1

Cambridge University Engineering Dept. Third year

Module 3F2: Systems and Control

LECTURE NOTES 1: STATE-SPACE SYSTEMS

1 Introduction 2
1.1 Pendulum Experiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2 State Space Descriptions Of Dynamical Systems 6

2.1 States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 General Guidelines For State-Space Modelling . . . . . . . . . . . . . . . . . . . 8
2.3 Ideal Operational Amplifier Circuits . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Velocity Fields In The State-Space . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3 Linearizing Nonlinear Dynamical Systems 12

3.1 Linearizing the standard form . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 Linearizing when the State Equations are Implicit . . . . . . . . . . . . . . . . . 14
3.3 Behaviour of Nonlinear Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3.1 Example: E4 Exam question 3, 1997 . . . . . . . . . . . . . . . . . . . . . 16

4 Solutions of Linear State Equations 18

4.1 Using Laplace Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Transfer function poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.3 Initial Condition Response of the State . . . . . . . . . . . . . . . . . . . . . . . 20
4.4 Example: Rotating Rigid Body . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.5 Convolution Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.6 Frequency Response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.7 Stability of ẋ = Ax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5 State Space Equations for Composite Systems 31

5.1 Cascade of Two Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2 Parallel combination of two systems . . . . . . . . . . . . . . . . . . . . . . . . . 32
K. Glover Revised by J.M. Maciejowski
November 2001 January 2009
 2

1 Introduction
By way of introduction to the material of these lectures on state space models of dynamical
systems and control the pendulum laboratory experiment will be first described and then some
basic matrix manipulation results will be reviewed.

1.1 Pendulum Experiment

The dynamic equations for this laboratory experiment are (assuming that the moments of
inertia of the pendulum about its centre of mass is negligible and ignoring friction):

Lθ̈(t) = ẍ(t) cos(θ(t)) − g sin(θ(t)) (1.1)

2 2
1 + M/m + I/(ma ) ẍ(t) = T /(ma) + Lθ̈(t) cos(θ(t)) − L sin(θ(t))θ̇ (t) (1.2)

Now define the states of this system to be the vector:

   
x1 x
   
x   ẋ 
 2  
x= = 
x3  Lθ
   
x4 Lθ̇

We can consider the input to be the amplifier input voltage, output current or the resulting
torque, T , which are just related by scaling factors. We will take the input to be

u = T /(aM + I/a),
 3

and also define the natural frequencies given by

ω12 = g/L
  −1 
ω02 = ω12 1 + M/m + I/(ma2 )
ω12
⇒ M/m + I/(ma2 ) =
ω02 − ω12
Recall that ω1 is the natural frequency of the pendulum when the cart is fixed, and it can be
shown that ω0 is the natural frequency of the pendulum when no force acts on the cart.
With this notation equations (1.1) and (1.2) become:
    2 
ω02 ω1 1 2
− cos(x3 /L) ẋ u − sin(x3 /L)x
 ω02 −ω12   2  =  ω02 −ω12 L 4
(1.3)
ẋ4 2
− cos(x3 /L) 1 −Lω1 sin(x3 /L)

Solving this equation for ẋ2 and ẋ4 together with the trivial identities ẋ1 = x2 and ẋ3 = x4 give
the nonlinear first order vector differential equation:

ẋ(t) = f (x, u) (1.4)

ω12
where defining α(x3 ) = sin2 (x3 /L) + ω02 −ω12
we have,
 
x2
   
1 ω12 1 2 2

 α(x3 ) ω 2 −ω 2
u − L sin(x3 /L)x4 − Lω 1 sin(x3 /L) cos(x3 /L) 

f (x, u) =  0 1


 x4 

 
1 ω12 2 2
ω ω
α(x3 ) ω02 −ω12
cos(x3 /L)u − L1 sin(x3 /L) cos(x3 /L)x24 − ω20−ω12 L sin(x3 /L)
0 1

The equation (1.4) is a standard form that is a common starting point for analysis.

