Stability: Dr. Wenjie Dong

Stability
Dr. Wenjie Dong
Department of Electrical and Computer Engineering,

The University of Texas Rio Grande Valley
W. Dong (UTRGV, Dept. of ECE) 1 / 84

Input-Output Stability of LTI Systems
Consider a SISO linear time-invariant (LTI) causal system described by

Z t Z t
y (t) = g (t − τ )u(τ )dτ = g (τ )u(t − τ )dτ (1)
0 0
where g (t) is the impulse response. The system must be initially relaxed
at t = 0.
Definition 1
A system is said to be BIBO stable (bounded-input bounded-output
stable) if every bounded input excites a bounded output.
Remark: This stability is defined for the zero-state response and is

applicable only if the system is initially relaxed.

Theorem 2
A SISO system described by (1) is BIBO stable if and only if g (t) is
absolutely integrable in [0, ∞), or
Z t
|g (t)|dt ≤ M < ∞
0
for some constant M.

Proof: ”If”: If g (t) is absolutely integrable,
Z t Z t

|y (t)| = g (τ )u(t − τ )dτ ≤ |g (τ )| · |u(t − τ )|dτ
0 0
Z ∞
≤ um |g (τ )|dτ ≤ um M
0
where |u(t)| ≤ um < ∞ for all t ≥ 0.

”only if”: Prove by contradiction. We assume g (t) is not absolutely
stable. Then for any arbitrarily large N, there exists a t1 such that
Z t1
|g (τ )|dτ ≥ N
0
We choose the input

1 if g (τ ) ≥ 0
u(t1 − τ ) =
−1 if g (τ ) < 0

then the output

Z t1 Z t1
y (t1 ) = g (τ )u(t − τ )dτ = |g (τ )|dτ ≥ N
0 0
because y (t1 ) can be arbitrarily large, we conclude that a similar bounded

input can excited an unbounded output.

Theorem 3
If a system with impulse response g (t) is BIBO stable, then, as t → ∞:
1. The output excited by u(t) = a, for t ≥ 0, approaches ĝ (0)a.
2. The output excited by u(t) = sin ωt, for t ≥ 0, approaches
|ĝ (jω0 )| sin(ω0 t + ∠ĝ (jω0 ))
where ĝ (s) is the Laplace transform of g (t) or

Z ∞
ĝ (s) = g (τ )e −sτ dτ (2)
0

Proof: 1. If u(t) = a for all t ≥ 0, then

Z t Z t
y (t) = g (τ )u(t − τ )dτ = a g (τ )dτ
0 0
which implies
Z ∞
lim y (t) = a g (τ )dτ = aĝ (0) as t → ∞
t→∞ 0
where we have used (2) with s = 0.

2. If u(t) = sin ωt, then

Z t
y (t) = g (τ ) sin ω0 (t − τ )dτ
0
Z t
= g (τ )[sin ω0 t cos ω0 τ − cos ω0 t sin ω0 τ ]dτ
0
Z t Z t
= sin ω0 t g (τ ) cos ω0 τ dτ − cos ω0 t g (τ ) sin ω0 τ dτ
Z0 ∞ Z0 ∞
lim y (t) = sin ω0 t g (τ ) cos ω0 τ dτ − cos ω0 t g (τ ) sin ω0 τ dτ
t→∞ 0 0
= sin ω0 t Re[ĝ (jω)] − cos ω0 t Im[ĝ (jω)]
= |ĝ (jω0 )| sin(ω0 t + ∠ĝ (jω0 ))

Theorem 4
A SISO system with proper rational transfer function ĝ (s) is BIBO stable
if and only if every pole of ĝ (s) has a negative real part or, equivalently,
lies inside the left-half s-plane.

Proof: If ĝ (s) has pole pi with multiplicity mi , then its partial fraction
expansion contains the factors
1 1 1
, ,...,
s − pi (s − pi )2 (s − pi )mi
Thus the inverse Laplace transform of ĝ (s) or the impulse response

contains the factors
e pi t , te pi t , . . . , t mi −1 e pi t
It is straightforward to verify that every such term is absolutely integrable

if and only if pi has a negative real part. Using this fact, we can establish
Theorem 4.

Example
For a a system with the transfer function

1
G (s) =
(s + 2)(s + 3)
is it BIBO stable or not?

Example
Solution: The impulse response of the system is

1
g (t) = L−1 [G (s)] = L−1 [ ]
(s + 2)(s + 3)
1 1
= L−1 [ − ] = e −2t − e −3t
s +2 s +3
It is obvious that g (t) is absolutely integrable. So, the system is BIBO
stable.
Also, G(s) has two poles: −2 and −3. They are negative.

