EEE 582

HW # 1 SOLUTIONS

Problem 1.2
Let i and p i be the i-th eigenvalue and eigenvector of A, so
Ap i = ip i

(1)

1.
A2 p i

=
..
.

AApi = A ( ip i) = 2i p i

Ak p i

AAk1 p i = ki p i

Hence the eigenvalues of Ak are ki for i = 1; :::; n.

2. Multiply both sides of (1) by A1 (all eigenvalues are dierent from zero), then
A1 Ap i = A1 ip i )
Hence the eigenvalues of A 1 are

1
i

1
p i = A1 p
i

for i = 1; :::; n.

3. The eigenvalues of A> are given by the roots of

h
i

>
>
det I A> = det (I A) = det [(I A)]
and for any square matrix X, det(X) = det(X > ), nally

det I A> = det (I A)

Hence the eigenvalues of A> are i for i = 1; :::; n.

4. Let AH = A> (conjugate transpose),

> = det I A
det I AH = det I A
then

A = det I A
det I AH = det I A = det I

i for i = 1; :::; n.
Hence the eigenvalues of AH are
5.

Ap i = (Ap i) = i pi
Hence the eigenvalues of A are i for i = 1; :::; n.
6. In general the eigenvalues of A> A does not relate nicely with the eigenvalues of A. For the special case
when A = A> (symmetric matrices), the eigenvalues of A>A are 2i for i = 1; :::; n.
Problem 1.10
Q symmetric (Q> = Q), Q> Q = Q 2 ) eigenvalues of Q2 are 2i for i = 1; :::; n.
p
kQk = max (Q2 ) = max ji j
i

From

> >
x Qx x Q kxk = kQxk kxk kQk kxk = max j ij x > x
i

hence x > Qx kQk for all unit-norm x. Pick x a as a unit-norm eigenvector of Q corresponding to the
eigenvalue that yields maxi ji j (possibly non-unique). Then

>

x a Qx a = x>
max
j
j
x a = max j ij
i
a
i

thus,

max x> Qx = kQk

kxk=1

Problem 1.15
Q symmetric with 1 , 2 such that

0 1 I Q 2 I
we know
0 < min (Q) x> x x> Qx max () x> x
Pick x as an eigenvalue corresponding to min (Q) and max (Q) respectively then
1 min (Q) ; max (Q) 2
Therefore

min Q 1 =

For Q 1 positive denite

1
1

max (Q)
2

1
1
max Q 1 =

min (Q)
1
0

1
1
I Q 1 I
2
1

Problem 1.16
W (t) I is symmetric and positive denite 8t, then for any x
x> (W (t) I) x 0 ) x> W (t)x x > Ix
Pick xt be an eigenvector corresponding to an eigenvalue t or W (t)
>
>
x>
t W (t)xt = t x x x x

That is t . This holds for any eigenvalue of W (t) and every t. Since the determinant is the product of the
eigenvalues then
det (W (t)) n > 0
Problem 1.17
Since A(t) is continuously dierentiable and invertible for each t, we can write A 1 (t)A(t) = I, then taking
the derivative with respect to time on both sides of the equation

d 1
A (t)A(t)
dt
_
A_ 1 (t)A(t) + A1 (t)(t)
A_ 1(t)

=
=
=

d
I
dt
0
1
_
A1 (t)A(t)A
(t)

EEE 582

HW # 2 SOLUTIONS

Problem 2.1
y (n) (t) + an 1 (t)y (n 1) (t) + + a0 (t)y(t) = b0 (t)u(t) + b1 (t)u (1) (t)

Let xn (t) = y ( n1) (t) b 1 (t)u(t) and x1 (t) = y(t) then

x_ 1 (t)

= y(t)
_ = x 2 (t)

x_ 2 (t)

= y(t) = x 3 (t)
..
.

x_ n2 (t)

= x n 1 (t)

x_ n1 (t)

= x n + b 1 (t)u(t)

x_ n (t)

(1)

= y (n) (t) b 1 u(t) b 1 (t)u( 1) (t)

=
=

n 1
X
i=0

n
X2
i=0

(1)

ai (t)y (i)(t) + b0 (t)u(t) b 1 (t)u(t)

(1)
ai (t)xi (t) + an1 (t) (x n (t) + b1 (t)u(t)) + b 0 (t) b 1 (t) u(t)

then we can write

A(t)

C(t)

6
6
6
= 6
6
4
=

0
0
..
.
0
a0 (t)

1
0
..
.
0
a1 (t)

0 ;

1 0

D(t) = 0

0
0
..
.
1
an1 (t)

7
7
7
7;
7
5

6
6
6
B(t) = 6
6
4

0
0
..
.
b 0 (t)

b 1 (t)
(1)
b 1 (t) + an 1 (t)b1 (t)

Problem 2.2
y (n) (t) + an 1 t 1 y (n 1) (t) + an2 t2 y (n2) (t) + + a1 tn+1 y (1) (t) + a0 t n y(t) = 0
Let x1 (t) = tn 1 y(t), then
with x 2 (t) = t n+2 y ( 1) (t), so
with x 3 (t) = t n+3 y ( 2) (t), ...

x_ 1(t) = (1 n)t 1 x1 (t) + t 1 x2 (t)

x_ 3(t) = (2 n)t 1 x2 (t) + t 1 x3 (t)
x_ n1 (t) = t1 x n1 (t) + xn (t)

with x n (t) = y (n)(t), nally

x_ n (t) = an1 t1 x n (t) an 1t 1 xn1 a1 t1 x 2 (t) a0 t1 x1 (t)

and we can write x(t)
_
= t1 Ax(t), with
2
6
6
6
A=6
6
4

1n
0
..
.

1
2n
..
.

0
a0

0
a1

0
0
..
.

0
..
.