Equilibrium States

A state, xe , and input ue are said to be an equilibrium state and input if the system remains in
this condition when released from this initial state, x(0) = xe and u(t) = ue for t ≥ 0. Hence in
(1.4) we need f (xe , ue ) = 0, so that ẋ(t) = 0 and hence remains zero.
Hence in this example x2e = x4e = 0 and (1.3) immediately gives that sin(x3e /L) = 0 and
ue = 0. There are therefore the two obvious equilibria with the pendulum vertically down,
x3e = 0, or vertically up, x3e = Lπ. In both cases the carriage position, x1 , can be arbitrary.
 4

Linearization

Given an equilibrium condition we can linearize the equations about this point and obtain a
good approximation for the system behaviour near the equilibrium. Linearizing about the
equilibrium with x3 = 0 and for example x1 = 0, gives
 
x2
 
u − (ω 2 − ω 2 )x 
 0 1 3
ẋ(t) ≈  

 x4 

2
u − ω0 x3
    
0 1 0 0 x1 0
    
0
 0 −(ω02 − ω12 ) 0 x2  1
   

=   +  u (1.5)
0 0 0 1  x3
  0
    
0 0 −ω02 0 x4 1
| {z } |{z}
A B
= Ax + Bu (1.6)

This is the standard form for a LINEAR dynamical system that we will study in some detail
using matrix techniques.

System Poles

To solve the state equations to get time histories we can take the Laplace transform of (1.6)
giving:

sX(s) − x(0) = AX(s) + BU (s)

⇒ (sI − A)X(s) = BU (s)
⇒ X(s) = (sI − A)−1 BU (s)
= [N (s)/d(s)] BU (s), where d(s) = det(sI − A).

Therefore the transfer functions from the input to the states have a common denominator,
d(s), whose roots will be the system poles. Moreover the roots of d(s) = 0 are the Eigenvalues
of the matrix A.

The eigenvalues for A given above are at 0, 0, ±jω0 .

 5

State Feedback

Suppose we can measure the four states and make the input, u, a linear function of the
measured states:

u = w − k1 x1 − k2 x2 − k3 x3 − k4 x4
h i
= w − kx, where k = k1 k2 k3 k4

The closed loop state equations will now be:

ẋ = Ax + B(w − kx) = (A − Bk)x + Bw

The closed loop poles will now be the eigenvalues of the matrix (A − Bk) which can in fact be
made arbitrary by suitable choice of the feedback gain vector.
 6

2 State Space Descriptions Of Dynamical Systems

2.1 States

The essence of a dynamical system is its memory, i.e. the present output, y(t) depends on past
inputs, u(τ ), for τ ≤ t.

Whereas in a static system y(t)is a function of u(τ ) at τ = t only.

Three sets of variables define a dynamical system -

1. The inputs, u1 (t), u2 (t), ...um (t) (input vector u(t))

2. The state variables, x1 (t), ...xn (t) (state vector x(t))
3. The outputs, y1 (t), y2 (t), ...yp (t), (output vector y(t))

The states may depend on u(τ ) and x(τ ) for τ ≤ t. However for any to < t1 , x(t1 ) can always
be determined from x(to ) and u(τ ), to ≤ τ ≤ t1 . [NB this is the definition of system state] That
is x(to ) summarizes the effect on the future of inputs and states prior to to .
The output, y at time t, is a memoryless function of x(t) and u(t).
 7

The most general class of dynamical system that we will consider is described by the set of first
order ordinary differential equations -


 dxi = fi (x1 (t), x2 (t), ...xn (t), u1 (t), ...um (t), t), i = 1, 2, ...n.
S dt
 y (t) = gj (x1 (t), ...xn (t), u1 (t), ...um (t), t) j = 1, 2, ...p.
j

S is the standard form for a state-space dynamical system model. Or a vector form -

 ẋ(t) = f (x(t), u(t), t)
S
 y(t) = g(x(t), u(t), t)

Note that S has only first order ode’s, but we can use a standard technique to convert high
order o.d.e’s to first order vector o.d.e.’s, by the use of auxiliary variables.

