You are on page 1of 9

1

Frequency responses for sampled-data systems


Robert Cloudt


Introduction
In this paper different notions of a frequency response are considered for the general sampled-data
interconnection below. The emphasis is on SISO systems, but many results may be generalised to
MIMO systems.









Figure 1. General sampled-data system

The sample rate in the system above is denoted by h. H represents a zero order hold;

( )
e
e
e
j
e
j H
h j

=
1
. (1)

K is a discrete-time controller.

Discrete-time frequency response
The discrete-time frequency response (DFR) is defined with respect to the system below.









Figure 2. Sampled-data system extended with extra sampler and hold

Let G
~
be the system with discrete-time inputs u w
~
,
~
and outputs y z
~
,
~
. G
~
is the ZOH discretisation of
the plant G. If the plant G has system matrices A, B, C and D, then the state space equations of the
discretised plant are given by:

| | | | | |
| | | | | | k Du k Cx k y
k Bu d e k x e k x
h
A Ah
+ =
+ = +
}
0
1 t
t
. (2)

The DFR is defined using the ratio of the z-transforms of w
~
and z
~
in the closed loop system in figure
2:


G
K H
w
u
z
y
y
~
u
~

G
K H
w
u
z
y
y
~
u
~

u
~

H
w
~

u
~

z
~

u
~

2
( )
( )
( )
h j
h j
zw
e W
e Z
DFR
e
e
e
~
~
= (3)

here

( ) ( )
( ) ( )
h j
h j
e z
h j
e z
h j
z W e W
z Z e Z
e
e
e
e

=
=
~ ~
~ ~
. (4)

The DFR is defined using sampled input and output signals. Therefore the DFR does not account for
intersample behaviour.

Fundamental frequency response
The system in figure 1 can be decomposed as in figure 3.











Figure 3. Decomposition

The transfer from v
~
to u
~
(denoted as R
d
) is in discrete time and can be expressed in terms of the pulse
transfer function K(z) and the ZOH discretisation of G
22
:

( )
( )
( ) ( )
h j h j
h j
h j
d
e G e K
e K
e R
e e
e
e
22
~
1
= . (5)

As a consequence of impulse sampling:

( ) ( )

=
=
n
s
h j
jn j V
h
e V e e
e
1 ~
. (6)

Using equations 5 and 6, the Fourier transform of the output z can be expressed as:

( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )

=
+ =
n
s s
h j
d
jn j W jn j G
h
e R j H j G j W j G j Z e e e e e e e e e
e
21 12 11
1
. (7)

The response to ( )
t j
e w t w
0
0
e
= can be obtained through inverse Fourier transform. The Fourier
transform of
t j
e w
0
0
e
is ( )
0 0
2 e e o t w .

K
H
w
v y
~
u
~

G
21
v
~

G
22
H
u
G
12
z
G
11
3
( ) ( )
( ) ( ) ( )
( )
( ) ( )
0 0 21 0 0 12 0 0 11
0 0 0
1
2
1
w j G e R e jn j H jn j G
h
e w j G
d e j Z t z
h j
d
t n j
n
s s
t j
t j
s
e e e e e e
e e
t
e e e e
e

+
= =

}
(8)

Equation 7 shows that the (asymptotic) response to a sinusoidal input of frequency
0
e is composed of
sinusoidal components with frequencies
s
ne e
0
, e n .
The fundamental frequency response (FFR) is defined as the quotient of components of frequency
0
e
in the input and output, analogous to the continuous-time case.


( ) ( ) ( ) ( ) ( ) ( ) e e e e e
e
j G e R j H j G
h
j G FFR
h j
d zw 21 12 11
1
+ = (9)

Performance frequency response
The performance frequency response (PFR) is defined as the square root of the maximum power gain
form input to output if the input is a (possibly multivariable) sinusoid. The mathematical definition of
power used here is:

( ) ( ) ( ) ( ) ( )
} }


= =
T
T
T
T
T
T
dt t z t z
T
dt t z
T
t z P
*
2
2
1
lim
2
1
lim . (10)

where * denotes the conjugate transpose.
The signal ( )
t j t j
e e
0 0
*
e e
= is orthogonal to
( )t n j
s
e
e e
0
which implies

