Solutions of Linear Systems

1/65
Analysis of autonomous linear systems
Stability
Lyapunov theory
Variation-of-Constants Formula
Transfer Matrices
Related Reading
[KK]: 2.5-2.6 and [AM]: 4.3-4.4, 5.1-5.3, 8.1-8.3
Autonomous Systems
2/65
A dynamical state-space system without inputs/outputs is described as
x = f(x) with f : X R
n
, X R
n
.
Understanding the response of such systems to initial conditions X
is a rich eld called dynamical systems theory. For us the main tool is
numerical simulation.
The system is linear if f is a linear map. All such maps are described as
f(x) = Ax with some matrix A R
nn
. Hence a linear autonomous
system is described as
x = Ax with A R
nn
.
The responses of such systems can be completely understood by
using tools from linear algebra.
Diagonal A
3/65
Suppose that A is a diagonal matrix:
A =
_
_

1
0
.
.
.
.
.
.
.
.
.
0
n
_
_
with
1
, . . . ,
n
R.
Then x = Ax splits up into the n scalar dierential equations
x
1
=
1
x
1
, . . . , x
n
=
n
x
n
.
These can be solved independently. All solutions are
x
1
(t) = e

1
t

1
, . . . , x
n
(t) = e

n
t

n
with
1
, . . . ,
n
R.
This can be compactly written as
x(t) =
_
_
e

1
t
0
.
.
.
.
.
.
.
.
.
0 e

n
t
_
_
with =
_
_

1
.
.
.

n
_
_
.
Observe that x(0) = .
State-Coordinate Change
4/65
In practice A will not be diagonal. Can we still exploit what we have
seen? For this this purpose we introduce the following key concept.
Denition 1 Any non-singular (i.e. invertible) matrix T R
nn
denes a state-coordinate transformation z = Tx.
Suppose that x(t) satises x(t) = Ax(t). Then z(t) = Tx(t) satises
z(t) = T x(t) = TAx(t) = (TAT
1
)z(t) =

Az(t).
In the new coordinates the system is described with

A = TAT
1
.
Conversely suppose that z(t) is any trajectory of z(t) =

Az(t). Then
x(t) = T
1
z(t) satises x(t) = Ax(t).
The solution set of z =

Az is linearly transformed by T
1
into the
solution set of x = Ax.
Diagonizable A
5/65
In many practical cases A can be transformed into a diagonal matrix

A
by a suitably chosen state-coordinate change.
Theorem 2 Suppose that TAT
1
= diag(
1
, . . . ,
n
) R
nn
.
Then the unique solution of x = Ax, x(0) = is given by
x(t) = [T
1
diag(e

1
t
, . . . , e

n
t
)T].
Proof. Just combine the two previous slides. x(t) satises
x(t) = Ax(t) with x(0) =
if and only if z(t) = Tx(t) and

= T satisfy
z(t) = diag(
1
, . . . ,
n
)z(t) with z(0) =

.
The latter ivp has the unique solution
z(t) = diag(e

1
t
, . . . , e

n
t
)

.
How to Diagonalize A?
6/65
Matlab does it for us with [S,La]=eig(A).
Linearize Segway in upright position: Leads to
x

= Ax

with A =
_
_
_
_
0 0 1 0
0 0 0 1
0 3.92 2 0.32
0 22.1 3.23 1.82
_
_
_
_
.
Get = S
1
AS with
S =
_
_
_
_
1 0.03 0.04 0.58
0 0.26 0.16 0.11
0 0.12 0.22 0.79
0 0.96 0.96 0.16
_
_
_
_
, =
_
_
_
_
0 0 0 0
0 3.72 0 0
0 0 6.16 0
0 0 0 1.38
_
_
_
_
.
Qualitative insight. Each component of any solution x

(.) of x

=
Ax

is a linear combination of the exponential functions

e
0 t
= 1, e
3.72 t
, e
6.16 t
, e
1.38 t
.
Example
7/65
We can now answer all kinds of questions
about all or particular solutions of the linearization.
If equals either the rst, the second, the third or
the fourth column of S, the respective solutions are
_
_
_
_
1
0
0
0
_
_
_
_
e
0 t
,
_
_
_
_
0.03
0.26
0.12
0.96
_
_
_
_
e
3.72 t
,
_
_
_
_
0.04
0.16
0.22
0.96
_
_
_
_
e
6.16 t
,
_
_
_
_
0.58
0.11
0.79
0.16
_
_
_
_
e
1.38 t
.
All other solutions are linear combinations thereof. Please note
that we have complete insight into how solutions behave for t !
Question: For which initial conditions does the solution not blow up?
Answer: For all in the subspace
_

_
S
_
_
_
_

1
0

4
_
_
_
_
|

1
,

3
,

4
R
_

_
.
Example
8/65
For the linearization in the downright position
suitable matrices S, T = S
1
and are given by
T =
_
_
_
_
1 0 0.7 0.12
0 0.03 + 2.58i 0.12 0.37i 0.54 + 0.08i
0 0.03 2.58i 0.12 + 0.37i 0.54 0.08i
0 0.36 1.33 0.21
_
_
_
_
=
_
_
_
_
r
1
r
2
r
3
r
4
_
_
_
_
S =
_
_
_
_
1 0.04i 0.04i 0.54
0 0.05 0.2i 0.05 + 0.2i 0.12
0 0.17 0.05i 0.17 + 0.05i 0.81
0 0.96 0.96 0.18
_
_
_
_
=
_
c
1
c
2
c
3
c
4
_
=
_
_
_
_
0 0 0 0
0 1.17 + 4.46i 0 0
0 0 1.17 4.46i 0
0 0 0 1.49
_
_
_
_
= diag(
1
,
2
,
3
,
4
).
Complex Transformations and Diagonal Matrices
9/65
Despite the fact that T and are complex, the main insight on
slide 5 stays true without change!
In fact observe that
2
=

3
, c
2
= c
3
, r
2
= r
3
. Therefore
T
1
diag(e

1
t
, e

2
t
, e

3
t
, e

4
t
)T =
=
_
c
1
e

1
t
c
2
e

2
t
c
2
e

2
t
c
4
e

4
t
_
_
_
_
_
r
1
r
2
r
2
r
4
_
_
_
_
=
= e

1
t
c
1
r
1
+ e

2
t
c
2
r
2
+ e

2
t
c
2
r
2
+ e

4
t
c
4
r
4
=
= e

1
t
c
1
r
1
+ 2Re[e

2
t
c
2
r
2
] + e

4
t
c
4
r
4
is always a real matrix. The response to the initial condition R
4
is
e