Ex:

...
y +6y ÿ + 5(ẏ)3 + 12 sin(y) = cos(t)

a state variable is given by: x1 = y, x2 = ẏ, x3 = ÿ, when,


 ẋ1 = x2


ẋ2 = x3


= −6x1 x3 − 5x32 − 12 sin(x1 ) + cos(t)

ẋ3

which is in the form

ẋ(t) = f (x(t), t)

where f is a vector valued function of a vector, i.e.

 
x2
 
f (x, t) = 
 x3


−6x1 x3 − 5x32 − 12 sin(x1 ) + cos(t)
 8

For linear time-invariant dynamical systems we use the standard form:


 ẋ(t) = Ax(t) + Bu(t)
S
 y(t) = Cx(t) + Du(t)

2.2 General Guidelines For State-Space Modelling

Method 1: Choose system states, x1 , x2 , ...xn by considering the independent ‘energy storage
devices’ or ‘memory elements’,

e.g.
electrical circuits - current in L or voltage on C.
mechanics - positions and velocities of masses (linear and angular).
chemical engineering - temperature, pressure, volume, concentration.

Then use the basic physical laws derive expressions involving ẋi , e.g.

• V = Ldi/dt, i = Cdv/dt
d(posn )
• m × accn = force; = velocity.
dt
d(volume)
• = flow = function of pressure.
dt

Solve for ẋ = f (x, u, t).

 9

Method 2: Choose state variables as successive time derivatives.

Example: Original ODE: J θ̈ + B θ̇ = M .
Let x1 = θ, x2 = θ̇. Then

In general for n’th-order ODE in θ define

dθ dn−1 θ
x1 = θ, x2 = , ..., xn =
dt dtn−1

2.3 Ideal Operational Amplifier Circuits

In a negative feedback configuration the amplifier acts so as to make the +ve and -ve inputs
have essentially equal voltages.


 C ẋ = (Vi − x1 )/R1 ; C2 ẋ2 = x1 /R2
1 1
 Vo = x1 + x2

Linear operation until the amplifier saturates when quite different equations may hold.
If the amplifier output saturates at ±Vs , then for |x1 + x2 | ≥ Vs
we will have: C2 ẋ2 = (±Vs − x2 )/R2 .
 10

2.4 Velocity Fields In The State-Space

Consider the free motion of the time invariant dynamical system -

ẋ = f (x) =⇒ x(t + δt) ∼

= x(t) + f (x(t))δt

This implies a velocity field in the state space which can give a good qualitative idea of the
system’s behaviour in simple cases.
Ex n=2
The velocity field can be sketched by drawing an arrow in the direction f (x) for many values of
x. Equilibrium points (or singular points) where f (xe ) = 0, may be stable or unstable.

Ex.

dx1
−L = x1 R + x2
dt
dx2
C = x1
dt
    
d  x1  −R/L −1/L x1
=   
dt x2 1/C 0 x2

i) Let R = 1/2, L = 1, C = 1 (in consistent units) then

 
− 12 −1
A =  
1 0
 11

2.5

1.5

0.5

−0.5

−1

−1.5

−2

−2.5
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5

Figure 1: State space trajectories: underdamped RLC network

 
−4 −1
(ii) Let R = 4, L = 1, C = 1 ⇒ A =  
1 0

2.5

1.5

0.5

−0.5

−1

−1.5

−2

−2.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

Figure 2: State space trajectories: overdamped RLC network

 12

The MATLAB code to produce these two sets of state-plane trajectories is as follows:
% A=[-.5 -1; 1 0]; % for underdamped example
A=[-4 -1; 1 0]; % for overdamped example

hold off

[x,y]=meshgrid(-2:.2:2,-2:.2:2);
xdot=A(1,1)*x+A(1,2)*y;
ydot=A(2,1)*x+A(2,2)*y;

quiver(x,y,xdot,ydot); drawnow; hold on

for x0=[-2 2],

for y0=[-2 -1.5 -1 -4+2*sqrt(3) 0 .5 1 1.5];
xx=[[x0;y0*sign(x0)],zeros(2,1000)];
for i=1:1000; xx(:,i+1)= (eye(2)+.02*A)*xx(:,i); end;
plot(xx(1,:)’,xx(2,:)’,’r’)
drawnow
end
end

3 Linearizing Nonlinear Dynamical Systems

3.1 Linearizing the standard form

Suppose we have a nonlinear (time invariant) dynamical system in state space form -


 ẋ = f (x, u)
S
 y = g(x)

Let xe be an equilibrium state for the system when u(t) = ue (constant)

i.e. f (xe , ue ) = 0 and also let ye = g(xe ).

