Consider a homogeneous linear system

x0 (t) = A(t) x(t) (H)

where A(t) is n × n matrix valued continuous function of t for t ≥ 0. All solutions have the same
stability characteristics as the equilibrium x = 0 (why?). We may thus speak of the stability of
the linear system (H) instead of that of a particular solution.

Theorem 3.1 Let X(t) be a f.m. of (H). Then (H) is:

(a) stable if and only if ∃ K > 0 such that

|X(t)| ≤ K, t ≥ 0.

3
 (b) uniformly stable if and only if ∃ K > 0 such that

|X(t) X −1 (s)| ≤ K, t ≥ s ≥ 0.

(c) asymptotically stable if and only if

|X(t)| → 0, as t → ∞.

(d) uniformly asymptotically stable if and only if ∃ K, α > 0 such that

|X(t) X −1 (s)| ≤ K e−α (t−s) , t ≥ s ≥ 0.

Proof:
(a). “ ⇐ ” : Since x(t) = X(t)X −1 (t0 )x0 , x0 = x(t0 ), we have

|x(t)| ≤ |X(t) |X −1 (t0 )| |x0 | ≤ K |X −1 (t0 )| |x0 | < ²

if |x0 | < δ = ²/(K |X −1 (t0 )|.

“ ⇒ ” : By definition, |x(t)| = |X(t)X −1 (t0 )x0 | < 1 for t ≥ t0 if |x0 | < δ(1, t0 ). Therefore
|X(t)X −1 (t0 )| < 1/δ(1, t0 ) It then follows that
1
|X(t)| ≤ |X(t)X −1 (t0 )X(t0 )| ≤ |X(t0 )|.
δ(1, t0 )

(b). “ ⇐ ” : Since x(t) = X(t)X −1 (t0 )x0 , the inequality

|X(t) X −1 (s)| ≤ K, t ≥ t0 ≥ 0

implies |x(t)| ≤ K |x0 | < ², if |x0 | < δ = ²/K.

“ ⇒ ” : By the definition of uniform stability,

|x(t)| = |X(t)X −1 (t0 )x0 | < 1, t ≥ t0 ≥ 0,

if |x0 | < δ(1). Thus |X(t)X −1 (t0 )| < 1/δ(1).

(c). |x(t)| = |X(t)X −1 (t0 )x0 | → 0 as t → ∞ if and only if |X(t)| → 0 as t → ∞.

(d). “ ⇐ ” : By (b), |X(t) X −1 (s)| ≤ K e−α (t−t0 ) , t ≥ t0 ≥ 0 implies uniform stability. Let
T (²) = − α1 log( K² ), 0 < ² < K. Then |x0 | < 1, t ≥ t0 + T (²) imply

|x(t)| = |X(t)X −1 (t0 )x0 | < K e−α (t−t0 ) |x0 | = K e−α T (²) |x0 | ≤ ².

This relation and the uniform stability imply that (H) is u.a.s.
“ ⇒ ” : Assume that (H) is u.a.s. Then ∃ δ0 > 0 such that ∀ δ0 > ² > 0 ∃ T (²) > 0 such that
|x0 | ≤ δ0 , t ≥ t0 + T (²) imply |x(t)| = |X(t)X −1 (t0 )x0 | < ².
Let θ = ²/δ0 . Then |x0 | ≤ δ0 , t ≥ t0 + T (²) imply |X(t)X −1 (t0 )x0 | ≤ ². Therefore

|X(t + T (²))X −1 (t)| ≤ θ < 1, t ≥ 0. (#)

4
Uniform stability implies

|X(t + h)X −1 (t)| < K, t ≥ 0, h ≥ 0. (##)

If t ≥ t0 , ∃ n such that t0 + nT (²) ≤ t ≤ t0 + (n + 1)T (²). Let tk = t0 + k T (²), k = 1, 2, · · · , n.

