From _Control System Design_

by Karl Johan Åström, 2002

7 Copyright 2002, Karl Johan Åström.

All rights reserved.
Do not duplicate or redistribute.

Specifications

7.1 Introduction
In this chapter we will discuss how the properties of a control system can
be specified. This is important for control design because it gives the goals.
It is also important for users of control so that they know how to specify,
evaluate and test a system so that they know it will have the desired
properties. Specifications on a control systems typically include: stability
of the closed loop system, robustness to model uncertainty, attenuation of
measurement noise, injection of measurement noise, and ability to follow
reference signals. From the results of Chapter 5 it follows that these
properties are captured by six transfer functions called the Gang of Six.
The specifications can be expressed in terms of these transfer functions.
Essential features of the transfer functions can be expressed in terms of
their poles and zeros or features of time and frequency responses.

7.2 Stability and Robustness to Process Variations

Stability and robustness to process uncertainties can be expressed in

terms of the loop transfer function L = PC, the sensitivity function and
the complementary sensitivity function

1 1 PC L
S= = , T= = .
1 + PC 1+ L 1 + PC 1+ L

Since both S and T are functions of the loop transfer function specifica-
tions on the sensitivities can also be expressed in terms of specifications on
the loop transfer function L. Many of the criteria are based on Nyquist’s

252
 7.2 Stability and Robustness to Process Variations

Figure 7.1 Nyquist curve of the loop transfer function L with indication of gain,
phase and stability margins.

stability criterion, see Figure 7.1. Common criteria are the maximum val-
ues of the sensitivity functions, i.e.

Ms = max h S(iω )h, Mt = max hT (iω )h

ω ω

Recall that the number 1/ Ms is the shortest distance of the Nyquist curve
of the loop transfer function to the critical point, see Figure 7.1. Also recall
that the closed loop system will remain stable for process perturbations
∆ P provided that
h∆ P(iω )h 1
≤ ,
h P(iω )h hT (iω )h
see Section 5.5. The largest value Mt of the complementary sensitivity
function T is therefore a simple measure of robustness to process varia-
tions.
Typical values of the maximum sensitivities are in the range of 1 to 2.
Values close to one are more conservative and values close to 2 correspond
to more aggressive controllers.

Gain and Phase Margins

The gain margin gm and the phase margin ϕ m are classical stability cri-
teria. Although they can be replaced by the maximum sensitivities it is
useful to know about them because they are still often used practically.

253
Chapter 7. Specifications

The gain margin tells how much the gain has to be increased before the
closed loop system becomes unstable and the phase margin tells how much
the phase lag has to be increased to make the closed loop system unstable.
The gain margin can be defined as follows. Let ω 180 be the lowest
frequency where the phase lag of the loop transfer function L(s) is 180○ .
The gain margin is then
1
gm = (7.1)
h L(iω 180 )h

The phase margin can be defined as follows. Let ω gc denote gain

crossover frequency, i.e. the lowest frequency where the loop transfer
function L(s) has unit magnitude. The phase margin is then given by

ϕ m = π + arg L(iω gc ) (7.2)

The margins have simple geometric interpretations in the Nyquist dia-

gram of the loop transfer function as is shown in Figure 7.1. Notice that
an increase of controller gain simply expands the Nyquist curve radially.
An increase of the phase of the controller twists the Nyquist curve clock-
wise, see Figure 7.1.
Reasonable values of the margins are phase margin ϕ m = 30○ − 60○ ,
gain margin gm = 2 − 5. Since it is necessary to specify both margins to
have a guarantee of a reasonable robustness the margins gm and ϕ m can
be replaced by a single stability margin, defined as the shortest distance
of the Nyquist curve to the critical point −1, this distance is the inverse
of the maximum sensitivity Ms . It follows from Figure 7.1 that both the
gain margin and the phase margin must be specified in order to ensure
that the Nyquist curve is far from the critical point. It is possible to
have a system with a good gain margin and a poor phase margin and
vice versa. It is also possible to have a system with good gain and phase
margins which has a poor stability margin. The Nyquist curve of the loop
transfer function of such a system is shown in Figure 7.2. This system has
infinite gain margin, a phase margin of 70○ which looks very reassuring,
but the maximum sensitivity is Ms = 3.7 which is much too high. Since
it is necessary to specify both the gain margin and the phase margin to
endure robustness of a system it is advantageous to replace them by a
single number. A simple analysis of the Nyquist curve shows that the
following inequalities hold.
Ms
gm ≥
Ms − 1
(7.3)
1
φm ≥ 2 arcsin
2Ms

