Chapter1: Linear Systems

1 Peano Baker Series

Solution of n dimensional system ẋ = A(t)x is given by x(t) = Φ(t, t0 )x(t0 ), where,
Z t Z t Z σ1
φ(t, t0 ) = I + A(σ1 )dσ1 + A(σ1 )A(σ2 )dσ2 dσ1 + . . . ...
t0 t0 t0

We show the above series converges. For σ ∈ [t0 , t], let a = max|Aij (σ)|. Then using
|BC| ≤ nbc, where b = max|Bij |, we get

1
|Φ(t, t0 )ij | ≤ 1 +( na(t − t0 ) + (na)2 (t − t0 )2 /2 + (na)3 (t − t0 )3 /3+ ...) (1)
n
1
≤ 1 + ( exp(na(t − t0 )) − 1 ) (2)
n

Z t+∆t
Φ(t + ∆t, t0 ) − Φ(t, t0 ) = B(σ)dσ (3)
t
Z σ
B(σ) = A(σ) + A(σ) A(σ1 )dσ1 + . . . (4)
t0
Φ(t + ∆t, t0 ) − Φ(t, t0 )
lim = B(t) = A(t)Φ(t, t0 ) (5)
∆t→0 ∆t
d
Φ(t, t0 ) = A(t)Φ(t, t0 ), Φ(t0 , t0 ) = I (6)
dt
Series is absolutely bounded then it converges absolutely and hence converges. Φ(t, t0 )
is called the state transition matrix.
Solution is unique, else there is another solution y(t), with y(t0 ) = x0 satifying ẏ(t) =
A(t)y(t) then for z(t) = z(t) − y(t) we have ż(t) = A(t)z(t), with z(0) = 0, then
dkzk2
= z T (AT + A)z ≤ λkzk2 ,
dt
for λ > 0. Then for η > λ, we have
d
exp(−ηt)kzk2 ≤ 0.
dt

1
Hence z(t) = 0 and y(t) = x(t). Therefore we have a solution that exists for all times and is
unique.
Observe
ẋ = (1 + x2 ), x(0) = 0
has solution x(t) = tan t that does not exist beyond π2 .
Similarly √
ẋ = x, x(0) = 0
has solution x(t) = (t − c)2 /4 for t ≥ c and x(t) = 0 for t < c. Family of solutions for c > 0,
not unique. For linear system we have existence and uniqueness.
By uniqueness, we can say that

Φ(t, t0 ) = Φ(t, τ )Φ(τ, t0 ).

In particular

I = Φ(t0 , t0 ) = Φ(t0 , τ )Φ(τ, t0 ).

Φ(t, t0 ) is invertible with inverse Φ(t0 , t).

When A is constant

2 2 k (t − t0 )k
Φ(t, t0 ) = I + A(t − t0 ) + A (t − t0 ) /2 + · · · + A + · · · = exp(A(t − t0 )).
k!
when A(t) = a(t) a scalar, then we have
Z tZ Z Rt
σ1 σk ( t0
a(σ)dσ)k
.. a(σ1 ) . . . a(σk )dσk ..dσ1 = .
t0 t0 t0 k!

Z t
Φ(t, t0 ) = exp( a(σ)dσ).
t0

As some examples

1.    
0 −a cos(at) − sin(at)
A= ; exp(At) =
a 0 sin(at) cos(at)

2
 2.    
0 a cosh(at) sinh(at)
A= ; exp(At) =
a 0 sinh(at) cosh(at)

3.    
0 1 1 t
A= ; exp(At) =
0 0 0 1

4.     Z t
0 −a(t) cos b(t) − sin b(t)
A= ; Φ(t, 0) = ; b(t) = a(σ)dσ
a(t) 0 sin b(t) cos b(t) 0

When A and B commute exp((A + B)t) = exp(At) exp(Bt). They both satify the Ẋ =
(A + B)X, X(0) = I. Therefore,
   
1 1 t 1 t
A= ; exp(At) = e .
0 1 0 1
If A(t), B(t) commute for all t then ΦA+B (t, t0 ) = ΦA (t, t0 )ΦB (t, t0 ).