Then

|X(t)X −1 (t0 )| = |X(t)X −1 (tn ) X(tn )X −1 (tn−1 ) · · · X(t1 )X −1 (t0 )|

≤ |X(t)X −1 (tn )| Πn−1
k=0 |X(tk+1 )X
−1
(tk )|
≤ K θn

by (#), (##). Let α = − T 1(²) log θ. Then

K n+1 K −(n+1)αT (²)

|X(t)X −1 (t0 )| ≤ θ = e
θ θ
K −α(t−t0 )
≤ e = C e−α(t−t0 )
θ
¤

3.3. Concrete conditions for stability using Lozinskiı̆ measure

Recall that a vector norm | · | in Rn induces a matrix norm for n × n matrices

|Ax|
|A| = sup {|Ax| : |x| = 1} = sup
x6=0 |x|

which has the following properties

(1) |Ax| ≤ |A| |x|

(2) |A| ≥ 0, |A| = 0 ⇔ A = 0

(3) | |A| − |B| | ≤ |A ± B| ≤ |A| + |B|

(4) |AB| ≤ |A| |B|

Examples: x = (x1 , · · · , xn )T , A = (aij ).

Pn
l∞ : |x| = maxi |xi | |A| = maxi j=1 |aij |

pP √
|A| = ρ,
l2 : |x| = i |xi |2
ρ largest eigenvalue of A∗ A

Pn Pn
l1 : |x| = i=1 |xi | |A| = maxj i=1 |aij |

5
Definition. Let | · | denote a vector norm and the matrix norm it induces. The Lozinskiı̆ measure
(or the logarithmic norm) µ(A) of an n × n matrix A is defined as
|I + hA| − 1
µ(A) = lim .
h→0+ h
(see W.A. Coppel, “Stability and Asymptotic Behaviors of ODEs,” or A.G. Kartsatos, “Advanced
ODEs.”)

Lemma 3.2 µ(A) exists for any n × n matrix A.

Proof. Exercise. Hint: use the convexity of the norm.

Proposition 3.3 The Lozinskiı̆ measure has the following properties.

(a) µ(αA) = αµ(A), α≥0

(b) µ(A + B) ≤ µ(A) + µ(B)

(c) A(t) continuous implies µ(A(t)) continuous.

(d) If λ is an eigenvalue of A, then −µ(−A) ≤ Reλ ≤ µ(A).

Proof. Exercise.
· ¸
2 −1
An Example. Let A = . Then A has two eigenvalues λ1 = 1, λ2 = 2, which are
0 1
contained in the interval [1, 2]. We will try to use Property (d) to get an estimate on λ1 and λ2
using Lozinskiı̆ measures with respect to the three common vector norms.

l∞ : µ∞ (A) = 3, −µ∞ (−A) = 1

l1 : µ1 (A) = 2,
√
−µ1 (A) = 0√
l2 : µ2 (A) = 2+2 2 , −µ2 (−A) = 3−2 2 .

Therefore the intervals

√ √
that contain λ1 , λ2 obtained by the three Lozinskiı̆ measures are [−1, 3],
3− 2 3+ 2
[0, 2], and [ 2 , 2 ]. The intersection of all three intervals is [1, 2].

This properties of the Lozinskiı̆ measure is closely related to the Gersgorin discs that are used
in locating eigenvalues of a matrix A = (aij ), real or complex. For 1 ≤ k ≤ n, set
n
X
rk = |aij |.
j=1,j6=k

Gersgorin discs are given by

Dk = {λ ∈ C : |λ − Reakk | ≤ rk }, k = 1, 2, · · · , n.

The following is standard.

6
Theorem 3.4 All eigenvalues of A lie in ∪nk=1 Dk .

Lemma 3.5 Let x(t) be a solution of (H). Then

D+ |x(t)| ≤ µ(A(t)) |x(t)| (3)

Proof. The differentiability of x(t) implies that, for h > 0,

x(t + h) = x(t) + x0 (t) h + o(h) = x(t) + A(t)x(t) h + o(h)

= (I + hA(t))x(t) + o(h)

Thus
|x(t + h)| − |x(t)| ≤ [|I + hA(t)| − 1] |x(t)| + o(h)
which implies that
|x(t + h)| − |x(t)| |I + hA(t)| − 1
≤ |x(t)| + o(1)
h h
which leads to the claim of the lemma. ¤

Theorem 3.6 Every solution x(t) of (H) satisfies

Rt Rt
− µ(−A) µ(A)
|x(t0 )| e t0 ≤ |x(t) ≤ |x(t0 )| e t0 , t ≥ t0 . (4)

Rt
− µ(A(s))ds
Proof. Lemma3.5 implies that |x(t)|e t0 is decreasing and thus the RH inequality. To
prove the LH inequality, let s = −t, y(s) = x(t). Then

dy dx
=− = −A(t)x(t) = −A(−s)y(s)
ds dt
and apply the RH inequality.