Example
Consider the positive feedback system shown in Fig. 1(a). Is it BIBO

stable or not?
Figure 1:

Example
Solution: For the positive feedback, it is a SISO system. Let r (t) be the
input of the feedback system. If r (t) = δ(t), then the output is
y (t) = aδ(t − 1) + ay (t − 1)
Since
ay (t − 1) = a2 δ(t − 2) + a2 y (t − 2)
a2 y (t − 2) = a3 δ(t − 3) + a3 y (t − 3)
..
.
then
∞
X
y (t) = aδ(t − 1) + a2 δ(t − 2) + · · · + = ai δ(t − i)
i=1
So, the impulse response of the feedback system is
∞
X
2 3
gf (t) = aδ(t − 1) + a δ(t − 2) + a δ(t − 3) + · · · = ai δ(t − i) (3)
i=1
Example
So,
∞
X
|g (t)| = |a|i δ(t − i)
i=1
and
∞
(
Z ∞ X ∞ if |a| > 1
i
|g (t)|dt = |a| = |a|
0 1−|a| <∞ if |a| < 1
i=1
We conclude that the system is BIBO stable if and only if |a| < 1.
The transfer function of the system is
ae −s
ĝ (s) =
1 − ae −s
It is an irrational function of s and Theorem 4 is not applicable.

Input-output Stability: MIMO Systems
For a LTI MIMO systems defined by

Z t
y (t) = G (t − τ )u(τ )dτ
0
where G (t − τ ) is the impulse matrix, we have the following results.

Theorem 5
A multivariable system with impulse response matrix G (t) = [gij (t)] is
BIBO stable if and only if every gij (t) is absolutely integrable in [0, ∞).
Theorem 6
A multivariable system with proper rational transfer matrix Ĝ (s) = [ĝij (s)]
is BIBO stable if and only if every pole of every ĝij (s) has a negative real
part.

Input-output Stability: MIMO Systems
Consider
ẋ = Ax + Bu (4)
y = Cx + Du (5)
Its transfer matrix is

1
Ĝ (s) = C (sI − A)−1 B + D = C [Adj(sI − A)]B + D
det(sI − A)
Since every pole of Ĝ (s) is an eigenvalue of A, Ĝ (s) is BIBO stable if

every eigenvalue of A has a negative real part.
: Remark: BIBO stability of (4)-(5) does not guarantee that the
eigenvalues of A are in the LHP.

Example
Consider the network shown in Fig. 2(b). Determine its BIBO stability.
Figure 2:

Example
Solution: Choose the the voltage over the capacitor to be x. V1 and V2
are the voltages of other two nodes. The ground is the middle point of the
right branch. By the nodal analysis, we have
V1 V1 − x
+ =u
1 1
x − V1 x − V2
+ + (−1)ẋ = 0
1 1
x − V2 0 − V2
+ =u
1 1
Substitute V1 and V2 into the second equation, we have
ẋ = x
The output is
y = V1 = 0.5x + 0.5u
The A-matrix is 1 and its eigenvalue is 1. It has a positive real part.
Example
The transfer function of the equation is
ĝ (s) = 0.5(s − 1)−1 × 0 + 0.5 = 0.5
The transfer function equals 0.5. It has no pole and no condition to meet.
Thus the system is BIBO stable even though it has an eigenvalue with a
positive real part.
Remark: BIBO stability does not say anything about the zero-input
response, which will be discussed later.

Input-output Stability: Discrete-time Case
Consider a discrete-time SISO Linear time-invariant causal system

described by
k
X k
X
y [k] = g [k − m]u[m] = g [m]u[k − m] (6)
m=0 m=0
where g [k] is the impulse response sequence of the system. This system
must be initially relaxed at k = 0.
Definition 7
A discrete-time system is said to be BIBO stable (bounded-input
bounded-output stable) if every bounded input sequence excites a
bounded-output sequence.
Remark:This stability is defined for the zero-state response and is

applicable only if the system is initially relaxed.
Discrete-time Case
Theorem 8
A discrete-time SISO system described by (6) is BIBO stable if and only if
g [k] is absolutely summable in [0, ∞) or
∞
X
|g [k]| ≤ M < ∞
k=0
for some constant M.

Its proof is similar to the proof of Theorem 2 and will not be repeated.

A discrete counterpart of Theorem 3 is in the following.

Theorem 9
If a discrete-time system with impulse response sequence g [k] is BIBO
stable, then, as k → ∞:
1. The output excited by u[k] = a, for k ≥ 0, approaches ĝ (1)a.
2. The output excited by u[k] = sin ωo k, for k ≥ 0, approaches
|ĝ (e jω0 )| sin(ω0 k + ∠ĝ (e jω ))
where ĝ (z) is the z-transform of g [k].

∞
X
ĝ (z) = g [m]z −m (7)
m=0

Proof: If u[k] = a for all k ≥ 0, then (6) becomes

k
X k
X
y [k] = g [m]u[k − m] = a g [m]
m=0 m=0
which implies
∞
X
lim y [k] = a g [m] = aĝ (1) as k → ∞
k→∞
m=0
where we have used (7) with z = 1. This establishes the first part of
Theorem 9.

