Discrete-time systems:

discretization, models and their

properties
Process controlled by a digital controller

u(tk) u(t)
e(tk) y(t)
r(tk) +
Controller D/A Process
_

A/D

y(tk) y(t)
Two main design approaches: a. discretize the
analog controller, b. discretize the process and
do the design totally in discrete time

a. r(t) u(t) y(t)

A/D Controller D/A Process
y(t)

r(tk) u(tk) y(tk)

b. Controller D/A Process A/D

y(tk)
Department of Automation
and Systems Technology
Let us consider the design approach b: A
discrete system from the controller viewpoint

r(tk) u(tk) y(tk)

Controller D/A Process A/D

y(tk)
 Sampling of continuous-time signals

A continuous function f(t) f ( t )  2e t

Set of integers Z l
Z   ,1,0,1, q
Sampling instants { tk : k  Z } l0,1,2,3,q
Sequence { f(tk) : k  Z } m r
2,2e,2e 2 ,2e 3 ,
l
 2.00 , 5.44 , 14.78 , 4017 q
. ,
Sampling of continuous-time signals

Periodic sampling

Sampling interval, h tk  k  h

Sampling frequency, fs f s  1/ h (Hz)

Sampling angular frequency s  s  2 / h  2  f s (rad / s)

Nyquist frequency, fN f N  1 / (2h) (Hz)

N  N  2 / (2h)   / h  2  f N (rad/s)
 Zero-order-hold ZOH and sampling

A continuous process is R|x (t )  Ax(t )  Bu(t ) dim{u}  r

described by a linear state-
space-representation.
S|y(t )  Cx(t )  Du(t ) dim{x}  n
T dim{y}  p

{u (tk)} u(t) y(t) {y (tk)}

D/A Process A/D

 The matrix exponential

Let A be a square matrix, define


1 1 1 i i
e  I  tA  t A  t A     t A
At 2 2 3 3

2 6 i 0 i !

which is always convergent.

de At
From the definition  Ae At  e At A
dt
The matrix exponential

The solution to the homogenous set of differential equations

x (t )  Ax(t ), x(0)  x0

is then x(t )  e At x(0)  e At x0

The term e At is called the state transition matrix

(in this context).
The matrix exponential

Solution by using the Laplace transformation

x (t )  Ax(t ), x(0)  x0 sX ( s )  x0  AX ( s )

It follows ( sI  A) X ( s )  x0
( sI  A) 1 ( sI  A) X ( s )  ( sI  A) 1 x0
X ( s )  ( sI  A) 1 x0

The solution is obtained by the inverse transformation

The matrix exponential

By comparing to the previous solution it follows that

e  L ( sI  A) 
At
 1 1

which is one way to solve the state-transition matrix.

 Solution of the state equation
Solution for the input-output representation

x (t )  Ax(t )  Bu (t ), x(0)  x0
y (t )  Cx(t )

is
t
x(t )  e At x0   e A( t  ) Bu ( ) d
0
t
y (t )  Ce At x0   Ce A(t  ) Bu ( ) d
0
Solution of the state equation
in which e At is the above state transition matrix.

To prove the solution check the initial condition and

differentiate the solution to see that the original
differential equation holds.
Zero- order- hold ZOH and sampling

Solution at any time t after

z
t
the sampling instant tk x(t )  e A ( t  t k ) x(t k )  e A ( t  s ') Bu( s' )ds'
tk
u(t) is constant between

z
t
sampling instants, ZOH
x(t )  e A ( t  t k ) x(t k )  e A ( t  s ') ds'Bu(t k )
tk
Change the integration
variable s’ = t - s c
x( t )  e A ( t tk )
F
h x(t )  GH z e
t tk
As
I
ds BJ u(t )
A state transition matrix 
k
0 K k

and control matrix  are x( t )   ( t  t k ) x( t k )   ( t  t k ) u ( t k )

obtained (independent of x
and u).
Zero- order- hold ZOH and sampling

At the next sampling instant

t  t k 1  RSx(t )   (t  t ) x(t )   (t
k 1 k 1 k k k 1  t k ) u(t k )
Ty(t )  Cx(t )  Du(t )
k k k