Examples:
P
Pn µ(A) = maxi {Re aii + j6=i |aij |}
l∞ : |A| = maxi j=1 |aij | P
−µ(−A) = mini {Re aii − j6=i |aij |}

√ µ(A) = λ1 − µ(−A) = λn
|A| = ρ,
l2 : λ 1 ≥ λ2 ≥ · · · ≥ λ n
ρ largest eigenvalue of A∗ A
eigenvalues of (A∗ + A)/2

P
Pn µ(A) = maxj {Re ajj + i6=j |aij |}
l1 : |A| = maxj i=1 |aij |
P
−µ(−A) = minj {Re ajj − i6=j |aij |}

7
Proof. ( l∞ case) Since X
|I + hA| = max{|1 + haii + h |aij |},
i
j6=i

and
1
|1 + haii | = {(1 + haii )(1 + haii )} 2
= [1 + h(aii + aii + h2 |aii |2 ]
h
= 1 + (aii + aii ) + o(h)
2
= 1 + h Re aii + o(h),

we have
|I + hA| − 1 1 X
= max {1 + hRe aii + h |aij | + o(h) − 1}
h h i
j6=i
X
= max {Re aii + |aij | + o(1)}.
i
j6=i

l2 and l1 cases: Exercise. ¤

Theorem 3.7 The linear system (H) is

Rt
(1) unstable if lim inf t→∞ 0 µ(−A) = −∞
Rt
(2) stable if lim supt→∞ 0 µ(A) < +∞
Rt
(3) asym. stable if lim supt→∞ 0 µ(A) = −∞
Rt
(4) unif. stable if ∃ K > 0 such that t0 µ(A) ≤ K for all t ≥ t0 ≥ 0.

(5) unif. asym. stable if ∃ α, β, α > 0 such that

Z t
µ(A) ≤ −α (t − t0 ) + β, ∀ t ≥ t0 ≥ 0.
t0

Proof. Exercise.

3.4. Stability of perturbed linear systems

We consider the following question.

Question: Assume that the linear system

x0 (t) = A(t)x(t) (H)

is unif. stable (unif. asym. stable). Under what condition of B(t), does the perturbed linear
system
x0 (t) = (A(t) + B(t))x(t) (PH)

8
remain unif. stable (unif. asym. stable)?

We have the following results.

Theorem 3.8 Suppose that (H) is unif. stable (for t ≥ 0 ). If B(t) satisfies
Z ∞
|B(t)| dt < ∞,
0

then (P H) is unif. stable.

Theorem 3.9 Suppose that (H) is unif. asym. stable (for t ≥ 0 ). If there exist γ > 0, β such
that B(t) satisfies
Z t
|B(t)| dt < γ (t − t0 ) + β, t ≥ t0 ≥ 0,
t0

then there exists γ > 0 such that (P H) is unif. asym. stable for γ < γ.

The following example shows that the integrability condition for |B(t)| in Theorem3.8 is nec-
essary.

Example. The simple harmonic oscillator

u00 + u = 0

is uniformly stable. This is clear from the equivalent linear system

µ ¶0 µ ¶µ ¶
u 0 1 u
= .
v −1 0 v

A perturbed linear equation

u00 − 2t−1 u0 + u = 0
is equivalent to µ ¶0 µ ¶µ ¶
u 0 1 u
= −1 .
v −1 2t v
with a “small” perturbation µ ¶
0 0
B(t) = ,
0 2t−1
in the sense that |B(t)| = 2t−1 → 0 as t → ∞. However, the perturbed system has two linearly
independent solutions
µ ¶ µ ¶ µ ¶ µ ¶
u1 sin t − t cos t u2 cos t + t sin t
= , = ,
v1 t sin t v2 t cos t

which are both unbounded as t → ∞. Thus the perturbed system is unstable! Note that
Z ∞ Z ∞
|B(t)| dt = 2t−1 dt = ∞
0 0

9
and B(t) does not satisfies the condition of Theorem3.8.

