5.

SAMPLED SIGNALS AND DIGITAL SYSTEMS

In a great number of modern control applications, digital computing equipment is used to evaluate
control algorithms. The general structure of a computer control system is

Computer disturbance
d(t)
reference pre-filter controller input plant + output
r(t) + u(t) + y(t)
F(w) G(w) P(s)
A/D D/A
-

A/D

The computer works with a sequence of numbers although you may read all sorts of confusing
things about comb functions and trains of impulses etc. We usually try to take samples of signals at
equidistant times, T. In the same way as we were able to define the Laplace transform of a signal
in the continuous time, we will define the Z transform of discrete signals. We need to be a little
careful with the relationship between Z transforms of signals when speaking about systems.

5.1. The Z transform of a signal

Consider the signal, y( t ) = e − t shown alongside and the sequence of numbers, yi=y(iT), T=0.2

ti y(ti) 0.9
0.0 1.0000
0.8
0.2 0.8187
0.4 0.6703 0.7
0.6 0.5488
0.6
0.8 0.4493
1.0 0.3679 0.5 y0
1.2 0.3012 y1
0.4
1.4 0.2466 y2
y3
1.6 0.2019 0.3 y4
1.8 0.1653 y5
0.2
2.0 0.1353 y6
0.1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

We can describe y(ti) = yi by means of a recurrence relationship (i.e a simple geometric series):

y i = e − T y i −1 , e −0.2 = 0.8187

In the same way as the continuous signal has a Laplace transform,

Systems & Simulation - ENEL3SS 5-1

SS1_5.DOC
 ∞
L{y( t )} = Y( s) = ∫ y( t ) e − st dt (5.1)
0−

the sequence, y(n) has a Z transform,

∞
Z{ y( n )} = Y( z ) = ∑ y i z − i (5.2)
i=0

For example, a sequence of ones has Z transform,

∞
1 z
Z{1} = ∑ 1 z − i = = (5.3)
i= 0 1 − z −1 z −1

5.2. Properties of the Z transform

5.2.1. Backward shift

∞ ∞
−i
Z{ y( n − 1)} = ∑ y i−1 z = ∑ y k z − k −1
i =0 k =−1
∞
= z −1 ∑ yk z−k (5.4)
k =−1
−1
= z Z{y( n )} + y ( −1)

5.2.2. Complex translation

∞
{
Z e − αn
}
y( n ) = ∑ e − αi y i z − i
i=0
∞
( )
−i
= ∑ yi eα z (5.5)
i=0
= Y (e α z )

Example 1, (from above samples of e-t)

{
Z e −0 .2n = } z
z − e −0 .2
, where y(n) is a sequence of ones and Y( z) =
z
z −1
(5.6)

In general,

{ }
Z e − anT = Z{1}
z = z e aT
=
z
z − e − aT
(5.7)

Systems & Simulation - ENEL3SS 5-2

SS1_5.DOC
Example 2
⎧⎪ e jβnT + e − jβnT ⎫⎪
Z{cos(βnT)} = Z⎨ ⎬
⎪⎩ 2 ⎪⎭
1⎛ z z ⎞
= ⎜ + ⎟
2 ⎝ z − e − jβT z − e jβT ⎠
z( z − cos(βT))
=
2
z − 2z cos(βT) + 1

These properties allow us to develop Z transform tables and we will use them in much the same
way as Laplace transform tables.

5.3. Inverse Z transform

5.3.1. Tables
One simple way to obtain the inverse Z transform is to write it out as a partial fraction expansion if
possible and use tables, as we have done with the Laplace transform.

Example
3z 2 − 2z
Find the inverse Z transform of Y( z ) = .
z 2 − 1.25z + 0.375
⎧ z 2z ⎫
y( n ) = Z −1 {Y( z )} = Z −1 ⎨ + ⎬
⎩ z − 0.75 z − 0.5 ⎭
= e − an + 2e − bn , e − a = 0.75, e − b = 0.5
y( n ) = e −0 .288n + 2e −0 .693n
= (3, 1.75, 1.06, 0.67, 0.44, K )

5.3.2. Long Division

Notice that we can also use a series expansion or long division in the case of rational functions.