( )
h
dt e
h
e
T
T
T
n
t n j t j
T
s
1 1
2
1
lim
0 0
}

=


=
e e e
. (11)

An expression for the output power as a result of a sinusoidal input can be obtained by combining
equations 8, 10 and 11:

( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) ( )
( ) ( )
( ) ( ) ( ) ( )
( ) ( )
( )
0 0
*
0
0 0 21
0 0 12 0
*
12 0
*
*
0
*
21
*
0 0 11 0
*
12 0
* *
0
*
21
*
0 0 21 0 0 12 0
*
11
*
0 0 11 0
*
11
*
0
0
0
0
1
1
1
1
w A w
w j G e R
jn j H jn j G jn j G jn j H
h
e R j G
h
w
w j G j G j H e R j G
h
w
w j G e R j H j G j G
h
w w j G j G w t z P
h j
d
s s
n
s s
h j
d o
h j
d o
h j
d o o
e
e
e e e e e e e e
e
e e e e
e e e e e e
e
e
e
e
=


+
+
+ =

=
(12)

For the SISO case ( )
0
e A is a positive real scalar. In general however, ( )
0
e A is a square matrix with
real eigenvalues. The maximal power gain is then given by the maximal eigenvalue of ( )
0
e A .

4
( ) ( ) ( ) e e A PFR
zw max
= . (13)

Note that the term in equation 12 containing the infinite summation describes the ZOH discretisation
of the system in figure 4.





Figure 4. System for computing infinite summation

Suppose G
12
has system matrices A, B, C and D. Its transfer function is:

( ) ( ) D B A I j C j G + = e e
12
(14)

The adjoint system has transfer function:

( ) ( ) ( ) ( ) ( )
T T T T T T T T T
D C A I j B D C A I j B j G j G + = + = = e e e e
12
*
12
. (15)

So
*
12
G has system matrices -A
T
, C
T
, -B
T
, D
T
. It is easy to derive that the series connection
12
*
12
G G has
state space representation:

( ) ( ) ( )
( ) ( ) ( ) ( ) t Du D t x B C D t y
t u
D C
B
t x
A CC
A
t x
T T T
T T T
+ =
|
|
.
|

\
|
+
|
|
.
|

\
|

=
0

. (16)

The adjoint of the zero order hold has transfer function:

( )
e e
e
e e
j
e
j
e
j H
h j h j
1 1
*

=

=

(17)

Let ( ) t h
*
denote the impulse response of the conjugated hold operator. It is obtained from equation 17
through inverse Fourier transform.

( )

s s
=
elsewhere
t -h
t h
0
0 1
*
(18)

If the system in figure 4 is prepended with a ZOH, then the signal y is described by the equations:

for h kh t kh + < s , e k :
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( )

+ =
|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

+
|
|
.
|

\
|

|
|
.
|

\
|

=
}
kh Du D t x B C D t y
kh u
D C
B
d t
A CC
A
kh x kh t
A CC
A
t x
T T T
t
kh
T T T T T
t t
0
exp
0
exp
(19)

where x is the internal state of
12
*
12
G G . ( ) t v is the convolution of ( ) t h
*
with ( ) t y which yields:

for h kh t kh + < s , e k :
12
*
12
G G
H
*

u y v
5
( ) ( ) ( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( ) ( ) kh Du D kh u B d d
A CC
A
B C D
kh x d kh
A CC
A
B C D
kh Du D d x B C D d y h t v
T
t
khkh
T T
T T
t
kh
T T
T T
T
t
kh
T T
t
kh
+
|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

=
+ = =
} }
}
} }
o
o t t o
o o
o o o o
0
exp
0
exp . (20)

for h kh t + = this yields:

( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) kh Du D kh u B d d
A CC
A
B C D
kh x d kh
A CC
A
B C D kh v
T
h kh
kh kh
T T
T T
h kh
kh
T T
T T
+
|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

=
} }
}
+
+
o
o t t o
o o
0
exp
0
exp
. (21)

Substitution of variables gives:

( ) ( ) ( ) ( )
( ) ( ) ( ) ( ) kh Du D kh u B d d
A CC
A
B C D
kh x d kh
A CC
A
B C D kh v
T
h
T T
T T
h
T T
T T
+
|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

|
|
.
|

\
|

=
} }
}
0 0
0
0
exp
0
exp
o
o t t o
o o
. (22)