R|(t  t )  e
k 1 k
A ( t k 1  t k )

S| (t  t )  e ds B
z
t k 1  t k
As

T k 1 k
0
 Zero- order- hold ZOH and sampling

By a periodic sampling the equations become

tk  k  h  tk 1  tk  h (constant)

RSx(kh  h)   (h) x(kh)   (h) u( kh) R|(h)  e Ah

Ty(kh)  Cx(kh)  Du(kh) S| (h)  z e

h
As
ds B
T 0
Usually this is written in the form (h is constant)

RSx(k  1)   x(k )   u(k )

Ty(k )  Cx(k )  Du(k )
 How to solve for  and  ?

• Symbolically
Eg. by using the Laplace transformation
By symbolic programs (Maple, Mathematica, …)
• Numerically
Eg. by the series expansion of the matrix exponential
function
By numeric software (Matlab)
 Example. Discretization by direct calculus
Sampling interval h = 0.1

RSx (t )  2 x(t )  u(t ) is of the RSx(t )  Ax(t )  Bu(t )

T y ( t )  3x ( t ) form
T y(t )  Cx(t )
R|  e  e  e
h 20.1 0.2

S|  e ds B  e
z z
h 0.1

d i d i
0.1
1 1 20.1 0 1 0. 2
As 2s
ds  / e  e  e  e 1
2s

T 0 0
2 0 2 2
RSx(kh  h)   x(kh)   u(kh) RSx(kh  h)  e x(kh)  0.5(e
0 .2 0 .2
 1) u( kh)
T y(kh)  Cx(kh) T y( kh)  3x( kh)
 Example. Discretization by using the series
expansion
State-space representation of the double integrator

R|x (t )  L0 1O x(t )  L0O u(t ) A 2h2

S| N Q N Q M 0 0P M 1P   e Ah  I  Ah 
2


T y ( t )  1 0 x( t )
M
L 1 0O L0 hO L0 0O
P M P M P   M
L 1 hO
P
N Q N Q N Q N Q
0 1 0 0 0 0 0 1

  z e ds B  z M
h
L
h
1 s O L 0O h
L sO L
ds M P  z M Pds  / M
s O L h O
h

P M
1

P
2 1 2

0
As

0
1P
N Q NQ NQ NsQ NhQ
0 1 0
1 0
2 2
 Example. Discretization by the series expansion

The corresponding discrete-time model becomes

R|x(kh  h)  L1 hO x(kh)  L h O u(kh)

2

M P
1

S| M
N Q
0 1 P 2

NhQ
T y(kh)  1 0 x(kh)
 Example. Discretization by using the Laplace
transformation
State-space representation of the DC motor
R|x (t )  L1 0O x(t )  L1O u(t )
S| M
N 1 0P
Q M
N 0P
Q   e Ah  e At
th
m
 L 1 ( s I  A )  1 r th

T y ( t )  0 1 x( t )
R|Ls  1 0O U| 1
RS 1 Ls OPUV
L 1
S|MN 1 sPQ V| L1

T M
N1
0
s  1Q W
T W s ( s  1)
t h
R|L 1 O0PU|
t h

MSM s  1 L t
OP L h
OP
1 PV M M
1 e 0 e 0
L
||MN s(s1 1) Ps Q|| N1  e t
1Q N1  e h
1Q
T W t h
t h
 Example. Discretization by the Laplace ….

L OP L O LM e OPds  LM 1 e OP
z ds B  z M
s

z s h
h h h
1
  e As
0 0
e
N1 e s
0
1
M0P
ds 
Q N Q N1 e Q Nh  1  e Q
0
s h

The discrete-time model is obtained

R|x(kh  h)  L e 0O x(kh)  L 1  e O u(kh)

h h

S| M P
N1  e 1Q
h M P
Nh  1  e Q h

T y( kh)  1 0 x(kh)
 Pulse transfer function, H(z)
A pulse transfer function H(z) can also be calculated directly
from G(s).

