Title | Notes#3 - Lecture 3 notes |
---|---|
Author | Ashish Mehta |
Course | Linear Systems Theory |
Institution | University of Pennsylvania |
Pages | 49 |
File Size | 1011.5 KB |
File Type | |
Total Downloads | 7 |
Total Views | 215 |
Lecture 3 notes...
Notes on Linear System Theory (ESE 500) Michael A. Carchidi September 24, 2015 Chapter 3 - State Equation Solutions
Linear
The following notes are written by Dr. Michael A. Carchidi for the Course (ESE 500) taught through the Department of Electrical
System Theory
and Systems Engineering at the University of Pennsylvania. These notes are based on the text entitled: by Wilson J. Rugh (2nd edition), and these can be viewed at
Linear System Theory
https://canvas.upenn.edu/ after you log in using your PennKey user name and Password.
1. Introduction The basic questions of existence and uniqueness of solutions are first addressed for the linear state equations unencumbered by inputs and outputs, which means we shall
first consider the equations x˙ (t) = A(t)x(t)
with initial conditions
×
x(t0 ) = x0
(1)
and solve these for x(t) for all t0 ≤ t. The n n matrix function A(t) is assumed to be continuous for all values of t. By definition, a solution to Equation (1) is a
×
continuously differentiable n 1 vector x(t) that satisfies the ordinary differential equation (ODE) x ˙ (t) = A(t)x(t) for all t0 < t, and the initial conditions (IC)
x(t) = x0 for t = t0 . To study the ODE and IC of Equation (1), it is convenient to write it as an integral equation (IE). Toward this end we take the ODE dx(t) = A(t)x(t) dt
and write it as
dx = A(t)x(t)dt.
After integrating, we get
Z
Z
x
A(τ )x(τ )dτ
dx = x0
Z
t
x
with x0
t0
Z
which leads to the IE
x − x0 =
dx = x − x0
t
A(τ )x(τ )dτ t0
Z
or
t
x(t) = x0 +
A(τ )x(τ )dτ.
(2)
t0
Note that this IE includes the initial condition
Z
x(t0 ) = x0
since
t0
A(τ )x(τ )dτ = 0, t0
and it should be clear after differentiating both sides of Equation (2) that
d d d x(t) = x0 + dt dt dt Thus we see that any
x(t)
Z
t
A(τ )x(τ )dτ = 0 + A(t)x(t). t0
that satisfies the ODE and IC of Equation (1) also
satisfies the IE of Equation (2) and conversely, which makes the ODE and IC of Equation (1) equivalent to the IE of Equation (2). 2. Existence of Solutions for the System
Given t0 ,
x˙ (t) = A(t)x(t)
x0 and an arbitrary time T > 0, we will use Equation (2) to construct n × 1 vector functions {xk (t)}∞ k=0 , defined on the interval t0 ≤ t ≤
a sequence of
t0 + T , that can be interpreted as a sequence of ”approximate” solutions to the IE of Equation (2). After proving that this sequence converges uniformly and
absolutely on the interval t0 ≤ t ≤ t0 + T , we shall then show that this limit function is continuously differentiable and satisfies the IE of Equation (2) and hence the ODE and IC of Equation (1). This will settle the question of existence of a solution by actually developing a method for constructing a solution to the ODE and IC of Equation (1).
2
To construct a sequence of approximating functions to the IE of Equation (2), we start with the initial condition x(t0 ) = x0 and define the starting function x0 (t) as the constant function
x0 (t) = x0 for all t0
≤ t ≤ t0 + T , and
x1 (t) = x0 + to get
x1 (t)
from
x0 (t).
Then we set
x2 (t) = x0 + to get
x2 (t)
from
Z
then we set
x1 (t), and
t
A(σ1 )x0 (σ1 )dσ1 t0
Z
Z
t
A(σ2 )x1 (σ2 )dσ2 t0
in general, we set
xk (t) = x0 +
t
A(σk )xk−1 (σk )dσk
(3)
t0
to get
xk (t)
from
xk−1 (t)
for all
k = 1, 2, 3, .....
To determine conditions for which the sequence
{xk (t)}∞ k=0
converges to a
solution of the IE of Equation (2), and hence the ODE and IC of Equation (1), we must first take a moment to discuss the idea of the Euclidean norm of an m × n matrix.
The Euclidean Norm of an m × n Matrix : A Mathematical Tangent Given the real
n×1
vector
⎡
⎢ ⎢ ⎢ x=⎢ ⎢ ⎣
x1 x2 x3 .. .
xn
⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦
vuuX t
we had already defined the Euclidean norm of
||x|| =
√ xT x =
x
as
n
i=1
3
xi2.