Given A(t) and its transition matrix Φ(t, t0 ), we have

d
detΦ(t, t0 ) = tr (a(t)) detΦ(t, t0 )
dt

d X ∂Φ(t, t0 )
detΦ(t, t0 ) = Φ̇ij
dt ∂Φij
X
detΦ(t, t0 ) = cij Φij
d X
detΦ(t, t0 ) = cij Φ̇ij = tr(AΦC T ) = tr(A)detΦ
dt Z t
detΦ(t, t0 ) = exp( tr(A(σ))dσ).
t0

2 Variation of Constant Formula

Now consider the linear control system,

ẋ = A(t)x + B(t) u(t) ; x(0) = x0 (7)

|{z} |{z}
n×m m×1

3
 The solution is Z t
x(t) = Φ(t, 0)x0 + Φ(t, τ )B(τ )u(τ )dτ. (8)
0
Observe
Z t Z t
d
Φ(t, τ )B(τ )u(τ )dτ = A(t) Φ(t, τ )B(τ )u(τ )dτ + B(t)u(t).
dt 0 0

When A and B are constant variation of constant formula is

Z t
x(t) = exp(At)x0 + exp(A(t − τ ))Bu(τ )dτ. (9)
0
Consider a force harmonic oscillator
d2
x(t) + x = u(t),
dt2
driven by a sinusoidal input u(t) = sin ωt, with x(0) = ẋ(0) = 0.
Writing x1 = x and x2 = ẋ, we have,
      
d x1 0 1 x1 0
= + u(t).
dt x2 −1 0 x2 1
| {z } | {z }
A B

     Z t  
x1 (t) cos t sin t x1 (0) cos(t − τ ) sin(t − τ ) 0
= +
x2 (t) − sin t cos t x2 (0) 0 − sin(t − τ ) cos(t − τ ) sin ωt
| {z }
exp(At)

  Z t 
x1 (t) 1 cos(−ωτ + t − τ ) − cos(ωτ + t − τ )
= dτ.
x2 (t) 2 0 sin(ωτ + t − τ ) − sin(−ωτ + t − τ )
when ω = 1,
   
x1 (t) 1 sin t − t cos t
= .
x2 (t) 2 t sin t
The solution grows with t. This is excitation on resonance. Linear systems are ubiquitous.
Electrical, mechanical systems.

4
3 Controllability
Consider Z t
ẋ = B(t)u(t), x(t) = x0 + B(τ )u(τ )dτ.
0
Where all can you drive the system starting from x0 . Let
Z t
W = B(τ )B T (τ )dτ.
0

x(t) is reachable iff y = x(t) − x(0) is in the range space of W denoted as R(W ).
Rt
Proof: If y ∈ R(W ) then y = W ξ = 0 B(τ )B T (τ )dτ ξ. Let u(t) = B T (τ )ξ.
/ R(W ), then y = z + a where z ∈ RR(W )⊥ and a ∈ R and z 6=R 0. Then ∃z 6= 0 such
If y ∈
t t
that hz, yi 6= 0 and z ∈ R(W )⊥ . Then 0 z T B(τ )u(τ )dτ 6= 0. But 0 z T B(τ )B T (τ )zdτ =
0 → z T B(τ ) = 0. Hence contradiction. Therefore y ∈ R(W ).

The theorem has at it essence that given matrix P, R(P ) is same as R(P P T ). Given
RT
y = 0 B(τ )u(τ )dτ , we can write this as y = P U , where define discrete time approximation
with ∆t = T /n, we have

 
u(n∆t)
√   √  u((n − 1)∆t) 
P = ∆t B(n∆t) B((n − 1)∆t) ... B(∆t) , U = ∆t  .
 ... 
u(∆t)

Now we come to the general linear system.

