Chapter 2

State Space Models

General format :
(valid for any nonlinear causal system)

CT : ẋ = f (x, u), DT : xk+1 = f (xk , uk ),

y = h(x, u). yk = h(xk , uk ).
where
x= the state of the system, n × 1-vector
u= the input of the system, m × 1-vector
y= the output of the system, p × 1-vector
f = state equation vector function
h= output equation vector function
n= number of states ⇒ n-th order system
m= number of inputs
p= number of outputs
Single-Input Single-Output system (SISO) : m = p = 1,
Multi-Input Multi-Output system (MIMO) : m, p > 1,
Multi-Input Single-Output system (MISO) : m > 1, p = 1,
Single-Input Multi-Output system (SIMO) : m = 1, p > 1.
& %

ESAT–SCD–SISTA CACSD pag. 34

Linear Time Invariant (LTI) systems
(only valid for linear causal systems)
Continuous-time system:

ẋ = Ax + Bu

y = Cx + Du
ẋ R x
u B C y

Discrete-time system:

xk+1 = Axk + Buk

yk = Cxk + Duk
xk+1 xk
uk B z −1
C yk

A: n × n system matrix
B: n × m input matrix
C: p × n output matrix
&
D: p × m direct transmission matrix %

ESAT–SCD–SISTA CACSD pag. 35

An example of linear state space modeling:

Example: Tape drive control - state space modeling

Process description:

p1 p3 p2
D D

β r r β
J K K J

θ1 θ2
i1 i2
+ +

Kt e1 Kt e2
− −

The drive motor on each end of the tape is independently

controllable by voltage resources e1 and e2. The tape is
modeled as a linear spring with a small amount of viscous
damping near to the static equilibrium with tape tension
6N. The variables are defined as deviations from this equi-
librium point.
& %

ESAT–SCD–SISTA CACSD pag. 36

The equations of motion of the system are:

dω1
capstan 1: J = T r − βω1 + Kti1,
dt
speed of x1: ṗ1 = rω1,
di1
motor 1: L = −Ri1 − Keω1 + e1,
dt
dω2
capstan 2: J = −T r − βω + Kti2,
dt
speed of x1: ṗ2 = rω2,
di2
motor 2: L = −Ri2 − Keω2 + e2,
dt
K D
Tension of tape: T = (p2 − p1) + (ṗ2 − ṗ1),
2 2
p1 + p2
Position of the head: p3 = .
2

& %

ESAT–SCD–SISTA CACSD pag. 37

Description of variables and constants:

D = damping in the tape-stretch motion

= 20 N/m·sec,
e1,2 = applied voltage, V,
i1,2 = current into the capstan motor,
J = inertia of the wheel and the motor
= 4×10−5kg·m2,
β = capstan rotational friction, 400kg·m2/sec,
K = spring constant of the tape, 4×104N/m,
Ke = electrical constant of the motors = 0.03V·sec,
Kt = torque constant of the motors = 0.03N·m/A,
L = armature inductance = 10−3 H,
R = armature resistance = 1Ω,
r = radius of the take-up wheels, 0.02m,
T = tape tension at the read/write head, N ,
p1,2,3 = tape position at capstan 1,2 and the head,
ṗ1,2,3 = tape velocity at capstan 1,2 and the head,
θ1,2 = angular displacement of capstan 1,2,
ω1,2 = speed of drive wheels = θ̇1,2.
& %

ESAT–SCD–SISTA CACSD pag. 38

With a time scaling factor of 103 and a position scaling fac-
tor 10−5 for numerical reasons, the state equations become:

ẋ = Ax + Bu,
y = Cx + Du,

where
   
p 0 2 0 0 0 0
 1  
 ω1   −0.1 −0.35 0.1 0.1 0.75 0
   
p   0 0 0 2 0 0
 2  
x =  , A= 
 ω2 
 
 0.1
 0.1 −0.1 −0.35 0 0.75 

i 
 1

 0 −0.03 0 0 −1 0 

i2 0 0 0 −0.03 0 −1
 
00
 
0 0
  " #
0 0 0.5 0 0.5 0 0 0

B=  , C = ,
0
 0 −0.2 −0.2 0.2 0.2 0 0
1 0
 
01
" # " # " #
0 0 p3 e1
D= , y= , u= .
0 0 T e2
& %

ESAT–SCD–SISTA CACSD pag. 39

State-space model, transfer matrix and impulse
response

Continuous-time system:
( Laplace
state-space ẋ = Ax + Bu x(0)=0 Y (s)
−→ G(s) = = D + C(sI − A)−1B
equations y = Cx + Du U (s)
| {z }
transfer matrix
in practice : D = 0
⇓
∞
Laplace X
G(t) = Ce B At
←→ G(s) = CAi−1Bs−i
| {z }
i=1
impulse response matrix | {z }
transfer matrix