H(z)

u(t) y(t)
{u(kh)} {y(kh)}
ZOH G(s)

zero order continuous sampling

hold process
 Pulse transfer function, H(z)
After ZOH u(t) consists of step functions. The output y(t) is a
set of step responses, which are sampled at constant intervals
h.

Discrete pulse transfer function and continuous transfer

function correspond to each other, if the outputs are equal at
the sampling instants.
The step response of the continuous system at the sampling
instants (t = kh)
l q
y ( kh)  y (t ) t  kh  L1 Y ( s)
t  kh
l q
 L1 G ( s)U ( s)
t  kh

L 1 RSG(s) 1UV
T sW t  kh
 Pulse transfer function, H(z)
The step response of the discrete system

y ( kh)  Z 1
lY ( z)q  Z lH (z)U (z)q  Z
1 1 RSH (z) z UV
T z  1W
Make these equal; the final result follows

1 RSG(s) 1UV R
 Z S H ( z)
z U
V
1
L
T sW T z  1W
t  kh

z 1 R R 1U U
H ( z)   Z S L SG ( s) V V
1

z T T sW W t  kh
 Can a pulse sequence be used to construct the
original continuous signal ?

samples y(t)=sin(0.2*pi*t)
Two different continuous 1 1
signals can be fitted to 0.5 0.5

the pulse train (h = 1) 0 0

-0.5 -0.5

y1(t) = sin(0.2π t) -1
0 5 10
-1
0 5 10
y2(t) = sin(1.8π t) y(t)=sin(1.8*pi*t)

1 1

Actually, an indefinite 0.5 0.5

number of continuous 0 0

-0.5 -0.5
signals can be fitted. -1 -1
0 5 10 0 5 10
Is it possible to find an ZOH-equivalent continuous
model for a given discrete model ?
Consider a simple example with one state variable only
RSx( k  1)   x( k )   u(k ) RSx(t )   x(t )   u(t )
T y(k )  Cx(k ) T y(t )  Cx(t )
R|  e Ah R| A  1 ln( )
S|  z e
h
As
ds B  c
B Ah
h
e 1
S| 1 h

T 0
A |T h   1
B  ln( )

If Φ is positive, a corresponding continuous model exists. For

negative Φ:s the logarithm is not real, and the model does not
exist.
 Is it possible to find an ZOH-equivalent continuous
model for a given discrete model ?
A more general case (multivariable system of nth order): The
result is known, but it is difficult to express. Sometimes you
can find the following in the literature:
“If matrix Φ has even one eigenvalue on the negative real
axis, the corresponding continuous-time system doesn’t exist.”
The above claim is incorrect however. (Take G(s) = 1/(s2+1),
do a state-space realization and discretize by using the sampling
interval h = .)
Moreover: the continous time model corresponding to a discrete
model is not always unique.
(Take the harmonic oscillator G(s) = 2/(s2+2),
do a realization and discretize by using  = +n2/h, n = 0,1,2,...)
 Discretization of systems with delay
With discretization an infinite dimensional (continuous) model
becomes finite dimensional (discrete). Analysis and controller
design of such systems then becomes easier.

Consider first the case, in which the delay is shorter than the
sampling interval (t < h)

RSx (t )  Ax(t )  Bu(t   )

Ty(t )  Cx(t )  Du(t )
The solution can be formed as
 Discretization of systems with delay

z
kh  h

x( kh  h)  e Ah x( kh)  e A ( kh  h  s ') Bu( s' )ds'

kh
Even though u(t) is constant over
the integration interval, u(t-t) is
not! u(t)