ẋ(t) = A(t)x(t) + B(t)u(t), x(0) = x0

and ask when is xf reachable at time T . Then it is reachable if we have u(τ ) such that
Z T
xf − Φ(T, 0)x0 = Φ(T, τ )B(τ )u(τ )dτ.
0

Then from the above results xf − Φ(T, 0)x0 is reachable at time T is it lies in rangle space
of W , where
Z T
W = Φ(T, τ )B(τ )B T (τ )Φ(T, τ )T dτ.
0

5
W is called controllability grammian. Everypoint can be reached at time T , if controllability
Grammian is full rank. When A and B are constant then

Z T Z T
T T
W = exp(A(T − τ ))BB exp(A(T − τ )) dτ = exp(At)BB T exp(At)T dt.
0 0

Let  
M= B AB ... An−1 B .
We claim
R(W ) = R(M M T ).
Recall by Cayley Hamilton theorem A satifies its nth order characteristic polynomial.
Which helps to express An in lower powers of A. Then we have
n−1
X
exp(At) = αi (t)Ai ,
i=0

and
Z T Z n−1
T X
T T
Wx = exp(At)BB exp(At) x dt = αi (t)Ai By(t) = M z.
0 0 i=0

for some z. Hence R(W ) ∈ R(M ).

Now we show that Null space of of W denoted N (W ) belongs to N (M T ). Suppose
x ∈ N (W ), then

Z T Z T
T T T
exp(At)BB exp(At) dt x = 0; x exp(At)BB T exp(At)T dt x = 0
0 0
Z T
kB T exp(At)T xk dt = 0
0

This means kB exp(At) xk = 0 for all t. It implies B T x = 0, else we make t small and
T T

kB T exp(At)T xk 6= 0. Similarly we have

(AB)T x = (A2 B)T x = · · · = (An−1 B)T x = 0.

Therefore x ∈ N (M T ) and hence x ∈ N (M M T ). Because given matrix P, N (P ) is same

as N (P T P ). Now given a symmetrix matrix S, N (S) and R(S) are orthogonal complement
of each other.

6
 To show this we remember given P , R(P )⊥ = N (P T ). Because if y ∈ R(P )⊥ , we have
y P x = 0, ∀x. Then (P T y)T x = 0, ∀x, hence P T y = 0 and y ∈ N (P T ). Converse follows
T

on same lines.
Thus we have
R(W ) = R(M M T ); N (W ) = N (M M T ).
 
Now if M = B AB . . . An−1 B is full rank, the system is controllable.
Consider the system
   
0 1 0 ... ... 0 0
 0 0 1 0 . . . 0   0 
   . 
 ...   . 
 0 0 1 ... 0   . 
ẋ =  . x +  . u
 0 0 ... .. . . . 0   .. 
   
 0 ... ... ... 0 1   0 
−αn −αn−1 . . . −αk . . . −α1 1
Then see M is full ranked. The above form is called controllable canonical form.
In general, given a controllable control system

ẋ = |{z} b u.
A x + |{z}
n×n n×1

Consider the full rank matrix

 
P = An−1 b ... Ab b ,
Let y = P −1 x.Then

ẏ = P −1 AP x + P −1 bu.
But note
 
0

 0 

 .. 
−1  . 
P b= .. .

 . 

 0 
1

7
  
−αn 1 0 ... ... 0

 −αn−1 0 1 0 ... 0 

 .. ...  n
 . 0 1 ... 0  X
−1
P AP =  ...  , An + αi Ai−1 = 0
 −α 0 ... ... 
0 
 k i=1
 .
..

 ... ... ... 0 1 
−α1 0 ... 0 ... 0
This is a canonical form.
Now let q be the first row of inverse of P . Then form the matrix
 
q
 qA 
Q=  ... .
 (10)
qAn−1

Let z = Qx.Then

ż = QAQ−1 x + Qbu.
But note
 
0

 0 

 .. 
 . 
Qb =  .. .

 . 

 0 
1
and
 
0 1 0 ... ... 0

 0 0 1 0 ... 0 

 ...  n
0 0 1 ... 0 n
X
αi Ai−1 = 0
−1  
QAQ = ... , A +

 0 0 ... ... 0 
 i=1
 0 ... ... ... 0 1 
−α1 −α2 . . . −αk ... −αn
The standard controllable canonical form.