1
t
c
1
(r
1
) + 2Re[e

2
t
c
2
(r
2
)] + e

4
t
c
4
(r
4
).
Complex Transformations and Diagonal Matrices
10/65
For a complex number = + i ( = Re(), = Im()) recall
e
t
= e
(+i)t
= e
t
[cos(t) + i sin(t)].
If c is a complex column and r a complex row then
cr = [Re(c) + iIm(c)][Re(r) + iIm(r)] = M + iN
with the real matrices
M = [Re(c)Re(r) Im(c)Im(r)], N = [Re(c)Im(r) + Im(c)Re(r)].
This leads to the explicit formula
Re[e
t
cr] = e
t
[cos(t)M sin(t)N] .
The components of Re[e
t
cr] hence are, qualitatively,
sustained ( = 0), growing ( > 0) or decaying ( < 0) oscillations.
Example: Linearization of Segway
11/65
Solution with = col(1, 0, 0, 0) is x

(t) = col(1, 0, 0, 0) for all t 0.

Solution with = col(0, 1, 0, 0) is
2 e
1.17 t
_

_
cos(4.46 t)
_
_
_
_
0.1
0.52
0.15
0.03
_
_
_
_
sin(4.46 t)
_
_
_
_
0.01
0.13
0.43
2.48
_
_
_
_
_

_
+e
1.49 t
_
_
_
_
0.2
0.04
0.29
0.06
_
_
_
_
and hence decaying in an oscillatory fashion:
0 1 2 3 4 5
0.5
0
0.5
1
p

and

Recap
12/65
The complex number C is an eigenvalue of A R
nn
if it is a
zero of the characteristic polynomial of A:
det(I A) = 0.
If is an eigenvalue of A then any non-zero vector v C
n
with
(I A)v = 0 or equivalently Av = v
is a corresponding eigenvector of A.
Some very simple facts
If (, v) is a pair of eigenvalue/eigenvector then so is (

, v).
Any matrix A R
nn
does have at most n dierent eigenvalues.
For each eigenvalue one can compute at most n linearly independent
eigenvectors (by computing a basis of the null space of I A).
Eigenvalues are invariant under a coordinate change.
How to Diagonalize A?
13/65
For each eigenvalue of A R
nn
we can hence compute a set of linearly
independent eigenvectors; if done for , take the complex conjugate
vectors for

; collect these vectors in the list v
1
, . . . , v
g
and denote by

1
, . . . ,
g
the corresponding (not necessarily dierent) eigenvalues.
Theorem 3 The vectors v
1
, . . . , v
g
are linearly independent and
hence g n. In case that g = n then A can be diagonalized:
S
1
AS = diag(
1
, . . . ,
n
) with S =
_
v
1
v
n
_
.
If g < n then A cannot be diagonalized by a coordinate change.
The formula for S
1
AS is very easy to see:
AS = A
_
v
1
v
n
_
=
=
_
Av
1
Av
n
_
=
_

1
v
1

n
v
n
_
=
=
_
v
1
v
n
_
diag(
1
, . . . ,
n
) = S diag(
1
, . . . ,
n
).
Example 1
14/65
Consider the matrix
A =
_
_
_
_
_
3 4 6 4 8
2 3 4 2 6
0 0 1 0 0
0 2 2 1 2
2 2 4 2 5
_
_
_
_
_
The characteristic polynomial of A is computed with poly(A) to
det(I A) = ( + 1)
2
( 1)
3
.
The LU-factorization (with command lu) of 1I A and 1I A give
U
1
=
_
_
_
_
_
2 4 6 4 8
0 2 2 2 2
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
_
_
_
_
_
, U
2
=
_
_
_
_
_
4 4 6 4 8
0 2 1 0 2
0 0 2 0 0
0 0 0 0 0
0 0 0 0 0
_
_
_
_
_
Since the dimensions of the null-spaces of U
1
and U
2
are 3 and 2, we
have g = 5 = n and hence A can be diagonalized.
Example 1
15/65
Linear independent vectors in the null-space of U
1
and U
2
are
_
_
_
_
_
1
1
1
0
0
_
_
_
_
_
,
_
_
_
_
_
0
1
0
1
0
_
_
_
_
_
,
_
_
_
_
_
2
1
0
0
1
_
_
_
_
_
and
_
_
_
_
_
1
0
0
1
0
_
_
_
_
_
,
_
_
_
_
_
1
1
0
0
1
_
_
_
_
_
.
Then
S =
_
_
_
_
_
1 0 2 1 1
1 1 1 0 1
1 0 0 0 0
0 1 0 1 0
0 0 1 0 1
_
_
_
_
_
satises S
1
AS =
_
_
_
_
_
1 0 0 0 0
0 1 0 0 0
0 0 1 0 0
0 0 0 1 0
0 0 0 0 1
_
_
_
_
_
.
Example 2
16/65
Consider the matrix
A =
_
_
_
_
_
1 7 7 8 6
1 5 5 5 5
1 0 2 1 1
0 3 3 3 2
1 4 5 5 4
_
_
_
_
_
.
The characteristic polynomial of A is computed with poly(A) to
det(I A) = ( + 1)
2
( 1)
3
.
The LU-factorization (with command lu) of 1I A and 1I A give
U
1
=
_
_
_
_
_
1 4 5 5 5
0 7 7 8 6
0 0 0 0.57 0.57
0 0 0 0 0
0 0 0 0 0
_
_
_
_
_
, U
2
=
_
_
_
_
_
2 7 7 8 6
0 3.5 0.5 3 2
0 0 2.57 0.57 0.29
0 0 0 0.89 0.44
0 0 0 0 0
_
_
_
_
_
.
Since the dimensions of the null-spaces of U
1
and U
2
are 2 and 1, we
have g = 3 < 5 = n and hence A is not diagonizable.
Summary
17/65
Let us be given x = Ax with A R
nn
:
We can check whether it can be diagonalized as shown on slide 13.
If existing one can compute a coordinate transformation T such that
TAT
1
has the eigenvalues of A on its diagonal.
If A can be diagonalized, we can get complete insight into the solution
set of the dierential equation as described on slide 5.
With the transformation as on slide 13 the formula on slide 5 reads as
x(t) =
_
e