Divide into two intervals, where

the
t
control signal is constant.
u(t-)

t1[kh , kh + t [ kh  kh +2h t
t2[kh + t , kh +h [
 Discretization of systems with delay

z
kh  h

e A ( kh  h  s ') Bu( s' )ds'

kh

z z
kh  kh  h

 e A ( kh  h  s ') Bu( s' )ds'  e A ( kh  h  s ') Bu( s' )ds'

kh kh 

F
G ze
kh 
A ( kh  h  s ')
I F
ds'BJ u( kh  h)  G z e
kh  h
I
ds'BJ u( kh)
A ( kh  h  s ')

H kh K H kh  K
F I F I
z dsBJ u( kh  h)  G z e dsBJ u( kh)
 h 

 Ge A ( h  )
e As As

H 0 K H 0 K
 Discretization of systems with delay

R|
z

A ( h  )
 e e As
dsB
  1u( kh  h)   0u( kh) |S 1
0

|| z
h 

 e As dsB
T 0
0

The system model becomes

x( kh  h)   x( kh)   0u( kh)   1u( kh  h)

 Discretization of systems with delay
To form a general state representation the delayed controls
can be chosen as additional state components (augmentation
of state)
x u ( kh)  u( kh  h)  x u ( kh  h)  u( kh)
L
x ( kh)  M
x( kh) O
P
*
N Q
x (
u kh )
The new state representation is
L
x ( kh  h)  M
 OP x (kh)  LM OPu( kh)
1 0
*
N0 0Q *
NIQ
x* (kh  h)  * x* (kh)  *u(kh) L
 M
 OP ,
1
*
L  O
 M Pu( kh)
0
*
N0 0Q NIQ
 Example. Double integrator with delay

The system matrix Φ was discretized earlier

R|x (t )  L0 1O x(t )  L0O u(t   )

M P M P L
(h)  M
1 hO
P
S| N Q N Q 0 0 1 N Q
0 1
T y ( t )  1 0 x( t )
R|  e L1 h   O L1 sO L0O L OP LMsOPds
z e dsB  MN0 1 PQz MN0 1PQ MN1PQds  MN0 h 
zQ N1Q
  
A( h ) As
1
|S1
0 0
1 0

||  z e L1 sO L0O LsO

dsB  z M ds  z M Pds
h h h

T
0
0
As

0
P M P
N0 1Q N1Q N1Q 0
 Example. Double integrator with delay
R| L 1 OP LM s OP  LM1
h   1 2
h  OP LM  OP  LM bh 
1 2 1
 gOP
 M
| 1 Q / N s Q N0
2 2 2

S|
1
N0 0 1 QN  Q N  Q
L s O L bh   g O
h  2

 /M PM P
2

||
1 1
2 2

T 0
N s Q MN h   PQ
0

The discrete model becomes

R|x(kh  h)  L1 hO x(kh)  L bh g Ou(kh)  Lbh  gOu(kh  h)
2

M P M P
1

M P
1
2 2

S| N Q MN h  PQ N
0 1  Q
Ty(kh)  1 0 x(kh)
 Example. Double integrator with delay
By using state augmentation, the general representation
follows
R|x (kh  h)  L  O x (kh)  L Ou( kh)
S|* M
N Q
0 0 P 1
*M
N Q
I P 0

Ty(kh)  C 0 x (kh) *

R| LML1 hO L bh   gOPOP LML bh   g OOP

2

M
1

M P M P
1
2

|S MMN Q N 
2
x ( kh  h )  0 1 QP x ( kh )  M MN h   PQPu(kh)
PQ MN 1 PQ
* *

|| N 0 0 0

T y(kh)  1 0 0 x (kh) *
 Example. Double integrator with delay

R| LM1 h  bh   gO LM bh   g OP
2

PP x (kh)  M h   Pu(kh)
1 1
2 2

|Sx (kh  h)  M0 1 
*

|| M
N0 0 0 PQ
*
MN 1 PQ
T y(kh)  1 0 0 x (kh)
*
 Long delay

If the delay is longer than the sampling interval (t > h), the
formulas must be modified a bit:

  (d  1)h   ' 0  ' h d  Z   {1,2,3,}

A similar result as in the previous case can be derived

x( kh  h)   x( kh)   0u( kh  (d  1) h)   1u( kh  dh)

where the coefficients are as earlier but  ' is used instead of

The delay terms can be taken as augmented state variables
 Long delay

LM 1 0  0 OP LM0OP
M
x ( kh  h)  M 
0 0

I  0
PP
   x* ( kh) 
MM0 PPu(kh)
*
MM 0 0 0  I
PP MM0PP
MN 0 0 0  0 PQ MN I PQ
LM x ( kh ) OP
x ( kh )  M
M u ( kh  dh )

P
P
*
MMu ( kh  2 h ) PP
MN u ( kh  h ) PQ
Example. “paper machine” model

h = 1, d = 3,  = 2.6,  = 0.6 x (t )   x (t )  u(t  2.6)

  e 1  0.3679
'
 0.6

 
A ( h  ') 0.4 s 0.4 1

 1  e e As
dsB  e e ds  e  e  0.3024
 0 0
 h  ' 0.4
  e As dsB  e  s ds 1  e 0.4  0.3297
 0  
 0 0

A simple model is obtained

x ( kh  h)  0.3679 x ( kh)  0.3297u( kh  2h)  0.3024u( kh  3h)

Example. “paper machine” model

R| LM0.3679 0.3024 0.3297 0 OP LM0OP

||x (kh  h)  M 0 0 1
P1P x (kh)  MM00PPu(kh)
0
S|
*
MM 0 0 0
P0Q
*

MN1PQ
N
|| y(kh)  1 0 0 0 x (kh)0 0 0

T *

LM x(kh) OP
* x ( kh)  M u( kh  3h)
MMu(kh  2h)PPP
N u(kh  h) Q
Discrete time systems

x(k  1)  x(k )  u(k ), initial value x(k0 )

Discrete system: 
y (k )  Cx(k )  Du(k )

Solution by x(k0  1)  x(k0 )  u(k0 )

direct x(k0  2)  x(k0  1)  u(k0  1)
recursive
calculus  2 x(k0 )  u(k0 )  u(k0  1)
starting from a 
given initial x(k )  k k0 x(k0 )  k k0 1u(k0 )    u(k  1)
state x(k0) up k 1
to any time k .  k  k0
x(k0 )   k  j 1u( j )
j  k0
 Input-Output-models, I/O-models, pulse response

Input and output signals as sequences

U  u(0) u(1)  u( N  1)
T

Y  y (0) y (1)  y ( N  1)
T

A linear model giving the relationship between Y and U is

Y  HU  YP
For a causal system H is a lower triangular matrix (element
h(k,m))
k
y ( k )   h ( k , m)u(m)  y p ( k )
m 0

h(k,m) is the pulse response or weighting function.

Input-Output-models, I/O-models, pulse response
For a time-invariant system the pulse response depends only
on k-m:
h ( k , m)  h( k  m)
For a state representation the pulse response is calculated
as
k 1
x( k )   k  k0
x(k0 )    k  j 1u ( j ) y ( k )  Cx( k )  Du( k )
j  k0
k 1
y (k )  C k  k0
x(k0 )   C k  j 1u ( j )  Du (k )
j  k0

0 k 0

h( k )   D k 0 pulse (assumek0  0)
C k 1 k  1 response

Linear time-invariant systems: The pulse transfer function,
weighting function, pulse response and convolution sum

Y ( z )  H ( z )U ( z ) Input-output: Z-domain
k k
y (k )   h(k  i )u (i )   h(i )u (k  i )
i 0 i 0
Input-output: time domain,
convolution sum
The z-transform of the weighting function h(k) is the
pulse transfer function H(z).
The (im)pulse response coincides with the weighting function
from the time that the pulse enters.
 Shift-operators
Corresponding to the differential operator p used in
continuous systems a (forward) shift operator q is defined for
discrete systems
q  f ( k )  f ( k  1)
q 1  f ( k )  f ( k  1)
By using the shift operator input-output relationships (difference
equations) can easily be described
y( k  na )  a1 y( k  na  1) ana y( k )  b0u( k  nb ) bnb u( k )
(q na  a1q na 1  ana ) y ( k )  (b0q nb  b1q nb 1 bnb )u( k )
 Shift-operators