1
an2

1
an 1

3
7
7
7
7
7
5

3
7
7
7
7
7
5

Problem 2.3
4
1
y(t) + y 3 (t) = u(t)
3
3
with initial conditions y(0) = 0, y(0)
_
= 1 and u(t) = u~(t) = sin(3t). We can write sin(3t) = 3 sin(t) 4 sin3 (t),
then the dierential equation is
4
4
y(t) + y 3 (t) = sin3 (t) sin(t)
3
3
Propose as a solution y(t) = A sin(t), substitute in the dierential equation
A sin(t) +

4 3 3
4
A sin (t) = sin3 (t) sin(t)
3
3

A=1

and it also satises the initial conditons

y(0) = 0
y(0)
_
=1
so y(t) = sin(t) is a solution to

)
)

x_ =

x 2 (t)
43 x 31 (t) 13 u(t)

The linearization around the nominal solution

@f (x; u)
0
A(t) =
=
4x 21
@x x ;u
=

@f (x; u)
@u

x ;u

Finally

cos(0) = 1

4
1
y(t) + y 3 (t) = u~ (t)
3
3
Let x1 = y(t), x 2 = y(t),
_

sin(0) = 0

0
1
3

1
0

= f(x; u)

x1=sin(t)

u = sin(3t)

0
4 sin2 (t)

1
0

x_ (t) = A(t)x (t) + Bu (t)

y (t) =
1 0 x (t)

sin(t)
0
with x (t) = x(t)
, x (0) = x(0)
, u (t) = u(t) sin(3t) and y (t) = y(t) sin(t)
cos(t)
1
Problem 2.8
u = 0, C ~x = u~ , this implies that the matrix
Identity
dc-gain means that for a given u~, 9~x, such that A~x + B~
A B
is invertible.
C 0
1. If K 2 IR mn is such that (A + BK) is invertible, then C(A + BK)1 B is invertible.

A B
A + BK B
Since
is invertible , for any K,
is invertible, this from
C 0
C
0

A + BK B
A B
I 0
=
C
0
C 0
K I
Then

A + BK
C

B
0

R1
R3

R2
R4

I
0

0
I

so
(A + BK)R1 + BR3

(A + BK)R2 + BR4

C R1

C R2

R2 = (A + BK) 1 BR4

C (A + BK) 1 BR4 = I

hence C (A + BK)1 B is invertible.

2. We need to show that there exits N such that
0

(A + BK)~
x + BN u~

u~

C x~

The rst equation gives x~ = (A + BK)1 BN u~ . Thus we need to choose N such that C(A +

1
BK) 1 BN = u~ . From part 1., we take N = C(A + BK) 1 B
.
Problem 2.10
For u(t) = u~, x~ is a constant nominal if and only if 0 = (A + D~
u )~
x + b u~ . This holds if and only if
b 2 Im[A + D~
u ], that is, if and only if rank(A + D~u) = rank A + D~
u b
If A + D u~ is invertible, then
x~ = (A + D~u) 1 b~
u
If A is invertible, then by continuity of the determinant det(A + B~
u) 6
= 0 for all u~ such that ju~ j is suciently
small, equation () denes a corresponding constant nominal. The linerized state equation is
(A + D u~)x (t) + [b D(A + D~
u )1 b u~]u (t)

x_ (t) =
y (t) =

Cx (t)

Problem 3.7
From
r(t) =

v()()d

t0

taking derivative with respect to time r(t)

_ = v(t)(t), and
(t) (t) +

v()()d

t0

(t) (t) + r(t)

multiplying by v(t) 0
(t)v(t) (t)v(t) + v(t)r(t)
| {z }

r(t)
_ r(t)v(t) (t)v(t)

r(t)
_

t
Multiply both sides by exp t0 v( )d ,

r(t)e
_

Rt

t0

v( ) d

Integrating both sides

r(t)e

Rt

v( )d
r(t)v(t)e t0
Rt

v( )d
t
0
r(t)e
dt

Rt

t0

v( )d

v(t)(t)e

v(t)(t)e

v()()e
t0

t0

v( )d

Rt

t0

Rt

t0

v( )d
v( )d

multiplying both sides by exp

t
t0

v( )d

r(t)
r(t)

Z
Z

v()()e
t0

t0

R t0

Rt
v( )d
d e t0

v( )d

Rt
v( )d
v()()e

From (t) (t) + r(t)

(t) (t) +

t0

Rt
v( )d
v()()e
d

Problem 3.12
x(t)
_
= A(t)x(t), x(t0 ) = x0 , integrating both sides and taking norms
x(t) x 0

kx(t)k

kx(t)k

A( )x( )d

t0

Z t

A( )x( )d + x0

t0
Z t

A( )x( )d

+ kx0 k
t0

Using Gronwall-Bellman inequality with (t) = kx(t)k, (t) = kx0 k, v(t) = kA(t)k,
kx(t)k kx0 k + kx0 k
|

Rt
kA( )kd
kA(t)k e
d
t0
{z
}
t

integrating by part s

kx(t)k kx0 k kx0 k + kx 0k e

so

Rt

t0

kA( )kd

Rt
kA( )kd
kx(t)k kx0 k e t0

EEE 582

HW # 3 SOLUTIONS

Problem 4.6
_
The unique solution for X(t)
= X(t)A(t), X(t0 ) = X0 is
X(t) = A (t; t0 )X0
on the other hand, take the transpose system X_ > = A> X > , this also has a solution
X > (t) = A> (t; t0 )X0> ) X(t) = X0 >
A> (t; t 0 )
For the second part, let 1 (t; ), 2 (t; ) be the transition matrices for A1 (t) and A2 (t), respectively. Propose
as a solution
Z t
X(t) = 1 (t; t0 )X0 2 (t; t0 ) +
1 (t; )F () 2(t; )d
t0

taking

d
dt

_
X(t)
= _ 1 (t; t0 )X0 2 (t; t0 ) + 1 (t; t0 ) X0 _ 2 (t; t0 ) + 1 (t; t) F (t) 2 (t; t)
|
{z
} | {z } | {z } |
{z
} | {z }
| {z }
A1 (t)

so

X (t)