1
t
v
1
e

n
t
v
n
_
S
1
.
The eigenvalues determine the dynamic characteristics of the solution.
The eigenvectors provide information how these inuence the state.
The eigenvalues of A are often called the modes of the system
x = Ax. The corresponding eigenvectors are the mode-shapes.
Matrix Exponential
18/65
If we recall the Taylor series expansion of the function x e
x
we infer
T
1
_
_
e

1
t
0
.
.
.
.
.
.
.
.
.
0 e

n
t
_
_
T = T
1
_
_
_
_
_
_
_
_

k=0
1
k!
(
1
t)
k
0
.
.
.
.
.
.
.
.
.
0

k=0
1
k!
(
n
t)
k
_
_
_
_
_
_
_
_
T =
=

k=0
1
k!
_
_
T
1
_
_

k
1
0
.
.
.
.
.
.
.
.
.
0
k
n
_
_
T
_
_
t
k
=

k=0
1
k!
_
_
T
1
_
_

1
0
.
.
.
.
.
.
.
.
.
0
n
_
_
T
_
_
k
t
k
.
In case that A = T
1
diag(
1
, . . . ,
n
)T we hence conclude that
T
1
_
_
e

1
t
0
.
.
.
.
.
.
.
.
.
0 e

n
t
_
_
T =

k=0
1
k!
(At)
k
.
Matrix Exponential
19/65
For any M R
nn
recall M
0
= I and M
k
= MM
k1
for k = 1, 2, . . . .
Theorem 4 For any A R
nn
and t R dene
e
At
:= lim
N
N

k=0
1
k!
(At)
k
=

k=0
1
k!
(At)
k
.
The series converges uniformly on [T, T] for any T > 0. Hence
t e
At
is a well-dened analytic function on R.
e
A0
= I and e
A(t+)
= e
At
e
A
and hence e
At
= [e
At
]
1
.

d
dt
e
At
= Ae
At
= e
At
A.
The third statement follows from
d
dt

k=0
1
k!
(At)
k
=

k=1
A
k
1
k!
_
d
dt
t
k
_
= A

k=1
A
k1
1
(k 1)!
t
k1
= Ae
At
and the observation that we can also pull A out to the right.
General A
20/65
Theorem 5 If A R
nn
the unique solution of x = Ax, x(0) =
is given by
x(t) = e
At
.
Proof. To show existence we just check that x(t) is a solution:
x(0) = e
A0
= and x(t) =
d
dt
e
At
= Ae
At
= Ax(t).
To show uniqueness let y(t) be another solution. Then
d
dt
[e
At
y(t)] = [
d
dt
e
At
]y(t) + e
At
[
d
dt
y(t)] =
= Ae
At
y(t) + e
At
Ay(t) = e
At
[A + A]y(t) = 0.
Since e
At
y(t) has zero derivative for all t R, it must be constant. If
we evaluate at t = 0, we observe that this constant must equal . We
conclude e
At
y(t) = or y(t) = e
At
= x(t) for all t R.
Computation of e
At
21/65
Example. The double integrator q = u has the state-space description
x =
_
0 1
0 0
_
x +
_
0
1
_
u = Ax + Bu.
Now note that (At)
2
= 0. Hence we have
e
At
= I + (At) =
_
1 0
0 1
_
+
_
0 t
0 0
_
=
_
1 t
0 1
_
.
All solutions of x = Ax are thus given by
x(t) =
_
1 t
0 1
_
=
_

1
+ t
2

2
_
.
This illustrates the following general structural insight.
All elements of e
At
are linear combinations of the terms
e

1
t
, te

1
t
, . . . , t
n1
e

1
t
, . . . , e

p
t
, te

p
t
, . . . , t
n1
e

p
t
if
1
, . . . ,
p
are the pairwise dierent eigenvalues of A.
Jordan Form
22/65
The key tool is the Jordan canonical form of a complex matrix.
Theorem 6 For A C
nn
there exists an invertible S C
nn
with
S
1
AS = J =
_
_
J
1
0
.
.
.
.
.
.
.
.
.
0 J
g
_
_
, J
l
=
_
_
_
_

l
1 0
.
.
.
.
.
.
.
.
.
.
.
.
0
l
1
0 0
l
_
_
_
_
.
J
l
are so-called Jordan blocks. Moreover:
Up to permutation of the Jordan blocks, the Jordan canonical
form diag(J
1
, . . . , J
g
) is uniquely determined by A.

1
, . . . ,
g
are the (not necessarily dierent) eigenvalues of A.
There are exactly g linearly independent eigenvectors of A.
A is diagonizable i all Jordan blocks have dimension one.
Example
23/65
Recall the example on slide 16. The command [S, J] = jordan(A) does
indeed return S
1
AS = J with
S =
_
_
_
_
_
3 1 1 5 3
2 1 1 4 3
0 0 1 1 1
1 0 1 2 2
2 1 1 3 2
_
_
_
_
_
and J =
_
_
_
_
_
_
1 1 0 0 0
0 1 0 0 0
0 0 1 1 0
0 0 0 1 0
0 0 0 0 1
_
_
_
_
_
_
.
A has three Jordan blocks of dimension 2, 2 and 1.
The rst, third, fth column of S are linearly independent eigenvec-
tors of A for the eigenvalues 1, 1, 1. Other columns: Generalized ei-
genvectors in the kernel of (AI)

for eig(A) and = 2, 3, . . ..

Warning. Computing the Jordan canonical is numerically unrelia-
ble. Do not thoughtlessly use the command jordan in Matlab! The
Schur-decomposition is a much more reliable computational tool.
Computation of e
At
24/65
Once we have computed the Jordan form of A we can determine e
At
.
Theorem 7 If A equals SJS
1
as in Theorem 6 then
e
At
= Se
Jt
S
1
= S diag(e
J
1
t
, . . . , e
J
g
t
)S
1
.
Moreover, if J
l
has dimension d, then
e
J
l
t
=
_
_
_
_
_
_
_
_
_
1 t
t
2
2!