Discrete-time system:
( z − transf
state-space xk+1 = Axk + Buk x0 =0 Y (z)
←→ G(z) = = D + C(zI − A)−1B
equations yk = Cxk + Duk U (z)
| {z }
transfer matrix
( ⇓
∞
D : k = 0 z − transf X
G(k) = k−1
←→ G(z) = CAi−1Bz −i + D
CA B : k ≥ 1 i=1
| {z } | {z }
impulse response matrix transfer matrix

In case of a SISO system, G(t) or G(k) is the impulse

response.
For MIMO G(t), G(k) are matrices containing the m × p
impulse responses (one for every input-output pair).
& %

ESAT–SCD–SISTA CACSD pag. 40

 Linearization of a nonlinear system
about an equilibrium point

Consider a general nonlinear system in continuous time :

dx
= f (x, u)
dt
y = h(x, u)
For small deviations about an equilibrium point (xe, ue, ye)
for which
f (xe, ue) = 0
ye = h(xe, ue)
we define

x = xe + ∆x, u = ue + ∆u, y = ye + ∆y,

and obtain

dx d∆x
= = f (x, u) = f (xe + ∆x, ue + ∆u)
dt dt

and

ye + ∆y = h(x, u) = h(xe + ∆x, ue + ∆u).

& %

ESAT–SCD–SISTA CACSD pag. 41

By first order approximation we obtain a linear state space
model from ∆u to ∆y :
d∆x
= f (xe + ∆x, ue + ∆u)
dt
⇓ f (xe, ue) = 0
 
d∆x (1) ∂f 
 ∂f 
= ∆x + ∆u
dt ∂x xe,ue ∂u xe,ue
| {z } | {z }
n×n n×m
⇓ ⇓
A B
and
ye + ∆y = h(xe + ∆x, ue + ∆u)
⇓ ye = h(xe, ue)
 
(1) ∂h  ∂h 
∆y =  ∆x +  ∆u
∂x xe,ue
 ∂u xe,ue
| {z } | {z }
p×n p×m
⇓ ⇓
C D
Conclusion : use a linear approximation about equilibrium.
& %

ESAT–SCD–SISTA CACSD pag. 42

Example :
Consider a decalcification plant which is used to reduce the
concentration of calcium hydroxide in water by forming a
calcium carbonate precipitate.

The following equations hold (simplified model) :

• chemical reaction : Ca(OH)2 + CO2 → CaCO3 + H2O

• reaction speed : r = c[Ca(OH)2][CO2]
• rate of change of concentration :
d[Ca(OH)2] k r
= −
dt V V
d[CO2] u r
= −
dt V V

k and u are the inflow rates in moles/second of calcium

hydroxide and carbon dioxide respectively. V is the tank
volume in liters.
Let the inflow rate of carbon dioxide be the input and the
concentration of calcium hydroxide be the output.

& %

ESAT–SCD–SISTA CACSD pag. 43

A nonlinear state space model for this reactor is :
d[Ca(OH)2] k c
= − [Ca(OH)2][CO2]
dt V V
d[CO2] u c
= − [Ca(OH)2][CO2]
dt V V
y = [Ca(OH)2]

The state variables are x1 = [Ca(OH)2] and x2 = [CO2].

In equilibrium we have :
k c k c
− [Ca(OH)2]eq[CO2]eq = − X1X2 = 0
V V V V
ueq c U c
− [Ca(OH)2]eq[CO2]eq = − X1X2 = 0
V V V V
Y = [Ca(OH)2]eq = X1
For small deviations about the equilibrium :
d∆x1 c c
= − X2∆x1 − X1∆x2
dt V V
d∆x2 c c 1
= − X2∆x1 − X1∆x2 + ∆u
dt V V V
∆y = ∆x1

so,
" # " #
− cX
V
2
− cX
V
1
0 h i
A= cX2 cX1 ,B = 1
,C = 1 0 ,D = 0
−V −V V
& %

ESAT–SCD–SISTA CACSD pag. 44

If
k = 0.1 mole U = 0.1 mole
2
sec , c = 0.5 sec· lmole , sec ,
X1 = 0.25 mole , X2 = 0.8 mole , V =5l
l l
then  
" # −0.08 −0.025 0
A B  
=  −0.08 −0.025 0.2 
C D
1 0 0

The corresponding transfer function is

∆y(s) −0.005
= 2
∆u(s) s + 0.105s
and its bode plot
Bode Diagrams

50

0
Phase (deg); Magnitude (dB)

−50

−250

−300

−350

−400
−3 −2 −1 0
10 10 10 10

& %
Frequency (rad/sec)

ESAT–SCD–SISTA CACSD pag. 45

One can also obtain a linear state space model for this
chemical plant from linear system identification.

For small deviations about the equilibrium point the dy-

namics can be described fairly well by a linear (A, B, C, D)-
model. Hence, a small white noise disturbance ∆u was
generated and was added to the equilibrium value U . We
applied this signal to the input of a nonlinear model of
the chemical reactor (Simulink model for instance), i.e.
u(t) = U + ∆u.