If u[k] = sin ω0 k, then (6) becomes
k
X
y [k] = g [m] sin ω0 [k − m]
m=0
Xk
= g [m](sin ω0 k cos ω0 m − cos ω0 k sin ω0 m)
m=0
k
X k
X
= sin ω0 g [m] cos ω0 m − cos ω0 k g [m] sin ω0 m
m=0 m=0
X∞ X∞
lim y [k] = sin ω0 g [m] cos ω0 m − cos ω0 k g [m] sin ω0 m
k→∞
m=0 m=0
= sin ω0 Re[ĝ (e jω0 )] − cos ω0 kIm[ĝ (e jω0
)]]
jω0 jω0
= |ĝ (e )| sin(ω0 k + ∠ĝ (e ))

Theorem 10
A discrete-time SISO system with proper rational transfer function ĝ (z) is
BIBO stable if and only if every pole of ĝ (z) has a magnitude less than 1
or, equivalently, lies inside the unit circle on the z-plane.

Proof: If ĝ (z) has pole pi with multiplicity mi , then its partial fraction
expansion contains the factors
1 1 1
, 2
,...,
z − pi (z − pi ) (z − pi )mi
Thus the inverse z-transform of ĝ (z) or the impulse response sequence

contains the factors
pik , kpik , . . . , k mi −1 pik
It is straightforward to verify that every such term is absolutely summable
if and only if pi has a magnitude less than 1. Using this fact, we can
establish Theorem 10.

Input-output Stability: MIMO Discrete-time Case
Theorem 11
A MIMO discrete-time system with impulse response sequence matrix
G [k] = [gij [k]] is BIBO stable if and only if every gij [k] is absolutely
summable.
Theorem 12
A MIMO discrete-time system with discrete proper rational transfer matrix
Ĝ (s) = [ĝij (z)] is BIBO stable if and only if every pole of every ĝij (z) has
a magnitude less than 1.

Input-output Stability: MIMO Discrete-time Case
Consider
x[k + 1] = Ax[k] + Bu[k] (8)

y [k] = Cx[k] + Du[k] (9)
Its discrete transfer matrix is

1
Ĝ (z) = C (zI − A)−1 B + D = C [Adj(zI − A)]B + D
det(zI − A)
• Thus Equation (8)-9) or, to be more precise, the zero-state response of

(8)-(9) is BIBO stable if and only if every pole of Ĝ (z) has a magnitude
less than 1.
• Ĝ (z) is BIBO stable if every eigenvalue of A has a magnitude less than 1.

Internal Stability for Continuous-time Systems
We study the stability of the zero-input response or the response of
ẋ(t) = Ax(t) (10)
excited by nonzero initial state x0 . The solution of (10) is
x(t) = e At x0 (11)
Definition 13
The equation ẋ = Ax is marginally stable or stable in the sense of
Lyapunov if every finite initial state x0 excites a bounded response. It is
asymptotically stable if every finite initial state excites a bounded
response, which, in addition, approaches 0 as t → ∞.
This definition is applicable only to linear systems.

Theorem 14
1. The equation ẋ = Ax is marginally stable if and only if all eigenvalues
of A have zero or negative real parts and those with zero real parts are
simple roots of the minimal polynomial of A.
2. The equation ẋ = Ax is asymptotically stable if and only if all
eigenvalues of A have negative real parts.

Proof: Consider x̄ = Px, where P is a nonsingular matrix. Then ẋ = Ax

is equivalent to x̄˙ = Āx̄ = PAP −1 x̄. Thus we may study the stability of A
by using Ā.
The response of x̄˙ = Āx̄ excited by x(0) equals x̄(t) = e Āt x̄(0). It is clear
that the response is bounded if and only if every entry of e Āt is bounded
for all t ≥ 0. If A is in Jordan form, then e Āt is of the form as follows
 λt
e 1 te λ1 t t 2 e λ1 t /2! t 3 e λ1 t /3! · · · t m e λ1 t /m!

Āt
 0 e λ1 t te λ1 t t 2 e λ1 t /2! · · · t m−1 e λ1 t /(m − 1)! 
e = .  (12)
 
. .. .. .. .. ..
 . . . . . . 
0 0 0 0 ··· e λ1 t

Using (12), we can show that if an eigenvalue has a negative real part,
then every entry of (12) is bounded and approaches 0 as t → ∞. If an
eigenvalue has zero real part and has no Jordan block of order 2 or higher,
then the corresponding entry in (12) is a constant or is sinusoidal for all t
and is, therefore, bounded. This establishes the sufficiency of the first part
of Theorem 14. If Ā has an eigenvalue with a positive real part, then every
entry in (12) will grow without bound. If Ā has an eigenvalue with zero
real part and its Jordan block has order 2 or higher, then (12) has at least
one entry that grows unbounded. This completes the proof of the first
part. To be asymptotically stable, every entry of (12) must approach zero
as t → ∞. Thus no eigenvalue with zero real part is permitted.

Remarks:
The asymptotic stability implies BIBO stability.
The BIBO stability does not imply the asymptotic stability.
Asymptotic stability is defined for the zero-input response, whereas
BIBO stability is defined for the zero-state response.
Equivalence transformation does not change the stability of a system.