A>
2 (t)

X (t)

_
X(t)
= A1 (t)X(t) + X(t)A>
2 (t) + F (t)

To prove uniqueness, pick \two solutions\ and assume them dierent

X_ 1 (t)
X_ 2 (t)

= A1 (t)X1 (t) + X1 (t)A>

2 (t) + F (t)
= A1 (t)X2 (t) + X2 (t)A>
2 (t) + F (t)

produce the dierence between the two

_
Z(t)
= A1 (t)Z(t) + Z(t)A>
2 (t);

Z(t0 ) = 0

with Z(t) = X1 (t) X2 (t). Integrating both sides, taking norms and using Gronwall-Bellman lemma we get
Rt
k A1( )+A>
2 ( )k d
kZ(t)k kZ0 k e t0
(1)
kZ(t)k

(2)

the last inequality imply that Z(t) = 0 for all t which in turn implies that X1 (t) = X2 (t). Hence there is just
one solution.
Problem 4.8
(() Assume A(t)A( ) = A( )A(t) 8t; then
A(t) A( ) A( )A(t) = 0 8t;
integrating both sides

(A(t)A() A()A(t)) d = 0

and since the dierence is zero for all t,

Z
A(t)

()) Assume A(t)

RT

A()d =

A()d =

A()d A(t)

A()d A(t), then

(A(t)A() A()A(t)) d = 0
1

suppose A(t)A( ) = A( )A(t) is false ) A(t)A( ) A( )A(t) 6

= 0, let v 2 IR n 6
=0
Z t

v>
(A(t)A() A()A(t)) d v = 0

and can be written as

v > [A(t)A() A()A(t)] vd = 0

but [A(t)A() A()A(t)] v = f (t; )8t; tau. From the \false assumption and continuity we know that there
exists a neighborhood around 0 (jx 0 j ) for which f (t; ) > . Let < 0 < 0, then
Z

f(t; )d =

and

f (t; )d +
{z
}

0 +

f (t; )d +

=0

0+

f (t; )d >
0

0 +

0+

f (t; )d
{z
}
=0

d > 2 6
=0

by contradiction we are done.

Problem 4.13
Using the fact that

@
@t A (t; )

= A(t) A(t; ), A ( ; ) = I,

@
A11 (t) A12 (t)
11 (t; ) 12 (t; )
A (t; ) =
0
A22 (t)
21 (t; ) 22 (t; )
@t

@
11 (t; )
@t
@
22 (t; )
@t
@
12 (t; )
@t
@
21 (t; )
@t

= A11 (t)11 (t; ); 11 (; ) = I

= A22 (t)22 (t; ); 22 (; ) = I
= A11 (t)12 (t; ) + A12 (t) 22 (t; ); 12 ( ; ) = 0
= A22 (t)21 (t; ); 21 (; ) = 0 ) 21 (t; ) = 0

so
A(t; ) =

11 (t; ) 12 (t; )
0
22 (t; )

writing the dierential equation

_ 12 (t; ) = A11 (t)12 (t; ) + A12 (t) 22 (t; )
The solution to the homogenous equation is 12 (t; ) = 11 (t; ) and the solution to the dierential equation is
12 (t; ) = 11 (t; ) 12(; ) +
| {z }
=0

then

12 (t; ) =

11 (t; )A12 () 22 (; )d

11 (t; )A12 ()22 (; )d

Problem 4.25
From the Peano-Baker formula
Z t+ )
1 Z
X
(t + ; ) = I +
( )d +

t+

A(1 )

k= 2

A(2 )

k1

A(k )dk d1

From the exponential matrix series representation

t ()t +
eAt ()t = I + A

1
X
1 k
A ( )tk
k! t
k=2

with At ()t =

RT

kR(t; )k

A( )d . Let R(t; ) = (t + ; ) eAt( )t , then kR(t; )k = (t + ; ) e At( )t

=

Z t+ )
Z 1
Z k1
1 Z t+

I +
( )d +
A(1 )
A(2 )
A(k )dk d1

k=2

X
1

k
k
t ()t +
I A
At ()t

k!
k= 2

using the triangle inequality

kR(t; )k
kR(t; )k

1 Z
X
k=2
1
X

t+

kA(1 )k

k1

kA(k )k dk d1 +

1
X
2 k k
2 k2 k2
t = 2 t2

t
k!
k!
k=2
k=2

changing variables and noting that

2
k!

1
,
(k2)!

1
X

tk
t()k
A
k!
k=2

for k 2 we get

1
X
1 m m
kR(t; )k t
t = 2 t2 et
m!
2 2

k=2

Problem 5.2a-b
1. The characteristic polynomial is given by det(I A) = 0, 2 + 2 + 1 = 0 hence 1;2 = 1. The
exponential matrix is given by: eAt = 0 (t)I + 1 (t)A. The functions 0 and 1 are given by:
et

= 0 + 1

tet

= 1

repeated eigenvalues

so 1 = te t and 0 = e t + te t. Then
At

t+1
t

t
1t

et

2. The characteristic polynomial is given by det(I A) = 0, ( +1)3 = 0 hence 1;2;3 = 1. The exponential
matrix is given by: eAt = 0 (t)I + 1 (t)A + 2 (t)A2 . The functions 0 , 1 and 2 are given by:
et
tet

= 1 + 22

2 t

= 22

t e
so 2 =

t2 t
2 e ,

= 0 + 1 + 2 2
repeated eigenvalues

1 = te t + t2e t and 0 = e t + te t + t2 e t. Then

2 t
3
e
0
0
eAt = 4 0
e t
0 5
t
te
0
et

Problem 5.7
Taking derivatives on both sides
Z t

d
A
A
e d
dt
0

d At
e I
dt

Ae At

= AeAt

Assume initial condition t = 0 then

Z

e A d = eAt I t=0 ) 0 = 0

hence the right side is equal to the left side and viceversa.
Assume A 1 exists, i.e., det(A) 6
= 0, then pre-multiply both sides by A 1 and post-multiply both sides by
At
1
e I
Z t

