EE547: Fall 2018

Lecture 10: Stability

Lecturer: L.J. Ratliff

Disclaimer: These notes have not been subjected to the usual scrutiny reserved for formal publications.
They may be distributed outside this class only with the permission of the Instructor.

Reference: C& D chapter 7; Hespanha chapter 8

10.0.1 Stability

x(t) = Φ(t, t0 )x0

where Φ(t, t0 ) is the state transition matrix. Note that

x0 = θn =⇒ x(t) = θn ∀t

xe = θn is called the equilibrium point.

Definition 0.1 (Stable Eq.) xe = θn is stable ⇐⇒ ∀ x0 ∈ Rn , ∀t0 ∈ Rn the map t → x(t) = Φ(t, t0 )x0
is bounded for all t ≥ t0 .

Definition 0.2 (Asymptotic) xe = θn is asymptotically stable ⇐⇒ x0 = θn is stable and t → x(t) =

Φ(t, t0 )x0 tends 0 as t → ∞.

Definition 0.3 (Exponential ) xe = θn is exponentially stable ⇐⇒ ∃M, α > 0 such that

kx(t0 )k ≤ M exp(−α(t − t0 ))kx0 k

Theorem 0.1 (necessary and sufficient conditions) x0 = 0 is asymptotically stable ⇐⇒ Φ(t, 0) → 0

as t → ∞

Proof. (⇐=)
x(t) = Φ(t, t0 )x0 = Φ(t, 0)Φ(0, t0 )x0
since Φ(t, 0) → 0 as t → ∞ then kΦ(t, 0)k → 0 as t → ∞ and

kx(t)k ≤ kΦ(t, 0)kkΦ(0, t0 )kkx0 k

thus x(t) → 0 as t → ∞.
(=⇒) By contradiction; assume that t → Φ(t, 0) does not tend to zero as t → ∞, i.e. ∃ i, j such that

Φij (t, 0) 6→ 0 as t → ∞

choose  
0
· · ·
 
0
 
1
x0 =  
0
 
· · ·
0

10-1
10-2 Lecture 10: Stability

with 1 in the j–th spot. Thus,

xi (t) = Φij (t, 0) 6→ 0 as t → ∞
contradicting the asymptotic stability of 0.

10.0.2 Two stability results

From the above derivation of functions of matrices, we can show that

k −1
σ mX
X
exp(tA) = t` exp(λk t)pk` (A)
k=1 `=0

This gives rise to the following stability condition:

Proposition 0.1 Consider the differential equation ẋ = Ax, x(0) = x0 . From the above expression:

{exp(At) → 0 as t → ∞} ⇐⇒ {∀λk ∈ σ(A), Re(λk ) < 0}

and  
∀λk ∈ σ(A), Re(λk ) < 0 &
{t 7→ exp(At) is bounded on R+ } ⇐⇒
mk = 1 when Re(λk ) = 0

We have a similar situation for discrete time systems:

k −1
σ mX
X
∀ ν ∈ N, Aν = ν(ν − 1) · · · (ν − ` + 1)λν−`
k pk` (A)
k=1 `=1

The above gives rise to the following stability condition:

Proposition 0.2 Consider the discrete time system x(k + 1) = Ax(k), k ∈ N, with x(0) = x0 . Then for
k ∈ N, x(k) = Ak x0 . From the above equation, we have that
 k
A → 0 as k → ∞ ⇐⇒ {∀ λi ∈ σ(A), |λi | < 1}

and  
∀λi ∈ σ(A), |λi | ≤ 1 &
k → Ak is bounded on N+ ⇐⇒

mi = 1 when |λi | = 1

10.0.3 Application to numerical integration of ẋ = Ax

Suppose we are given ẋ = Ax, x(0) = x0 , A ∈ Cn×n , x ∈ Cn . Call t 7→ x(t) the exact solution
x(t) = exp(At)x0 . Note that t 7→ x(t) is analytic in t. Call (ξ0 , ξ1 , . . .) the sequence of computed val-
ues.

Forward Euler Method.

For small h > 0, we have for any tk ∈ R+ ,
x(tk + h) = x(tk ) + hẋ(tk ) + O(h2 )
= x(tk ) + hAx(tk ) + O(h2 )
In other words, we have approximately
x(tk + h) ' (I + hA)x(tk )
Lecture 10: Stability 10-3

So if we perform repeatedly this step starting at t0 = 0, we have the computed sequence {ξi }∞
0 by

ξm = (I + hA)m x0 , m = 0, 1, 2, . . .