Proof of Theorem3.8. Let X(t) be a f.m. of (H). Write (P H) as

x0 (t) = A(t)x(t) + B(t)x(t).

By the variation of constant formula, any solution x(t) of (P H) can be written as

Z t
−1
x(t) = X(t)X (t0 )x(t0 ) + X(t)X −1 (s)B(s)x(s) ds.
t0

Uniform stability of (H) implies ∃ K > 0 such that

|X(t)X −1 (s)| ≤ K, t ≥ s ≥ 0.

Thus
Z t
−1
|x(t)| ≤ |X(t)X (t0 )| |x(t0 )| + |X(t)X −1 (s)| |B(s)| |x(s)| ds
t0
Z t
≤ K|x(t0 )| + K |B(s)| |x(s)| ds.
t0

Gronwall’s lemma implies

Rt
K |B(s)|ds|
|x(t)| ≤ K |x(t0 )|e t0
R∞
≤ K eK 0 |B(s)|ds
|x(t0 )|,

which impels the unif. stability of (P H). ¤

Exercise. Prove Theorem3.9.

3.5. Stability of nonlinear systems

We can generalize the results in the previous section to stability of systems of form

x0 = A(t)x + F (t, x) (NPH)

where F (t, x) is a continuous nonlinear function. The following theorem can be proved in the same
way as Theorems 3.8, 3.9, using the variation of constant formula and the Gronwall lemma.

Theorem 3.10 Suppose

(a) (H) is unif. asym. stable, namely, a f.m. X(t) of (H) satisfies

|X(t)X −1 (t0 )| ≤ K e−α (t−s) , t≥s≥0

for K, α > 0.

(b) |F (t, x)| ≤ m |x|.

10
Then every solution x(t) of (N P H) satisfies

|x(t)| ≤ K |x(t0 )| eKm−α) (t−s) , t ≥ s ≥ 0. (5)

In particular, if 0 < Km < α, then the solution x = 0 of (N P H) is unif. asym. stable

(exponentially stable).

Proof. Exercise.

As an application of Theorem 3.10, consider a nonlinear autonomous system

x0 = f (x) (NL)
∂f
where f is C 1 in an open set D ⊂ Rn , and f (0) = 0. Let A = ∂x (0) be the Jacobian matrix of
f at x = 0. Rewrite (N L) as
x0 = Ax + F (x)
where F (x) = f (x)−Ax. Differentiability of f implies F (x) = o(|x|), as |x| → 0. Therefore F (x)
satisfies the condition (b) of Theorem3.10 with m can be chosen as small as we wish by restricting
x in a sufficiently small neighbourhood of 0. Applying Theorem3.10 to (N L), we arrived the
following result, which is often referred to as the method of linearization.

Theorem 3.11 Suppose

(a) f (x0 ) = 0.
∂f
(b) The eigenvalues λi of ∂x (x0 ) satisfy Reλi < 0.
Then the equilibrium x = x0 of (N L) is (uniformly) asym. stable.

Similarly, we have the following result on unif. stability.

Theorem 3.12 Suppose

(a) (H) is uniformly stable.
Rt
(b) |F (t, x)| ≤ γ(t) |x| and 0 |γ(t)|dt < ∞.
Then x = 0 of (N P H) is unif. stable.

Proof. Exercise.

A technical problem in the application of the method of linearization is to verify that all eigen-
values of the Jacobian matrix, ∂f
∂x (0), have negative real parts. A set of necessary and sufficient
conditions for this to hold are the well-known Routh-Hurwitz conditions (see Coppel’s book, or
Horn and Johnson, Matrix Analysis). For n = 2, 3, the Routh-Hurwitz conditions are,

n=2: tr(A) < 0, det(A) > 0;

n=3: tr(A) < 0, det(A) < 0, tr(A) a2 < det(A),

11
where a2 is the sum of 2 × 2 principal minors of A.