(4a)
Using this, we may now define the Euclidean norm of an
⎡
a11 a12 a13 ⎢ a21 a22 a23 ⎢ ⎢ A = ⎢ a31 a32 a33 ⎢ .. .. ... ⎣ . . am1 am2 am3
as
A|| = max
||
all
··· ···
..
...
.
amn
···
Ax||
||
x6=0 ||x||
p(Ax) (Ax)
which is the same as
a1n a2n a3n
···
A|| = max all
x6=0
√ xT x
matrix
⎥ ⎥ ⎥ ⎥ ⎥ ⎦
,
T
||
m×n ⎤
(4b)
r (Ax) (Ax) T
= max all
xT x
x6=0
.
We remind the student that a norm defined on a real vector space valued function from is denoted by
V
to the real numbers
v∈V
R,
i.e.,
f :V →R
V
is a real-
whose value at
f(v) = ||v|| ∈ R,
and which has the following four properties:
v|| ≥ 0
N1
||
N2
||
N3
||
N4
v|| = 0
for all
v ∈ V,
if and only if
v = 0,
αv|| = |α|||v|| for any scalar α ∈ R and any vector v ∈ V , ||u + v|| ≤ ||u|| + ||v|| for all vectors u, v ∈ V .
Note that a norm on
V
defines a metric on
V
and
via
d(u, v) = ||u − v|| where
u − v = u + (−v), || • || on V .
and this metric is called the metric induced on
V
by
the norm
Going back to Equation (4b), we note that since all x 6= 0, we may write
1 Ax|| || Ax|| = = ||x|| ||x||
||
°° 1 °° x ||
||
4
x||
||
is a positive scalar for
°° °° x °° Ax ° ° = °°A x °° = ||
||
Au||
||
where
x u= ||x||
satisfies
u||
||
°° x °° =° ° x °° = ||
||
1 ||x|| = 1. ||x||
Thus we may write Equation (4b) as
A|| = max
||
all
||u||=1
p(Au) (Au) = T
Au|| = max
||
all
||u||=1
Example #1: The Euclidean Norm of a 2 Consider the
2×2
¸∙ u ¸
Then
Au = with
2 5 2 2
u12 + u22 = 1, and
1
u2
hence
Au|| =
||
=
A|| = 2max 2
||
u1 +u2 =1
(u1 , u2 )
∙
p(2u
and so
Since the point
2
matrix
A=
∙
×
1
∙
2 5 2 2
all
||u||=1
.
¸
for any
1
p(2u
1
+ 5u2 )2 + (2u1 + 2u2 )2 .
+ 5u2 )2 + (2u1 + 2u2 )2
is a maximum is the same in which
(2u1 + 5u2 )2 + (2u1 + 2u2 )2 is a maximum, we may
u=
+ 5u2 )2 + (2u1 + 2u2 )2
in which
p(2u
√ uT AT Au.
Matrix
¸
2u1 + 5u2 2u1 + 2u2
max
first determine the maximum of
f(u1 , u2 ) = (2u1 + 5u2 )2 + (2u1 + 2u2 )2
5
∙
u1 u2
¸
(4c)
subject to the condition that
u21 + u22 = 1,
and then take its square root at the
end of the calculation. Of course, computing
∂f(u1 , u2 ) ∂u1
and setting equal to zero is not valid since
2 constraint u1
u22
∂f(u1 , u2 ) ∂u2
and
u1
and
u2
are dependent through the
= 1.
To get around this we may use the method of Lagrange
Using the method of
Lagrange multipliers, we write the constraint as u21 + u22 −
+
multipliers.
1 = 0 and
consider the function
L(u1 , u2 , λ) = f(u1 , u 2 ) + λ(u21 + u22 − 1) or
L(u1 , u2 , λ) = (2u1 + 5u2 )2 + (2u1 + 2u2 )2 + λ(u21 + u22 − 1) which, through the extra multiplier variable and
λ
µ,
allows one to treat all of
u1 , u2
as independent variables and hence we may now set
∂L = (16 + 2λ)u1 + 28u2 = 0 ∂u1 and
∂L = 28u1 + (58 + 2λ)u2 = 0 ∂u2
and
The
∂L = u21 + u22 − 1 = 0. ∂λ
first two equations lead to ∙
16 + 2λ 28 28 58 + 2λ
¸∙
u1 u2
while the third equation leads to the constraint the solution
u1 = u2 = 0 is
¸ =
∙
0 0
¸ .
u21 + u22 = 1, and since u21 + u22 = 1,
unacceptable, and thus we must require that
∙
16 + 2λ 28 28 58 + 2λ 6
¸−1
does not exist, because otherwise, we would get
∙
u1 u2
¸ =
∙
16 + 2λ 28 28 58 + 2λ
¸−1 ∙
0 0
¸ =
∙
0 0
¸ .