The following input-output set was obtained :

u(t)
0.104

0.102

0.1

0.098

0.096
0 100 200 300 400 500 600 700 800 900 1000
time (s)

y(t)
0.2505

0.25

0.2495

0.249
0 100 200 300 400 500 600 700 800 900 1000
time (s)

& %

ESAT–SCD–SISTA CACSD pag. 46

By applying a linear system identification algorithm
(N4SID), the following 2nd-order model was obtained :
 
" # 0.0015 −0.0589 −0.0867
A B  
=  0.0027 −0.1066 0.1713 
C D
0.2546 0.129 0

The corresponding transfer function is

y(s) −1.299.10−7s − 0.005
= 2
u(s) s + 0.1051s + 1.346.10−8
It has 2 poles at −1.281.10−7 and −1.051.

As −1.281.10−7 lies close to 0, and taking in account the

properties of the manually derived linear model of page 45,
we conclude that the plant has one integrator pole. Hence,
it might be better to fix one pole at s = 0. In this way
it is guaranteed that the linear model obtained by system
identification is stable.
The transfer function which was obtained using this mod-
ified identification procedure is
y(s) −1.68.10−8s − 0.005
=
u(s) s2 + 0.1051s

& %

ESAT–SCD–SISTA CACSD pag. 47

The different models are validated by comparing their re-
sponse to the following input :
∆u(t) = 0.003 if 100 < t < 200
= 0 elsewhere
Four responses are shown :
• the nonlinear system (—)
• the linearised model obtained by hand (page 45) (- -)
• the 2nd-order linear model obtained from N4SID (· · ·)
• the linear model obtained from N4SID having a fixed
integrator pole by construction (-.)
y(t)
0.25

0.245
mole/l

0.24

0.235
0 100 200 300 400 500 600 700 800 900 1000
y(t) : detail 1
0.2364

0.2362
mole/l

0.236

0.2358

0.2356
900 905 910 915 920 925 930 935 940 945 950
y(t) : detail 2
0.23574

0.23573
mole/l

0.23572

900 905 910 915 920 925 930 935 940 945 950

& %
time(s)

ESAT–SCD–SISTA CACSD pag. 48

 Digitization

in this course we want to control physical plants using a

digital computer :
+ actuators Plant sensors

D/A A/D
clock
digital
controller
computer

• analog-to-digital converter :

lowpass sampler quantizer

filter
x(t) x[k] x(t) x[k]
A/D Q

clock

– the analog anti-aliasing filter filters out high frequent

components (> Nyquist frequency)
– the filtered analog signal is sampled and quantized
in order to obtain a digital signal. For quantization
10 to 12 bits are common.
& %

ESAT–SCD–SISTA CACSD pag. 49

• digital-to-analog converter :
– zero-order hold
ZOH approximation
3

−1

−2

−3

−4
0 10 20 30 40 50 60 70 80 90 100
time

∗ introduces a delay
∗ frequency spectrum deformation
– first-order hold
– n-th order polynomial (more general case) : fit a n-
th order polynomial through the n+1 most recent
samples and extrapolate to the next time instance

& %

ESAT–SCD–SISTA CACSD pag. 50

Digitization of state space models
1. discretization by applying numerical integration rules :
a continuous-time integrator

ẋ(t) = e(t)

⇔
sX(s) = E(s)
can be approximated using
• the forward rectangular rule or Euler’s method

xk+1 − xk
= ek with xk = x(kTs)
Ts e(t)

⇔
z−1
X(z) = E(z)
Ts
k-1 k k+1 t
• the backward rectangular rule

xk+1 − xk
= ek+1 e(t)
Ts
⇔
z−1
X(z) = E(z)
zTs k-1 k k+1 t
& %

ESAT–SCD–SISTA CACSD pag. 51

 • the trapezoid rule or bilinear transformation

xk+1 − xk ek+1 + ek
= e(t)
Ts 2
⇔
2 z−1
X(z) = E(z)
Ts z + 1 k-1 k k+1 t

Change a continuous model G(s) into a discrete model

Gd(z) by replacing all integrators with their discrete equiv-
alents :
= projection of the left half-plane
Euler backward rect. bilinear transf.
s → z−1
Ts s → z−1
zTs s → T2s z−1
z+1
Im
Im Im

Re Re Re

Except for the forward rectangular rule stable continuous

poles (grey zone) are guaranteed to be placed in stable
discrete areas, i.e. within the unit circle.
The bilinear transformation maps stable poles → stable
poles and unstable poles → unstable poles.
& %

ESAT–SCD–SISTA CACSD pag. 52

The following continuous-time model is given :

ẋ = Ax + Bu sX = AX + BU
⇔
y = Cx + Du Y = CX + DU
z−1
following Euler’s method s is replaced by Ts , so
z−1
X = AX + BU
Ts
or
zX = (I + ATs)X + BTsU
the output equation Y = CX + DU remains. A similar
calculation can be done for the backward rectangular rule
and the bilinear transformation resulting in the following
table :
Euler backward rect. bilinear transf.
Ad I + ATs (I − ATs )−1 (I − AT2 s )−1(I + AT2 s )
Bd BTs (I − ATs )−1BTs (I − AT2 s )−1BTs
Cd C C(I − ATs)−1 C(I − AT2 s )−1
Dd D D + C(I − ATs )−1BTs D + C(I − AT2 s )−1 BT2 s

& %

ESAT–SCD–SISTA CACSD pag. 53

2. discretization by assuming zero-order hold ...