Remark. It is important to note that the method of linearization does not apply if x = 0 is a
center for the linearized system or ∂f
∂x (0) has an eigenvalue with zero real part (in this case, we
say x = 0 is not hyperbolic), as shown in the following example.

Example. Consider two nonlinear systems

p
x0 = −y ± x x2 + y 2
p
y 0 = x ± y x2 + y 2 ,

where the signs are either all plus or all minus. The linearization at (0, 0) for both systems is the
same
x0 = −y
y 0 = x,
and (0, 0) is a center for the linearized system. However, in polar coordinates, the nonlinear
systems can be written as
1
r0 = ± r 2
θ0 = 1.
Therefore, (0, 0) is a sink when the sign is minus, and a source when the sign is plus. Thus, a
center is easily destroyed by the higher order terms.

3.6. Lyapunov’s direct method – the method of Lyapunov functions

This is an important method of investigating stability properties, especially when the method of
linearization won’t yield any information, such as in the case of non-hyperbolic equilibria.

Let V ∈ C 1 (Rn+1 → R), V = V (t, x). Consider an ODE

x0 = f (t, x) (E)

with f (t, 0) = 0. Define

X ∂V n
∗ ∂V
V (t, x) = (t, x) + (t, x)fi (t, x) (6)
∂t ∂xi
i=1

Lemma 3.13 Suppose W ∈ C(Rn+1 → R) and

∗
V (t, x) ≤ W (t, x), ∀ (t, x) ∈ Rn+1 .

If x(t) is a solution of (E), then

Z t2
V (t2 , x(t2 )) − V (t1 , x(t1 )) ≤ W (s, x(s)) ds
t1

12
Proof. We can verify that
d ∗
V (t, x(t)) = V (t, x(t)) ≤ W (t, x(t)).
dt
The integration of the above relation leads to the claim of lemma. ¤
Regarding the stability of x = 0, we have the following results.

Theorem 3.14 Suppose there exists V (t, x) that is defined and C 1 for all t ≥ 0 and |x| ≤ δ0
such that

(a) V (t, 0) = 0.

(b) V (t, x) ≥ θ(|x|), θ ∈ C([0, ∞) → R) is increasing, θ(0) = 0.

∗
(c) V ≤ 0 for t ≥ 0, |x| ≤ δ0 .

Then the equilibrium x = 0 is stable.

Proof. Continuity of V and assumption (a) imply that, ∀0 < ² < δ0 , ∃ δ(t0 , ²) > 0 such that
|x0 | < δ(t0 , ²) implies V (t0 , x0 ) < θ(²). Lemma3.13 and assumption (c) imply

θ(|x(t)|) ≤ V (t, x(t)) ≤ V (t0 , x(t0 )) < θ(²), t ≥ t0 .

The monotonicity of θ then implies |x(t)| < ² for t ≥ t0 . (Why does the solution x(t) exist for
all t ≥ t0 ? ) ¤

Theorem 3.15 Suppose there exists V (t, x) that is defined and C 1 for all t ≥ 0 and |x| ≤ δ0
such that

(a) θ1 (|x|) ≤ V (t, x) ≤ θ2 (|x|), θi ∈ C([0, ∞) → R) are increasing, θi (0) = 0, i = 1, 2.

∗
(2) V ≤ 0 for t ≥ 0, |x| ≤ δ0 .

Then the equilibrium x = 0 is unif. stable.

Proof. The assumptions (a), (b) imply

θ1 (|x(t)|) ≤ V (t, x(t)) ≤ V (t0 , x(t0 )) ≤ θ2 (|x0 |).

Therefore |x(t)| ≤ θ1−1 (θ2 (|x0 |)). For ² > 0, choose δ(²) = θ2−1 (θ1 (²)). Then |x0 | ≤ δ(²) implies
|x(t)| < ², t ≥ t0 . ¤

Examples.