From our knowledge of linear algebra, we know that the nonexistence of
∙
16 + 2λ 28 28 58 + 2λ
¸−1
is equivalent to requiring that
det
∙
¸
16 + 2λ 28 28 58 + 2λ
= 0.
This leads to the quadratic equation
4λ2 + 148λ + 144 = 0 = 4(λ + 1)(λ + 36) resulting in
For
λ = −1
or
λ = −36.
λ = −1, the expression ∙ ¸∙ ¸ ∙ ¸ 16 + 2λ 28 u1 0 = 28 58 + 2λ u2 0
becomes
∙
or
yielding
16 − 2 28 28 58 − 2 ∙
u1 = −2u2 , and
14 28 28 56
¸∙
¸∙ u1 u2
coupling this with
1 u2 = ± √ 5
and hence
u1 u2
¸ =
¸ = ∙
0 0
∙
0 0
¸
¸
u21 + u22 = 1,
we
2 u1 = −2u2 = ∓ √ , 5
and this gives
f(u1 , u2 ) = (2u1 + 5u2 )2 + (2u1 + 2u2 )2
7
find that 5u22 = 1, or
µ
or
2 1 f ∓√ , ± √ 5 5
¶ µ =
2 ∓2 √ 5
µ
or
1 ± 5√ 5
1 2 f ±√ , ±√ 5 5
For
λ = −36,
the expression
becomes
∙
16 + 2λ 28 28 58 + 2λ
∙
16 − 72 28 28 58 − 72
or
∙
yielding
u2 = 2u1 , and
−56 28 28 −14
¸∙
coupling this with
1 u1 = ±√ 5
¶2 µ
2 + ∓2√ 5
¶
¶2
= 1.
¸∙
u1 u2
¸∙
u1 u2
u1 u2
1 ± 2√ 5
¸ =
∙
0 0
=
∙
0 0
¸
¸
=
∙
0 0
u21 + u22 = 1,
¸ ¸
¸
we
find that 5u21 = 1, or
2 u2 = 2u1 = ± √ , 5
and
and this gives
f(u1 , u2 ) = (2u1 + 5u2 )2 + (2u1 + 2u2 )2
µ
or
2 1 f ±√ , ± √ 5 5 or
Thus we
find that
¶ µ =
µ
1 ±2 √ 5
2 ± 5√ 5
¶2 µ
¶
+
u22
2 ± 2√ 5
¶2
1 2 = 36 = fmax . f ±√ , ± √ 5 5 √ √ ||A|| = fmax = 36 = 6.
Note that we may also solve this by setting
2 that u1
1 + ±2√ 5
= 1 is
fi
u1 = cos(θ)
and
u2 = sin(θ)
automatically satis ed. Then
f(u1 , u2 ) = (2u1 + 5u2 )2 + (2u1 + 2u2 )2 = (2 cos(θ) + 5 sin(θ))2 + (2 cos(θ) + 2 sin(θ))2 8
so
which reduces to
f(θ) = 8 cos2 (θ) + 28 cos(θ) sin(θ) + 29 sin 2 (θ) or
f(θ) = 8 + 14 sin(2θ) + 21 or
f(θ) =
µ 1 − cos(2θ) ¶ 2
21 37 + 14 sin(2θ) − cos(2θ) 2 2
Then
f 0 (θ) = 28 cos(2θ) + 21 sin(2θ) = 0 leads to
tan(2θ) = −
4 3
sin(2θ)
or
which says that
sin(2θ) = ∓
4 5
3 cos(2θ) = ± . 5
and
Using the “upper signs” for these gives
µ
4 37 + 14 − f (θ) = 2 5
¶
U
µ ¶
3 21 + − 5 2
and using the “lower signs” for these gives
µ ¶
4 37 + 14 + f (θ) = 5 2 L
showing that
fmax = 36 and
A|| =
hence ||
21 − 2
√ fmax
=1
µ 3¶
= 36 5 √ = 36 = 6. ¥ −
It should be noted that the Euclidean norm of a matrix
A
fi
satis es properties
N1 - N4 for all norms so that:
A|| ≥ 0 for all A ∈ Rm×n , m × n matrices
N1 ||
A|| = 0 if
N2 ||
and only if
αA|| = |α|||A||
N3 ||
where
Rm×n
is the vector space of all real
A = 0,
for any scalar
9
α∈R
and any matrix
A ∈ Rm×n ,
and
N4
A + B|| ≤ ||A|| + ||B||
||
for all matrices
A, B ∈ Rm×n .