254
 7.2 Stability and Robustness to Process Variations

Figure 7.2 Nyquist curve of the loop transfer function for a system with good gain
and phase margins but with high sensitivity and poor robustness. The loop transfer
0.38(s2 +0.1s+0.55)
function is L(s) = s(s+1)(s2 +0.06s+0.5
.

A controller with Ms = 2 thus has a gain margin of at least 2 and a phase

margin of at least 30○ . With Ms = 1.4 the margins are gm ≥ 3.5 and
φ m ≥ 45○ .

Delay Margin
The gain and phase margins were originally developed for the case when
the Nyquist curve only intersects the unit circle and the negative real axis
once. For more complicated systems there may be many intersections and
it is more complicated to find suitable concepts that capture the idea of a
stability margin. One illustration if given in Figure 7.3. In this case the
Nyquist curve has a large loop and the Nyquist curve intersects the circle
h Lh = 1 in three points corresponding to the frequencies 0.21, 0.88 and
1.1. If there are variations in the time delay the Nyquist curve can easily
enclose the critical point. In the figure it is shown what happens when
the time delay is increased from 3 to 4.5 s. This increase corresponds to a
phase lag of 0.3 rad at the crossover frequency 0.21 rad/s, the phase lag
is however 1.6 rad at the frequency 1.1 rad/s which is marked A in the
figures. Notice that the point A becomes very close to the critical point.
A good measure of the stability margin in this case is the delay margin
which is the smallest time delay required to make the system unstable.
For loop transfer functions that decay quickly the delay margin is closely
related to the phase margin but for systems where the amplitude ratio of

255
Chapter 7. Specifications

−2

y
−4

−6

−8

−10
−6 −4 −2 0 2 4 6 8 10
x

0.2
Figure 7.3 Nyquist curve of the loop transfer function L(s) = s(s2 +0.025s+1)
e−3s .

the loop transfer function has several peaks at high frequencies the delay
margin is a much more relevant measure.

7.3 Disturbances

In the standard system, Figure 5.1, we have used in this book there are
two types of disturbances, the load disturbances that drive the system
away from its desired behavior and the measurement noise that corrupts
the information about the process obtained by the sensors.

Response to Load Disturbances

The response of the process variable to a load disturbance is given by the
transfer function
P T
G xd = = PS = (7.4)
1 + PC C
Since load disturbances typically have low frequencies it is natural that
the specifications should emphasize the behavior of the transfer function
at low frequencies. The loop transfer function L = PC is typically large
for small s and we have the approximation

T 1
G xd =  . (7.5)
C C

256
 7.3 Disturbances

0
10

−1
10

−2
10
−2 −1 0 1
10 10 10 10

Figure 7.4 Typical gain curve for the transfer function Gxd from load disturbance
to process output. The gain curve is shown in full lines and the transfer function
ki /s in dotted lines and the process transfer function in full lines.

If P(0) = 0 and the controller with integral action control we have the
following approximation for small s

s
G xd  .
ki

Since load disturbances typically have low frequencies this equation im-
plies that integral gain ki is a good measure of load disturbance atten-
uation. Figure 7.4 shows the magnitude curve of the transfer function
(7.4) for a PI control of the process P = (s + 1)−4 . The transfer function
G xd has typically the form shown in Figure 7.4. The curve can typically
be characterized by the low frequency asymptote ( ki ), the peak ( Mxd ),
the frequency (ω xd )where the peak occurs and the high frequency roll-
off. It follows from (7.4) that the high frequency behavior is essentially
determined by the process and the maximum sensitivity.
Attenuation of load disturbances can also be characterized in the time
domain by showing the time response due to a representative disturbance.
This is illustrated in 7.5 which shows the response of the process output to
a unit step disturbance at the process input. The figure shows maximum
error emax , the steady state error ess , the error of the open loop system eol ,
the time to maximum tmax and the settling time ts .