Now it is possible to write the infinite summation in equation 11 as a transfer function of a discrete-
time system;

( )
d d d
h j
d
D B A I e C +
e
(23)

where

|
|
.
|

\
|
|
|
.
|

\
|

= h
A CC
A
A
T T d
0
exp (24)
} |
|
.
|

\
|

|
|
.
|

\
|
|
|
.
|

\
|

=
h
T T T d
D C
B
d
A CC
A
B
0
0
exp t t (25)
( ) ( )
} |
|
.
|

\
|

|
|
.
|

\
|

=
h
T T
T T
d
d kh
A CC
A
B C D C
0
0
exp o o (26)
( ) ( ) D D B d d
A CC
A
B C D D
T
h
T T
T T
d
+
|
|
.
|

\
|

|
|
.
|

\
|

=
} }
0 0
0
exp
o
o t t o . (27)

The integral expressions in equations 25, 26 and 27 might be computed as matrix exponentials using
van Loans results [1].
This result allows the exact calculation of the performance frequency response.

6
Alias response power
Suppose the input of a sampled-data control system is excited with a sinewave of frequency . The
asymptotic response consists of a series of sinewaves with frequencies
s
ne e , e n . The alias
response power (ARP) is the square root of the power in the response due to the aliasing components
of frequency
s
ne e , e n . It is the difference between the PFR and the FFR.

( ) ( ) ( ) e e e
zw zw zw
FFR PFR ARP = . (28)

Robustness frequency response
Suppose the exogenous input signal is composed as:

( )
( )

=
+
=
n
t n j
n
s
e w t w
e e
0
, e
n
w . (29)

The asymptotic response to this kind of signal has the same structure:

( )
( )

=
+
=
n
t n j
n
s
e z t z
e e
0
, e
n
z . (30)

The robustness frequency response (RFR) is now defined as:

( )
( )
( ) ( )
( ) ( ) t w P
t z P
RFR
t w
zw
0
max
=
= e . (31)

The supremum of the RFR over frequency is equal to the power-induced norm and the L
2
-induced
norm of a sampled-data system. Therefore it is suited for robustness analysis. Note that

( )

>
11
G RFR
zw
e . (32)

The computation of the RFR for the general case is quite complicated [2]. However, for SISO systems
with 0
11
= G and 1
21
= G , the RFR equals the PFR.

Proof:
Equation 7 reduces to:

( ) ( ) ( ) ( ) ( )

=
=
n
s
h j
d
jn j W
h
e R j H j G j Z e e e e e
e
1
12
. (33)

Note that the PFR and RFR both can be expressed as:

( ) ( )
( ) ( ) t z P
t w P 1
max
=
(34)

but for the RFR w has the structure of equation 31 and in the PFR case ( )
t j
e w t w
0
0
e
= , e
0
w .
Consider the PFR case first. 1
0
= w such that ( ) ( ) 1 = t w P . ( ) ( ) t z P can be obtained through equation
12:

7
( ) ( ) ( )
( ) ( ) ( ) ( )
( )
( ) ( ) ( )

=
=


=
n
s s
h j
d
h j
d
s s
n
s s
h j
d o
jn j H jn j G e R
h
w
w e R
jn j H jn j G jn j G jn j H
h
e R
h
w t z P
2
0 0 12
2
2
2
0
0
0 0 12 0
*
12 0
*
* *
0
0
0
1
1
1
e e e e
e e e e e e e e
e
e
e
(35)

and the maximum:

( ) ( )
( ) ( ) ( ) ( ) ( )

=
=
=
n
s s
h j
d
t w P
jn j H jn j G e R
h
t z P
2
0 0 12
1
0
1
max e e e e
e
. (36)

Now consider the RFR case. Let

=1
2
n
w such that ( ) ( ) 1 = t w P . The Fourier transform of (29) is:

( ) ( )

=
=
n
n
n w j W
0
2 e e o t e . (37)

The output power can be obtained by applying Parsevals theorem to (37) substituted in (7):

( ) ( ) ( ) ( )
( ) ( ) ( )

}



|
|
.
|

\
|
=
=
n
s s
h j
d
n
n
jn j H jn j G e R
h
w
d j Z j Z t z P
2
0 0 12
2
2
2
*
0
1
2
1
e e e e
e e e
t
e
(38)

and again the maximum is:

( ) ( )
( ) ( ) ( ) ( ) ( )

=
=
=
n
s s
h j
d
t w P
jn j H jn j G e R
h
t z P
2
0 0 12
1
0
1
max e e e e
e
. (39)

Equations 36 and 39 are equal, so in this case the PFR is equal to the RFR.