It also satisfies the properties
Ax|| ≤ ||A||||x||
||
and if
A
is
m×n
with entries
,
(4f)
AB|| ≤ ||A||||B||
||
aij ,
max |aij | ≤ ||A|| ≤ (i,j)
√ mn max |aij |.
(4g)
(i,j)
Q is a symmetric n × n matrix (so QT = Q) then all of its eigenvalues are real and Q can be diagonalized by an orthogonal matrix P whose columns are simply the corresponding eigenvectors of Q. This says that if λ1 , λ2 , λ3 , ..., λn (with possible repetition) are the eigenvalues of Q, then setting ⎡ ⎤ ⎡ ⎤ λ1 0 0 · · · 0 v11 v21 v31 · · · vn1 ⎢ 0 λ2 0 · · · 0 ⎥ ⎢ v12 v22 v32 · · · vn2 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ D = ⎢ 0 0 λ3 · · · 0 ⎥ and P = ⎢ v13 v23 v33 · · · vn3 ⎥ ⎢ .. .. .. . . ⎥ ⎢ .. .. ⎥ . .. ... ... ⎣ . ⎣ .. . . . . . ⎦ . ⎦ 0 0 0 · · · λn v1n v2n v3n · · · vnn We also know from matrix algebra that if
that
with
⎡
⎢ ⎢ ⎢ Q⎢ ⎢ ⎣
we have
vk1 vk2 vk3 .. .
vkn
⎤
⎡
⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ = λk ⎢ ⎥ ⎢ ⎦ ⎣
PT = P−1 ,
and
vk1 vk2 vk3 .. .
vkn
⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦
⎡ and
PT QP = D.
⎢ ⎢ ⎢ ⎢ ⎢ ⎣
vk1 vk2 vk3 .. .
vkn
⎤T ⎡ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦
⎢ ⎢ ⎢ ⎢ ⎢ ⎣
vk1 vk2 vk3 .. .
vkn
⎤
⎥ ⎥ ⎥ ⎥ = 1, ⎥ ⎦
An Important Inequality Using the fact that
λmax
PT = P−1
are smallest and largest
PT QP = D, let us show that eigenvalues of Q, respectively, then and
λmin xT x ≤ xT Qx ≤ λmax xT x 10
if
λmin
and
(4h)
n × 1 vectors x. Toward this end, P QP = D, we have Q = PDP and so for all real
we note that since
P = P−1 and T
T
T
T
T
T
T
T
T
T
T
T
T
x Qx = x (PDP )x = x PDP x = (x P)D(P x) = (P x) D(P x) or T
T
x Qx = u Du
T
u = P x.
with
But
X
⎡
⎢ ⎢ ⎢ D=⎢ ⎢ ⎣
n
T
u Du =
λi u2i
since
i=1
and so we
λ1 0 0 0 λ2 0 0 0 λ3
···
0 0 0
. . .
.. .
.. .
..
.. .
0
0
0
···
find that
X
··· ···
.
λn
⎡
⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦
and
⎢ ⎢ ⎢ u=⎢ ⎢ ⎣
u1 u2 u3 . . .
un
n
T
x Qx =
λi ui2.
i=1 But clearly
X
λmin ≤ λi ≤ λmax
for all
n
i=1 or
n
λmin u2i
≤
i=1
and so
≤
i=1
n
u2i ≤ u2i
i=1
λmax ui2
X X n
λi ui2 ≤ λmax
i=1
u2i .
i=1
n
n
λmin
X
X
and so
n
λi ui2
i=1
X X X n
λmin
X
i = 1, 2, 3, ..., n
T
≤ x Qx ≤ λmax
u2i .
i=1
But since
n
u2i = u u = (P x) (P x) = x PP x = x (PP )x = x (I)x = x x T
T
T
T
T
T
T
T
i=1 we see that T
T
T
λmin x x ≤ x Qx ≤ λmax x x 11
T
T
⎤ ⎥ ⎥ ⎥ ⎥ ⎥ ⎦
and the proof is complete. Note also that if if xmax satisfies Qxmax = λmaxxmax , then
xmin
satisfies
Qxmin = λmin xmin
and