t
d2
(d2)!
t
d1
(d1)!
0 1 t
t
d3
(d3)!
t
d2
(d2)!
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0 0 0 1 t
0 0 0 0 1
_
_
_
_
_
_
_
_
_
e

l
t
for l = 1, . . . , g.
This leads to an explicit formula for e
At
. In Matlab use expm(A*t).
C. Moler, Ch. Van Loan, Nineteen dubious ways to compute the exponential
of a matrix, twenty-ve years later, SIAM Review, 45: 3-49, 2003.
Complex Eigenvalues
25/65
Partition the columns of S and the rows of T = S
1
according to J as
S =
_
C
1
C
g
_
and T =
_
_
R
1
.
.
.
R
g
_
_
.
Then the next formula reveals how the modes of A contribute to e
At
:
e
At
= C
1
e
J
1
t
R
1
+ C
2
e
J
2
t
R
2
+ C
g
e
J
g
t
R
g
If
k
is real we can make sure that C
k
and R
k
are real.
If
k
is complex and
k
=

l
, we can enforce C
k
=

C
l
and R
k
=

R
l
.
For example if
1
=

2
are complex and all other eigenvalues are real,
we then obtain a representation which is analogous to that on slide 9:
e
At
= C
1
e
J
1
t
R
1
+

C
1
e

J
1
t

R
1
+ C
3
e
J
3
t
R
3
+ + C
g
e
J
g
t
R
g
=
= 2Re[C
1
e
J
1
t
R
1
] + C
3
e
J
3
t
R
3
+ + C
g
e
J
g
t
R
g
.
Relation to Eigenvectors
26/65
Another way to read the formulas is
A
_
C
1
C
g
_
=
_
C
1
C
g
_
diag(J
1
, . . . , J
g
)
e
At
_
C
1
C
g
_
=
_
C
1
C
g
_
diag(e
J
1
t
, . . . , e
J
g
t
)
or AC
k
= C
k
J
k
and e
At
C
k
= C
k
e
J
k
t
for all k = 1, . . . , g.
If c
k
denote the rst columns of C
k
we infer from these two relations:
Ac
k
= c
k

k
=
k
c
k
. Hence the collection c
1
, . . . , c
g
is a (largest) set
of linearly independent eigenvectors of A (as on slide 13).
e
At
c
k
= c
k
e

k
t
. Hence with
k
=
k
+ i
k
we infer
e
At
Re(c
k
) = Re(c
k
e

k
t
) = e

k
t
[Re(c
k
) cos(
k
t) Im(c
k
) sin(
k
t)]
e
At
Im(c
k
) = Im(c
k
e

k
t
) = e

k
t
[Im(c
k
) cos(
k
t) + Re(c
k
) sin(
k
t)].
Example
27/65
Consider again the matrix on slide 23. We infer that
e
At
_
_
_
_
_
3 1 1 5 3
2 1 1 4 3
0 0 1 1 1
1 0 1 2 2
2 1 1 3 2
_
_
_
_
_
. .
S
=
=
_
_
_
_
_
3 1 1 5 3
2 1 1 4 3
0 0 1 1 1
1 0 1 2 2
2 1 1 3 2
_
_
_
_
_
_
_
_
_
_
_
e
t
te
t
0 0 0
0 e
t
0 0 0
0 0 e
t
te
t
0
0 0 0 e
t
0
0 0 0 0 e
t
_
_
_
_
_
_
.
If is a linear combination of the rst two columns of S then e
At

converges to zero for t . This follows from

lim
t
e
t
= 0 and lim
t
te
t
= 0.
Asymptotic Stability
28/65
The stability analysis of autonomous systems is related to examining the
asymptotic behavior of x(t) if t . For linear systems the explicit
representation of the solutions clearly allows to answer such questions.
As an illustration we formulate the most important concept as follows.
Denition 8 The (equilibrium 0 of the) linear system x = Ax is
asymptotically stable if all solutions satisfy lim
t
x(t) = 0.
Since convergence should hold for all initial conditions, this is often
called global asymptotic stability. The dynamic property of asymptotic
stability can be veried by a simple algebraic test.
Theorem 9 The system x = Ax is asymptotically stable if and only
if all the eigenvalues of A have a negative real part. Matrices A with
this property are called Hurwitz.
Proof
29/65
1) Suppose that all eigenvalues of A have negative real parts. If =
+i is such an eigenvalue we have < 0. For k N
0
, t 0 we infer
|t
k
e
t
| = |e
t
[t
k
e
it
]| = e
t
t
k
0 for t .
Since all elements of e
At
are linear combinations of such terms we get
lim
t
e
At
= 0 and hence lim
t
e
At
= 0 for all R
n
.
2) Conversely suppose that = +i is an eigenvalue of A with 0
and with eigenvector v = 0. Then e
At
v = ve
t
and hence
x(t) = e
At
Re(v) = e
t
[Re(v) cos(t) Im(v) sin(t)].
Since either Re(v) = 0 or Im(v) = 0 and since 0, we infer x(t) 0
for t . Hence we have found a real initial condition (either Re(v)
or Im(v)) for which the solution does not converge to zero for t .
Example
30/65
Consider the mechanical system ([AM] pages 142-145) depicted in
and described by
d
dt
_
_
_
_
q
1
q
2
q
1
q
2
_
_
_
_
=
_
_
_
_
0 0 1 0
0 0 0 1

2k
m
k
m

c
m
0
k
m

2k
m
0
c
m
_
_
_
_
_
_
_
_
q
1
q
2
q
1
q
2
_
_
_
_
.
The coordinate change z
1
=
1
2
(q
1
+ q
2
), z
2
= z
1
, z
3
=
1
2
(q
1
q
2
),
z
4
= z
4
leads to
_
_
_
_
z
1
z
2
z
3
z
4
_
_
_
_
=
_
_
_
_
0 1 0 0

k
m

c
m
0 0
0 0 0 1
0 0
3k
m

c
m
_
_
_
_
=
_
A
1
0
0 A
2
_
. .
A
_
_
_
_
z
1
z
2
z
3
z
4
_
_
_
_
.
Example
31/65
The modes of the system are the eigenvalues of
A
1
=
_
0 1

k
m

c
m
_
and A
2
=
_
0 1

3k
m

c
m
_
.
If all constants m, c, k are positive these are Hurwitz. Therefore the
system on the previous slide is asymptotically stable.
A real 2 2 matrix A is Hurwitz i trace(A) < 0 and det(A) > 0.
From the denition of the matrix exponential we also infer
e
At
=
_
e
A
1
t
0
0 e
A
2
t
_
.
If z
3
(0) = 0, z
4
(0) = 0 then z
3
(t) = 0, z
4
(t) = 0 as well as
_
z
1
(t)
z
2
(t)
_
= e
A
1
t
_
z
1
(0)
z
2
(0)
_
.
Since q
1
(t) = q
2
(t) this is the movement in Mode 1 (see gure).
Example
32/65
If there is no damping we have c = 0. With
1
=
_
k
m
and
2
=
_
3k
m
we then infer
e
A