Measurement Noise
An inevitable consequence of using feedback is that measurement noise is
fed into the system. Measurement noise thus causes control actions which
in turn generate variations in the process variable. It is important to keep
these variations of the control signal at reasonable levels. A typical re-
quirement is that the variations are only a fraction of the span of the
control signal. The variations in the control variable are also detrimental

257
Chapter 7. Specifications

Figure 7.5 Errors due to a unit step load disturbance at the process input and
some features used to characterize attenuation of load disturbances. The curves
show the open-loop error (dashed lines) and the error (full lines) obtained using a
controller without integral action (upper) and with integral action (lower).

by themselves because they cause wear of actuators. Since measurement

noise typically has high frequencies the high high frequency gain of the
controller is a relevant measure. Notice however that the low frequency
gain of the controller is not essential since measurement noise is high fre-
quency. Frequencies above the gain crossover frequency will be regarded
as high.
To get a feel for the orders of magnitude consider an analog system
where the signal levels are 10V. A measurement noise of 1 mV then satu-
rates the input if the gain is 104 . If it is only permitted that measurement
noise gives control signals of 1V the high frequency gain of the controller
must be less than 103 .
As an other illustration we consider a digital control system with 12
bit AD- and DA-converters. A change of the input of one bit saturates
the DA-converter if the gain is 4096. Assume that we permit one bit to
give a variation of 0.4% of the output range. The high frequency gain of
the controller must then be less than 500. With converters having lower
resolution the high frequency gain would be even lower.
High precision analog systems with signal ranges of 1 to 104 have been

258
 7.4 Reference Signals

designed. For digital systems the signal ranges are limited by the sensors
and the actuators. Special system architectures with sensors and actua-
tors having multiple signal ranges are used in order to obtain systems
with a very high signal resolution. In these cases it is possible to have
signal ranges up to 1 to 106 .
The effects of measurement noise can be evaluated by the transfer
function from measurement noise to the control signal, i.e.,

C T
Gun = = CS = . (7.6)
1 + PC P

Recall that P and C and are the transfer functions of the process and
the controller, and that S is the sensitivity function. Notice that when
L = PC is large we have approximately Gun  1/ C. Since measurement
noise typically has high frequencies and since the sensitivity function is
one for high frequencies we find that the response to measurement noise
is essentially determined by the high frequency behavior of the transfer
function C. A simple measure is given by

Mc = max h Gun (iω )h ≤ Ms max h C (iω )h

ω ≥ω g c ω ≥ω g c

where Mc is called the maximum high frequency gain of the controller.

When there is severe measurement noise it is advantageous to make sure
that the transfer function C goes to zero for large frequencies. This is
called high frequency roll-off.

7.4 Reference Signals

The response to set point changes is described by the transfer functions

F PC FC
G yr = = FT , Gur = = FCS
1 + PC 1 + PC

Compare with (5.1). A significant advantage with controller structure

with two degrees of freedom is that the problem of set point response can
be decoupled from the response to load disturbances and measurement
noise. The design procedure can then be divided into two independent
steps.
• First design the feedback controller C that reduces the effects of
load disturbances and the sensitivity to process variations without
introducing too much measurement noise into the system

259
Chapter 7. Specifications

1.2

0.8

0.6

0.4

0.2

0
0 2 4 6 8 10 12 14 16

2.5

1.5

0.5

−0.5
0 2 4 6 8 10 12 14 16 18 20

Figure 7.6 Specifications on reference following based on the time response to a

unit step in the reference.

• Then design the feedforward F to give the desired response to set

points.

Specifications on reference following are typically expressed in the time

domain. They may include requirements on rise time, settling time, de-
cay ratio, overshoot, and steady-state offset for step changes in reference.
These quantities are defined as follows, see 7.6. These quantities are de-
fined in different ways and there are also different standards.

• The rise time tr is either defined as the inverse of the largest slope
of the step response or the time it takes the step to pass from 10%
to 90% of its steady state value.

• The settling time ts is the time it takes before the step response
remains within p percent of its steady state value. The value p = 2
% is commonly used.