Now consider small perturbations from this equilibrium, let

x(t) = xe + δx(t), u(t) = ue + δu(t), y(t) = y e + δy(t)

 13

A Taylor Series expansion of the i-th equation of ẋ = f (x, u) gives

ẋi = ẋei + δx˙ i

  
∂fi  ∂fi  ∂fi 
= fi (xe , ue ) + δx1 + δx2 + ... + δxn
∂x1 x ∂x2 x ∂xn x
e ,ue e ,ue e ,ue

 
∂fi  ∂fi 
+ δu1 + ... + δum + Remainder
∂u1 x ∂um x
e ,ue e ,ue

Thus for small δx and δu (⇒ v. small remainder) we get

˙ ∼
δx = Aδx + Bδu

where

 
∂f1 ∂f1 ∂f1
∂x1 ∂x2 ··· ∂xn
 
 ∂f2 ∂f2 
∂f  ∂x1 ··· ∂xn 
A= (xe , ue ) =  .. ..

∂x 
 . .


 
∂fn ∂fn
∂x1 ··· ∂xn x=xe ,u=ue

 
∂f1 ∂f1 ∂f1
∂u1 ∂u2 ··· ∂um
∂f 
 .. ..


B= (xe , ue ) =  . . 
∂u  
∂fn ∂fn
∂u1 ··· ∂um x=xe ,u=ue

∂g(xe )
Similarly δy = Cδx where C= ∂x .

The linearized system equations are thus


˙
 δx = Aδx + Bδu
 δy = Cδx

which will accurately predict the system behaviour for x and u close to xe and ue respectively.
 14

3.2 Linearizing when the State Equations are Implicit

Quite often a system’s equations cannot easily be written as ẋ = f (x, u) but can be written as

F (ẋ, x, u) = 0 (n equations in the n unknowns ẋ1 ...ẋn .)

The linearized model can be derived without solving for ẋ as follows. Since (xe , ue ) gives an
equilibrium we have,

F (0, xe , ue ) = 0

Now linearize F about (0, xe , ue ) to get

  
∂F  ˙ ∂F  ∂F 
δx + δx + δu ≃ 0
∂ ẋ (0,x ,u ) ∂x (0,x ,u ) ∂u (0,x ,u )
e e e e e e
| {z } | {z } | {z }
L M N

˙ + M δx + N δu ≃ 0
Lδx =⇒ ˙ ≃ −L−1 M δx − L−1 N δu
δx

(N.B. no nonlinear equations to solve except to obtain the equilibrium point)

 15

3.3 Behaviour of Nonlinear Systems

As mentioned above the linearized equations will accurately predict the behaviour of the
nonlinear system for x and u close to their equilibrium values. When the states and inputs are
far away from the equilibrium values then the behaviour can be quite different, e.g.
• many equilibria with some stable and some unstable (e.g. inverted pendulum).
• limit cycles.
• divergence.
Example: Van der Pol oscillator

 ẋ = x
1 2
 ẋ2 = −x1 + (1 − x2 )x2
1

This has a stable Limit Cycle and an unstable equilibrium at x = 0.