Example above:
3 + 1.75z −1 + 1.06z −2 + 0.66z −3 +K
z 2 − 1.25z + 0.375 3z 2 − 2z etc
3z 2 −3.75z +1125
.
1.75z −1125
.
1.75z −21875
. +0.6563z −1
1.0625 −0.6563z −1
def ∞
Since Z{ y( n )} = ∑ y i z − i , coefficients of the expansion are the values of yi required.
i=0

Systems & Simulation - ENEL3SS 5-3

SS1_5.DOC
5.3.3. Residue theorem
We can also use Cauchy’s residue theorem to obtain the inverse Z transform. This will not be
covered in this course.

5.4. Relationship between poles of a signal in the Laplace and Z domains

Consider
1
y( t ) = e − at Y (s ) = pole at s=-a (5.8a)
s+a
z
y( n ) = e − anT Y( z ) = pole at z= e − aT (5.8b)
− aT
z−e
The poles of a signal are related by z = e sT . Be careful of placing too much importance on this
relationship as it applies to the poles and not the zeros and it also does not apply to the Z domain
“transfer function”. The mapping is illustrated below.

(ref. Franklin & Powell, p.656)

5.3.1. Mapping has the following consequences:

1) Stability is determined by z ≤ 1 (Radius of convergence of geometric series.)
2) z=1 corresponds to s=0
3) z plane pole locations depend on T
Systems & Simulation - ENEL3SS 5-4
SS1_5.DOC
4) Negative real axis represents frequency ωs /2 where ωs = 2π/T is the sampling rate.
5) Vertical lines in s plane (constant time constant, τ) map to circles in the z plane.
6) Horizontal lines in the s plane map to radial lines in the z plane.
7) No location on the z plane represents frequencies above ωs/2 (Called aliasing or “wagon-wheel
effect”).
8) The mapping is conformal. (Angle preserving.)

Initial Value Theorem

def ∞
As Z{ y( n )} = ∑ y i z − i , y0 = lim Y( z) (5.9)
i=0 z→∞

(Recall the initial value theorem for Laplace transforms,

y(0+) = lim sY( s) )
s→∞

Final Value Theorem lim y N = lim ( z − 1) Y( z) (5.10)

N →∞ z→1

∞
Z{y ( n + 1) − y ( n )} = ∑ ( y i +1 − y i ) z− i
i=0
= ( z − 1) Y( z) − z y 0

let z → 1,
∞
lim Z{y ( n + 1) − y ( n )} =
z→1
∑ ( y i +1 − y i )
i=0
= lim y N − y 0
N →∞
lim y N = lim ( z − 1) Y( z)
N →∞ z→1

(Recall the final value theorem for Laplace transform,

lim y ( t ) = lim sY( s)

t→∞ s→ 0

Systems & Simulation - ENEL3SS 5-5

SS1_5.DOC
5.4. Systems described by difference equations

Real world systems are described by continuous differential equations as their signals exist for all
time. Assume for the moment that we can somehow get a system that is described by a difference
equation. (An example would be the adding of interest to a bank account at the end of a month.)