At
1
eA d
e I
= A1
0

from the assumption that A

taking the limit as t ! 1

exists, it implies that eAt I 6

= 0 8t and also that 1 < eAt I < 1 8t,
lim

t!1

eA d

i
lim e At I = A 1

t!1

we need that limt!1 eAt I be nite this implies that the eigenvalues of A have negative real part.
Under this condition we can write
Z 1
Z 0
A
1
1

e d = A ) A =
eA d
0

Problem 5.14
R

t
Since A(t) is diagonal then ii (t; ) = exp aii()d , so
(t; ) =

A(t) has period T = , then R =

1
T

e2 + 2 sin 2
0

ln (T; 0) =

P (t) = (t; 0)e

Rt

2
0
=

0
1

e 3 + 2 sin 2
0
3

and

e 2 sin 2t
0

0
1

e 2 sin 2t

Problem 5.16
From the Floquet decomposition, A(t; ) = P (t)eR(t )P 1 ( ) the solution to the dierential equation is
x(t) = P (t)eR(t ) P 1 ( )x( )
with x( ) the initial condition, pre-multiplying both sides of () by P 1 (t) and let z(t) = P 1 (t)x(t), then we
can write () as
z(t) = e R(t ) z( )
which is the solution to the dierential equation
z(t)
_ = Rz(t)
where R is a constant matrix and the change of variables is given by x(t) = P (t)z(t).
4

Problem 5.17
Since A(t) is T -periodic we can write
(t; t0 ) = P 1 (t)eR(t t0) P (t0 )
where P (t) is continuous, T -periodic and invertible at every t. Let S = P 1 (t0 )RP (t0 ), Q(t; t0 ) = P 1 (t)P (t0 ),
then Q(t; t0 ) is continuous, T -periodic and invertible at every t, so Q(t +T; t0 ) = Q(t; t0 ) and R = P (t0 )SP 1 (t0 )
(notice that P (t0 ) is a similarity transformation between R and S
(t; t0 ) = P 1 (t)eR( tt0) P (t0 ) = P 1 (t)P (t0 )e S(tt0 P 1 (t0 )P (t0 ) = Q(t; t0 )eS(t t0)

EEE 582

HW # 4 SOLUTIONS

Problem 6.3
1. a(t) = 0, A(t) is constant and its eigenvalues are 0 and 1, hence the system can not be uniformly
exponentially stable.
2. a(t) = 1, A(t) is constant and its eigenvalues are 1; since both of them have negative real part, the
system is uniformly exponentially stable.
3. a(t) = t
x1 (t) = x01 e

12 (t2 t20 )

t R
t

a()d

x2 (s)ds

t0

x2 (t) = x02 e(tt0 )

The rst term of x1 (t) is not boundded uniformly with respect to t0 (take t0 ! 1), therefore the system
is not UES.
4. a(t) = et
11 (t; ) = exp

es ds

et e

=e

For = 0 we get
lim 11 (t; 0) =

t!1

1
e

This implies that the system is not asymptotically stable.

1; t < 0
5. a(t) =
et ; t 0

For t0 0, this case is as in Exercise 4, and hence the system can not be UES.

Problem 6.7
From A(t) = A> (t), 8t 2 R, x(t)
_
= A(t)x(t) = A> (t)x(t). Let A (t; ) be the state transition matrix
>
of A(t). Then A (; t) is the state transition matrix of A> (t) (Property 4.5). So, for any x0 we have
>
x(t) = A (t; t0 )x0 = >
A (t0 ; t)x0 . Hence, A (t; t0 ) = A (t0 ; t). Multiplying both sides from the left with
>
>
>
A (t; t0 ) we get A (t; t0 )A (t; t0 ) = [A (t; t0 )A (t0 ; t)] = I. So, A (t; t0 ) is uniformly bounded and, in fact,
kA (t; t0 )k = 1. This implies that x(t)
_
= A(t)x(t) is uniformly stable.
Next, for P (t) to be a Lyapunov transformation we need that

kP (t)k < 1 ; P 1 (t) < 2

Since P (t) = A (t; 0), we have kP (t)k = 1. On the other hand, P 1 (t) = A (0; t) and, from the previous
expression, we have again that kP 1 (t)k = 1. Hence P (t) is a Lyapunov transformation.

Problem 6.8
())
Assume x_ = A(t)x(t) is UES, then 9; > 0 such that kA (t; )k e(t ) . The state transition matrix
for z(t)
_ = A> (t)z(t) is >
A (; t). Then

>
(; t) = kA (; t)k e( (t)) )
A

>
(; t) e(t )
A
1

and this implies that the linear state equation z(t)

_ = A> (t)z(t) is UES.
(()

Assume z(t)
_ = A> (t) is UES, then 9; > 0 such that A> (t) (t; ) e(t ) . But the state transition
matrix for x(t)
_
= A(t)x(t) is >
A> (t) (; t). Then

>
A> (t) (; t) = A> (t) (; t) e( (t))

A> (t) (; t) e(t )

and this implies that the linear state equation x(t)

_
= A(t)x(t) is UES.
Problem 6.11
We know that

1
>
max A + A
kx(t)k = kx0 k exp
2
so if A + A> < 0 implies that < fi (A)g < 0 for i = 1; :::; n, |n being the dimension of the matrix A| and
the system is UES.
1
1
1
1>
1
1
Since F = F > > 0 we can factorize F = F 2 F 2 such that F 2 = F 2 > 0, let z = F 2 x, then x = F 2 z and
the dierential equation becomes
1

F 2 z_ = F AF 2 z
1

z_ = F 2 AF 2 z
Next,
1
1
1
1
1
1
1
1
F 2 AF 2 + F 2 A> F 2 = F 2 A + A> F 2 = F 2 A + A> (F 2 )> < 0
| {z }
<0

Therefore < fi (F A)g < 0 and the system is UES.