Difference equations can be described by polynomials

A( z )  z na  a1z na 1  ana

A( q ) y ( k )  B ( q ) u ( k )
nb 1
B ( z )  b0 z  b1z
nb
bnb

A backward shift operator q-1 can also be used

y( k )  a1 y( k  1) ana y( k  na )  b0u( k  d ) bnb u( k  d  nb )

d  na  nb

(1  a1q 1  ana q  na ) y ( k )  (b0  b1q 1 bnb q  nn ) u( k  d )

Shift-operators

A* ( z )  1  a1z  ana z na A * ( q  1 ) y ( k )  B * ( q 1 ) u ( k  d )

B * ( z )  b0  b1z bnb z nb A * ( q  1 ) y ( k )  q  d  B * ( q 1 ) u ( k )

Resiprocal polynomials A* , B*

Polynomial representations are used, because in some

control design methods they are more natural to use than
state-space representations.
 Pulse-transfer operator, H(q)
The pulse transfer operator is an I/O-representation obtained
by eliminating internal variables. E.g. from state
representation

RSx(k  1)  x(k )  u(k ) RSq  x(k )  x(k )  u(k )

Ty(k )  Cx(k )  Du(k ) Ty(k )  Cx(k )  Du(k )
RS(qI  )x(k )  u(k ) R|Sx(k )  bqI  g u(k )
1

Ty(k )  Cx(k )  Du(k ) |Ty(k )  Cx(k )  Du(k )

y( k )  eCbqI   g   Dju( k )  H(q )u( k )
1
 Pulse-transfer operator, H(q)
From polynomial representations
B(q )
A( q ) y ( k )  B ( q ) u ( k ) H ( q ) 
A(q )
or
d * 1
1 d 1 q B ( q )
A (q ) y(k )  q  B (q )u(k )
* * * 1
H (q ) 
A* ( q 1 )

Representation by reciprocal polynomials

H * ( q 1 )  H ( q )
 Pulse-transfer operator, H(q)

Consider a simple example of a pulse transfer function.

RSx(k  1)  0.5x(k )  0.5u(k )

T y( k )  2 x( k )
We obtain
2  0.5
b
H (q )  C q   g 1
 D 
1
q  0.5 q  0.5
Starting from IO-difference equation the calculation is equally
easy
 Pulse-transfer operator, H(q)

RSx(k )  y(k )
1
2
0.5 y ( k  1)  0.25 y ( k )  0.5u( k )
Tx(k  1)  0.5x(k )  0.5u(k )
Form A- and B-polynomials

(0.5q  0.25) y ( k )  0.5u( k ) (q  0.5) y ( k )  u( k )

A(q )  q  0.5
A( q ) y ( k )  B ( q ) u ( k )
B(q )  1
B(q ) 1
The pulse transfer operator is H (q )  
A(q ) q  0.5
 Pulse-transfer operator, H(q)
Let us calculate the same with the q-1-operator

. y(k  1)  025
05 . y(k )  05
. u(k ) 0.5 y ( k )  0.25 y ( k  1)  0.5u( k  1)

. q1 ) y(k )  05
.  025
(05 . q1u(k ) (1  0.5q 1 ) y ( k )  1 u( k  1)

A* (q 1 )  1  0.5q 1
A* (q 1 ) y( k )  q  d  B* (q 1 )u( k ) B * ( q 1 )  1
The result becomes d 1

* 1
H (q ) 
q d *
B (q 1
)

q 1
1 FG q 1

1 IJ
* 1
A (q ) . q 1
1  05 H 1  0.5q 1
q  0.5 K
 Pulse-transfer operator, H(q)

All methods lead to the same pulse transfer operator

FG q 1

1 IJ
H 1  0.5q 1
q  0.5 K
H * ( q 1 )  H ( q )
 Pulse-transfer operator, H(q)
•The characteristic polynomial of the system is the
denominator A(q) of the pulse transfer operator.