t
=
_
cos(

t)
1

sin(

t)

sin(

t) cos(

t)
_
for = 1, 2.
All system motions are sustained oscillations. We can as well distinguish
the two modes of movement as indicated in the gure:
0 5 10
1
0.5
0
0.5
1
0 5 10
1
0.5
0
0.5
1
Systems whose solutions do not decay to zero but only stay bounded
for future times, as in this example, have their own interest.
Lyapunov Stability
33/65
Often we are only interested in analyzing whether the solutions stay
bounded for t and do not explode. This is related to what is
called Lyapunov stability or neutral stability.
Theorem 10 All solutions of the system x = Ax are bounded for
t i all eigenvalues of A have a non-positive real part, and all
Jordan blocks of eigenvalues with real part zero have dimension 1.
Lyapunov stability boils down to e
At
being bounded for t .
Hence solutions that start close to zero stay close to zero.
Asymptotic stability just means lim
t
e
At
= 0. Note that the decay is
actually exponential.
All other more rened stability questions (decay rates, frequency of os-
cillations) can be obtained by directly analyzing e
At
.
Nonlinear Systems
34/65
From now on let us consider
x = f(x) with f C
1
(G, R
n
) and open G R
n
.
We denote by (., ) the solution of the ivp with x(0) = G.
Denition 11 The equilibrium x
e
G of x = f(x) is said to be
stable, if for any > 0 there exists > 0 such that
G, x
e
(t, ) x
e
for all t [0, ).
unstable if not stable.
attractive, if there exists some > 0 such that
G, x
e
lim
t
(t, ) = x
e
.
asymptotically stable, if it is stable and attractive.
Note that these are all local concepts!
Geometric Visualization
35/65
Stability:

Asymptotic Stability:

Lyapunov Functions
36/65
Denition 12 A continuously dierentiable function V : G R is
said to be a Lyapunov function for the nonlinear ode x = f(x) if

V (x) :=
x
V (x)f(x) 0 for all x G.
This is an algebraic property which is purely expressed in terms of the
function f and the gradient of V without involving any system trajectory.
As the key feature, this property implies for any state-trajectory of the
nonlinear ode in G, by the chain-rule, that
d
dt
V (x(t)) =
x
V (x(t)) x(t) =
x
V (x(t))f(x(t)) 0.
Therefore along all state-trajectories of x = f(x) in G, the function
t V (x(t)) is monotonically non-increasing for increasing times.
This feature turns V into an potential-function. Intuitively, state-
trajectories should hence converge to points in which V is minimal.
Lyapunov Stability Theorem
37/65
This is made precise in the following Lyapunov theorem.
Theorem 13 Let V be a Lyapunov function for x = f(x) with
equilibrium x
e
G.
1. If V (x) > V (x
e
) for all x G\ {x
e
} then x
e
is stable.
2. If V (x) > V (x
e
) and

V (x) < 0 for all x G \ {x
e
} then x
e
is asymptotically stable.
One can assume w.l.o.g. that V (x
e
) = 0. Functions with V (x) > 0
for all x G\ {x
e
} are called positive denite on G.
Guaranteeing stability by searching for a Lyapunov function is the
so-called Direct Method of Lyapunov. There are no general recipes
for doing this. Dont forget: The stability properties are local!
In practice one often reduces G to an open ball around x
e
, which
facilitates to fulll the conditions.
37/65
Proof
Notation: U
r
:= {x R
n
| x x
e
< r} and

U
r
:= {x R
n
| x x
e
r}.
Suppose r > 0 is taken with

U
r
G. If we choose (0, r), we infer that
m := min{V (x) | x

U
r
, x x
e
} satises m > V (x
e
).
If we hence x any (V (x
e
), m) we conclude that
M := {x R
n
| x U
r
, V (x) } is compact in R
n
and M U

.
Proof: x M x

U
r
and V (x) < m xx
e
< x U

. Hence M
is bounded. We also infer {x R
n
| x

U
r
, V (x) } = M; hence M is closed.
This implies that M is positively invariant:
M (t, ) M for all t 0.
Proof: For M let (t

, t
+
) be the interval of existence of (., ) w.r.t. U
r
. Due
to V () we infer V ((t, )) for all t [0, t
+
) and thus (t, ) M for
all t [0, t
+
). Since this implies (t, ) U

for t [0, t
+
) and since < r, (t, )
cannot approach the boundary of U
r
which implies t
+
= .
Since V (x
e
) < , there exists > 0 with

U

M

U

. This implies stability:

U

M (t, ) M

U

for t 0.
Example
38/65
Consider the unactuated mass-spring-damper system from Lecture 1:
_
x
1
x
2
_
=
_
x
2

k
m
x
1

1
m
c(x
2
)
_
= f(x
1
, x
2
) with c(.) C
1
(R).
Suppose that c(x
2
) = 0 exactly for x
2
= 0 such that x
e
= (0, 0) is the
unique equilibrium. Motivated by considering the total energy dene
V (x
1
, x
2
) =
1
2
kx
2
1
+
1
2
mx
2
2
.
Note that V (x) > V (0) = 0 for all x = 0. Moreover

x
V (x)f(x) =
_
kx
1
mx
2
_
_
x
2

k
m
x
1

1
m
c(x
2
)
_
= x
2
c(x
2
).
Suppose that c() satises, in addition, x
2
c(x
2
) 0 for all x
2
R.
This implies that x
e
= 0 is Lyapunov stable.
We are not able to conclude asymptotic stability from our result.
Lasalles Invariance Principle
39/65
Theorem 14 Let V be a Lyapunov function for x = f(x) with
equilibrium x
e
G.
1. Let V (x) > V (x
e
) for all x G\ {x
e
}.
2. Let G and

V ((t, )) = 0 for all t (t

, t
+
) imply = x
e
.
Then x
e
is asymptotically stable.
Statement 2. trivially implies Statement 2. in Theorem 13.
Example continued
Suppose that c() satises x
2
c(x
2
) > 0 for all x
2
R \ {0}. Then
x
e
= 0 is asymptotically stable.
Indeed x
2
(t)c(x
2
(t)) = 0 for t (t