Problem 6.13
x(t)
_
= A(t)x(t) US implies that 9 > 0 such that k(t; )k .
Z t
x(t) = (t; t0 )x0 +
(t; )f ( )d
t0

kx(t)k k(t; t0 )k kx0 k + inttt0 k(t; )k kf ( )k d

Z t
kf ( )k d
kx(t)k kx0 k +
t0

Thus, if
lim

t!1

t0

kf (t)k dt < 1

then x(t) is bounded.

In general, (arbitrary A; f ) this condition is also necessary: Let A(t) be a constant matrix equal to zero and
f (t) having the same sign for all t. Then if this condition is violated, x(t) is unbounded.
Problem 7.1
For US wee need
I Q(t) I

_
0
A (t)Q(t) + Q(t)A(t) + Q(t)
>

Pick Q(t) = I, then I Q(t) I with = = 1, and

A> (t) + A(t) +0 0
{z
}
|
=0

00

There is not a Q(t) that results in UES. As a counter-example, pick A(t) = 0 that satises the hypothesis
but is not UES.
Problem 7.8
For US the set of conditions is derived from
I Q(t) I

_
A> (t)Q(t) + Q(t)A(t) + Q(t)
0
which are
a1 (t)
a_ 1 (t) 0
0 a2 (t)

For UES the set of conditions is derived from

I Q(t) I
_
A (t)Q(t) + Q(t)A(t) + Q(t)
I
>

which are
a1 (t)
a_ 1 (t)

a2 (t)
2
Problem 7.9 For UES the set of conditions is derived from
I Q(t) I
_
A (t)Q(t) + Q(t)A(t) + Q(t)
I
>

which are
1
0<< p
2
2 + 1
=

1
a(t)
2
a2 (t) 2a3 (t) a(t)
_
a(t)
Problem 7.11
We can write the equation as

>

A + I Q + Q (A + I) = M

By Theorem 7.11 we conclude that all eigenvalues of A + I have negative real part, that is, if
0 = det (I (A + I)) = det (( ) I A)
3

then < [] < 0. Since > 0 ) < [ ] < , that is, all the eigenvalues of A have real parts strictly less than
.
Suppose all eigenvalues of A have real parts strictly less than . Then, as above, all eigenvalues of A + I
have negative real part. Then, by Theorem 7.1, given asymmetric
positive denite matrix M , there exists a
unique, symmetric, positive denite matrix Q such that A> + I Q + Q (A + I) = M holds. This implies
that A> Q + QA + 2Q = M holds.
Problem 7.12
Substitute Q in A> Q + QA = M such that
Z t

Z t
At
A> t
At
A>
A
A>
A
> A> t
>
Qe + e
Qe A + A
e
M e d +
e
M e d A = M
A e
0
{z 0
}
|
eA> M eA

>

>

>

A> eA t QeAt + eA t QeAt A + eA t M eAt M = M

>
>
>
A> eA t QeAt + eA t QeAt A + eA t A> Q QA eAt = 0
h
i

>
>
>
A> eA t eA t A QeAt + eA t Q eAt A AeAt = 0
|
{z
}
{z
}
|
=0

=0

0=0

Problem 7.16
For an arbitrary but xed t 0, let xa be such that

kxa k = 1; eAt xa = eAt

By Theorem 7.11 the unique solution of QA + A> Q = M is the symmetric, positive denite matrix
Z 1
>
Q=
eA M eA d
0

Then we can write

Z 1

A>
x>
M eA xa d
ae

A
x>
ae

>

M eA xa d = x>
a Qxa max (Q) = kQk

Making a change of integration variable from to = t,

Z

A>
x>
M eA xa d
ae

A> (t+ )
x>
M eA(t+ ) xa d
ae

Hence,

and, since t is arbitrary,

A> t
x>
QeAt xa
ae

At 2
e
kQ1 k

kQk

p
max eAt kQk kQ1 k
t0

At 2 eAt 2
min (Q) e xa =
kQ1 k

EEE 582

HW # 5 SOLUTIONS

Problem 8.6
The solution to the dierential equation for any x0 , t0 0 is
Z t
A (t; )F ()x()d
x(t) = A+F (t; t0 )x0 = A (t; t0 )x0 +
T0

since x(t)
_
= A(t)x(t) is UES, there are > 0, > 0 such that
kx(t)k e(tt0 ) kx0 k +
T

kx(t)k e

t0

t0

kx0 k +

e(t) kF ()k kx()k d

t

t0

kF ()k e kx()k d

Using Gronwall-Bellman lemma

t

e kx(t)k e kx0 k exp

t0
(tt0 )

kx(t)k e

kF ()k d

kx0 k

then
kx(t)k 1 e(tt0 ) kx0 k

with 1 = e

Problem 8.7
Since F (t) is continuous, we can partition the interval [t0 ; t] such that t0 < t1 < t and kF ()k <
and kF ()k < for t0 < < t1 . Then
kx(t)k kA (t; t0 )k kx0 k +

t1
t0

kA (t; )k kF ()k kx()k d +

kx(t)k e(tt0 ) kx0 k +

t1

t0

e(t) kx()k d +

for > t1

t1
Z t
t1

kA (t; )k kF ()k kx()k d

e(t) kx()k d

and using Exercise 8.6

kx(t)k e(tt0 ) e(+)
Hence, lim kx(t)k = 0 ) x(t) ! 0.
t!1

Problem 8.8
R1 >
Using Theorem 8.7 it follows that the solution of A> (t)Q(t) + Q(t)A(t) = I is Q(t) = 0 eA (t) eA(t) d
which is continously-dierentiable and satises I Q(t) I, for all t, where and are positive constants.
_
Then with F (t) = A(t) 12 Q1 (t)Q(t)
_
= A> (t)Q(t) + Q(t)A(t) = I
F > (t)Q(t) + Q(t)F (t) + Q(t)
Thus, using Theorem 7.4, x(t)
_
= F (t)x(t) is UES.
Problem 9.1
The controllability matrix is
Wc =