8
4 Least square theory
For m ≤ n consider n × m matrix A of rank m, and the problem of minimizing kAu − bk for
choice of u. If we decompose b as b̂ + b⊥ , where b̂ is projection of b on subspace spanned by
A, then

kAu − bk = kAu − b̂k + kb⊥ k

Then choose u such that

Au = b̂, AT Au = AT b̂, AT Au = AT b , u = (AT A)−1 AT b.

AT A invertible as A rank m. Now consider a second problem
For m ≤ n consider m × n matrix A of rank m, and the problem of minimizing kuk such
that Au = b. u = u0 + n where Au0 = b and n ∈ N (A). kn + u0 k = kn − (−u0 )k, then
as above n = −uˆ0 and n + u0 = u⊥ 0 , thus u lies perpendicular to null space of A and recall
⊥ T T
N (A) = R(A ). Then u = A ξ and we get
AAT ξ = b; u = AT (AAT )−1 b.
Again AAT invertible as A rank m.
Now consider the problem of steering the control system

ẋ = A(t)x(t) + B(t)u(t), x(0) = x0

RT
to x(T ) = xf and minimize 0 kuk2 dt.
Then,
Z T
z = xf − Φ(T, 0)x0 = Φ(T, τ )B(τ ) u(τ )dτ.
0 | {z }
c(τ )

As before define descrete time approximation with ∆t = T /n, we have

 
u(n∆t)
√   √  u((n − 1)∆t) 
C= ∆t c(n∆t) c((n − 1)∆t) ... c(∆t) , U = ∆t  ,
 ... 
u(∆t)
We want to solve minimize U T U subject to z = CU . This is what we just solved. The
answer is U = C T (CC T )−1 z. which is expressed as
Z T
′ ′ −1
u(τ ) = B (τ )Φ (T, τ )W z, W = Φ(T, τ )B(τ )B ′ (τ )Φ′ (T, τ )dτ.
0

9
5 Misclleneous
let X be a n × n matrix that satifies

Ẋ = AX + XB T
Then X(t) = ΦA (t, 0)X(0)ΦB (t, 0)T . Gust verify by differentiation.
Note
ΦA (T, t)ΦA (t, 0) = Φ(T, 0)
Differentiating both sides
d
ΦA (T, t) = −ΦA (T, t)A.
dt
or
d ′
Φ (T, t) = −A′ Φ′A (T, t).
dt A

Feedback Control Systems

6 Propotional Controller
Consider the servo control system

ẍ + αẋ = u
we want to stabilize the system to x = 1. Let y = x − 1, then

ÿ + αẏ = u, α > 0
Let y1 = y and y2 = ẏ, then
      
d y1 0 1 y1 0
= +u (11)
dt y 2 0 −α y 2 1
        
d y1 0 1 y1 0   y1
= + −K 0 (12)
dt y2 0 −α y2 1 y2
    
d y1 0 1 y1
= (13)
dt y2 −K −α y2
When K > 0 both eigenvalues have negative real part and we stabilize the system.

10
7 PI controller and Speed Control
Consider the control system

ẋ + αx = u
we want to stabilize the system to x = 1. Let y = x − 1, then

ẏ + αy + α = u,
Rt
Let u(t) = −Ky(t) − Ki 0
y(σ)dσ
Then

ÿ + αẏ = −K ẏ − Ki y,
Let y1 = y and y2 = ẏ, then
    
d y1 0 1 y1
= (14)
dt y2 −Ki −(α + K) y2
When Ki > 0 and K + α > 0 both eigenvalues have negative real part and we stabilize
the system.

7.1 speed control

Consider the servo
θ̈ + αθ̇ = u + T,
where T is unknown torque and we want to control to a certain speed θ̇. Let x = θ̇ then

ẋ + αx = u + T
we want to stabilize the system to x = 1. Let y = x − 1, then

ẏ + αy + α = u + T,
Rt
Let u(t) = −Ky(t) − Ki 0
y(σ)dσ
Then

ÿ + αẏ = −K ẏ − Ki y,
and unknown torque disappears. Now everything is same as above. Suppose torque
is changing with time say in discrete steps, then when we differentiate we get impulses at
discrete times. These are called disturbances. We don’t care now, we have a stable system.