, t
+
) implies x
2
(t) = 0 and hence
x
1
(t) = 0 for all t (t

, t
+
) due to the (second) dierential equation.
39/65
Proof
Take M from the proof of Theorem 13. If M it suces to show lim
t
(t, ) = x
e
.
Suppose not. Since M is compact, there exists t

and x M U
r
G with
lim

(t

, ) = x = x
e
.
Since V is continuous on M and hence bounded from below, V ((t, )) is for t 0
bounded from below. By monotonicity, the following limit hence exists:
l := lim
t
V ((t, )) R.
Now consider (t, x) for some t (t

, t
+
). Because of the continuous dependence
on initial conditions and the semi-group property we infer
(t, x) = lim

(t, (t

, )) = lim

(t +t

, ).
We thus infer with the continuity of V und and the denition of l:
V ((t, x)) = lim

V ((t +t

, )) = l.
Here is the crux: Since V ((t, x)) = l for all t (t

, t
+
), we infer
d
dt
V ((t, x)) = 0
and thus

V ((t, x)) = 0 for all t (t

, t
+
). By the hypothesis in 2. we conclude
x = x
e
, a contradiction.
Estimation of Region of Attraction
40/65
Theorem 15 Let V be a Lyapunov function for x = f(x) with
equilibrium x
e
G.
1. For R let M := {x G | V (x) } be compact in R
n
.
2. Let M and

V ((t, )) = 0 for t (t

, t
+
) imply = x
e
.
For all M we can then conclude lim
t
(t, ) = x
e
.
The sublevel-set M comprises points that are attracted by x
e
. In other
words, M is a subset of the region of attraction of x
e
, dened as
{ G | lim
t
(t, ) = x
e
}.
Note that M can be large and that this is not a local result any more.
Also note that we do not assume/claim stability of x
e
.
Proof
40/65
M is positively invariant:
M (t, ) M for all t 0.
Proof. If M then V () and thus V ((t, )) , i.e. (t, )
M for all t [0, t
+
).
We claim t
+
= : If t
+
< and since M is compact, there exists t

t
+
with (t

, ) x
+
M; by the EES (t
+
, x
+
) is on the boundary of
R G; hence x
+
G; on the other hand, since x
+
M =

M G
and since G is open, we get a contradiction.
Now the proof is concluded exactly as that of Theorem 14.
Global Attractivity
41/65
If the Lyapunov function is a barrier-function for G we can infer global
attractivity of the equilibrium.
Theorem 16 Let V be a Lyapunov function for x = f(x) with
equilibrium x
e
G.
1. For any sequence x

G with x

x G or x

suppose that V (x

) for .
2. Let G and

V ((t, )) = 0 for t (t

, t
+
) imply = x
e
.
Then lim
t
(t, ) = x
e
for all G.
If G = R
n
then 1. reduces to the condition V (x) for x ;
such Lyapunov functions are said to be radially unbounded.
Proof
41/65
For G we trivially have M := {x G | V (x) V ()}. Let us
show that M is compact in R
n
.
If M is unbounded, there exists a sequence x

M with x

; by hypothesis, V (x

) such that there exists some mit

V (x

) > V (); this contradicts x

M.
If M is not closed, there exists a sequence x

M with x

x R
n
and x M. In case of x G, we could conclude by continuity of V
that V (x) = lim

V (x

) V () and thus x M. Therefore, x

is not contained in G but in

G, which implies x G. By hypothesis,
we have V (x

) , which again contradicts V (x

) V ().
Hence M is closed and bounded and thus compact. We can apply Theo-
rem 15 and conclude (t, ) x
e
for t .
Indirect Method of Lyapunov
42/65
The linearization of x = f(x) at x
e
was given by
x

= Ax

with A =
x
f(x
e
).
Recall that we hope for x(t) x
e
+ x

(t): A solution of the linear

system leads to a good approximation of that of the nonlinear system.
This hope is reality, at least for the property of local asymptotic stability.
Theorem 17 Suppose that
x
f(x
e
) is Hurwitz. Then x
e
is a locally
asymptotically stable equilibrium of x = f(x).
Colloquially, asymptotic stability of the linearization also implies local
asymptotic stability of the original non-linear system around the point
of linearization. The converse is not true!
This is a major motivation why to study linear systems in such depth!
Example
43/65
Consider again the example on slide 38. The linearization at x
e
= 0 is

x
f(x) =
_
0 1

k
m

1
m
c

(x
2
)
_
and thus A =
_
0 1

k
m

1
m
c

(0)
_
.
Recall that
_
0 1

_
is Hurwitz i > 0 and > 0.
Hence the matrix A of the linearization is Hurwitz if and only if c

(0) > 0.
x
e
= 0 is locally asymptotically stable in case that c

(0) > 0.
Note that c

(0) > 0 implies x

2
c(x
2
) > 0 for x
2
= 0 close to zero.
The converse is not true. This shows that Lasalles principle is a more
powerful tool to prove asymptotic stability.
Variation-of-Constants-Formula
44/65
Lets come back to linear systems with inputs and outputs described as
x = Ax + Bu and y = Cx + Du
with A R
nn
, D R
km
. There exists an explicit formula for the
state- and output-response of such systems.
Theorem 18 For the input u C([a, b], R
m
) and the initial condi-
tion x() = R
n
, [a, b], the unique state-response equals
x(t) = e
A(t)
+
_
t

e
A(ts)
Bu(s) ds.
The output-response hence is
y(t) = Ce
A(t)
+
_
t
0
[Ce
A(ts)
B]u(s) ds + Du(t).
Hence the responses are given by a convolution. W.l.o.g. we take = 0.
Proof
45/65
Important Trick: Search for solution x(t) = e
At
z(t) with suitable z(t).
Dierentiation implies
x(t) = Ae
At
z(t) + e
At
z(t) = Ax(t) + e
At
z(t).
This equals Ax(t) + Bu(t) as desired if
e
At
z(t) = Bu(t) or z(s) = e
As
Bu(s).
Hence for any constant vector c the function
z(t) = c +
_
t

