EE547: Fall 2018

Lecture 11: Lyapunov 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 8/9

11.1 Two Notions of Stability

Remark. Norms are important for stability. So, reach back to the induced matrix norm definitions we
discussed in the linear algebra review. Here is where they will come in handy.

For some context, there are two essential notions of stability for linear system: (a) Bounded Input/Output
(BIBO) Stability and (b) Lyapunov Stability. The first may be one that you are familiar with and the
second may be a new concept. In essence, the follow is what these two notions mean:
BIBO stability. This notion of stability concerns the input output relationship of a system. That is, given
a system such that the output y is related to the input u through some map say G—i.e. y = G(u)—then
this system is BIBO stable if
kuk ≤ Ku =⇒ kyk ≤ Ky

For example, for a linear time varying system

ẋ = A(t)x + B(t)u
y = C(t)x + D(t)u

the forced response—i.e. the response when the initial condition x(t) is zero—is given by
Z t
yf (t) = C(t)Φ(t, τ )B(τ )u(τ ) dτ + D(t)u(t)
0

Hence, formally we define BIBO as follows.

Definition 1.1 (BIBO Stability ) The continuous LTV system is said to be (uniformly) BIBO stable
if there exists a finite constant K such that, for every input u(·), its forced response yf (·) satisfies

sup kyf (t)k ≤ K sup ku(t)k

t∈[0,∞) t∈[0,∞)

Lyapunov stability. This notion of stability is characterized only in terms of the state dynamics. That
is, the matrices B(·), C(·), and D(·) play no role in this definition; only A(·) matters because this
matrix completely defines the state transition matrix Φ(t, t0 ). Therefore one often simply talks about the
Lyapunov stability of the homogeneous system

ẋ(t) = A(t)x(t)

Hence, formally we define Lyapunov stability as follows.

11-1
11-2 Lecture 11: Lyapunov Stability

Definition 1.2 (Lyapunov stability ) The continuous LTV system is said to be

1. (marginally) stable in the sense of Lyapunov or internally stable if, for every initial condition x(t0 ) =
x0 , the homogeneous state response

x(t) = Φ(t, t0 )x0 , ∀ t ≥ 0

is uniformly bounded,
2. asymptotically stable (in the sense of Lyapunov) if, in addition, for every initial condition x(t0 ) = x0 ,
we have
x(t) → 0, as t → ∞
3. exponentially stable if, in addition, there exists constants c, λ > 0 such that, for every initial condition
x(t0 ) = x0 , we have
kx(t)k ≤ ceλ(t−t0 ) kx(t0 )k, ∀ t ≥ 0

4. unstable if it is not marginally stable in the sense of Lyapunov.

Chapter 9 of Hespanha does a great job of comparing the two notions. One important comparison is that a
system that is exponentially stable (in the sense of Lyapunov) is always BIBO stable, however the converse
is not true. Indeed, for a linear time invariant (LTI) system this happens when the premultiplication of eAt
by C and/or the postmultiplication by B cancel terms in eAt that are not converging to zero exponentially
fast.

Example. Consider the system

   
1 0 1  
ẋ = x+ u, y = 1 1 x
0 −2 0

The state transition matrix is  t 

At e 0
e =
0 e−2t
is unbounded and thus not Lyapunov unstable, but
 et
  
0 0
CeAt B = 1 1 = e−2t

0 e−2t 1

so that the system is BIBO stable.

11.2 Lyapunov Stability LTI Systems

Consider an LTI system

ẋ = Ax, x(t0 ) = x0

The results in previous lectures about matrix exponentials provides us with simple conditions to classify
the continuous-time homogeneous LTI system—since the state transition matrix is eA(t−t0 ) —in terms of its
Lyapunov stability, without explicitly computing the solution to the system.
The following result (see, e.g., Theorem 8.1 Hespanha) summarizes how to characterize Lyapunov stability
in terms of the eigenvalues of A.

Theorem 2.1 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
Lecture 11: Lyapunov Stability 11-3

2. asymptotically stable if and only if all the eigenvalues of A have strictly negative real parts (i.e. in open
left half plane),
3. exponentially stable if and only if all the eigenvalues of A have strictly negative real parts (i.e. in open
left half plane), 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 implies the following fact.

Fact. For the time-invariant case, asymptotic stability is equivalent to exponential stability.

The proof of the above results can be found in C&D page 185.

Remark. These conditions do not generalize to time-varying systems, even if the eigenvalues of A(t)
do not depend on t. One can find matrix-valued signals A(t) that are stability matrices for every fixed
t ≥ 0, but the time-varying system ẋ = A(t)x is not even stable.

11.2.1 Lyapunov Equation for Characterizing Stability

Recall the definition of a positive definite matrix.

Definition 2.1 (Positive Definite Matrix ) A symmetric matrix Q ∈ Rn×n is positive definite if

xT Qx ≥ 0, ∀ x ∈ Rn /{0}

The Lyapunov stability theorem provides an alternative condition to check whether or not the continuous-
time homogeneous LTI system
ẋ = Ax (11.1)
is asymptotically stable.