-1

-2

-3
-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

Figure 3: State space trajectories for Van der Pol Oscillator

Note that if we were to change the signs on the ẋ1 and ẋ2 terms then this will just change the
directions of the arrows in the state space. Hence the origin would then be a stable equilibrium
and the limit cycle would be unstable (i.e. if perturbed from the limit cycle then it would
either decay to the origin or diverge to infinity).
 16

3.3.1 Example: E4 Exam question 3, 1997

17
 18

4 Solutions of Linear State Equations

4.1 Using Laplace Transforms

Taking Laplace Transforms of

ẋ(t) = Ax(t) + Bu(t); x(0) = x0

y(t) = Cx(t) + Du(t)

gives

sX(s) − x0 = AX(s) + BU (s)

(sI − A)X(s) = x0 + BU (s)
X(s) = (sI − A)−1 x0 + (sI − A)−1 BU (s)
Y (s) = C(sI − A)−1 x0 + (D + C(sI − A)−1 B)U (s)
| {z } | {z }
initial condition response input response

For x0 = 0, Y (s) = (D + C(sI − A)−1 B) U (s)

| {z }
G(s)

and
G(s) = D + C(sI − A)−1 B

is called the transfer function matrix.

The i, j th entry of G(s) gives the transfer function from uj to yi .
 19

4.2 Transfer function poles

Poles are values of s at which the transfer function becomes infinite:

kG(p)k = ∞ ⇒ p is a pole of G(s)

This can only happen when the matrix (sI − A) becomes singular, ie when

det(sI − A) = 0

namely at the eigenvalues of A.

Later we will see that eigenvalues of A are not always poles of G(s).
Hence we have the important result:

Poles of G(s) ⊂ eigenvalues of A

Analytical expression for transfer function matrix

It can be shown that, for any matrix M ,
 
M11 M21 . . . Mn1
 
 
1  M12 
M −1 =  
det M  ...
 

 
M1n ... Mnn

where Mij is called the cofactor of mij given by

Mij = (−1)i+j det(M with i-th row and j-th column deleted)

−1 1
Hence (sI − A) = N (s) where α(s) = det (sI − A)
α(s)
= sn + α1 sn−1 + · · · + αn−1 s + αn
and N (s) = N1 sn−1 + N2 sn−2 + · · · + Nn−1 s + Nn
1
The transfer function can therefore be written as G(s) = α(s) (CN (s)B + Dα(s)) with
(CN (s)B + Dα(s)) being a matrix of polynomials in s.
 20

Cayley-Hamilton Theorem

An + α1 An−1 + · · · + αn I = 0
since


α(s)I = (sI − A) N1 sn−1 + N2 sn−2 + · · · + Nn

Equating coefficients of sk and premultiplying by Ak for k = n, .., 1, 0 gives

sn : An I = An N1
sn−1 : An−1 α1 I = An−1 N2 − An−1 AN1
.. ..
. .
s : Aαn−1 I = ANn − A2 Nn−1
s0 : αn I = −ANn

Adding these equalities gives the result.

Hence any power of A is a linear combination of I, A, A2 , . . . , An−1 — only!

4.3 Initial Condition Response of the State

For u(t) = 0 we have ẋ(t) = Ax(t) and hence,

−1
X(s) = (sI − A) x0

=⇒
 
x(t) = L−1 (sI − A)−1 x0 = Φ(t)x0

where
n o
−1
Φ(t) = L−1 (sI − A)
 
X 
= L−1 Ak s−(k+1)
 
k≥0

= I + At + A2 t2 /2! + · · · + Ak tk /k! + · · ·
def
= eAt
 21

Note that
d
Φ(t) = AeAt = eAt A
dt
(check by differentiating series expansion of Φ(t)).
Hence with x(t) = Φ(t)x0 ,
d d
x(t) = {Φ(t)x0 } = AeAt x0 = A {Φ(t)x0 } = Ax(t)
dt dt
Φ(0) = I
x(0) = x0

and the differential equation and initial condition are satisfied as required.
Φ(t) is called the state transition matrix.

Properties of eAt

(1) Change of state coordinates

If

A = T −1 ĀT
then A2 = T −1 ĀT T −1ĀT = T −1 Ā2 T
Ak = T −1 Āk T
X∞
hence eAt = Ak tk /k!
k=0
X∞
= T −1 Āk T tk /k!
k=0
∞
!
X
−1 k k
= T Ā t /k! T
k=0

⇒ eAt = T −1 eĀt T
 22

Why is this ‘change of coordinates’ ?