11
8 PD controller
Consider the control system

ẍ = u
we want to stabilize the system to x = 1. Let y = x − 1, then

ÿ = u.
Let y1 = y and y2 = ẏ, then
      
d y1 0 1 y1 0
= +u (15)
dt y2 0 0 y2 1
        
d y1 0 1 y1 0   y1
= + −K −Kd (16)
dt y2 0 0 y2 1 y2
    
d y1 0 1 y1
= (17)
dt y2 −K −Kd y2
When K > 0 and Kd > 0 both eigenvalues have negative real part and we stabilize the
system.

9 PID controller and Servo

Consider the servo control system

ẍ + αẋ + ω 2 x = u
we want to stabilize the system to x = 1. Let y = x − 1, then

ÿ + αẏ + ω 2 y + ω = u,
Rt
Let u(t) = −Ky(t) − Kd ẏ(t) − Ki 0 y(σ)dσ
...
y + αÿ + ω 2 ẏ = −K ẏ(t) − Kd ÿ(t) − Ki y(t),
Let y1 = y and y2 = ẏ, and y3 = ÿ then
    
y1 0 1 0 y1
d 
y2  =  0 0 1   y2  (18)
dt 2
y3 −Ki −(K + ω ) −(α + Kd ) y3
We choose K, Ki , Kd so all eigenvalues have negative real part and we stabilize the system.
That is Ki > 0 and α + Kd > 0 and K + ω 2 > 0.

12
9.1 Servo
Consider the servo control system

ẍ + αẋ = u + T,
where T is unknown torque.
we want to stabilize the system to x = 1. Let y = x − 1, then

ÿ + αẏ = u + T,
Rt
Let u(t) = −Ky(t) − Kd ẏ(t) − Ki 0 y(σ)dσ
...
y + αÿ + ω 2 ẏ = −K ẏ(t) − Kd ÿ(t) − Ki y(t),
everything is same as above , unknown torque disappears.

10 Pole Placement
Now consider a controllable single input system

ẋ = Ax + bu
Then we can always stabilize the system and place the poles of the closed loop system
as desired. We can write the system in canonocal controllable form using y = P −1 x
   
0 1 0 ... ... 0 0
 0 0 1 0 ... 0   0. 
  

 ...   . 
 0 0 1 ... 0   . 
ẏ =  ...  y +  . u
 0 0 . . . . . . 0   .. 
   
 0 ... ... ... 0 1   0 
−α1 −α2 . . . −αk . . . −αn 1
Let  
y1

 y2 

 .. 
  . 
u= β1 β2 . . . βk . . . βn  
 yk 
 .. 
 . 
yn
Then closed loop evolution is

13
  
0 1 0 ... ... 0
 0 0 1 0 ... 0 
 
 ... 
 0 0 1 ... 0 

ẏ =  ... 
y
 0 0 ... ... 0 

 0 . . . . .. ... 0 1 

 β1 − α1 β2 − α2 . . . βk − αk . . . βn − αn 
| {z } | {z } | {z } | {z }
−γ1 −γ2 −γk −γn

The characteristic polynomial of the system takes the form

p(s) = sn + γn sn−1 + · · · + γ1
We can choose γk and hence βk are place the poles as we like. Going back to original
coordinates,

   
y1 x1

 y2 


 x2 

 −1  ..    .. 
  .   . 
u= β1 β2 . . . βk . . . βn P P   = β̃1 β̃2 . . . β̃k . . . β̃n  
 yk   xk 
 ..   .. 
 .   . 
yn xn

10.1 Observers
Our feedback in previous section depended on all the state variables. In practice all state
variables may not be measured. However we can be in a situation that although we don’t
see all state variables our system is still observable, i.e., given the controllable system

ẋ = Ax + bu; y = cx, (19)

we have
 
c

 cA 

P = .. ,
 . 
cAn−1
is full rank. h i
Let P −1 = ... ..  
. . . . q . Let Q = An−1 q An−2 q ... q .