1. Consider u00 + a(t)u = 0. Assume a0 (t) ≤ 0 and limt→∞ a(t) = a∞ > 0. The equivalent system
is µ ¶0 µ ¶
x1 x2
= .
x2 −a(t)x1

13
Let V (t, x) = a(t)x21 + x22 . Then
∗
V = a0 (t)x21 + 2a(t)x1 x2 + 2x2 (−a(t)x1 ) ≤ 0.
We can choose
θ1 (|x|) = c(x21 + x22 ) = c|x|, c = min{a∞ , 1}
θ2 (|x|) = k(x21 + x22 ) = k|x|, k = max{a(0), 1}
Then θ1 (|x|) ≤ V (t, x) ≤ θ2 (|x|). Thus V satisfies all assumptions of Theorem3.15. Therefore
x = 0 is uniformly stable.

2. Consider u00 + g(u) = 0, g ∈ C 1 , ug(u) > 0 for 0 < |u| < δ0 . For the equivalent system
µ ¶0 µ ¶
x1 x2
= ,
x2 −g(x1 )
define Z x1
1
V (x1 , x2 ) = x22 + g(u)du.
2 0
∗
( V (x) represents the total energy of the nonlinear oscillator.) We can verify that V (x1 , x2 ) = 0.
Thus
V (x1 (t), x2 (t)) = V (x1 (0), x2 (0)), t ≥ 0.
Define
θ1 (r) = inf{V (x1 , x2 ) : x21 + x22 = r2 }
θ2 (r) = sup{V (x1 , x2 ) : x21 + x22 = r2 }
Then θ1 (|x|) ≤ V (t, x) ≤ θ2 (|x|). Therefore, (0, 0) is a uniformly stable equilibrium.

Theorem 3.16 Suppose there exists V (t, x) that is defined and C 1 for all t ≥ 0 and |x| ≤ δ0
such that
(a) V (t, 0) = 0.
(b) V (t, x) ≥ θ(|x|), θ ∈ C([0, ∞) → R) is increasing, θ(0) = 0.
∗
(c) V ≤ −ϕ(V (t, x)), for t ≥ 0, |x| ≤ δ0 . ϕ is continuous, increasing, ϕ(0) = 0.
Then the equilibrium x = 0 is asymptotically stable.

∗
Proof. V ≤ −ϕ(V (t, x)) ≤ 0 implies the V (t, x(t)) is nonincreasing, and assumptions (a), (b)
imply that x = 0 is stable by Theorem3.14. Moreover, the monotonicity and non-negativeness
of V (t, x(t)) imply limt→∞ V (t, x(t)) = V0 ≥ 0. We want to show V0 = 0. Suppose otherwise,
d
V0 > 0. Then dt V (t, x(t)) ≤ −ϕ(V (t, x(t))) ≤ −ϕ(V0 ) < 0. Thus
V (t, x(t)) − V (t0 , x0 ) ≤ −ϕ(V0 )(t − t0 ) → −∞,
as t → ∞. This contradicts V (t, x(t) ≥ 0. We thus have limt→∞ V (t, x(t)) = 0. Therefore
θ(|x(t)|) → 0 by (b), and thus |x(t)| → 0 as t → ∞, by the monotonicity of θ and θ(0) = 0. ¤

14
Theorem 3.17 Suppose there exists V (t, x) that is defined and C 1 for all t ≥ 0 and |x| ≤ δ
such that

(a) θ1 (|x|) ≤ V (t, x) ≤ θ2 (|x|), θi ∈ C([0, ∞) → R) are increasing, θi (0) = 0, i = 1, 2

∗
(b) V ≤ −ϕ(V (t, x)), for t ≥ 0, |x| ≤ δ. ϕ is continuous, increasing, ϕ(0) = 0.

Then the equilibrium x = 0 is uniformly asymptotically stable.