If ẋ(t) = Ax(t) and z = T x, then

ż(t) = T ẋ(t)
= T Ax(t)
= T AT −1 z(t)
= Āz(t)
z(t) = eĀt z(0)
x(t) = T −1 eĀt T x(0)

Special case: Eigenvectors as coordinate axes

Recall that for W the matrix of eigenvectors of A, then the defining relations,

Aw 1 = w 1 λ1
Aw 2 = w 2 λ2
..
.
Aw n = w n λn .
 
λ1 0 ... 0
 
 
 0 λ2 ... 
can be written as: A [w 1 , w 2 , ..., wn ] = [w1 , w2 , ..., wn ]  .. ..

| {z } | {z }  . .


W W  
0 . . . λn
| {z }
Λ

Hence for non-defective A we have the eigenvalue/eigenvector decomposition:

A = W ΛW −1
 23

So that for T = W −1 we have,

 
λ1
 
 
 λ2 
Ā = T AT −1 −1
= W AW = Λ = 
 ..
 = diag {λi }

 . 
 
λn

∞
X X
eΛt = Λk tk /k! = diag { λki tk /k!}
k=0

= diag {eλi t }.

This gives one way of evaluating eAt .

(Another way is to evaluate L−1 (sI − A)−1 .)

2) Semigroup property — Don’t worry about the fancy name.

eA(t1 +t2 ) = eAt1 eAt2 = eAt2 eAt1

since x(t1 + t2 ) = Φ(t1 + t2 )x(0)
= Φ(t2 )x(t1 )
= Φ(t2 )Φ(t1 )x(0) for all x(0)

NB: This only works because (At1 )(At2 ) = (At2 )(At1 ).

For arbitrary matrices A and B, eA+B 6= eA eB .

3) Inverse

I = eA.0 = eA(t−t) = eAt e−At ⇒ (eAt )−1 = e−At

 24

4) Derivative — Repeated here for completeness.

d At
e = AeAt = eAt A
dt

5) Integral
Z t ∞
Z tX
Aτ
e dτ = Ak τ k /k! dτ
0 0 k=0
∞
X
k k+1
t
= A τ /(k + 1)! 0
k=0
( ∞
)
X
−1 k+1 k+1
= A A t /(k + 1)! − 0
k=0
−1 At
= A e − A−1 if det(A) 6= 0.

If det(A) = 0 then the above formula is not valid and the integration needs to be done directly.
e.g. A = 0.

4.4 Example: Rotating Rigid Body

Let I1 , I2 , I3 be the moments of inertia of a rigid body rotating in free space, about its 3
principal axes, and w1 , w2 , w3 the corresponding angular velocities. Then in the absence of
externally applied torques EULER’S EQUATIONS OF MOTION are:-


I1 ẇ1 = (I2 − I3 )w2 w3 


I2 ẇ2 = (I3 − I1 )w3 w1 nonlinear state space equations



I3 ẇ3 = (I1 − I2 )w1 w2

This is a lossless system since the Kinetic Energy,

1 1 1
V (w1 , w2 , w3 ) = I1 w12 + I2 w22 + I3 w32
2 2 2
dV
= I1 w1 ẇ1 + I2 w2 ẇ2 + I3 w3 ẇ3
dt
= w1 w2 w3 [(I2 − I3 ) + (I3 − I1 ) + (I1 − I2 )]
= 0

⇒ if the trajectory starts on an ellipsoid V (w1 , w2 , w3 ) = constant, it stays on it.

In addition conservation of angular momentum implies that,
d  2 2
I1 w1 + I22 w22 + I32 w32 = 2w1 w2 w3 [I1 (I2 − I3 ) + I2 (I3 − I1 ) + I3 (I1 − I2 )]
dt
= 0
 25

Suppose I1 = 6, I2 = 2, I3 = 5 then


ẇ1 = − 21 w2 w3 


1
ẇ2 = − 2 w3 w1


4 
ẇ3 = 5 w1 w2

Equilibrium Solutions satisfy: ẇ1 = ẇ2 = ẇ3 = 0 ⇒ w2 w3 = w1 w3 = w1 w2 = 0 ⇒

either (a) w2 = w3 = 0 & w1 = w̄1

or (b) w3 = w1 = 0 & w2 = w̄2
or (c) w1 = w2 = 0 & w3 = w̄3
or (d) w1 = w2 = w3 = 0.