E.g. y(n) + a1 y(n-1) +a0 y(n-2) = b2 u(n) + b1 u(n-1) + b0 u(n-2)

We do not know what happens between sampling instants. Examine Y(z) = Z{y(n)}:

Y( z) + a1z−1Y( z) + a 0z−2 Y( z) = b 2 U( z) + b1z−1U( z) + b 0z−2 U( z)

b + b1z−1 + b 0z−2
Y ( z) = 2 U ( z)
1 + a1z−1 + a 0z−2
b z2 + b1z1 + b 0
= 22 U ( z)
z + a1z1 + a 0

5.4.1. Obtaining sequences from a sampled continuous plant

Consider plant/system described by

x = Ax + Bu (5.11a)

y = Cx + Du (5.11b)

(
A t − to )x t A ( t − τ) B u( τ) dτ
x=e o + ∫t o − e (5.12a)
y = Cx + Du
(5.12b)

Recall the transition matrix, Φ = eAt

Φ = e At = L-1{(sI -A)-1} (5.13)

t A ( t − τ ) B u( τ) dτ → convolution integral
∫o − e

(φ(t) * B u(t) = L-1{ (sI - A)-1 B u(s)}

The transfer function is,

Y(s) = [C(sI -A)-1 B + D]U(s) (for xo = 0) (5.14)

For discrete times, ti, ti+1

Systems & Simulation - ENEL3SS 5-6
SS1_5.DOC
 A( t i + 1 − t i ) t i +1 A ( t i − τ)
x ( t i + 1) = e x( t i ) + ∫ e B u( τ) dτ
ti
(5.15)
xi + 1 = e ATxi + ∫ e A (T − τ) B u( t i + τ) dτ
T
0

yi = Cxi + Dui

The behaviour of {xi} and {yi} depend on how u(t) behaves over all time. To make {xi}, {yi}
unique, we specify how u(t) will behave (e.g. ramp/staircase etc.) between samples.
1

0.9 Staircase
(Zero order hold)
0.8 First order hold

0.7

0.6

0.5 y0
y1
0.4 y2
y3
0.3 y4

0.2
y5
y6
0.1
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

xi + 1 = A d xi + B d ui
(5.18)
yi = Cxi + Du i

For a staircase input, u(t) is constant over [ti, ti+1) and

( ( ) ( )
A d = e AT = φ t i + 1 − φ t i = φ (T) )
T A ( T − τ ) dτ B
Bd = ∫0 e
= ⎛⎜⎝ e AT − I⎞⎟⎠ A −1 B
for A regular (i.e. non-singular)

If λ is eigenvalue of A, eλT is eigenvalue of Ad

e(
σ + jω )T
= eσT e jωT = eσT < 1 ifσ < 0 (5.19)

Systems & Simulation - ENEL3SS 5-7

SS1_5.DOC
Taking Z transform of (18)

z X ( z ) − z x ( 0) = A d x ( z ) + B d U ( z )
Y ( z ) = Cd x ( z ) + DU ( z ) (5.20)
Y ( z ) = (C Iz − A −1B + D) U ( z )
d ( d ) d x ( 0) = 0

It can be shown that,

z − 1 ⎧⎪ -1⎧ P(s) ⎫ ⎪⎫
P( z ) = Z⎨ L ⎨ ⎬ ⎬ (5.21)
z ⎪⎩ ⎩ s ⎭ t = iT ⎪⎭

5.5 The modified z transform - see e.g. Gupta & Hasdoff

To examine the output of a continuous plant between sampling instances, add a delay, 0 < Δ < to
the input staircase. See back of table. Can be used to locate oscillations in pure discrete designs.

5.6. The Tustin (Trapezoidal) T transform

Numerical integration (approximate) for y = ∫ udt

T
(
yi + 1 ≈ yi + ui + 1 + ui
z
)
T
( z − 1)Y (z) = (z + 1)U(z)
2

2 z −1
s≈ (5.22)
T z +1

1 + sT / 2
z≈ (5.23)
1 − sT / 2

, consider
1 1 − e− aT ⎛ 2 z − 1⎞ T ( z + 1)
Ps (s) = , Pz ( z ) = Ps ⎜ ⎟=
s/ a +1 z − e− aT ⎝ T z + 1⎠ 2 ( z − 1) / a + T ( z + 1)