Proof. By Theorem3.15, x = 0 is uniformly stable. For ² > 0, there exists δ(²) > 0 such that
|x(t0 )| < δ(²) implies |x(t)| < ², t ≥ t0 . It remains to show that there exists δ0 > 0 such that,
if ² > 0, there exists T (²) > 0 such that |x(t0 )| ≤ δ0 , t > t0 + T (²) imply |x(t)| < ². Set
δ0 = δ(²), T (²) = θ2 (δ0 )/ϕ(θ1 (δ0 )).
Claim: there exists t1 ∈ [t0 , t0 + T (²)] such that |x(t1 )| < δ0 = δ(²). Under the claim, we have
|x(t)| < ² for t ≥ t1 , and thus for t ≥ t0 + T (²), and proof is complete.

To show the claim, we assume otherwise, i.e., |x(t)| ≥ δ0 for all t ∈ [t0 , t0 + T (²)]. This would
imply
d ∗
V (t, x(t)) = V (t, x(t)) ≤ −ϕ(V (t, x(t)) ≤ −ϕ(θ1 (|x(t)|))
dt
≤ −ϕ(θ1 (δ0 )).

Thus

V (t, x(t)) ≤ V (t0 , x(t0 )) − ϕ(θ1 (δ0 )) (t − t0 )

≤ θ2 (δ1 ) − ϕ(θ1 (δ0 )) (t − t0 ), t0 ≤ t ≤ t0 + T (²).

At t = t0 + T (²), this implies

V (t0 + T (²), x(t0 + T (²))) ≤ θ2 (δ0 ) − ϕ(θ1 (δ0 )) T (²) = 0

by our choice of T (²). But |x(t0 + T (²))| ≥ δ0 > 0 implies V (t0 + T (²), x(t0 + T (²))) > 0. This
contradiction establishes the claim, completing the proof. ¤

Theorem 3.18 (Chetaev Instability Criteria) Suppose there exists V (t, x) that is defined and
C 1 for all t ≥ 0 and |x| ≤ δ0 such that

(a) |V (t, x)| ≤ θ(|x|), θ ∈ C([0, ∞) → R) is increasing, θ(0) = 0.

(b) ∀δ ∈ (0, δ0 ) and t0 > 0, there exists x0 , |x0 | < δ such that V (t0 , x0 ) < 0.
∗
(c) V ≤ −ϕ(|V (t, x)|), for t ≥ 0, |x| ≤ δ0 . ϕ is continuous, increasing, ϕ(0) = 0.

Then the equilibrium x = 0 is unstable.

15
Proof. Suppose x = 0 is stable. Then, ∀² > 0, t0 ≥ 0, there exists δ(², t0 ) > 0 such that |x(t0 )| <
δ(², t0 ) implies |x(t)| < ², t ≥ t0 . Choose x(t0 ) such that |x(t0 )| < δ(², t0 ), V (t0 , x(t0 )) < 0. Then,

|V (t, x(t))| < θ(|x(t)|) < θ(²), t ≥ t0 . (7)

But (c) implies V (t, x(t)) ≤ V (t0 , x0 ) < 0, thus |V (t, x(t))| ≥ |V (t0 , x0 )| > 0. Moreover,

V (t, x(t)) ≤ V (t0 , x0 ) − ϕ(|V (t0 , x0 )|)(t − t0 ) → −∞

as t → ∞. This contradicts relation (7). Therefore x = 0 is unstable.

3.7. Autonomous systems

In this section we consider the special case of autonomous system

x0 = f (x) (8)

where f ∈ C 1 (Rn → Rn ), f (0) = 0. Let V ∈ C 1 (U → R), U ⊂ Rn a neighbourhood of 0.

Define
∗
V (x) = gradV (x) · f (x).
Since uniform (asymptotic) stability is equivalent to (asymptotic) stability for autonomous systems,
we have the following three cases:
Case I. )
V (0) = 0, V (x) > 0, x 6= 0
∗ ⇒ x = 0 is stable.
V (x) ≤ 0, x 6= 0
Case II. )
V (0) = 0, V (x) > 0, x 6= 0
∗ ⇒ x = 0 is asym. stable.
V (x) < 0, x 6= 0
Case III. 
V (0) = 0, V (x) not identically 0, 

∃ xn → 0 such that V (xn ) < 0, ⇒ x = 0 is unstable.
∗ 

V (x) < 0, x 6= 0

Examples.