Theorem 2.2 The following five conditions are equivalent:

1. The system (11.1) is asymptotically (equiv. exponentially) stable
2. All the eigenvalues of A have strictly negative real parts
3. For every symmetric positive definite matrix Q, there exists a unique solution P to the following
Lyapunov equation
AT P + P A = −Q
Moreover, P is symmetric and positive-definite.
4. There exists a symmetric positive-definite matrix P for which the following Lyapunov matrix inequality
holds:
A∗ P + P A < 0

Proof. Proof Sketch. We have already seen that 1 and 2 are equivalent. To show that 2 implies 3, we need
 11-4 Lecture 11: Lyapunov Stability

to show that Z ∞
T
P = eA t QeAt dt
0
is the unique solution to AT P + P A = −Q. This can be done by showing that i) the integral is well-defined
(i.e. finite), ii) P as defined solves AT P + P A = −Q, iii) P as defined is symmetric and positive definite,
and lastly, iv) no other matrix solves the equation. Indeed,
i) This follows from exponential stability—i.e.
T
keA t QeAt k → 0
exponentially fast as t → ∞. Hence, the integral is absolutely convergent.
ii) We simply need to compute by direct verification
Z ∞
T T
T
A P + PA = AT eA t QeAt + eA t QeAt A dt
0
But,
d  AT t At  T T
e Qe = AT eA t QeAt + eA t QeAt A
dt
so that
Z ∞
T d  AT t At 
A P + PA = e Qe dt
dt
0 T  ∞
= eA t QeAt 

t=0
 T
 T
= lim eA t QeAt − eA 0 QeA0
t→∞

And, the right-hand side is equal to −Q since limt→∞ eAt = 0 (by asymptotic stability) and eA0 = I.
iii) this follows by direct computation. Symmetry easily follows:
Z ∞ Z ∞  T T Z ∞
T
T T
PT = (eA t QeAt )T dt = eAt QT eA t dt = eA t QeAt dt = P
0 0 0

To check positive definiteness, pick an arbitrary vector z and compute:

Z ∞ Z ∞
T T AT t At
z Pz = z e Qe z dt = w(t)T Qw(t) dt
0 0

where w(t) = e z. Since Q is positive definite, we get that z T P z ≥ 0. Moreover

At
Z ∞
T
z P z = 0 =⇒ w(t)T Qw(t) dt = 0
0

which only happens if w(t) = e z = 0 for all t ≥ 0, from which one concludes that z = 0, because eAt is
At

non-singular (recall all state transition matrices are!). Thus P is positive definite.
iv) prove by contradiction. assume there is some P̄ that solves it:
AT P + P A = −Q, and AT P̄ + P̄ A = −Q
Then
AT (P − P̄ ) + (P − P̄ )A = 0
T
Multiplying by eA t
and eAt on the left and right, respectively, we conclude that
T T
eA t AT (P − P̄ )eAt + eA t (P − P̄ )AeAt = 0, ∀ t ≥ 0
Yet,
d  AT t  T T
e (P − P̄ )eAt = eA t AT (P − P̄ )eAt + eA t (P − P̄ )AeAt = 0
dt
AT t
implying that e (P − P̄ )eAt must be constant. But,because of stability, this quantity must converge to
zero as t → ∞, so it must be always zero. Since eAt is nonsingular, this is possible only if P = P̄ .
Lecture 11: Lyapunov Stability 11-5

The implication that condition 3 =⇒ condition 4 follows immediately, because if we select Q = −I in

condition 3, then the matrix P that solves the Lyapunov equation also satisfies AT P + P A < 0.
Now to complete the proof we need to show that condition 4 implies condition 2. Let P be a symmetric
positive-definite matrix for which
AT P + P A < 0 (11.2)
holds and define
Q = −(AT P + P A)
Consider an arbitrary solution to the LTI system and define the scalar time-dependent map

v(t) = xT (t)P x(t) ≥ 0

Taking derivatives, we have

v̇ = ẋT P x + xT P ẋ = xT (AT P + P A)x = −xT Qx ≤ 0

Thus, v(t) is nonincreasing and we conclude that

v(t) = xT (t)P x(t) ≤ v(0) = xT (0)P x(0)

But since v = xT P x ≥ λmin (P )kxk2 we have that

xT (t)P x(t) v(t) v(0)

kxk2 ≤ = ≤
λmin (P ) λmin (P ) λmin (P )
which means that the system is stable. To verify that it is actually exponentially stable, we go back to the
derivative of v, and using the facts that xT Qx ≥ λmin (Q)kxk2 and v = xT P x ≤ λmax (P )kxk2 , we get that

λmin (Q)
v̇ = −xT Qx ≤ −λmin (Q)kxk2 ≤ − v, ∀ t ≥ 0
λmax (P )
Applying the fact that for some constant µ ∈ R,

v̇ ≤ µv(t), ∀t ≥ t0 =⇒ v(t) ≤ eµ(t−t0 ) v(t0 ), ∀t ≥ t0

we get that
λmin (Q)
v(t) ≤ e−λ(t−t0 ) v(t0 ), ∀t ≥ 0, λ = −
λmax (P )
which shows that v(t) converges to zero exponentially fast and so does kx(t)k.

There are discrete time versions of the above results. Indeed, consider

xk+1 = Axk

then we have a similar theorem.

Theorem 2.3 The following four conditions are equivalent:

1. The DT LTI system is asymptotically (exponentially) stable.
2. All the eigenvalues of A have magnitude strictly smaller than 1.
3. For every symmetric positive definite Q, there exists a unique solution P to the following Stein equation
(aka the discrete-time Lyapunov equation):

AT P A − P = −Q

Moreover, P is symmetric and positive-definite.

4. There exists a symmetric positive-definite matrix P for which the following Lyapunov matrix inequality
11-6 Lecture 11: Lyapunov Stability

holds:
AT P A − P < 0