ẋ = Ax + Bu

⇒ Z t
x(t) = eAtx(0) + eA(t−τ )Bu(τ )dτ
|0 {z }
convolution integral

Let the sampling time be Ts, then

Z t+Ts
x(t + Ts) = eA(t+Ts)x(0) + eA(t+Ts−τ )Bu(τ )dτ
Z0 t+Ts
= eATs x(t) + eATs eA(t−τ )Bu(τ )dτ
|t {z }
approximated
ZOH t=kTs
⇒ xk+1 = eATs xk + eATs A−1(I − e−ATs )Buk
ATs −1 ATs
= e
|{z} x k + A (e − I)B uk
| {z }
Ad Bd

One can prove that Gd(z) can be expressed as

  
ZOH −1 −1 G(s)
Gd(z) = (1 − z )Z L
s
For this reason this discretization method is sometimes
called step invariance mapping.
& %

ESAT–SCD–SISTA CACSD pag. 54

3. discretization by zero-pole mapping :
following the previous method the poles of Gd(z) are
related to the poles of G(s) according to z = esTs . If
we assume by simplicity that also the zeros undergo this
transformation, the following heuristic may be applied:
(a) map poles and zeros according to z = esTs
(b) if the numerator is of lower order than the denomi-
nator, add discrete zeros at -1 until the order of the
numerator is one less than the order of the denomi-
nator. A lower numerator order corresponds to zeros
at ∞ in continuous time. By discretization they are
put at -1.
(c) adjust the DC gain such that

lim G(s) = lim Gd(z)

s→0 z→1

& %

ESAT–SCD–SISTA CACSD pag. 55

Example :
given the following SISO system
 
" # −0.2 −0.5 1
A B  
= 1 0 0
C D
1 0.4 0
−1 s + 0.4
⇒ G(s) = C(sI − A) B + D = 2 .
s + 0.2s + 0.5
Now compare the following discretization methods (Ts = 1
sec.) :
1. bilinear transformation :
(a) Ad, Bd, Cd and Dd are calculated using the conver-
sion table
(b) Gbilinear(z) = Cd(zI − Ad)−1Bd + Dd.
0.4898 + 0.1633z −1 − 0.3265z −2
=
1 − 1.4286z −1 + 0.8367z −2
2. ZOH :
(a) Ad = eATs , Bd = A−1 (eATs − I)B, Cd = C and
Dd = D
(b) GZOH(z) = Cd(zI − Ad)−1Bd + Dd.
1.0125z −1 − 0.6648z −2
=
1 − 1.3841z −1 + 0.8187z −2
& %

ESAT–SCD–SISTA CACSD pag. 56

 3. pole-zero mapping :
(a) the poles of G(s) are −0.1 + j0.7 and −0.1 − j0.7
(b) there is one zero at −0.4
(c)
z − e−0.4
Gpz = K
(z − e−0.1+j0.7)(z − e−0.1−j0.7)
z −1 − 0.6703z −2
=K
1 − 1.3841z −1 + 0.8187z −2
lims→0 G(s)
hence K = = 1.0546
limz→1 Gpz(z)
Compare the bode plots :
1
10 50

0
frequency amplitude spectrum

phase spectrum (degr.)

−50
0
10

−100
__ : continuous system __ : continuous system

−− : bilinear approximation −− : bilinear approximation

−. : zero−order hold assumption
−. : zero−order hold assumption
−150 .. : pole−zero mapping
.. : pole−zero mapping

−1
10 −2 −1 −2 −1
10 10 10 10
frequency (Hz) frequency (Hz)

& %

ESAT–SCD–SISTA CACSD pag. 57

Sampling rate selection :
• the lower the sampling rate, the lower the implementa-
tion cost (cheap microcontroller & A/D converters), the
rougher the response and the larger the delay between
command changes and system response.
• an absolute lower bound is set by the specification to
track a command input with a certain frequency :

fs ≥ 2BWcl

fs is the sampling rate and BWcl is the closed-loop sys-

tem bandwidth.
• if the controller is designed for disturbance rejection,

fs > 20BWcl

• when the sampling rate is too high, finite-word effects

show up in small word-size microcontrollers (< 10 bits).
• for systems where the controller adds damping to a
lightly damped mode with resonant frequency fr ,

fs > 2fr

In practice, the sampling rate is a factor 20 to 40 higher

than BWcl.
& %

ESAT–SCD–SISTA CACSD pag. 58

Advantages of state space models

• More general models: LTI and Nonlinear Time Varying

(NTV).
• Geometric concepts: more mathematical tools (linear
algebra).
• Internal and external descriptions: “divide and con-
quer” strategy.
• Unified framework: the same for SISO and MIMO.