Example
Consider  
0 0 0
ẋ =  0 0 0  x
0 0 −1
Its characteristic polynomial is ∆(λ) = λ2 (λ + 1). The index (i.e., the
highest order of the Jordan blocks) of the eigenvalue λ = 0 is 1. So, the
minimal polynomial of A is ψ(λ) = λ(λ + 1). The matrix has eigenvalues
0, 0, and −1. The eigenvalue 0 is a simple root of the minimal polynomial.
Thus the equation is marginally stable.
The equation  
0 1 0
ẋ =  0 0 0  x
0 0 −1
is not marginally stable because its minimal polynomial is λ2 (λ + 1) and
λ = 0 is not a simple root of the minimal polynomial.
Internal Stability for Discrete-Time Systems
Consider the discrete-time system
x[k + 1] = Ax[k] (13)
with nonzero initial state x0 . The solution of (13) is
x[k] = Ak x0 (14)
Definition 15
Equation (13) is said to be marginally stable or stable in the sense of
Lyapunov if every finite initial state x0 excites a bounded response. It is
asymptotically stable if every finite initial state excites a bounded
response, which, in addition, approaches 0 as k → ∞.

Internal Stability for Discrete-Time Systems
Theorem 16
1. The equation x[k + 1] = Ax[k] is marginally stable if and only if all
eigenvalues of A have magnitudes less than or equal to 1 and those equal
to 1 are simple roots of the minimal polynomial of A.
2. The equation x[k + 1] = Ax[k] is asymptotically stable if and only if all
eigenvalues of A have magnitudes less than 1.
The proof is similar to the continuous-time case and will not be repeated.
Remark:
Asymptotic stability implies BIBO stability but not the converse.
Equivalence transformation does not change the stability of a system.

Summary
BIBO stability:
For a continuous-time system with impulse response G (t) or transfer
function G (s), BIBO stability ⇔ each element of G (t) is absolutely
integrable over [0, ∞) ⇔ the poles of each element of G (s) are in the
LHP.
For a discrete-time system with impulse response G [k] or transfer
function G (z), BIBO stability ⇔ each element of G [k] is absolutely
summable over [0, ∞) ⇔ the poles of each element of G (z) are inside
the unit circle.
Internal stability:
For a continuous-time system ẋ = Ax, the system is asymptotically
stable ⇔ all eigenvalues of A are in the LHP.
For a discrete-time system x[k + 1] = Ax[k], the system is
asymptotically stable ⇔ all eigenvalues of A are inside the unit circle.

Lyapunov Theorem for Continuous-time Systems
Look at a different method of checking asymptotic stability of
ẋ = Ax.
For convenience, we call A stable if every eigenvalue of A has a negative

real part.
Theorem 17
All eigenvalues of A have negative real parts if and only if for any given
positive definite symmetric matrix N, the Lyapunov equation
A0 M + MA = −N (15)
has a unique symmetric solution M and M is positive definite.

Proof: Necessity: Equation (15) is a special case of Lyapunov equation

with A = A0 and B = A. Because A and A0 have the same set of
eigenvalues. If A is stable, A has no two eigenvalues such that
λi + λj = 0. Thus the Lyapunov equation is nonsingular and has a unique
solution M for any N. We claim that the solution can be expressed as
Z ∞
0
M= e A t Ne At dt (16)
0
This is because
Z ∞ Z ∞
0 0 A0 t At 0
A M + MA = Ae Ne dt + e A t Ne At Adt
Z0 ∞ 0
d A0 t At 0
= e Ne dt = e A t Ne At |∞
t=0 = −N
0 dt
where we have used the fact e At = 0 at t = ∞ for stable A.

This shows that the M in (16) is the solution. It is clear that if N is

symmetric, so is M. Let us decompose N as N = N̄ 0 N̄, where N is
nonsingular (Theorem ??) and consider
Z ∞ Z ∞
0 0 A0 t 0 At
x Mx = x e N̄ N̄e xdt = kN̄e At xk22 dt (17)
0 0
Because both N and e At are nonsingular, for any nonzero x, the integrand
of (17) is positive for every t. Thus x 0 Mx is positive for any x 6= 0. This
shows the positive definiteness of M.
Sufficiency: We show that if N and M are positive definite, then A is
stable. Let λ be an eigenvalue of A and v 6= 0 be a corresponding
eigenvector: that is, Av = λv . Even though A is a real matrix, its
eigenvalue and eigenvector can be complex. Taking the complex-conjugate
transpose of Av = λv yields v ∗ A∗ = v ∗ A0 = λ∗ v ∗ , where the asterisk
denotes complex-conjugate transpose.

Pre-multiplying v ∗ and post-multiplying v to (15) yields
−v ∗ Nv = v ∗ A0 Mv + v ∗ MAv (18)
∗ ∗ ∗
= (λ + λ)v Mv = 2Re(λ)v Mv (19)
Because v ∗ Mv and v ∗ Nv are both real and positive, (19) implies

Re(λ) < 0. This shows that every eigenvalue of A has a negative real part.