1. Consider
dx
dt = −x + 2y 2
dy (9)
dt = −2y + x3 y.
Show that (0, 0) is asym. stable.

Method I: Linearize the system at (0, 0), we have the Jacobian matrix diag(−1, −2), which has
two negative eigenvalues. Thus, (0, 0) is asymptotically stable for the linearized system, and for
the nonlinear system by our perturbation theorem (Theorem3.11.

16
Method II: Let V (x, y) = x2 + y 2 . Then V is positive definite, and
∗
V (x, y) = 2x(−x + 2y 2 ) + 2y(−2y + x3 y)
= −2x2 − 2y 2 (2 − 2x − 2x3 )
1
< −2(x2 + y 2 ), if x < .
2
Therefore (0, 0) is asym. stable.

2. Consider
dx
dt = 2xy + x3
dy (10)
dt = −x2 + y 5 .
Show that (0, 0) is unstable.

Remark. The method of linearization won’t work here since there are no linear terms at (0, 0).

Consider V (x, y) = −x2 − 2y 2 . Then

∗
V (x, y) = −2x(2xy + x3 ) − 4y(−x2 + y 5 )
= −2x4 − 4y 6 < 0, (x, y) 6= (0, 0).

Thus (0, 0) is unstable.

3. Consider
dx
dt = −2x3 − y 5
dy (11)
dt = x5 − 2y 3 .
Show that (0, 0) is asymptotically stable.

Set V (x, y) = x2 + y 2 . Then

∗
V (x, y) = 2x(−2x3 − y 5 ) + 2y(x5 − 2y 3 )
= −4(x4 + y 4 ) + 2xy(x4 − y 4 ) ≤ −4(x4 + y 4 ) + 2|xy|(x4 + y 4 )
≤ −2(x4 + y 4 ) < 0, if |xy| < 1, (x, y) 6= (0, 0).

Thus (0, 0) is asym. stable.

4. Consider
dx
dt = y − x2
dy (12)
dt = x − y2.
Find all its equilibria and discuss their stability.

Solve the equilibrial equation

0 = y − x2
0 = x − y2,
we obtain two equilibria of (12): (0, 0) and (1, 1).

17
 ¶ µ
0 1
Method I: By linearization. The Jacobian matrix of (12) at (0, 0) is , which has two
1 0
eigenvalues λ1 = 1, λ2 = −1. Therefore, (0, 0) is unstable (a saddle point) for the linearized
system, and for the nonlinear system, by a modification of Theorem 3.11.

Set u = x − 1, v = y − 1. We have
du
dt = −2u + v − u2
dv (13)
dt = u − 2v − v 2 ,

in which the equilibrium (u, v) = (0, 0) µcorrespond

¶ to (x, y) = (1, 1). The linearization of (13)
−2 1
at (0, 0) leads to the Jacobian matrix , which has eigenvalues λ1 = −1, λ2 = −3.
1 −2
Therefore (0, 0) is asym. stable for (13) and (1, 1) is asym. stable for (12).

Method II:

For (x, y) = (0, 0), set V (x, y) = −xy. Then

∗
V (x, y) = −x(x − y 2 ) − y(y − x2 )
= −(1 − y)x2 − (1 − x)y 2 < 0, if |x| < 1, |y| < 1, (x, y) 6= (0, 0).

Therefore (0, 0) is unstable for (12).

For (x, y) = (1, 1), consider (u, v) = (0, 0) for system (13). Set V (u, v) = u2 + v 2 . Then
∗
V (u, v) = 2u(−2u + v − u2 ) + 2v(u − 2v − v 2 )
= −4(u2 + v 2 ) + 4uv − 2(u3 + v 3 ≤ −2(u2 + v 2 ) − 2(u3 + v 3 ) (2uv ≤ u2 + v 2 )
= −(2 + u)u2 − (2 + v)v 2 < 0, if |u| < 2, |v| < 2, (u, v) 6= (0, 0).

Therefore (0, 0) is asym. stable for (13) and (1, 1) is asym. stable for (12).

18