& %

ESAT–SCD–SISTA CACSD pag. 59

 Geometric properties of linear state
space models
Canonical forms
Control canonical form (SISO):

ẋ = Acx + Bcu, y = Ccx.

 
−a1 −a2 · · · · · · −an  
  1
 1 0 ··· ··· 0   
  0
Ac =  0 . .
.. .. .
.  , Bc = 
 ..  ,
  
 . . . . . . . . . . .   
 . . 
0
0 ··· 0 1 0
h i
Cc = b1 b2 · · · b n
Corresponding transfer function:
b(s)
G(s) =
a(s)
b(s) = b1sn−1 + b2sn−2 + · · · + bn
a(s) = sn + a1sn−1 + a2sn−2 + · · · + an

& %

ESAT–SCD–SISTA CACSD pag. 60

Block diagram for the control canonical form

b1

b2
y

bn−1

x1 x2 xn−1 xn bn
u R R R R

−a1

−a2

−an−1

−an

& %

ESAT–SCD–SISTA CACSD pag. 61

Modal canonical form:

ẋ = Amx + Bmu, y = Cmx.

   
−s1 1
   
 −s2  1
Am = 

...  , Bm =  .. 
  
   
−sn 1
h i
C m = r1 r2 · · · rn
Corresponding transfer function:
n
X ri
G(s) =
i=1
s + si
Block diagram for the modal canonical form:

R x1
r1

−s1

R x2
r2
u y
−s2

R xn
rn

−sn
& %

ESAT–SCD–SISTA CACSD pag. 62

Similarity transformation
A state space model (A, B, C, D) is not unique for a phys-
ical system. Let

x = T x̄, det(T ) 6= 0,

Ā = T −1AT, B̄ = T −1B, C̄ = CT, D̄ = D

⇒
x̄˙ = T −1AT x̄ + T −1Bu = Āx̄ + B̄u,
y = CT x̄ + Du = C̄ x̄ + Du
T is chosen to give the most convenient state-space descrip-
tion for a given problem (e.g.control or modal canonical
forms).
Example: Eigenvalue decomposition of A:
 
α1 β1
 
 −β1 α1 
 
−1
A = XΛX , Λ =   . .. 

 
 λ1 
...

ẋ = Ax + Bu X −1 x = z ż = Λz + X −1Bu
=⇒
y = Cx + Du y = CXz + Du
& %

ESAT–SCD–SISTA CACSD pag. 63

Controllability
An n-th order system is called controllable if one can reach
any state x from any given initial state x0 in a finite time.

Controllability matrix:
h i
C = B AB · · · An−1B

rank(C) = n ⇔ (A, B) controllable

Explanation:
Consider a linear (discrete) system of the form

xk+1 = Axk + Buk

Then
xk+1 = Ak+1x0 + Ak Bu0 + · · · + Buk
such that
 
uk
h i


uk−1 
k+1 k
xk+1 − A x0 = B AB · · · A B  . 

 . 

u0

& %

ESAT–SCD–SISTA CACSD pag. 64

Note that
h i h i
rank B AB · · · Ak B = rank B AB · · · An−1B

for k ≥ n − 1 (Cayley-Hamilton
 Theorem).

uk
 
 uk−1 
 ..  if rank(C) = n.
There exists always a vector  
 
u0
Remarks: Controllability matrices for a continuous time
system and a discrete time system are of the same form.

& %

ESAT–SCD–SISTA CACSD pag. 65

Observability
An n-th order system is called observable if knowledge of
the input u and the output y over a finite time interval is
sufficient to determine the state x.

Observability matrix:
 
C
 
 CA 
O=
 ..


 
CAn−1
rank(O) = n ⇔ (A, C) observable
Explanation:
Consider an autonomous system

xk+1 = Axk

yk = Cxk
Then we find easily

yk = CAk x0

& %

ESAT–SCD–SISTA CACSD pag. 66

and    
y0 C
   
 y1   CA 
 =  x0
 ..   .. 
   
yk CAk
Note that
   
C C
   
 CA   CA 
rank 
 · · ·  = rank  · · · 
  
   
CAk CAn−1
for k ≥ n − 1.
One can always determine x0 if rank(O) = n.
Remarks: Observability matrices for a continuous time sys-
tem and a discrete time system are of the same form.