Corollary 18
All eigenvalues of an n × n matrix A have negative real parts if and only if
for any given m × n matrix N̄ with m < n and with the property
 
N̄
 N̄A 
rank(O) := rank   = n ( full column rank) (20)
 
..
 . 
N̄An−1
where O is an nm × n matrix, the Lyapunov equation
A0 M + MA = −N̄ 0 N̄ =: −N (21)

Proof: Necessity: It is shown that the solution of (21) is

Z ∞
0
0
So, Z ∞ Z ∞
0
x 0 Mx = x 0 e A t N̄ 0 N̄e At xdt = kN̄e At xk22 dt (23)
0 0
Suppose the integrand of (23) is identically zero or N̄aAt x ≡ 0. Then its

derivative with respect to t yields N̄Ae At x = 0.

Proceeding forward, we can obtain
 
N̄
 N̄A 
 At
e x = 0 (24)

 ..
 . 
N̄An−1
This equation implies that, because of (20) and the nonsingularity of e At ,

the only x meeting (24) is 0. Thus the integrand of (17) cannot be
identically zero for any x 6= 0. Thus M is positive definite under the
condition in (20).
Sufficient: Let λ be an eigenvalue of A and v 6= 0 be a corresponding
eigenvector: that is, Av = λv . Even though A is a real matrix, its
eigenvalue and eigenvector can be complex. Taking the complex-conjugate
transpose of Av = λv yields v ∗ A∗ = v ∗ A0 = λ∗ v ∗ , where the asterisk
denotes complex-conjugate transpose.
Pre-multiplying v ∗ and post-multiplying v to (21) yields
−v ∗ Nv = v ∗ A0 Mv + v ∗ MAv (25)
∗ ∗ ∗
= (λ + λ)v Mv = 2Re(λ)v Mv = −kN̄v k22 (26)
We show that N̄v is nonzero under (20). Because of Av = Av , we have

A2 v = λAv = λ2 v ,. . . An−1 v = λn−1 v . Consider
   
N̄   N̄v
 N̄A  N̄v
..  =  λN̄v 
 
v =
  
 ..   N̄Av .   .. 
 .  n−1
 . 
n−1 N̄A v n−1
N̄A λ N̄v
If N̄v = 0, the rightmost matrix is zero; the leftmost matrix, however, is

nonzero under the conditions of (20) and v 6= 0. This is a contradiction.
Thus N̄v is nonzero and (26) implies Re(A) < 0.
Remark: 1. For any N̄, the matrix N in (21) is positive semidefinite.

Theorem 17 and its corollary are valid for any given N; therefore we shall
use the simplest possible N. Even so, using them to check stability of A is
not simple.
2. It is much simpler to compute, using MATLAB, the eigenvalues of A
and then check their real parts.
3. The importance of Theorem 17 and its corollary is not in checking the
stability of A but rather in studying the stability of nonlinear systems.
They are essential in using the so-called second method of Lyapunov.

In the proof of Theorem of 17, we have established the following result.

For easy reference, we state it as a theorem.
Theorem 19
If all eigenvalues of A have negative real parts, then the Lyapunov equation
A0 M + MA = −N
has a unique solution for every N, and the solution can be expressed as
Z ∞
0
0

Proof: This theorem has been proved in the proof of Theorem 17. Next,
we give another proof of the uniqueness of M. Suppose there are two
solutions M1 and M2 . Then we have
A0 (M1 − M2 ) + (M1 − M2 )A = 0
which implies
0 d A0 t
e A t [A0 (M1 − M2 ) + (M1 − M2 )A]e At = [e (M1 − M2 )e At ] = 0
dt
Its integration from 0 to ∞ yields
0
[e A t (M1 − M2 )e At ]|∞
0 =0
or, using e At → 0 as t → ∞, 0 − (M1 − M2 ) = 0. This shows the

uniqueness of M.

Lyapunov Theorem for Discrete-Time Systems
Consider the discrete-time Lyapunov equation
M − AMB = C (28)
where A and B are, respectively, n × n and m × m matrices, and M and C

are n × m matrices.
Theorem 20
The eigenvalues of A are λi (1 ≤ i ≤ n) and the eigenvalues of B are µj
(1 ≤ j ≤ m). If there are no i and j such that λi λj = 1, then for any
C ∈ R n×m there exists a unique solution M.

Proof:
vec(M) − vec(AMB) = vec(C )
then
vec(M) − (B > ⊗ A)vec(M) = vec(C ) ⇒ (I − B > ⊗ A)vec(M) = vec(C )
If there are no i and j such that λi λj = 1, then (I − B > ⊗ A) is

nonsingular. Therefore, vec(M) = (I − B > ⊗ A)−1 vec(C ) and is unique.