AB

A2 B

1
=4 1
1
1

3
+ 2 2 + 2
5
1
1
0
0

For controllability we need that det(Wc ) 6

= 0. But det(Wc ) = , hence the system is controllable for all 6
= 0.
The observability matrix is
2
3
0 1 0
3 6 0 2 1 7
2
6
7
C
6 0 0 7
6
7
Wo = 4 CA 5 = 6
0 2 0 7
2
6
7
CA
4 0 0 5
0 2 0
notice that rank(Wo ) = 2 < 3 for all 2 IR, hence the system is not observable for any
Problem 9.4

W (t; tf ) =

tf

(t; )B( )B > ( )> (t; )d

d
W (t; tf ) =
dt

Z t
tf

W (tf ; tf ) = 0

Z t
d
d
>
>
(t; )B( )B>( )
(t; ) B( )B ( ) (t; )d
>(t; ) d B(t)B > (t)
dt
dt
tf

d
W (t; tf ) = A(t)W (t; tf ) + W (t; tf )A> (t) B(t)B > (t)
dt
Using P_ 1 (t) = P 1 (t))P_ (t)P 1 (t), we can write
d 1
W (t; tf ) = W 1 (t; tf )A(t) A> (t)W 1 (t; tf ) + W 1 (t; tf )B(t)B > (t)W 1 (t; tf )
dt
W (to ; tf ) =

tf

(t0 ; )B( )B > ( )> (to ; )d

t0

Then, for t0 < t < tf

W (to ; tf ) =

(t0 ; )B( )B>( )d +

tf

(t0 ; )B( )B > ( )d

t0

W (to ; tf ) = W (t0 ; t) + (t0 ; t)W (t; tf )> (t0 ; t)

Problem 9.7
Rt
A(t )
()) [A; B] controllable , WcA = t0f A (t; )BB > >
. Since (A)(I) =
A (t; )d > 0, but A (t; ) = e
(AI)(tt0 )
(I)(A) ) (AI) (t; ) = e
then
Wc(AI) =

Wc(AI) =

t0
Z tf

tf

e(AI)(tt0 ) BB > e(A

>

I)(tt0 )

e2(t ) eA(t ) BB > eA

>

(t )

t0

The function e2(t ) is bounded above and below for any 2 [t0 ; tf ] by 0 < e2(t ) < 1;
then
0 < WcA Wc(AI) WcA < 1
so Wc(AI) > 0 , [(A I); B] controllable.

(() [(A I); B] controllable , Wc(AI) =

R tf
t0

WcA =
WcA =

e(AI)(tt0 ) BB > e(A

tf

eA(t ) BB > eA

>

>

I)(tt0 )

(t )

d > 0, and

t0
tf

e2(t ) eA(t ) BB > eA

>

(t )

t0

The function e2(t ) is bounded above and below for any 2 [t0 ; tf ] such that 0 < e2(t ) < 1;
then
0 < Wc(AI) WcA Wc(AI) < 1
so WcA > 0 , [A; B] controllable.
Problem 9.13
From the PBH test, a system is controllable if and only if
)
p> A = p>
) p> 0
p> B = 0
then the problem is equivalent to show that N (B > ) = N (BB > )
()) Let (A,B) be controllable then p> B = 0 ) p 0, for (A; BB > ), p> BB > = 0 then N (B > ) N (BB > ).
(() By contradiction, let p 2 N (BB > ) ) BB > p = 0, assume B > p 6
= 0 and let m = B > p, then m> m =
>
>
>
>
= 0, which the contradiction, so N (BB ) N (B ).
p BB p 6
The two set inclusions imply N (B > ) = N (BB > ).
Notice that this property is a fundamental one and is not limited to the time-invariant case (the use of PBH
is TI-specic). More generally, we can use the denition to show this equivalence:
Suppose x_ = Ax + BB > u is controllable and let u be the input that transfers the arbitrary initial state
x0 to the origin. Then the input v = B > u applied to the system x_ = Ax + Bv also transfers the same
initial state to the origin. Since this argument holds for any initial state, x_ = Ax + Bv is controllable.
Suppose x_ = Ax + Bu is controllable and let u be the input that transfers the arbitrary initial state x0
to the origin. Then the input v : BB > v = Bu applied to the system x_ = Ax + BB > v also transfers
the same initial state to the origin. This input always exists since R(B) = R(BB > ), even if BB > is not
invertible (a variation of the proof given above, e.g., using null-range properties). Since this argument
holds for any initial state, x_ = Ax + BB > v is controllable.

EEE 582

HW # 6 SOLUTIONS

Problem 10.3
Controllable & observable:

0
1
1 2

y= 1 0 x

x_ =

x+

9
>
0
u=
1
>
;

Controllable & not observable

2

3 9
0
1
0
0
>
>
x_ = 4 1 2 0 5 x + 4 1 5 u=
0
0 3
1
>
>

;
y= 1 0 0 x
Not controllable & observable
2

3 9
0
0
1
0
>
>
x_ = 4 1 2 0 5 x + 4 1 5 u=
0
0 3
0
>
>

;
y= 1 0 1 x

Qo =

det(Qc ) = 4
det(Qo ) = 0

3
0 1 2
Qc = 4 1 2 3 5 ; det(Qc ) = 0
0 0
0
2
3
1
0
1
1 3 5 ; det(Qo ) = 4
Qo = 4 0
1 2 9

3
0 1 2
Qc = 4 1 2 3 5 ;
0 0
0
2
3
1
0 1
1 0 5;
Qo = 4 0
1 2 0
2

Problem 11.3
Let P (t) be a change of variables such that
x(t)
_
= A(t)x(t) + B(t)u(t)
y(t) = C(t)x(t)
is equivalent to

z(t)
_ = A(t)z(t)
+ B(t)u(t)

y(t) = C(t)z(t)

= P 1 (t) P (t)A(t) + P_ (t)