& %

ESAT–SCD–SISTA CACSD pag. 67

The Popov-Belevitch-Hautus tests (PBH)

PBH controllability test:

(A, B) is not controllable if and only if there exists a left
eigenvector q 6= 0 of A such that

AT q = qλ

BT q = 0
In other words, (A, B) is controllable if and only if there is
no left eigenvector q of A that is orthogonal to B.
If there is a λ and q satisfying the PBH test, then we
say that the mode corresponding to eigenvalue λ is uncon-
trollable (uncontrollable mode), otherwise it is controllable
(controllable mode).
How to understand (for SISO) ? Put A in the modal canon-
ical form:
     
∗ ∗ 0
     
A =  λi  , B =  0  , ⇒ q =  1 
∗ ∗ 0
Uncontrollable mode: ẋi = λixi.

& %

ESAT–SCD–SISTA CACSD pag. 68

PBH observability test:
(A, C) is not observable if and only if there exists a right
eigenvector p 6= 0 of A such that

Ap = pλ

Cp = 0
If there is a λ and p satisfy the PBH test, then we say that
the mode corresponding to eigenvalue λ is unobservable
(unobservable mode), otherwise it is observable (observable
mode).

How to understand for SISO ? Put A in a modal canonical

form.

& %

ESAT–SCD–SISTA CACSD pag. 69

Stability/Stabilizability/Detectability

Stability:
A system with system matrix A is unstable if A has an
eigenvalue λ with real(λ) ≥ 0 for a continuous time system
and |λ| ≥ 1 for a discrete time system.

Stabilizability:
(A, B) is stabilizable if all unstable modes are controllable.

Detectability:
(A, C) is detectable if all unstable modes are observable.

& %

ESAT–SCD–SISTA CACSD pag. 70

Example

xk+1 = Axk + Buk

yk = Cxk + Duk
 
1.1 0 0 0 1
" #  0 −0.5 0.6 0 2

A B  
= 0 −0.6 −0.5 0 −1 

C D  
 0 0 0 2 2
0 1 2 3 1

Modes Stable? Stabilizable? Detectable?

1.1 no yes no
−0.5 ± 0.6i yes yes yes
2 no yes yes

& %

ESAT–SCD–SISTA CACSD pag. 71

Kalman decomposition and minimal realization

Kalman decomposition
Given a state space system [A, B, C, D]. Then we can
always find an invertible similarity transformation T such
that the transformed matrices have the structure

 r1 r2 r3 r4 
r1 Aco 0 A13 0
 
r2
 A21 Aco A23 A24 
T AT −1 = 
r3
 0 0 Aco 0 
r4 0 0 A43 Aco

 m 
r1 Bco
   r1 r2 r3 r4 
r2 Bco 
  −1
TB = , CT = p Cco 0 Cco 0
r3 0 
 
r4 0
where
r1 = rank (OC), r2 = rank (C) − r1,
r3 = rank (O) − r1, r4 = n − r1 − r2 − r3.

& %

ESAT–SCD–SISTA CACSD pag. 72

The subsystem
[ Aco, Bco, Cco, D ]
is controllable and observable. The subsystem
! !
Aco 0 Bco  
[ , , Cco 0 , D ]
A21 Aco Bco
is controllable. The subsystem
! !
Aco A13 Bco
[ , , ( Cco Cco ), D ]
0 Aco 0
is observable. The subsystem

[ Aco , 0, 0, D ]

is neither controllable nor observable.

& %

ESAT–SCD–SISTA CACSD pag. 73

 Kalman decomposition

u y
CO

C Ō

C̄O

C̄ Ō

& %

ESAT–SCD–SISTA CACSD pag. 74

Minimal realization:
A minimal realization is one that has the smallest-size A
matrix for all triplets [A, B, C] satisfying

G(s) = D + C(sI − A)−1B

where G(s) is a given transfer matrix.

h i
A B C D is minimal ⇔ controllable and observable.
h i
Let Ai Bi Ci D , i = 1, 2, be two minimal realizations
of a transfer matrix, then
h i
A1 B1 C1 D

T ⇓ ⇑ T −1
h i
A2 B2 C2 D
with
T = C1C2T (C2C2T )−1

& %

ESAT–SCD–SISTA CACSD pag. 75

Input/output properties of state-space
models
Transfer matrix
General transfer matrix for a state space model :

G(ξ) = D + C(ξI − A)−1B

ξ can be s (CT) or z (DT).

Poles
Characteristic polynomial of matrix A:

α(ξ) = det(ξI − A)
= αn ξ n + αn−1ξ n−1 + · · · + α1ξ + α0

Characteristic equation:

α(ξ) = 0

The eigenvalues λi, i = 1, · · · , n of the system matrix A

are called the poles of the system.
The pole polynomial is defined as

π(ξ) = Πni=1(ξ − λi)

(= characteristic equation up to within a scalar)

& %

ESAT–SCD–SISTA CACSD pag. 76

Physical interpretation of a pole :
Consider the following second order system :
 
" # α β b1
A B  
=  −β α b2 
C D
c1 c2 0
The transfer matrix is given by (try to verify it) :