14
 Then observe The transformation z = Q−1 x gives the system

ż = Q−1 AQz + Q−1 bu; y = cQz, (20)

Note
 
c̃ = cQ = 1 0 ... 0 .
and
 
−αn 1 0 ... ... 0

 −αn−1 0 1 0 ... 0 

 .. ...  n
 n   . 0 1 ... 0  X
−1
Q AQ = Q −1 n−1
A q A q . . . Aq =  ... , An + αi Ai−1 = 0.
 −α 0 ... ... 
0 
 k i=1
 .
..

 ... ... ... 0 1 
−α1 0 ... 0 ... 0
| {z }
R

Then

ż = Rz + b̃u; y = c̃z, (21)

Let  
β1
 β2 
 c̃(o − z);
ȯ = Ro + b̃u +  (22)
 ... 
βn
| {z }
β

We are free to choose β, we call o observer. It reconstructs the state z, with inputs as what
can be observed about z.
Then  
β1
d  β2 
(z − o) = R(z − o) +   . . .  c̃(z − o);
 (23)
dt
βn
| {z }
β

Now choose R + βc̃ so that we can place the poles and stabilize, then o aproaches z.
| {z }
R̄
Now note z equation is a controllable equation so there exists a control u = dz such that

15
 ż = Rz + b̃dz = R̃z (24)
is stable. We instead use the feedback

ż = Rz + b̃do = Rz + b̃dz + b̃d(o − z) = R̃z + b̃d(o − z) (25)

Now writing a equation
    
d z R̃ b̃d z
= (26)
dt z−o 0 R̄ z−o

This is a stable system z goes to 0.

Let q = Qo and β̃ = Qβ, and d˜ = dQ−1 then note from Eq. 22 we have

q̇ = Aq + bu + β̃c(q − x); (27)

and

˜
ẋ = Ax + bdq; (28)
As we have shown z goes to zero and x goes to zero.

11 Excercises
1. Let A be a n × n matrix,
 
0 1 0 ... ... 0

 0 0 1 0 ... 0 

 .. ... 
 . 0 1 ... 0 
A=
 ... ... .
 0 0 ... 0 

 .. 
 . ... ... ... 0 1 
0 0 ... 0 ... 0

Find exp(At).

2. Let A be a n × n matrix,

16
  
1 1 0 ... ... 0

 0 1 1 0 ... 0 

 .. ... 
 . 0 1 ... 0 
A=
 ... ... .
 0 0 ... 0 

 .. 
 . ... ... ... 1 1 
0 0 ... 0 ... 1
Find exp(At).
3. Let A be a constant matrix, find the state transtion matrix Φ(T, 0) for the system
ẋ = f (t)A x, where f is a continuous function.
4. Let Ω(t) = −ΩT (t). Show that ΦΩ (t, 0)ΦTΩ (t, 0) = I.
5. Find
d
exp(A + tB)|t=0
dt
.
6. Given n × n matrices A, B, define
 
A11 B . . . ... A1n B

 A21 B . . . ... A2n B 

A⊗B = .. .. .. .. .
 . . . . 
An1 B . . . ... Ann B
Show
exp(A ⊗ I + I ⊗ B) = exp(A) ⊗ exp(B).
7. Let A be an n by n matrix. We say that a linear subspace of Rn is invariant under
A if every vector x in that subspace has the property that Ax also belongs to that
subspace. Show that
b u.
ẋ = Ax + |{z}
n×1
is controllable if and only if b does not belong to an invariant subspace of A.
8. Consider the scalar system ẋ = x + u. Given the constraint u(t + T ) = u(t), find the
R 2T
control that drives x(0) = 1 to x(2T ) = 0 and minimizes 0 u2 (t)dt.
9. Show that Q in Eq. 10 is invertible.
10. Let A(t + T ) = A(t), express ΦA (t, 0) for t > T in terms of ΦA (σ, 0) for σ ≤ T .

17