Theorem 21
All eigenvalues of an n × n matrix A have magnitudes less than 1 if and
only if for any given positive definite symmetric matrix N or for N = N̄ 0 N̄,
where N̄ is any given m × n matrix with m < n and with the property
 
N̄
 N̄A 
rank(O) := rank   = n ( full column rank)
 
..
 . 
N̄A n−1
the discrete Lyapunov equation
M − A0 MA = N (29)

Proof: We sketch briefly its proof for N > 0. If all eigenvalues of A and,
consequently, of A0 have magnitudes less than 1, then we have |λi λj | < 1
for all i and j. Thus λi λj 6= 1 and (29) is nonsingular. Therefore, for any
N, a unique solution exists in (29). We claim that the solution can be
expressed as
X∞
M= (A0 )m NAm (30)
m=0
Because |λi | < 1 for all i, this infinite series converges and is well defined.
Substituting (30) into (29) yields
∞ ∞ ∞
!
X X X
(A0 )m NAm − A0 (A0 )m NAm A = N + (A0 )m NAm
m=0 m=0 m=1
∞
X
− (A0 )m NAm = N
m=1

Thus (30) is the solution. If N is symmetric, so is M. If N is positive

definite, so is M. This establishes the necessity.
To show sufficiency, let A be an eigenvalue of A and v 6= 0 be a
corresponding eigenvector, that is, Av = λv . Then we have
v ∗ Nv = v ∗ Mv − v ∗ A0 MAv = v ∗ Mv − λ∗ v ∗ Mv λ
= (1 − |λ|2 )v ∗ Mv
Because both v ∗ Nv and v ∗ Mv are real and positive, we conclude

(1 − kλk)2 > 0 or |λ|2 < 1. This establishes the theorem for N > 0. The
case N ≥ 0 can similarly be established.

Theorem 22
If all eigenvalues of A have magnitudes less than 1, then the discrete
Lyapunov equation
M − A0 MA = N
has a unique solution for every N and the solution can be expressed as
∞
X
M= (A0 )m NAm .
m=0
Remark: It is important to mention that even if A has one or more

eigenvalues with magnitudes larger than 1, a unique solution still exists in
the discrete Lyapunov equation if λi λj 6= 1 for all i and j. In this case, the
solution cannot be expressed as in (30) but can be computed from a set of
linear algebraic equations.
Let us discuss the relationships between the continuous-time and
discrete-time Lyapunov equations. The stability condition for
continuous-time systems is that all eigenvalues lie inside the open left-half
s-plane. The stability condition for discrete-time systems is that all
eigenvalues lie inside the unit circle on the z-plane. These conditions can
be related by the bilinear transformation
z −1 1+s
s= , z= (31)
z +1 1−s
which maps the left-half s-plane into the interior of the unit circle on the
z-plane and vice versa. To differentiate the continuous-time and
discrete-time cases, we write
A0 M + MA = −N (32)
and
Md − A0d Md Ad = −Nd (33)
Following (31), these two equations can be related by
A = (Ad + I )−1 (Ad − I ), Ad = (I + A)(I − A)−1 , (34)

0
M = Md , N = 0.5(I − A )Nd (I − A) (35)
The MATLAB function lyap computes the Lyapunov equation in (32) and
dlyap computes the discrete Lyapunov equation in (33). The function
dlyap transforms (33) into (32) by using (35) and then calls lyap. The
result yields M = Md .

Stability of LTV Systems
Consider a SISO linear time-varying (LTV) system described by

Z t
y (t) = g (t, τ )u(τ )dτ (36)
t0
• The system is said to be BIBO stable if every bounded input excites a

bounded output.
• The necessary and sufficient conditions for (36) to be BIBO stable is
that there exists a finite constant M such that
Z t
|g (t, τ )|dτ ≤ M < ∞ (37)
t0
for all t and t0 with t ≥ t0 .

For the multivariable case, (36) becomes

Z t
y (t) = G (t, τ )u(τ )dτ (38)
t0
• The necessary and sufficient conditions for (38) to be BIBO stable is

that every entry of G (t, τ ) meets the condition in (37).
• The necessary and sufficient condition for (38) to be BIBO stable is that
there exists a finite constant M such that
Z t
kG (t, τ )kdτ ≤ M < ∞
t0

The impulse response matrix of
ẋ = A(t)x + B(t)u (39)

y = C (t)x + D(t)u (40)
is
G (t, τ ) = C (t)Φ(t, τ ) + D(t)δ(t − τ )
Thus (39) or, more precisely, the zero-state response of (39) is BIBO
stable if and only if there exist constants M1 and M2 such that
kD(t)k ≤ M1 < ∞
and Z t
kG (t, τ )kdτ ≤ M2 < ∞
t0

Definition 23
The equation ẋ = A(t)x is marginally stable if every finite initial state
excites a bounded response. The equation ẋ = A(t)x is asymptotically
stable if the response excited by every finite initial state is bounded and
approaches zero as t → ∞.
Theorem 24
Marginal and asymptotic stabilities of ẋ = A(t)x are invariant under any
Lyapunov transformation.