e
As
Bu(s) ds
does the job. c = leads to satisfaction of x(0) = .
Note that this trick of varying the constants can often be directly
applied for nding responses to explicitly given input signals.
Example
46/65
Consider a constant input: u(t) = u
e
for all t R. Then
x(t) = e
At
+
__
t
0
e
A(ts)
ds
_
Bu
e
= e
At
+
__
t
0
e
A
d
_
Bu
e
.
If A is Hurwitz then it is invertible and
_
t
0
e
A
d =
_
t
0
d
d
e
A
A
1
d = e
At
A
1
A
1
.
The state-response can hence be written with the matrix exponential as
x(t) = e
At
[ + A
1
Bu
e
] A
1
Bu
e
.
For t we conclude x(t) A
1
Bu
e
=: x
e
. Therefore the state
converges to the unique solution of Ax
e
+ Bu
e
= 0. Moreover
y(t) = Ce
At
[ + A
1
Bu
e
] + [D CA
1
B]u
e
.
This nicely displays the transient and the steady-state response. The
matrix D CA
1
B is called steady-state gain of the system.
Superposition Principle
47/65
Clearly the matrix exponential e
At
determines the state-response both to
non-zero initial conditions and to forcing inputs. The response depends
linearly on both and u(). This is called superposition principle.
If the system has multiple inputs (m > 1) we can partition
u(t) =
_
_
u
1
(t)
.
.
.
u
m
(t)
_
_
, B =
_
B
1
B
m
_
, D =
_
D
1
D
m
_
.
By the superposition principle the full output response is the sum of
Ce
At
and
_
t
0
Ce
A(ts)
B
k
u
k
(s) ds + D
k
u
k
(t), k = 1, . . . , m
and each of these contributions can be analyzed separately.
Step and Impulse Response
48/65
Let u
k
() be equal to the step function s(t) = 1 for t 0. Then
_
t
0
Ce
A(t)
B
k
s() d + D
k
s
k
(t) =
_
t
0
Ce
A
B
k
d + D
k
is the step response of the system (in the k-th input).
If u
k
() is equal to the impulse () at zero then
_
t
0
Ce
A(t)
B
k
() d + D
k
(t) = Ce
At
B
k
+ D
k
(t)
is the impulse response of the system (in the k-th input). Observe
that it is the derivative of the step response.
Denition 19 The step- and impulse-response of the system are
_
t
0
[Ce
A
B] d + D and Ce
At
B + D(t)
and can be obtained by applying m steps/impulses for each input.
Step and Impulse Response
49/65
Steps and impulses should be viewed as test signals that allow to gather
information about the dynamical behavior of a system.
Both responses can be obtained by exiting the system with m well-
dened input signals from initial condition zero. Knowledge of the im-
pulse response
H(t) = Ce
At
B + D(t)
then allows to determine the response for any other signal by
_
t
0
H(t )u() d.
This is a principle path in order to determine a system description
from a nite number of (experimental) responses. This is the topic
in courses on system identication.
Sinusoidal Inputs
50/65
For a complex = + i C and u
e
R
m
consider the input
u(t) = u
e
e
t
= u
e
e
t
[cos(t) + i sin(t)].
Suppose I A is invertible. As on slide 46 we then have with = t:
y(t) = C
_
e
At
+
__
t
0
e
A(t)
e

d
_
Bu
e
_
+ D(u
e
e
t
) =
= C
_
e
At
+ e
t
__
t
0
e
(AI)
d
_
Bu
e
_
+ D(u
e
e
t
) =
= C
_
e
At
+ e
t
_
e
(AI)t
I

(A I)
1
Bu
e
_
+ D(u
e
e
t
).
Reordering leads to the most important formula
y(t) = Ce
At
_
+ (A I)
1
Bu
e

+
_
C(I A)
1
B + D

(u
e
e
t
)
which displays the transient and steady-state response. (Terminology
sensible if A is Hurwitz - then the transient decays to zero for t .)
Sinusoidal Inputs: Summary
51/65
Theorem 20 For exponentially weighted sinusoidal complex input
u(t) = u
e
e
t
= u
e
e
t
[cos(t) + i sin(t)] ( = + i)
such that I A is invertible, the state-response is
x(t) = e
At
_
(I A)
1
Bu
e

+
_
(I A)
1
B

(u
e
e
t
)
and the output-response is
y(t) = Ce
At
_
(I A)
1
Bu
e

+
_
C(I A)
1
B + D

(u
e
e
t
).
Since A, B, C, D and are real, the responses to the inputs
v(t) = u
e
e
t
cos(t) and w(t) = u
e
e
t
sin(t)
are just obtained by taking the real and imaginary parts.
Remark. For = 0 (pure sinusoidal inputs) the steady-state respon-
se involves the frequency response matrix C(iI A)
1
B + D.
The Laplace Transform
52/65
Suppose that f : R C is measurable and of exponential type: There
exists a constant such that
e
t
f(t) is bounded for t 0.
Denition 21 For all s in the half-plane {s C : Re(s) > }, the
one-sided Laplace transform of f is given by

f(s) = L(f)(s) =
_

0
e
st
f(t) dt.
It is easy to show that

f = L(f) is analytic on the above given half-plane
(that can often be extended to much larger regions of C).
We rather consider this integral transform as an map which takes f into

f := L(f). L is obviously linear. Moreover L is injective. We only hint

at very few more properties of the Laplace transform.
The Laplace Transform
53/65
The following elementary facts suce for our purposes.
Suppose

f is the Laplace transform of f. Then
1. f

has the Laplace transform s

f(s) f(0).
2.
_
t
0
f() d has the Laplace transform
1
s

f(s).
3. e
pt
f(t) has the Laplace transform

f(s + p).
By iteratively applying these rules one can nd pairs of functions that
are related through the Laplace transformation. An example is
1
(m1)!
t
m1
e
pt
and
1
(s + p)
m
.
Computing Responses with Laplace Transform
54/65
Taking the transforms of all system signals on slide 44 leads to
s x(s) x(0) = A x(s) + B u(s), y(s) = C x(s) + D u(s).
We can hence algebraically solve for x(s) to get with x(0) = that
x(s) = (sI A)
1
+ (sI A)
1
B u(s),
y(s) = C(sI A)
1
+ [C(sI A)
1
B + D] u(s).
This is the so-called frequency-domain analogue of the time-domain
formulas on slide 44. It is a key observation that the convolution integral
(in the time-domain) is just replaced by a simple multiplication (in the
frequency-domain).
The inverse Laplace transform often allows to determine x(t) and y(t).
Transfer Matrices
55/65
Given the system on slide 44 we have seen that the matrix
G(s) = C(sI A)
1
B + D
depending the the complex variable s C plays a major role.
For s C not being an eigenvalue of A we can compute (sI A)
1
.
The elements of (sI A)
1
are rational functions of s, since
(sI A)
1
=
1
(s)
adj(sI A) with (s) = det(sI A)
where adj is the algebraic adjoint. In fact each element can be written
as
n
ij
(s)
(s)
with a polynomial n
ij
(s) of degree strictly smaller than that
of (s). Such rational functions are called strictly proper.
The elements of C(sI A)
1
B+D are linear combinations of those
of (sI A)
1
plus a constant matrix D.
Transfer Matrices
56/65
Denition 22 C(sI A)
1
B + D is the transfer matrix of the
system with state-space description x = Ax + Bu, y = Cx + Du.
All elements of G(s) = C(sI A)
1
B + D are rational functions of
s whose numerator degree is not larger than that of the denominator.
Such rational functional are called proper and G(s) is said to be a
proper rational matrix. In Matlab G(s) is computed by tf.
The poles of the transfer matrix G(s) are determined as follows:
Write each element of G(s) in the form
n
ij
(s)
d
ij
(s)
where the numerator
and denominator polynomials have no common zeros.
The union of all zeros of d
ij
(s) for all elements are the poles of G(s).
Denition 23 G(s) is stable if all its poles have negative real parts.
Example
57/65
Consider the electrical network from [F] p.21
and described by
_
v
1
v
2
_
=
_
1
C
1
R
1
+
1
C
1
R
3