• The delay time is the time required for the step response to reach
50 % of its steady state value for the first time.

• The decay ratio d is the ratio between two consecutive maxima of the
error for a step change in reference or load. The value d = 1/4, which

260
 7.4 Reference Signals

is called quarter amplitude damping, has been used traditionally.

This value is, however, too high as will be shown later.
• The overshoot o is the ratio between the difference between the first
peak and the steady state value and the steady state value of the
step response. In industrial control applications it is common to spec-
ify an overshoot of 8%–10%. In many situations it is desirable, how-
ever, to have an over-damped response with no overshoot.
• The steady-state error ess is the value of control error e in steady
state. With integral action in the controller, the steady-state error
is always zero.
Classical specifications have been strongly focused on the behavior of the
process output. It is however important to also consider the control signal.
Analogous quantities can be defined for the control signal. The overshoot
of the control signal is of particular importance, see Figure 7.4.
Step signals are often used as reference inputs In motion control sys-
tems it is often more relevant to consider responses to ramp signals or
jerk signals. Specifications are often given in terms of the the value of the
first non-vanishing error coefficient.

Tracking Slowly Varying Signals - Error Coefficients

Step signals is one prototype of reference signals. There are however sit-
uations when other signals are more appropriate. One example is when
the reference signal has constant rate of change, i.e.

r(t) = v0 t

The corresponding Laplace transform is R(s) = v0 / s2 .

For a system with error feedback the error e = r − y has the Laplace
transform
v0
E ( s) = S ( s) V ( s) = S ( s) 2 (7.7)
s
The steady state error obtained depends on the properties of the sensitiv-
ity function at the origin. If S(0) = e0 the steady state tracking error is
asymptotically e(t) = v0 e0 t. To have a constant tracking error it must be
required that S(0) = 0. With S(s)  e1 s for small s we find that the steady
state error is e(t) = v0 e1 as t goes to infinity. To have zero steady state
error for a ramp signal the function S(s) must go to zero faster than s for
small s. If S(s)  e2 s2 for small s we find that the error is asymptotically
zero. Since
1
L(s) = −1
S ( s)

261
Chapter 7. Specifications

it follows that the condition S(s)  e2 s2 implies that L(s)  s−2 for small
s. This implies that there are two integrations in the loop. Continuing
this reasoning we find that in order to have zero steady state error when
tracking the signal
t2
r(t) =
2
it is necessary that s(s)  e3 s3 for small s. This implies that there are
three integrals in the loop.
The coefficients of the Taylor series expansion of the sensitivity s(s)
function for small s,

S(s) = e0 + e1 s + e2 s2 + . . . + en sn + . . . (7.8)

are thus useful to express the steady state error in tracking low frequency
signals. The coefficients ek are called error coefficients. The first non van-
ishing error coefficient is the one that is of most interest, this is often
called the error coefficient.

7.5 Specifications Based on Optimization

The properties of the transfer functions can also be based on integral
criteria. Let e(t) be the error caused by reference values or disturbances
and let u(t) be the corresponding control signal. The following criteria are
commonly used to express the performance of a control system.
Z ∞
IE = e(t)dt
Z0 ∞
I AE = h e(t)hdt
Z0 ∞
IT AE = th e(t)hdt
0
Z ∞
IQ = e2 (t)dt
0
Z ∞
WQ = ( e2 (t) + ρ u2 (t))dt
0

They are called, IE integrated error, IAE integrated absolute error, ITAE
integrated time multiplies absolute error, integrated quadratic error and
WQ weighted quadratic error. The criterion WQ makes it possible to
trade the error against the control effort required to reduce the error.

262
 7.6 Properties of Simple Systems

7.6 Properties of Simple Systems

It is useful to have a good knowledge of properties of simple dynamical
systems. In this section we have summarize such data for easy reference.