The linearized equations are

 
0 − 12 w3 − 12 w2
∂f 
1

δ̇w ≃ δw = 
 − 2 w3 0 − 12 w1 
 δw.
∂w
4 4
5 w2 5 w1 0

case (a)
 
0 0 0
 
˙ ∼
⇒ δw − 12 w̄1 
= 0 0  δw.
4
0 5 w̄1 0

    
d  δw2  ∼  0 − 21 w̄1 δw2
⇒ δw1 (t) ≃ δw1 (0)& =  
dt δw3 4
5 w̄1 0 δw3

p
and the state trajectories are ellipses (with the ratio of the principal axes = 8/5 ≃ 1.26.
 26

case (b)
 
0 0 − 12 w̄2
 
˙ ≃
δw 0 0 0  δw
 
4
5 w̄2 0 0

    
d  δw1   0 − 12 w̄2 δw1
δw2 ≃ δw2 (0)& ≃  
dt δw3 4
5 w̄2 0 δw2

case (c)

 
0 − 12 w̄3 0
 
˙ ≃  − 1 w̄3
δw  2 0  δw
0 
0 0 0

    
d  δw1  1  0 1   δw1 
δw3 (t) ≃ δw3 (0), = − w̄3
dt δw2 2 1 0 δw2

and
   
 
δw1 1 δw1 (0) + δw2 (0)
  =   e− 21 w̄3 t
δw2 1 2
 
 
1 δw1 (0) − δw2 (0)
+  e 12 w̄3 t
−1 2
 27

The trajectories in the 3-dimensional state space can thus be sketched as follows on a
particular ellipsoid V (w1 , w2 , w3 ) = const.

0.6

0.4

0.2

−0.2

−0.4

−0.6

0.5
0.6
0.4
0 0.2
0
−0.2
−0.5 −0.4
−0.6

Figure 4: State space trajectories of a rotating rigid body

4.5 Convolution Integral

Consider
ẋ − Ax = Bu (∗)
Comparing with the scalar case note
d  −At d  −At dx
e x(t) = e x(t) + e−At
dt dt dt
−At −At dx
= −e Ax(t) + e
dt
now premultiply (*) by e−At to give
dx
e−At − e−At Ax = e−At Bu
dt
d  −At
⇒ e x(t) = e−At Bu(t)
dt
Z t
⇒ e−At x(t) − x0 = e−Aτ Bu(τ )dτ
0
Z t
At At
⇒ x(t) = e x0 + e e−Aτ Bu(τ )dτ
0
Z t
⇒ x(t) = eAt x0 + eA(t−τ ) Bu(τ )dτ
0
 28

and
Z t
y(t) = CeAt x0 + Du(t) + CeA(t−τ ) Bu(τ ) dτ
| {z } 0
| {z }
initial condition response
input response


 Dδ(t) + CeAt B t≥0
Let H(t) =
 0 t<0
then if x0 = 0,
Z t
y(t) = H(t − τ )u(τ )dτ = H(t) ∗ u(t)
0

H(t) is called the impulse response matrix, since for multiple-input/multiple-output

systems, if an impulse is applied at time 0+ to the j-th input, with the other inputs at zero,
then the i-th output will be
Z t
yi (t) = hij (t − τ )uj (τ )dτ = hij (t)
0

Note that the transfer function, G(s) = L (H(t)).

4.6 Frequency Response

Consider a linear time-invariant system that is asymptotically stable. What is the response
due to sinusoidal input at each of the inputs? Let

uj (t) = Aj cos(ωo t + θj ), j = 1, 2, · · · , m

then

yi (t) → Bi cos(ωo t + φi ) as t → ∞

where

Bi ejφi = gi1 (jωo )A1 ejθ1 + gi2 (jωo )A2 ejθ2 + · · · + gim (jωo )Am ejθm

i.e. the sum of the sinusoidal responses from each input. The rate at which the steady state is
achieved depends on how quickly the impulse response tends to zero, which in turn depends on
the pole positions.
 29