A(t)
= P 1 (t)B(t)
B(t)
= C(t)P (t)
C(t)
x(t) = P (t)z(t)
A (t; ) = P (t)A (t; )P 1 ( )
1

1
0

1
2

0
1

3
0 1 2
Qc = 4 1 2 3 5 ;
1 3 9
2
3
1
0 0
1 0 5;
Qo = 4 0
1 2 0

3 9
0
1
0
0
>
>
x_ = 4 1 2 0 5 x + 4 1 5 u=
0
0 3
0
>
>

;
y= 1 0 0 x

with

0
1

Not controllable & not observable

2

Qc =

det(Qc ) = 0

det(Qo ) = 0

We know that for instantaneous controllability

Qc = K0 (t) K1 (t)
K0 (t) = B(t)

Kn1 (t)

rank (Qc ) = n

Kj (t) = A(t)Kj1 (t) + K_ j1 (t); j = 1; :::; n

For the transformed system

1 (t)
0 (t) K
c = K
Q
0 (t) = B(t)

n1 (t)
K

c = n
rank Q

_ j1 (t); j = 1; :::; n
K
j1 (t) + K
j (t) = A(t)
K
but
0 = P (t)B(t) = P (t)K0
K
0 + K
1 = AK
_ 0 (t) = P (t)K1 (t)
K
_ j1 (t) = P (t)Kj (t);
K
j1 (t) + K
j (t) = A(t)
K

c = P (t)Qc , and since P (t) is invertible it follows that rank Q
c = n
Hence Q
Similarly for the observability case.
Problem 12.5
Z

(t; ) = 1 )

1 2t 3
t (t )3
e
3
t
3
2
p !2
2
p
1

3
e2t 4
3t
+ 5 > 0 8t

3
2
4

W (t ; t) =

2 e2 d

On the other hand, for any > 0, limt!1 W (t ; t) = 0 and, consequently, there does not exist an > 0 such
that W (t ; t) > 0 for all t.
Problem 13.11
For the time invariant case
p> A = p> ;

p> B = 0 ) p = 0

Then
p> (A + BK) = p> ;

p> B = 0 ) p = 0

and, therefore, controllability of the open-loop state equation implies controllability of the closed-loop state
equation.
In the time-varying case, suppose the open-loop state equation is controllable on [t0 ; tf ]. Then given x(t0 ) =
xa , 9ua (t) such that the solution xa (t)jtf = 0. Next, the closed-loop equation
z(t)
_ = [A(t) + B(t)K(t)] z(t) + B(t)v(t)
with initial state z(t0 ) = xa and input va (t) = ua (t) K(t)xa (t) has the solution z(t) = xa (t). Thus, z(tf ) = 0.
Since the argument is valid for any x0 , the closed-loop state equation is controllable on [t0 ; tf ].
(See also Pr. 9.13 in HW 5.)
Problem 14.7
Using the hint,

K = 1 0 0 0 p(A)
2

Also, for the controller canonical form

det(I A bK) = n + (an1 + kn1 )n1 + + (a0 + k0 )

So, given p() ) k = c0 + a0 c1 + a1 cn1 + an1 , and p(A) = An + cn1 An1 + c1 A+c0 I,
then

1 0 0 c0 I = c0 0 0

1 0 0 c1 A = 0 c1 0
..
.

n1
1 0 0 cn1 A
= 0 0 cn1

1 0 0 An = a0 a1 an1

so k = 1 0 0 p(A) = a0 c0 a1 c1 an1 cn 1
For the general case, using a similarity transformation (z = T x), it is possible to express the system
x_ = Ax + bu in controllable
z = Ac z + bcu, with Ac = T AT 1 , bc = T b, kc = kT 1 Qcz = T Qcx
canonical form
n1
1
b , Qcz = bc Ac bc An1
bc then
and T = Qcz Qcx , Qcx = b Ab A
c

k = kc T = 1 0 0 p(Ac )T = 1 0 0 T p(A)T 1 T

k = 1 0 0 Qcz Q1
cx p(A)
2

but Qcz = 4

0
1

1
%

k=

5 hence

1 0

Q1
cx p(A) =

b Ab An1 b

Problem 15.2
v(t) = Cx(t) + CLz(t)
u(t) = M z(t) + NCx(t) + N CLz(t)
(

x(t)
_
= (A + BNC)x(t) + (BM + BN CL)z(t)
z(t)
_ = GCx(t) + (F + GCL)z(t)

Multiply z(t)
_
by L, add to x(t)
_
and simplify
x(t)
_ + Lz(t)
_ = (A HC)(x(t) + Lz(t))
Let w(t) = x(t) + Lz(t); then the closed-loop system is

w(t)
_
A HC
=
z(t)
_
GC

0
F

w(t)
z(t)

The eigenvalues of the system are given by the eigenvalues of the diagonal elements.

p(A)

EEE 482 HW#4

Problem 1.
Show that controllability is invariant under similarity transformations and under state feedback.
The first of these basic properties states that if [A1, B1] and [A2, B2] are related by a similarity
transformation then controllability of one is equivalent to controllability of the other.
The second property concerns a system [A, B] to which the feedback u = Kx + v is applied. The new
system, from v to x, is [A+BK, B]. Again controllability of one is equivalent to controllability of the other.
Problem 2.
Consider the system with transfer function G(s) = 1/(10s+1)(0.2s+1). We would like to design a statefeedback-plus-observer type controller to achieve closed-loop bandwidth around 5 (e.g., closed-loop poles
with magnitude 5).
1. Design the controller using pole-placement techniques to compute the state-feedback and observer gains.
2. Design the controller using linear quadratic regulator techniques to compute the state feedback and
observer gains. (You will need to try different weights to achieve the desired bandwidth.)
3. Use integrator augmentation to achieve integral action and repeat the designs 1&2.
Problem 3.
In the control of practical systems, the ubiquitous nonlinearities translate in an output offset that depends on
the operating conditions (justify this from the linearization of a nonlinear system). To account for such an
offset, one may design an observer with integral action. For example, suppose that the state equations are

x! = Ax + Bu
y = Cx + Du + v

where v is the offset (constant for a fixed operating point). Design an observer that estimates both the states
x and the offset v. Are the required conditions for observability satisfied?
Problem 4.
In designing an output feedback controller with integral action, it is necessary that the plant has no zeros at
the origin. Otherwise, the controller integrator will cancel the plant zero and cause internal stability
problems. Consider the SISO system (A,b,c,0) for which the control input is defined as follows

z! = r y
u = kx + k z z

Assume that the state x is available for measurement. Show that if (A,b) is c.c. and c(sI-A)-1b has no zero at
s=0, then all the eigenvalues of the closed-loop system can be assigned arbitrarily. (For simplicity, assume
that A is nonsingular.)
Hint: You need to show that the augmented system is c.c., that is