G(ξ) = C(ξI − A)−1B + D

ξ(b1c1 + b2c2) + β(b2c1 − b1c2) − α(b1c1 + b2c2)
=
ξ 2 − 2αξ + α2 + β 2
There are 2 poles at α ± jβ.
There is a zero at
α(b1c1 + b2c2) − β(b2c1 − b1c2)
b1 c 1 + b2 c 2

& %

ESAT–SCD–SISTA CACSD pag. 77

In continuous time the impulse response takes on the form:

g(t) = L−1 {G(s)}

 
s(b1c1 + b2c2) + β(b2c1 − b1c2) − α(b1c1 + b2c2)
= L−1
s2 − 2αs + α2 + β 2
= (Am cos(βt) + Bm sin(βt))eαt = Cmeαt cos(βt + γ)

with Am = b1c1 + b2c2 and Bm = b2c1 − b1c2

q
⇒ Cm = (b21 + b22)(c21 + c22)
 
−Bm
⇒ γ = atan
Am
Define the damping ratio ζ and natural frequency ωn:
p
α = −ζωn, β = ωn 1 − ζ 2
jω

α + jβ
β
θ ζ = sin θ
ωn
r
α

stability degree
α − jβ

& %

ESAT–SCD–SISTA CACSD pag. 78

For a second order system
ωn2
G(s) = 2
s + 2ζωns + ωn2
the time response looks like :

rise time tr ' 1.8

ωn
4.6
settling time ts = ζω n
π
p
peak time tp = ω , ωd = ωn 1 − ζ 2
d √
−πζ/ 1−ζ 2
overshoot Mp = e , 0≤ζ<1
In general : if a continuous-time system (A, B, C, 0) has
poles at α ± jβ ⇒ the impulse response will have time
modes of the form

Ameαt cos(βt + γ)
)
Am: amplitude
determined by B and C.
γ: phase
& %

ESAT–SCD–SISTA CACSD pag. 79

In discrete time the impulse response matrix Gd(k) takes
on the form :

Gd(k) = CdAk−1
d Bd , k≥1

which can be proven to be a sum of terms of the form :

Cmbk−1 cos(ω(k − 1) + γ)

each of which satisfies a second order linear system

 
" # α β b1
A B  
=  −β α b2 
C D
c1 c2 0
Parameterize A as
" #
b cos ω b sin ω
A=
−b sin ω b cos ω
then " #
k k cos ωk sin ωk
A =b
− sin ωk cos ωk

& %

ESAT–SCD–SISTA CACSD pag. 80

Now,

CAk B = (Am cos ωk + Bm sin ωk)bk

= Cmbk cos(ωk + γ)

with Am = Cm cos γ = b1c1 + b2c2

and Bm = −Cm sin γ = b2c1 − b1c2
q
⇒ Cm = (b21 + b22)(c21 + c22)
p
⇒ b = α2  + β2

β
⇒ ω = atan
α 
−Bm
⇒ γ = atan
Am

& %

ESAT–SCD–SISTA CACSD pag. 81

Transmission zeros

Definition: The zeros of a LTI system are defined as those

values ζ ∈ C for which the rank of the transfer matrix G(ξ)
is lower than its normal rank (=rank of G(ξ) for almost all
values of ξ):

rank(G(ζ)) < normal rank

Property: Let ζ be a zero of G(ξ) (p × m), then

rank (G(ζ)) < normal rank,

⇓ if ζ is not a pole of G(ξ)

 −1

∃v =
6 0 s.t. D + C(ζI − A) B v = 0,
∆
⇓ define w = (ζI − A)−1Bv
Dv + Cw = 0, (ζI − A)w − Bv = 0,
⇓
" #" #
ζI − A −B w
=0
C D v

& %

ESAT–SCD–SISTA CACSD pag. 82

 How to find zeros for square MIMO systems (p = m) with
invertible D ?

(ζI − A)w − Bv = 0, Cw + Dv = 0
⇓
Aw + Bv = wζ, v = −D−1Cw
⇓
(A − BD−1C)w = wζ

ζ = eigenvalue of A − BD−1C,
w = corresponding eigenvector !

For other cases :

• Generalized eigenvalue problem.

• Use Matlab function “tzero”.

Minimum and non-minimum phase systems:

If a system has an unstable zero (in the right half–plane
(RHP) or outside the unit circle), then it is a non-minimum
phase (NMP) system, otherwise it is a minimum phase
(MP) system.
& %

ESAT–SCD–SISTA CACSD pag. 83

Physical explanation (continuous time) :
Let ζ be a real zero, then there will exist vectors x0 and u0
such that: " #" #
ζI − A −B x0
=0
C D u0
This means if we have an input u0eζt, there exists an initial
state x0 such that the response is

y(t) = 0.