First Order Systems

Consider a system where the transfer function from reference to output
is
a
G ( s) = (7.9)
s+a
The step and impulse responses of the system are

h(t) = 1 − e−at = 1 − e−t/T

1
g(g) = ae−at = e−t/T
T

where the parameter T is the time constant of the system. Simple calcu-
lations give the properties of the step response shown in Table 7.1. The
2% settling time of the system is 4 time constants. The step and impulse
responses are monotone. The velocity constant e1 is also equal to the time
constant T. This means that there will be a constant tracking error of
e1 v = v0 T when the input signal is a ramp r = v0 t.
This system (7.9) can be interpreted as a feedback system with the
loop transfer function
a 1
L( s ) = =
s sT
This system has a gain crossover frequency ω gc = a. The Nyquist curve
is the negative imaginary axis, which implies that the phase margin is
90○ . Simple calculation gives the results shown in Table 7.1. The load
disturbance response of a first order system typically has the form

s
G xd =
s+a

The step response of this transfer function is

h xd = e−at

The maximum thus occurs when the disturbance is applies and the set-
tling time is 4T. The frequency response decays monotonically for increas-
ing frequency. The largest value of the gain is a zero frequency.
Some characteristics of the disturbance response are given in Table 7.2.

263
Chapter 7. Specifications

Table 7.1 Properties of the response to reference values for the first order system
Gxr = a/(s + a).

Propety Value
Rise time Tr = 1/ a = T
Delay time Td = 0.69/ a = 0.69T
Settling time (2%) Ts = 4/ a = 4T
Overshoot o=0
Error coefficients e0 = 0, e1 = 1/ a = T
Bandwidth ωb = a
Resonance peak ωr = 0
Sensitivities Ms = Mt = 1
Gain margin gm = ∞
Phase margin ϕ m = 90○
Crossover frequency ω gc = a
Sensitivity frequency ω sc = ∞

Table 7.2 Properties of the response to disturbances for the first order system
Gxd = s/(s + a).

Property Value
Peak time Tp = 0
Max error emax = 1
Settling time Ts = 4T
Error coefficient e1 = T
Largest norm hh G xd hh = 1
Integrated error I E = 1/ a = T
Integrated absolute error I AE = 1/ a = T

Second Oder System without Zeros

Consider a second order system with the transfer function

ω 02
G ( s) = (7.10)
s2 + 2ζ ω 0 s + ω 02

264
 7.6 Properties of Simple Systems

The system has two poles, they are complex if ζ < 1 and real if ζ > 1.
The step response of the system is


 e−ζ ω 0 t

 1− p sin(ω d t + φ ) for hζ h < 1

 1−ζ2

h(t) = 1 − (1 + ω 0 t) e−ω 0 t forζ = 1

  

 ζ

 sinh ω d t e−ζ ω d t
1 − cosh ω d t + p 2 for hζ h > 1
ζ −1
p
where ω d = ω 0 h1 − ζ 2 h and φ = arccos ζ . Whenpζ < 1 the step response
is a damped oscillation, with frequency ω d = ω 0 1 − ζ 2 . Notice that the
step response is enclosed by the envelopes

e−ζ ω 0 t ≤ h(t) ≤ 1 − e−ζ ω 0 t

This means that the system settles like a first order system with time
constant T = ζ ω1 0 . The 2% settling time is thus Ts  ζ ω4 0 . Step responses
for different values of ζ are shown in Figure 4.9.
The maximum of the step response occurs approximately at Tp  π /ω d ,
i.e. half a period of the oscillation. The overshoot depends on the damping.
The largest overshoot is 100% for ζ = 0. Some properties of the step
response are summarized in Table 7.3.
The system (7.10) can be interpreted as a feedback system with the
loop transfer function
ω 02
L(s) =
s(s + 2ζ ω 0 )
This means that we can compute quantities such as sensitivity functions
and stability margins. These quantities are summarized in Table 7.3.

Second Oder System with Zeros

The response to load disturbances for a second order system with integral
action can have the form
ω 0s
G ( s) =
s2 + 2ζ ω 0 s + ω 02

The frequency response has a maximum 1/(2ζ ) at ω = ω 0 . The step

response of the transfer function is

e−ζ ω 0 t
h(t) = sinω d t
sqrt1 − ζ 2

265
Chapter 7. Specifications

Table 7.3 Properties of the response to reference values of a second order system.