A 0 b
,

c 0 0

is a c.c. pair. For this, use the PBH test. Notice that the condition that c(sI-A)-1b has no zero at the origin
means that cA-1b is non-zero.

HW#MinimalRealization,Solutions
1.Theminimalrealizationhasdimension4.
2.TheGramiansofthebalancedrealizationarediagonalswithentries
9.5135e0011.5584e0018.9122e0021.4599e002
Wecan,therefore,eliminatethelastbalancedstatewithadditiveerroratmost
2*1.4599e002=0.029
andthelasttwostateswithadditiveerroratmost
2*(8.9122e002+1.4599e002)=0.207
LetHbetheoriginalsystem,Hmtheminimalrealization,Hbthebalancedrealization,Hb2the
balancedtruncationwithtwostatesandHb3thebalancedtruncationwiththreestates.The
maximumsingularvalueofthedifferencetransferfunctionistheinducedL2normofthe
difference(error)system.WecanplotthesingularvaluesinMATLABusingsigma.
Hb2=ss(a(1:2,1:2),b(1:2,:),c(:,1:2),d),Hb3=ss(a(1:3,1:3),b(1:3,:),c(:,1:3),d)
sigma(HHm,HmHb,HbHb2,HbHb3)
Thefirsttwoarenumericallythesamesothesingularvaluesofthedifferencearenumerically
zero.HbandHb3aredifferentonlyinonestatesoonlyonesingularvalueisessentiallydifferent
fromzero.Zoomingin,weobservethatthepeaksigma(HbHb2)is15dB~0.178(<=0.207)and
thepeaksigma(HbHb3)is30.5dB~0.029(<=0.029),asexpected.

EEE 582

HW # SVD

% SVD application in the modeling of a noisy oscillatory signal

%
as the output of an autoregressive model:
%
y(n)=[y(n-1),y(n-2),...]*q
% Define the signals
t=(0:.01:10)';
% time
n=rand(size(t))-.5;
% noise
y=sin(2*t+1)+sin(10*t+1)+.02*n/2;
% measurement
N0=10:length(t);
% fitting window
F=tf(.01,[1 -.99],.01);y=lsim(F,y); % Optional Filtering for frequency-weighted fit
% Form the regressor by taking lags of the output
W=[y(N0-1),y(N0-2),y(N0-3),y(N0-4),y(N0-5),y(N0-6),y(N0-7),y(N0-8),y(N0-9)];
NN=0.02*[n(N0-1),n(N0-2),n(N0-3),n(N0-4),n(N0-5),n(N0-6),n(N0-7),n(N0-8),n(N0-9)];
q=W\(y(N0)), % least squares fit
plot(t,y,t(N0),W*q,t(N0),W*q-y(N0)); pause
% check the fit
% Autoregressive transfer function: resonance at the oscillation frequency
g=tf(1,[1 -q'],.01), bode(g)
% check the t.f.
% Model order: How many columns of W do you need?
s=svd(W)
% svd of regressor
s2=svd(W'*W)
% svd of gramian
%
%
%
%
%
%
%
%

Questions:
1. What is the relationship between s and s2? How many lags do you need
in the model?
2. The singular values of W appear to reach a floor related to the noise.
Derive this value analytically and verify with an example.
3. What is ithe effect of the noise amplitude?
4. What happens when the signal is composed of two frequencies, say 10
and 2?

Answers p
1. s = s2; we need at least two lags (2nd order dierence equation) to describe a sinusoidal solution.
2. Letting W0 be the deterministic component and n be the noise, W = W0 +n, W T W = (W0 +n)T (W0 +n) =
W0T W0 + nT n since n ? W (noise uncorrelated with signal). Furthermore, since each sample of the noise is
independent, nT n = N I, where is the noise variance and N is the number of points. For the uniform
Rb
1
distribution in the interval [a; b], with mean n
= (a + b)=2, var(n) = ba
(n n
)2 dn = (b a)2 =12. In our
a
case the distribution is symmetric, zero mean. Let r denote the maximum amplitude. Then, = r2 =3. In the
program, r = 0:01, N = 1001, so rho = 0:033. The small eigenvalues of s2 range in 0.0355 to 0.0316, which
agrees with the theoretical value.
3. Increasing the noise amplitude increases the singular values of W and introduces a bias. Denoting by ']'
the left inverse (LS solution), we have
q = W ] y = (W0 + n)] (y0 + nk+1 ) = (W0T W0 + I)1 W0T y0
The higher the , the more the solution deviates from the nominal one (W0T W0 )1 W0T y0 For example, we nd
by trial and error that when the noise amplitude is 0.45, the resonance is smeared and is bearly recognizable.
4. When two frequencies are present, the identication is more dicult. With noise amplitude 0.01, the
frequency 2 is not recognized (because it does not possess enough energy-cycles in the data interval). Reducing
the noise amplitude by more than a factor of 50 allows for the second peak to appear in the tted model.
Alternatively, introducing a low-pass lter to pre-process the data has a similar eect since it attenuates the
noise at high frequencies and eectively reduces the variance entering the regressor matrix.
1