•The poles of a discrete system are the zeros of the

characteristic polynomial.

•The zeros of a discrete system are the zeros of the

numerator polynomial B(q).

•More delay means more pole excess d.

•The order (dimension) of the system is the same as the
dimension of the state-space representation or the number of
poles.
 Poles and zeros
Some basics from matrix calculus. For the pulse transfer
operator we can write

1 C adj(qI ) D det(qI )

H(q)  C(qI )  D 
det(qI )
where ’adj’ means the adjungate matrix. For the matrix
 the eigenvalues λi and eigenvectors ei are defined as follows

ei  i ei  (i I  )ei  0

There exist non-zero eigenvectors when det(i I  )  0

 Poles and zeros

The poles are the roots of the characteristic polynomial. They

Belong to the set of eigenvalues of the system matrix Φ.

Consider the polynomial

f ( )   0  m  1 m 1     m
The corresponding function for the square matrix A is defined as
f ( A)   0 Am  1 Am 1     m I

Result: Let the eigenvalues of A be λi and the corresponding

eigenvectors ei. It holds
 Poles and zeros

f ( A)ei  f (i )ei

meaning that f(λi) is the eigenvalue of f(A) and the

Corresponding eigenvector is ei.

The proof is straightforward by writing the above formula and

using repeatedly the definitions of eigenvalue and eigenvector
 Exercise.
 Poles and zeros
The poles of a discrete system are the zeros of the
denominator of H(z). The zeros are the zeros of the
numerator. The poles are also eigenvalues of the the system
matrix Φ.

From the location of poles in the complex plane stability,

oscillations and speed of the system can be deduced.

The poles of a continuous n:th order system are mapped to

the discrete system poles according to:

RSx(t )  Ax(t )  Bu(t ) b g

poles:  i A , i  1,  , n
Ty(t )  Cx(t )
Poles and zeros

Discrete system RSx(kh  h)   x(kh)   u(kh)

Ty(kh)  Cx(kh)
b g
poles:   , i  1,  , n
i

e Ah
b g
=>  i   e  i b A gh

A simple relationship for the mapping of zeros does not hold.

Even the number of zeros does not necessarily remain
invariant. The mapping of zeros is a complicated issue.
Poles and zeros

Mapping z  e sh  e( h  j h )  e h e j h
for s    j
 Poles and zeros
Second order system: Mapping of poles as functions of the damping ratio 
and natural frequency  (dashed lines denote constant values of 0 h

02
G ( s)  2
s  20 h  02
 Unstable inverse, non-minimum phase systems

•A continuous time system is non-minimum phase, if has

zeros on the right half plane (RHP) or if it contains a delay.

•A discrete system has an unstable inverse, if it has zeros

outside the unit circle.

•Zeros are not mapped in a similar way as poles, so a

minimum phase continuous system may have a discrete
counterpart with an unstable inverse and a non-minimum
phase continuous system may have a discrete counterpart
with a stable inverse.
 Selection of the sampling rate
• The proper choice of the sampling interval is very
important. Too low sampling frequency may lose so much
information that the control performance deteriorates and
the system dynamics is lost.

• Too high sampling rate increases the burden of the

processor; also, it may lead to discrete representation with
bad numerical properties.

• For oscillating systems the sampling interval is often tied to

the frequency of the dominating oscillation. For damped
systems the sampling interval is usually chosen to be in
relation to the time constant.
 Selection of the sampling rate

Nr means the amount of samples during the rise time.

Tr
Nr 
h
For a self-oscillating system (2nd order, damping ratio z and
natural frequency w0) the rise time is:

1
Tr   0 e tan ,   cos
Tr
A usual sampling rate isN r   4  10
h
Selection of the sampling rate

Sampling examples for

a sinusoidal and
exponential signal.

a. Nr = 1
b. Nr = 2
c. Nr = 4
d. Nr = 8