1
C
1
R
3

1
C
2
R
3
1
C
2
R
2
+
1
C
2
R
3
__
v
1
v
2
_
+
_
1
R
1
C
1
1
R
2
C
2
_
u.
With y = v and C
1
= 1, C
2
= 2, R
1
= 1, R
2
= 2, R
3
= 3 we have
G(s) =
_
s+0.5
s
2
+1.75s+0.5
0.25s+0.5
s
2
+1.75s+0.5
_
.
The poles are the zeros of s
2
+ 1.75s + 0.5 which are 1.39, 0.36.
G(s) is stable. The steady-state gain is G(0) = DCA
1
B = (1 1)
T
.
Example
58/65
Lets change C
1
in steps from 1 to 4 and consider the step responses
from zero initial conditions:
0 10 20 30
1
0.5
0
0.5
1
v
1
0 10 20 30
1
0.5
0
0.5
1
v
2
For C
1
= 4 the transfer matrix is
G(s) =
_
0.25s+0.125
s
2
+0.75s+0.125
0.25s+0.125
s
2
+0.75s+0.125
_
=
_
0.25
s+0.25
0.25
s+0.25
_
.
We had two dierent poles for C
1
= 1 which reveal themselves in the
dierent time-constants of the response. For C
1
= 4 only one pole
{0.25} remains related to the red response. Why? Well explain later!
Relation to Classical Techniques and Identication
59/65
Lecture 1
First-order
state-space
system
description
System of
higher order
dierential
equations
Transfer
function/matrix
description
Laplace & Inverse
(Classical)
Laplace
(Just Done)
???
Realization
Theory
Identication
Experiments
Working with Transfer Matrices: Summary
60/65
In view of slide 54 the transfer matrix is most useful in case that the
initial condition is zero. For a given input the transfer matrix fully
determines the output response.
In addition, if two systems have transfer matrices G
1
(s) and G
2
(s) then
the parallel and series interconnections have transfer matrices
G
1
(s) + G
2
(s) and G
1
(s)G
2
(s)
with the usual matrix operations. Note that the dimensions must be
compatible, and the order of the multiplication is important to be kept!
Moreover if A is Hurwitz and is not an eigenvalue of A then
G(0)u
e
, G(i)e
it
u
e
, G()e
t
u
e
are steady-state responses for constant ( = 0), purely sinusoidal ( =
i) and exponentially weighted sinusoidal inputs in the direction u
e
.
Block-Diagrams
61/65
Block-diagrams for systems translate into block-diagrams for transfer
matrices. Let us stress the similarity to diagrams for linear maps in
linear algebra or functional analysis:
Series interconnection: y(s) = [G
1
(s)G
2
(s)] u(s) is depicted as
u(s) y(s)
G
2
(s) G
1
(s)
u
1
(s) = y
2
(s)
Notice the order of the blocks!
Parallel interconnection: y(s) = [G
1
(s) + G
2
(s)] u(s) is depicted as
y(s) + u(s)
G
2
(s)
G
1
(s)
y
1
(s)
y
2
(s)
Realizations
62/65
A state-space system (uniquely) determines its transfer matrix by direct
computation. The converse question reads as follows.
Problem: Let us be given a k m-matrix G(s) whose elements
consist of proper rational functions. Can we compute matrices A
R
nn
, B R
nm
, C R
kn
and D R
km
such that
G(s) = C(sI A)
1
B + D ?
Let us introduce the following important terminology.
Denition 24 If G(s) = C(sI A)
1
B + D then (A, B, C, D) is
said to be (state-space) realization of G(s).
The above formulated question is called the realization problem. In
Matlab just use ss. The theory will be discussed in detail in Lecture 5.
Coordinate Change
63/65
The matrices (A, B, C, D) of a realization of G(s) are never unique.
One reason (there are many more!) is as follows.
A state-coordinate change modies the describing matrices of a
state-space system but it does not change the transfer matrix.
More precisely, for the state-space system
x = Ax + Bu, y = Cx + Du
perform the coordinate change z = Tx. This leads to
z =

Az+

Bu, y =

Cz+

Du with
_

A

B

C

D
_
=
_
TAT
1
TB
CT
1
D
_
while we have

C(sI

A)
1

B +

D = C(sI A)
1
B + D.
Proofs
64/65
If we perform for x = Ax + Bu, y = Cx + Du the coordinate change
z = Tx, we infer x = T
1
z and we can thus conclude
z = T x = TAx+TBu = [TAT
1
]z+[TB]u and y = [CT
1
]z+Du.
For

A = TAT
1
,

B = TB,

C = CT
1
,

D = D this is nothing but
z =

Az +

Bu, y =

Cz +

Du.
Invariance of the transfer matrix is shown as follows:

C(sI

A)
1

B +

D =
= [CT
1
]
_
sTT
1
[TAT
1
]
_
1
[TB] + D =
= CT
1
T (sI A)
1
T
1
TB + D =
= C(sI A)
1
B + D.
Covered in Lecture 2
65/65
Analysis of autonomous linear systems
coordinate change, diagonalization, Jordan form, modes, all solutions
matrix exponential
Stability and Lyapunov functions
Hurwitz matrices, Lyapunov functions, Lyapunov stability theorems,
indirect method
Analysis of LTI systems
Variation-of-constants formula, computation of responses (steady-
state, transient), Laplace
Transfer matrices
denition, working with transfer matrices, realizations, behavior under
state coordinate-change