From the spectral mapping theorem and the above equation for ξm we have the following.

Example. Consider
ẋ(t) = λx(t)
with λ ∈ C. Then the equation is stable if Re(λ) ≤ 0. In this case the system is exponentially decaying

lim x(t) = 0
t→∞

When is the numerically solution xi also decaying, limi→∞ xi = 0?

xi+1 = xi + hλxi ⇐⇒ xi+1 = (1 + hλ)i+1 x0

The solution is decaying (stable) if |1 + hλ| ≤ 1

Fact. Suppose σ(A) ⊂ C◦− (equiv., the origin is exponentially stable). Let h0 be the largest positive h such
that
max |1 + hλi | = 1
i
Under these conditions,
1. {ξm }∞
0 → 0 exponentially for all ξ0 iff h ∈ (0, h0 ).
2. if h > h0 , then for almost all x0 , the sequence of computed values {ξk }∞ ∞
0 is such that {kξm }0 grows
exponentially.
Interpretation. Thus even if σ(A) ⊂ C◦− (and hence the exact solution x(t) → 0 exponentially), for h > h0 ,
for almost all x0 , the sequence of computed vectors {kξm k}∞
0 blows up. It is for this reason that in practice
we often prefer the backward Euler method.

Backward Euler:
For small h > 0, we have that for any tk ∈ R,

x(tk ) = x(tk + h) − hẋ(tk + h) + O(h2 )

= x(tk + h) − hAx(tk + h) + O(h2 )
10-4 Lecture 10: Stability

Thus we have approximately

x(tk + h) ' (I − hA)−1 x(tk )
So if we perform repeatedly this step, starting from t0 = 0, we get the computed sequence {ξi }∞
0 given by

ξm = (I − hA)−m x0 , m = 0, 1, 2, . . .

Now, the spectrum of (I − hA)−1 is {(1 − hλi )−1 }σi=1 . Hence by the above expression for ξm , we have

ξm → 0 as m → ∞
⇐⇒ ∀λi ∈ σ(A), |(1 − hλi )−1 | < 1
⇐⇒ ∀λi ∈ σ(A), |1 − hλi | > 1

Note that if Re(λi ) < 0, then |1 − hλi | > 1, since h > 0. Thus we have shown the following result.

Example. Consider
ẋ(t) = λx(t)
with λ ∈ C. Then the equation is stable if Re(λ) ≤ 0. In this case the system is exponentially decaying

lim x(t) = 0
t→∞

When is the numerically solution xi also decaying, limi→∞ xi = 0?

 i+1
1
xi+1 = xi + hλxi+1 ⇐⇒ xi+1 = x0
1 − hλ

The solution is decaying (stable) if |1 + hλ| ≥ 1

Fact. If σ(A) ⊂ C◦− , then for all h > 0, for all x0 ∈ Cn , the computed sequence {ξm }∞
0 obtained via
backward Euler goes to zero exponentially.

This is very important in practice, because if h is unfortunately chosen too large the computed sequence
may lose accuracy but at least it will never blow up!

10.0.4 LTI

remark: for the time-invariant case, asymptotic stability is equivalent to exponential stability.
Lecture 10: Stability 10-5

Theorem 0.2 (exponential stability ) the system ẋ = Ax is exponentially stable iff all of the eigen-
values of A are in the open left half plane.

Proof. Follows from properties of matrix exponential. (proof pg 185 of C& D; proof not provided in hespanha)

Theorem 0.3 (Theorem 8.1 Hespanha) The system ẋ = Ax is

1. marginally stable if and only if all the eigenvalues of A have negative or zero real parts and all the
Jordan blocks corresponding to eigenvalues with zero real parts are 1 × 1
2. asymptotically stable if and only if all the eigenvalues of A have strictly negative real parts
3. exponentially stable if and only if all the eigenvalues of A have strictly negative real parts, or
4. unstable if and only if at least one eigenvalue of A has a positive real part or zero real part, but the
corresponding Jordan block is larger than 1 × 1.

Note: When all the eigenvalues of A have strictly negative real parts, all entries of eAt converge to zero
exponentially fast, and therefore eAt converges to zero exponentially fast (for every matrix norm); i.e., there
exist constants c, λ > 0 such that
keAt k ≤ ce−λt , ∀t ∈ R
In this case, for a sub-multiplicative norm, we have

kx(t)k = keA(t−t0 ) x0 k ≤ keA(t−t0 ) kkx0 k ≤ ce−λ(t−t0 ) kxk0 , ∀t ∈ R

This means asymptotic stability is equivalent to exponential stability for LTI systems.