Proof: Let x̄ = P(t)x is a Lyapunov transformation. The fundamental

matrix X (t) of ẋ = A(t)x and the fundamental matrix X̄ (t) of x̄˙ = Ā(t)x̄
are related by
X̄ (t) = P(t)X (t)
which implies
Φ̄(t, τ ) = X̄ (t)X̄ −1 (τ ) = P(t)Φ(t, τ )P −1 (τ )
Because both P(t) and P −1 (t) are bounded, if kΦ(t, τ )k is bounded, so is

kΦ̄(t, τ )k; if kΦ(t, τ )k → 0, so is kΦ̄(t, τ )k. This establishes Theorem 24.

Let Φ(t, t0 ) be the state transition matrix of the system.

• The system is marginally stable if and only if kΦ(t, t0 )k ≤ M < ∞ for
any t0 and t ≥ t0 .
• The system is asymptotically stable if and only if kΦ(t, t0 )k ≤ M < ∞
for any t0 and t ≥ t0 and kΦ(t, t0 )k → 0 as t → ∞.

• For a linear time-varying system, asymptotically stability of zero-input

response does not imply BIBO stability of zero-state
Rt response. This is
because kΦ(t, τ )k → 0 does not guarantee t0 kC (t)Φ(t, τ )B(τ )kdτ < ∞.
• For a linear time-varying system, if the zero-input response is
exponentially stable and C (t) and B(t) are bounded for all t, then the
zero-state response is BIBO.

z Transform
Definition: If a signal has discrete values {e(k)}∞

k=0 , the z-transform of
the signal is defined as
∞
X
E (z) = Z{e(k)} = ek z −k , r0 < |z| < R0 (41)
k=0
and we assume we can find values of r0 and R0 as bounds on the

magnitude of the complex variable z for which the series eqn. (41)
converges.

Example
ek = e −akT 1(kT ), find the z-transform of this signal.

Example
Solution:
∞ ∞
X X 1 z
ek z −k = e −akT z −k = = , |z| > e −aT
1− e −aT z −1 z − e −aT
k=−∞ 0

Example
Given
1, if k = 0
e1 (k) = δ(k) =
0, 6 0
if k =
∞
X
E1 (z) = e1 (k)z −k = 1
k=−∞

Example
Given
1, if k ≥ 0
e2 (k) = 1(k) =
0, if k < 0
∞ ∞
X X 1
E1 (z) = e2 (k)z −k = z −k = , (|z| > 1)
1 − z −1
k=−∞ k=0

z transform
Figure 3: z table

Properties of z transform
1 Linearity:
Z[αf1 (k) + βf2 (k)] = αZ[f1 (k)] + βZ[f2 (k)]
2 Real translation:
Z[f (k − n)u(k − n)] = z −n [Zf (k)]
n−1
" #
X
n −k
Z[f (k + n)u(k)] = z Z[f (k)] − f (k)z
k=0
3 Time shifting:
Z[f (k − n)u(k − n)] = z −n F (z), n ≥ 0
n−1
" #
X
Z[f (k + n)u(k)] = z n F (z) − f (l)z −l , n ≥ 1
l=0
4 Differentiation:
dF (z)
Z[kf [k]] = −z
dz
5 Complex translation
Z[e ak f (k)] = F (ze −a )
6 Scaling in the z-plane
Z[r −k f (k)] = F (rz)
where F (z) = Z[f (k)].
7 Initial value: f (0) = limz→∞ F (z).
8 Final-value Theorem: If F (z) converges for |z| > 1 and all poles of
(z − 1)F (z) are inside the unit circle, then
lim f (k) = lim (z − 1)F (z)
t→∞ z→1
9 Convolution of Time Sequences:
( ∞ )
X
Z f1 (l)f2 (k − l) = F1 (z)F2 (z)
l=−∞

Figure 4: z properties

Example
Given a discrete-time system
x[k + 1] = Ax[k] + Bu[k], y [k] = Cx[k] + Du[k]
Take Z transform on both sides, we have
z(X (z) − x[0]) = AX (z) + BU(z), Y (z) = CX (z) + DU(z)
If x[0] = 0, then
X (z) = (zI − A)−1 BU(z), Y (z) = [C (zI − A)−1 B + D]U(z)
So,
Y (z)
= C (zI − A)−1 B + D
U(z)
is called the transfer matrix in Z domain.

The inverse z-transform
• Inverse formula method: Given E (z),
√
I
1
e(k) = E (z)z k−1 dz, j = −1
2πj Γ
where Γ is any path that encloses all the finite poles of E (z)z k−1 .
Using the theorem of residues, we have
X
e(k) = [residues of E (z)z k−1 ] (42)
at poles of E (e)z k−1
if z = a is a simple pole
(residue of E (z)z k−1 )z=a = (z − a)E (z)z k−1 |z=a
if z = a is an m repeated pole
1 d m−1
(residue of E (z)z k−1 )z=a = [(z − a)m E (z)z k−1 ]|z=a
(m − 1)! dz m−1
Example
z
E (z) =
(z − 1)(z − 2)
find ek .

Example
Solution: By the formula
zz k−1 zz k−1
e(k) = (z − 1) |z=1 + (z − 2) |z=2 = −1 + 2k
(z − 1)(z − 2) (z − 1)(z − 2)

Example
a
E (z) =
(z − 1)2
find e(k).