4.7 Stability of ẋ = Ax

Stability of systems is concerned with whether as t → ∞,

x(t) → 0
→ ∞
or remains bounded ?

but since x(t) = eAt x0 the question becomes whether the elements of eAt → 0, → ±∞ or
remain bounded as t → ∞?
Consider

(sI − A)−1 = N (s)/α(s)

where α(s) is the characteristic polynomial of A,

α(s) = det(sI − A) = sn + α1 sn−1 + .. + αn

and

N (s) = N1 sn−1 + N2 sn−2 + .. + Nn

where Ni are n × n constant matrices. (Recall section 4.2.)

If we factor α(s) as

α(s) = (s − λ1 )n1 (s − λ2 )n2 ..(s − λr )nr

where λi will be the eigenvalues of A and n1 + n2 + .. + nr = n = dim (A), then the partial
fraction expansion gives
ni
r X
−1
X Ci,k
(sI − A) =
i=1 k=1
(s − λi )k

for suitable constant matrices Ci,j .

The inverse transform then gives
r Xni
X Ci,k tk−1 eλi t
eAt =
i=1
(k − 1)!
k=1
 30

   
Let λi = σi + jωi then eλi t  = eσi t ejωi t  = eσi t
3 cases
 
(i) σi < 0, tk−1 eλi t  → 0, k = 1, 2, 3, ...
 
(ii) σi > 0, tk−1 eλi t  → ∞, k = 1, 2, 3, ...
(iii) σi = 0
 
(a) tk−1 eλi t  = 1 k=1
 k−1 λ t 
(b) t e i  → ∞ k = 2, 3,.

Hence x(t) → 0 as t → ∞ for all x0 if and only if all λi satisfy Re(λi ) < 0. (i.e. case i)).

Also x(t) remains bounded as t → ∞ for all x0 if and only if Re(λi ) ≤ 0 and in partial fraction
expansion of (sI − A)−1 there are no terms of the form Ci,k /(s − jωi )k with k ≥ 2. (i.e. case (i)
or (iii)(a).).
 31

5 State Space Equations for Composite Systems

5.1 Cascade of Two Systems

Let G1 (s) be realized by the state equation:


 ẋ (t) = A x (t) + B u(t)
1 1 1 1
 w(t) = C1 x (t) + D1 u(t)
1

and G2 (s) be realized by, (input w, output y)


 ẋ (t) = A x (t) + B w(t)
2 2 2 2
 y(t) = C2 x (t) + D2 w(t)
2

then Y (s) = G2 (s)G1 (s)U (s) is realized by

      
d  1 x (t) A 1 0 x (t) B 1
=  1 +  u(t)
dt x (t) B2 C1 A2 x2 (t) B2 D1
2
| {z } | {z } | {z } | {z }
ẋ(t) A x(t) B
h i
y(t) = D2 C1 C2 x(t) + D2 D1 u(t)
| {z } | {z }
D
C
The cascade realization of a single-input single-output system can hence be obtained if G(s) is
factored into second order factors such as
(s2 + c1 s + d1 ) (s2 + c2 s + d2 ) . . .
G(s) = A
(s2 + a1 s + b1 ) (s2 + a2 s + b2 ) . . .
then the overall system can be realized as the cascade of these factors.
 32

5.2 Parallel combination of two systems

Let G1 (s) be realized by the state equation:


 ẋ (t) = A x (t) + B u(t)
1 1 1 1
 y (t) = C1 x (t) + D1 u(t)
1 1

and G2 (s) be realized by, (input u, output y2 )


 ẋ (t) = A x (t) + B u(t)
2 2 2 2
 y (t) = C2 x (t) + D2 u(t)
2 2

then Y (s) = (G1 (s) + G2 (s))U (s) is realized by

      
d  1 x (t) A 0 x (t) B1
= 1  1 +  u(t)
dt x (t) 0 A2 x2 (t) B2
2
| {z } | {z } | {z } | {z }
ẋ(t) A x(t) B
h i
y(t) = C1 C2 x(t) + (D1 + D2 ) u(t)
| {z } | {z }
C D