with a = 1, T = 5 × 10−3

4.99 x10−3
Pz ( z ) =
z − 0.995

Systems & Simulation - ENEL3SS 5-8

SS1_5.DOC
 ⎛ 2 ⎛ z − 1⎞ ⎞ (T / 2)( z + 1)
Ps ⎜ ⎜ ⎟⎟ =
⎝ T ⎝ z + 1⎠ ⎠ ( z − 1) + (T / 2)a ( z + 1)
T ( z + 1) / 2
=
1 / a + T / 2 z − (1 − T / 2) / (1 / a + T / 2)
( z + 1) / 2
= 4.99 × 10−3
z − 0.995
i.e. The same except input is averaged over [ui, ui+1]

5.7. The w-transform

The above (and other considerations) motivate definition of the w transform pair (bilinear)

2 z −1
w=
T z +1
(5.24)
1 + wT / 2
z=
1 − wT / 2

In older books the scaling by T/2 is often left out. This is an exact transform, defined on z-
transform i.e.: we need to know T before starting.

1) The stability boundary is the Im{w} axis.

2) If we can transform t-domain specifications into w domain and T has been fixed, we design in
w domain using familiar s domain tools.

Note that sampling too fast causes problems: (+ extra cost of fast hardware) (Middleton &
Goodman “Improved Finite Word Length characteristics in Digital Control Using Delta Apparatus”
IEEE Transactions on Automatic Control, Vol. AC-31 No 11, November 1986).

For example, implement a sampled version of

1
P ( s) = with finite precision
s/ a +1
1 − e− aT
P( z ) =
z − e− aT
yi + 1e− aT yi + ⎜⎛⎝ 1 − e− aT ⎟⎞⎠ ui eg a = 1, T = 5x10−3 two digits

If T is small,

yi + 1 = yi + 0 - i.e. do nothing (at the precision used)

We can develop discrete domain models based on changes in variables over the step,
Systems & Simulation - ENEL3SS 5-9
SS1_5.DOC
 z −1
δ= ; z = δT + 1 (5.25)
T

In above,
yi + 1 − yi 1 − e− aT 1 − e− aT 1 − e− aT
= yi + ui P(δ ) =
T T T 1 − e− aT + δT

Δy = T (0.999 yi + 0.998 ui)

↑this can be accumulated separately

Another view is obtained by writing Ps(s) as partial fraction expansion and then taking the z-
transform (transfer function with “zero-order hold”, not signal) and w-transform. Assuming no
repeated roots (for convenience),
n
Ps ( s ) = ∑
ri
, and Pz ( z ) = ∑ i
(
n r 1 − e − pi T p )
i , giving,
−
i =1 s + pi
p T
i =1 z −e i

⎡n
Pz ( w) = ⎢ ∑
( )
ri 1 − e − piT pi
⎤
(
⎥ 1− wT )
⎣
( T
2
) ( )
⎢i =1 1 + w − 1 − w e T
2
− pi T ⎥
⎦
2

= ⎢∑ T
( ) (
⎡ n ri 2 tanh pi T pi ⎤
⎥ 1− wT )
( )
2 (5.26)
⎢i =1 w + tanh pi
2 T ⎥ 2
⎣ T 2 ⎦
⎡n
≈ ⎢∑
ri ⎤
( )
⎥ 1− w 2
T
⎣i =1 w + pi ⎦

The approximations require that piT << 1, meaning that the sampling rate, 1/T must be more than
the highest plant corner frequency that has significant residual, ri. Eq. (2.1.7) illustrates the effect of
frequency warping when translating between the s- and w- domains.

Systems & Simulation - ENEL3SS 5-10

SS1_5.DOC
 Warping of poles from s- to w-domain
1

0.8

0.6

0.4

0.2
imag

-0.2

-0.4

-0.6

-0.8

-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
real

Warping of poles from s- to w-domains, p to 2 tanh p T with T=1

T
( 2)

Systems & Simulation - ENEL3SS 5-11

SS1_5.DOC