Property Value
Rise time Tr = ω 0 eφ / tan φ  2.2Td
Delay time Td
Peak time Tp  π /ω D = Td /2
Settling time (2%) Ts  4/(ζ ω 0 )
√ 2
Overshoot o = e−π ζ / 1−ζ
Error coefficients e = 0, e1 = 2ζ /ω 0
q0 p
Bandwidth ω b = ω 0 1 − 2ζ 2 + (1 − 2ζ 2 )2 + 1
r √
8ζ 2 +1+(4ζ ĺ2+1) 8ζ 2 +1
Maximum sensitivity Ms = √
8ζ 2 +1+(4ζ ĺ2−1) 8ζ 2 +1
√ 2
1+ 8ζ +1
Frequency
( wmsp= ω0
2 √
1/(2ζ 1 − ζ 2 ) if ζ ≤ 2/2
Max. comp. sensitivity Mt = √
1( p if ζ ≤√ 2/2
ω 0 1 − 2ζ 2 if ζ ≤ 2/2
Frequency ω mt = √
1 if ζ ≤ 2/2
Gain margin gm = ∞
Phase margin ϕ m = 90 − arctan ω c /(2ζ ω 0 )
○
qp
Crossover frequency ω gc = ω 0 4ζ 4 + 1 − 2ζ 2
√
Sensitivity frequency ω sc = ω 0 / 2

This could typically represent the response to a step in the load distur-
bance. Figure 7.7 shows the step response for different values of ζ . The
step response has its maximum

max h(t) = ω 0 e−ζ //sqrt1−ζ

2
(7.11)
t

for
arccos ζ
t = tm =
ω0

266
 7.7 Poles and Zeros

0.5

h 0

−0.5

−1
0 5 10 15
ω 0t

Figure 7.7 Step responses of the transfer function (7.11) for ζ = 0 (dotted), 0.1,
0.2, 0.5, 0.7 (dash-dotted), 1, 2, 5, 10 (dashed).

Figure 7.8 Typical configuration of poles and zeros for a transfer function describ-
ing the response to reference signals.

Systems of Higher Order

7.7 Poles and Zeros

Specifications can also be expressed in terms of the poles and zeros of
the transfer functions. The transfer function from reference value to the
output of a system typically has the pole zero configuration shown in
Figure 7.8. The behavior of a system is characterized by the poles and
zeros with the largest real parts. In the figure the behavior is dominated
by a complex pole pair p1 and p2 and real poles and zeros. The dominant
poles are often characterized by the relative damping ζ and the distance
from the origin ω 0 . Robustness is determined by the relative damping and
the response speed is inversely proportional to ω 0 .

267
Chapter 7. Specifications

• Dominant poles
• Zeros
• Dipoles

7.8 Relations Between Specifications

A good intuition about the different specifications can be obtained by in-
vestigating the relations between specifications for simple systems as is
given in Tables 7.1, 7.2 and 7.3.

The Rise Time Bandwidth Product

Consider a transfer function G (s) for a stable system with G (0) = 0.
We will derive a relation between the rise time and the bandwidth of a
system. We define the rise time by the largest slope of the step response.

G (0)
Tr = (7.12)
maxt g(t)

where g is the impulse response of G, and let the bandwidth be defined

as R∞
h G (iω )h
ωb = 0 (7.13)
π G (0)
This implies that the bandwidth for the system G (s) = 1/(s + 1) is equal
√
to 1, i.e. the frequency where the gain has dropped by a factor of 1/ 2.
The impulse response g is related to the transfer function G by
Z i∞ Z ∞
1 1
g(t) = e G (s)ds =
st
eiω t G (iω )dω
2π i − i∞ 2π −∞

Hence
Z ∞  Z
1  iω t  1 ∞
max g(t) ≤  e G (iω )dω = h G (iω )hdω
t 2π −∞ π 0

Equations (7.12) and (7.13) now give

Tr ω b ≥ 1

This simple calculation indicates that the product of rise time and band-
width is approximately constant. For most systems the product is around
2.

268
 7.9 Summary

7.9 Summary
It is important for both users and designers of control systems to un-
derstand the role of specifications. The important message is that it is
necessary to have specifications that cover properties of the Gang of Six,
otherwise there is really no guarantee that the system will work well.
This important fact is largely neglected in much of the literature and in
control practice. Some practical ways of giving reasonable specifications
are summarized.

269