The requirements 0
11
= G and 1
21
= G seem very restrictive, but they are satisfied when additive and
multiplicative plant uncertainties are considered.

The connection between DFR
yr
and PFR
yr
in a classical SISO feedback control structure
Consider the classical sampled-data SISO feedback control system in the figure below.








Figure 5. classical sampled-data SISO feedback control system

K(z)
H(s) P(s)
r
v w
n
y
u e
h
-
x
8
Let P
d
bet the ZOH discretisation of the plant P:

( ) ( ) ( )
s
n
s
h j
d
jn j P jn j H
h
e P e e e e
e
=

=
1
(40)

DFR
yr
can be expressed as:

( ) ( ) ( ) ( ) ( ) ( )
s
n
s
h j
d
h j
d
h j
d yr
jn j P jn j H e R
h
e R e P DFR e e e e e
e e e
= =

=
1
(41)

with P G =
22
. PFR
yr
can be obtained form (12) with P G G P G G = = = =
22 21 12 11
, 1 , , 0 .

( )
( ) ( ) ( ) ( ) ( ) ( )

=
h j
d s s
n
s s
h j
d
yr
e R jn j H jn j P jn j P jn j H
h
e R
h
PFR
e e
e e e e e e e e
e
* * * *
2
1 1
(43)
( ) ( ) ( ) ( )

=
=
n
s s
h j
d yr
jn j P jn j H e R
h
PFR
2 1
e e e e e
e
(44)

Due to the Schwarz inequality:

( ) ( ) ( ) ( )


=

=
s
n
s s s
n
s
jn j P jn j H jn j P jn j H
2
2
e e e e e e e e . (45)

From this one can conclude that DFR
yr
is a lower bound on PFR
yr
.
Define the fidelity
d
u as:

( )
( ) ( )
( ) ( )
1
2
2
>


= u

=
s
n
s
n
s s
h j
d
jn j P jn j H
jn j P jn j H
e
e e e e
e e e e
e
(46)

then
( ) ( ) ( ) e e
e
yr
h j
d yr
DFR e PFR u = (47)

The fidelity is a measure for how close the PFR and DFR are, i.e. a measure for intersample activity.
Note that the fidelity is independent of the controller.

Literature
[1] VanLoan, C.
COMPUTING INTEGRALS INVOLVING THE MATRIX EXPONENTIAL.
IEEE Transactions on Automatic Control, Vol. 23 (1978), No. 3, p. 395-404.


9
[2] Yamamoto, Y. and P.P. Khargonekar
FREQUENCY RESPONSE OF SAMPLED-DATA SYSTEMS.
IEEE Transactions on Automatic Control, Vol. 41 (1996), No. 2, p. 166-176.

[3] Lindgrde, O. and B. Lennartson
FREQUENCY ANALYSIS FOR SINGLE AND MULTIRATE SAMPLED-DATA
CONTROL.
Proc. of the 34
th
Conference on Decision & Control, New Orleans, USA, 13-15 december
1995.
New York: IEEE, 1995, Vol 2, p. 1770-1775.

[4] Lindgrde, O. and B. Lennartson
COMPARING FREQUENCY ANALYSIS METHODS FOR SAMPLED-DATA
CONTROL.
Proc. of the 37
th
Conference on Decision & Control, Tampa, Florida USA, 16-18 december
1998.
New York: IEEE, 1998, Vol 1, p. 829-834.

[5] Braslavsky, J.H. and R.H. Middleton, J.S. Freudenberg
L
2
-INDUCED NORMS AND FREQUENCY GAINS OF SAMPLED-DATA SENSITIVITY
OPERATORS.
IEEE Transactions on Automatic Control, Vol. 43 (1998), No. 2, p. 252-258.

You might also like