For complex zeros :

if ζ is a complex zero, then also its complex conjugate ζ ∗
is a zero. Try to prove that if
" #" #
ζI − A −B x0
=0
C D u0
the output y(t) will be exactly zero if the input is

u(t) = |u0|eRe{ζ}t · cos(Im{ζ}t + φu0 )

and the initial state is chosen to be Re{x0}. “·” is the

elementwise multiplication, both “cos()” and “e()” are as-
sumed to be elementwise operators, Re{x0} is the real part
of x0, Im{ζ} is the imaginary part of ζ and u0 = |u0|·ejφu0 .
Try to derive equivalent formulas for the discrete-time case.
& %

ESAT–SCD–SISTA CACSD pag. 84

Why zeros are important :
• Limited control system performance
k
C G

If G is NMP, k can not go to ∞, since unstable zeros

(which become unstable poles) put a limit on high gains.
• Stability of a stabilizing controller, Parity Interlacing
Property (PIP) :

Let G be unstable. Then, G can be stabilized by a

controller C which itself is stable ⇔ between every pair
of real RHP zeros of G (including at ∞), there is an
even number of poles.
Example:
s
G(s) = 2
s −1
pole zero pole zero
× ◦ × ◦
-∞ −1 0 1 ∞

G(s) can not be stabilized by a stable controller.

& %

ESAT–SCD–SISTA CACSD pag. 85

RHP zeros can cause undershoot !

Let
Πni=1(1 − ζsi )
G(s) = s ,
Πn+r
i=1 (1 − λi
)
where r = relative degree of G(s). Let y(t) be the step
response.

 y(0) and its first r − 1 derivative are 0;

Fact : y (r) is the first non-zero derivative;


y(∞) = G(0).
The system has undershoot ⇔ the steady state value
y(∞) has a sign opposite to y (r)(0), i.e.

y (r)(0)y(∞) < 0.

It can be proven that the system has undershoot ⇔ there

is an odd number of RHP zeros.

& %

ESAT–SCD–SISTA CACSD pag. 86

An example of state-space analysis

Example: Tape drive control - state space analysis

Controllability/observability

Direct rank test using SVD:

The singular values of C are

4.5762, 4.3861, 1.2940, 1.2758, 0.5737, 0.4787,

so the system is controllable (rank=6).

The singular values of O are

2.5695, 1.4504, 0.7145, 0.7071, 0.5171, 0.2356,

so the system is observable (rank=6).

& %

ESAT–SCD–SISTA CACSD pag. 87

PBH test:
There are 6 (independent) left eigenvectors associated with
the 6 eigenvalues of A :
   
−0.1480 ± 0.0826i 0.0766
   
 0.0703 ± 0.5380i   0.5624 
   
 0.1480 ∓ 0.0826i   0.0766 
   
 ,  
 −0.0703 ∓ 0.5380i   0.5624 
   
 0.2994 ± 0.2954i   0.4218 
   
−0.2994 ∓ 0.2954i 0.4218
     
0 −0.0046 0.0000
     
 0.4892   −0.0227   −0.0295 
     
 0.0000   0.0046   0.0000 
     
 ,  ,  
 0.4892   0.0227   −0.0295 
     
 0.5105   −0.7067   −0.7065 
     
0.5105 0.7067 −0.7065
It can be easily checked that none of them is orthogonal to
B. The result is the same as the rank test above.

The same can be done to investigate the observability.

& %

ESAT–SCD–SISTA CACSD pag. 88

Poles and zeros:

Poles: the eigenvalues of A are

−0.2370 ± 0.5947i, 0, −0.2813, −0.9760, −0.9687

The system is not stable since there is a pole at zero.

The first two oscillatory poles are from the spring-mass

system consisting of the tape and the motor/capstan in-
ertias. The zero pole reflects the pure integrative effect of
the tape on the capstans.

& %

ESAT–SCD–SISTA CACSD pag. 89

Zeros: it can be checked in Matlab using function “tzero”
that there is only one zero at -2. The eigenvector of the
matrix " #
−2I − A −B
C 0
associated with the eigenvalue 0 is
 
−0.1959
 
 0.1959 
" #   " #
x0  
 0.1959  0.4630
with x0 =   , u0 = .
u0  −0.1959  −0.4630
 
 −0.4571 
 
0.4571
Thus, let
x(0) = x0, u(t) = u0e−2t,
then the output y(t) will be always zero, ∀t ≥ 0.

& %

ESAT–SCD–SISTA CACSD pag. 90

Physical explanation of the zero:

Since the initial values i1(0) = −i2(0) and the inputs

e1(t) = −e2(t), and the motor drive systems in both sides
are symmetric, the two capstans will rotate at the same
speed, but in opposite directions. So the tape position over
the read/write head will remain at 0 if the initial value is
at 0.

The tension in the tape is the sum of the tensions in the

tape spring and the tape damping. The input voltages vary
such a way that the tensions in the tape spring and the tape
damping cancel each other out.

& %

ESAT–SCD–SISTA CACSD pag. 91

 Matlab Functions

c2d
c2dm
canon
expm
tf2ss
ss2tf
(d)impulse
(d)step
minreal
tzero
ctrb
obsv
ctrbf
obsvf
(d)lsim
ss
ssdata

& %

ESAT–SCD–SISTA CACSD pag. 92