Example
Solution:
1 d 2−1 z
e(k) = [(z − 1)2 z k−1 ]|z=1 = k
(2 − 1)! dt 2−1 (z − 1)2

The Inverse z Transform
• Power Series Method: Write E (z) as
E (z) = e0 + e1 z −1 + e2 z −2 + · · · +
we can find {ek }. The power series may be obtained by long division.

Example
z
E (z) =
z2 − 3z + 2
find ek .

Example
Solution: Using long division (Fig. 5), we have ek = 2k − 1.

The Inverse z Transform
• Partial-fraction expansion method: If
N(z)
E (z) =
D(z)
1 Decompose E (z)/z as a sum of simple terms

2 Obtain E (z)
3 find the inverse z transform with the aid of the z transform table.

Example
z
E (z) =
(z − 1)(z − 2)
Find ek .

Example
Solution:
E (z) 1 1 1
= = −
z (z − 1)(z − 2) z −2 z −1
then by z table
z z
Z −1 [E (z)] = Z −1 [ ] − Z −1 [ ] = 2k − 1
z −2 z −1

Stability: Dr. Wenjie Dong

Uploaded by

Document Information

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: Dr. Wenjie Dong

Uploaded by

Copyright:

Available Formats

Stability

Dr. Wenjie Dong

Department of Electrical and Computer Engineering,

W. Dong (UTRGV, Dept. of ECE) 1 / 84

Consider a SISO linear time-invariant (LTI) causal system described by

Remark: This stability is defined for the zero-state response and is

W. Dong (UTRGV, Dept. of ECE) 1 / 84

for some constant M.

W. Dong (UTRGV, Dept. of ECE) 2 / 84

where |u(t)| ≤ um < ∞ for all t ≥ 0.

We choose the input

W. Dong (UTRGV, Dept. of ECE) 3 / 84

then the output

because y (t1 ) can be arbitrarily large, we conclude that a similar bounded

W. Dong (UTRGV, Dept. of ECE) 4 / 84

|ĝ (jω0 )| sin(ω0 t + ∠ĝ (jω0 ))

where ĝ (s) is the Laplace transform of g (t) or

W. Dong (UTRGV, Dept. of ECE) 5 / 84

Proof: 1. If u(t) = a for all t ≥ 0, then

where we have used (2) with s = 0.

W. Dong (UTRGV, Dept. of ECE) 6 / 84

2. If u(t) = sin ωt, then

W. Dong (UTRGV, Dept. of ECE) 7 / 84

W. Dong (UTRGV, Dept. of ECE) 8 / 84

Thus the inverse Laplace transform of ĝ (s) or the impulse response

It is straightforward to verify that every such term is absolutely integrable

W. Dong (UTRGV, Dept. of ECE) 9 / 84

For a a system with the transfer function

is it BIBO stable or not?

W. Dong (UTRGV, Dept. of ECE) 10 / 84

Solution: The impulse response of the system is

W. Dong (UTRGV, Dept. of ECE) 11 / 84

Consider the positive feedback system shown in Fig. 1(a). Is it BIBO

W. Dong (UTRGV, Dept. of ECE) 12 / 84

W. Dong (UTRGV, Dept. of ECE) 14 / 84

For a LTI MIMO systems defined by

where G (t − τ ) is the impulse matrix, we have the following results.

W. Dong (UTRGV, Dept. of ECE) 15 / 84

Its transfer matrix is

Since every pole of Ĝ (s) is an eigenvalue of A, Ĝ (s) is BIBO stable if

W. Dong (UTRGV, Dept. of ECE) 16 / 84

W. Dong (UTRGV, Dept. of ECE) 17 / 84

The transfer function of the equation is

ĝ (s) = 0.5(s − 1)−1 × 0 + 0.5 = 0.5

W. Dong (UTRGV, Dept. of ECE) 19 / 84

Consider a discrete-time SISO Linear time-invariant causal system

Remark:This stability is defined for the zero-state response and is

for some constant M.

W. Dong (UTRGV, Dept. of ECE) 21 / 84

A discrete counterpart of Theorem 3 is in the following.

|ĝ (e jω0 )| sin(ω0 k + ∠ĝ (e jω ))

where ĝ (z) is the z-transform of g [k].

W. Dong (UTRGV, Dept. of ECE) 22 / 84

Proof: If u[k] = a for all k ≥ 0, then (6) becomes

W. Dong (UTRGV, Dept. of ECE) 23 / 84

W. Dong (UTRGV, Dept. of ECE) 24 / 84

W. Dong (UTRGV, Dept. of ECE) 25 / 84

Thus the inverse z-transform of ĝ (z) or the impulse response sequence

W. Dong (UTRGV, Dept. of ECE) 26 / 84

W. Dong (UTRGV, Dept. of ECE) 27 / 84

x[k + 1] = Ax[k] + Bu[k] (8)

Its discrete transfer matrix is

• Thus Equation (8)-9) or, to be more precise, the zero-state response of